2016-02-10 00:12:00 +01:00
|
|
|
// Copyright (c) 2011-present, Facebook, Inc. All rights reserved.
|
2017-07-16 01:03:42 +02:00
|
|
|
// This source code is licensed under both the GPLv2 (found in the
|
|
|
|
// COPYING file in the root directory) and Apache 2.0 License
|
|
|
|
// (found in the LICENSE.Apache file in the root directory).
|
2013-10-16 23:59:46 +02:00
|
|
|
//
|
2011-03-18 23:37:00 +01:00
|
|
|
// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
|
|
|
|
// Use of this source code is governed by a BSD-style license that can be
|
|
|
|
// found in the LICENSE file. See the AUTHORS file for names of contributors.
|
|
|
|
|
2014-09-05 08:14:37 +02:00
|
|
|
#ifndef __STDC_FORMAT_MACROS
|
2014-03-05 02:08:05 +01:00
|
|
|
#define __STDC_FORMAT_MACROS
|
2014-09-05 08:14:37 +02:00
|
|
|
#endif
|
2014-05-09 02:25:13 +02:00
|
|
|
|
2016-02-16 21:16:47 +01:00
|
|
|
#ifdef GFLAGS
|
2014-07-07 19:53:31 +02:00
|
|
|
#ifdef NUMA
|
|
|
|
#include <numa.h>
|
|
|
|
#include <numaif.h>
|
|
|
|
#endif
|
2015-07-02 01:13:49 +02:00
|
|
|
#ifndef OS_WIN
|
2015-04-07 02:16:40 +02:00
|
|
|
#include <unistd.h>
|
2015-07-02 01:13:49 +02:00
|
|
|
#endif
|
2015-04-03 03:14:49 +02:00
|
|
|
#include <fcntl.h>
|
2014-03-05 02:08:05 +01:00
|
|
|
#include <inttypes.h>
|
Generate mixed workload with Get, Put, Seek in db_bench (#4788)
Summary:
Based on the specific workload models (key access distribution, value size distribution, and iterator scan length distribution, the QPS variation), the MixGraph benchmark generate the synthetic workload according to these distributions which can reflect the real-world workload characteristics.
After user enable the tracing function, they will get the trace file. By analyzing the trace file with the trace_analyzer tool, user can generate a set of statistic data files including. The *_accessed_key_stats.txt, *-accessed_value_size_distribution.txt, *-iterator_length_distribution.txt, and *-qps_stats.txt are mainly used to fit the Matlab model fitting. After that, user can get the parameters of the workload distributions (the modeling details are described: [here](https://github.com/facebook/rocksdb/wiki/RocksDB-Trace%2C-Replay%2C-and-Analyzer))
The key access distribution follows the The two-term power model. The probability density function is: `f(x) = ax^{b}+c`. The corresponding parameters are key_dist_a, key_dist_b, and key_dist_c in db_bench
For the value size distribution and iterator scan length distribution, they both follow the Generalized Pareto Distribution. The probability density function is `f(x) = (1/sigma)(1+k*(x-theta)/sigma))^{-1-1/k)`. The parameters are: value_k, value_theta, value_sigma and iter_k, iter_theta, iter_sigma. For more information about the Generalized Pareto Distribution, users can find the [wiki](https://en.wikipedia.org/wiki/Generalized_Pareto_distribution) and [Matalb page](https://www.mathworks.com/help/stats/generalized-pareto-distribution.html)
As for the QPS, it follows the diurnal pattern. So Sine is a good model to fit it. `F(x) = sine_a*sin(sine_b*x + sine_c) + sine_d`. The trace_will tell you the average QPS in the print out resutls, which is sine_d. After user fit the "*-qps_stats.txt" to the Matlab model, user can get the sine_a, sine_b, and sine_c. By using the 4 parameters, user can control the QPS variation including the period, average, changes.
To use the bench mark, user can indicate the following parameters as examples:
```
-benchmarks="mixgraph" -key_dist_a=0.002312 -key_dist_b=0.3467 -value_k=0.9233 -value_sigma=226.4092 -iter_k=2.517 -iter_sigma=14.236 -mix_get_ratio=0.7 -mix_put_ratio=0.25 -mix_seek_ratio=0.05 -sine_mix_rate_interval_milliseconds=500 -sine_a=15000 -sine_b=1 -sine_d=20000
```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4788
Differential Revision: D13573940
Pulled By: sagar0
fbshipit-source-id: e184c27e07b4f1bc0b436c2be36c5090c1fb0222
2019-01-22 19:40:44 +01:00
|
|
|
#include <math.h>
|
2011-03-18 23:37:00 +01:00
|
|
|
#include <stdio.h>
|
|
|
|
#include <stdlib.h>
|
2016-10-19 01:30:57 +02:00
|
|
|
#include <sys/types.h>
|
db_bench periodically writes QPS to CSV file
Summary:
This is part of an effort to better understand and optimize RocksDB stalls under high load. I added a feature to db_bench to periodically write QPS to CSV files. That way we can nicely see how our QPS changes in time (especially when DB is stalled) and can do a better job of evaluating our stall system (i.e. we want the QPS to be as constant as possible, as opposed to having bunch of stalls)
Cool part of CSV files is that we can easily graph them -- there are a bunch of tools available.
Test Plan:
Ran ./db_bench --report_interval_seconds=10 --benchmarks=fillrandom --num=10000000
and observed this in report.csv:
secs_elapsed,interval_qps
10,2725860
20,1980480
30,1863456
40,1454359
50,1460389
Reviewers: sdong, MarkCallaghan, rven, yhchiang
Reviewed By: yhchiang
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D40047
2015-06-12 23:31:53 +02:00
|
|
|
#include <atomic>
|
|
|
|
#include <condition_variable>
|
2016-10-19 01:30:57 +02:00
|
|
|
#include <cstddef>
|
2017-07-29 01:23:50 +02:00
|
|
|
#include <memory>
|
db_bench periodically writes QPS to CSV file
Summary:
This is part of an effort to better understand and optimize RocksDB stalls under high load. I added a feature to db_bench to periodically write QPS to CSV files. That way we can nicely see how our QPS changes in time (especially when DB is stalled) and can do a better job of evaluating our stall system (i.e. we want the QPS to be as constant as possible, as opposed to having bunch of stalls)
Cool part of CSV files is that we can easily graph them -- there are a bunch of tools available.
Test Plan:
Ran ./db_bench --report_interval_seconds=10 --benchmarks=fillrandom --num=10000000
and observed this in report.csv:
secs_elapsed,interval_qps
10,2725860
20,1980480
30,1863456
40,1454359
50,1460389
Reviewers: sdong, MarkCallaghan, rven, yhchiang
Reviewed By: yhchiang
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D40047
2015-06-12 23:31:53 +02:00
|
|
|
#include <mutex>
|
|
|
|
#include <thread>
|
2015-12-08 02:33:18 +01:00
|
|
|
#include <unordered_map>
|
db_bench periodically writes QPS to CSV file
Summary:
This is part of an effort to better understand and optimize RocksDB stalls under high load. I added a feature to db_bench to periodically write QPS to CSV files. That way we can nicely see how our QPS changes in time (especially when DB is stalled) and can do a better job of evaluating our stall system (i.e. we want the QPS to be as constant as possible, as opposed to having bunch of stalls)
Cool part of CSV files is that we can easily graph them -- there are a bunch of tools available.
Test Plan:
Ran ./db_bench --report_interval_seconds=10 --benchmarks=fillrandom --num=10000000
and observed this in report.csv:
secs_elapsed,interval_qps
10,2725860
20,1980480
30,1863456
40,1454359
50,1460389
Reviewers: sdong, MarkCallaghan, rven, yhchiang
Reviewed By: yhchiang
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D40047
2015-06-12 23:31:53 +02:00
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
#include "db/db_impl.h"
|
2018-10-24 20:36:19 +02:00
|
|
|
#include "db/malloc_stats.h"
|
2011-03-18 23:37:00 +01:00
|
|
|
#include "db/version_set.h"
|
2016-03-03 20:20:25 +01:00
|
|
|
#include "hdfs/env_hdfs.h"
|
2017-04-06 04:02:00 +02:00
|
|
|
#include "monitoring/histogram.h"
|
|
|
|
#include "monitoring/statistics.h"
|
2018-10-24 20:36:19 +02:00
|
|
|
#include "options/cf_options.h"
|
2016-03-03 20:20:25 +01:00
|
|
|
#include "port/port.h"
|
|
|
|
#include "port/stack_trace.h"
|
2013-08-23 17:38:13 +02:00
|
|
|
#include "rocksdb/cache.h"
|
|
|
|
#include "rocksdb/db.h"
|
|
|
|
#include "rocksdb/env.h"
|
2014-09-08 19:37:05 +02:00
|
|
|
#include "rocksdb/filter_policy.h"
|
2016-03-03 20:20:25 +01:00
|
|
|
#include "rocksdb/memtablerep.h"
|
|
|
|
#include "rocksdb/options.h"
|
|
|
|
#include "rocksdb/perf_context.h"
|
2017-02-27 21:59:34 +01:00
|
|
|
#include "rocksdb/persistent_cache.h"
|
2015-05-13 02:40:27 +02:00
|
|
|
#include "rocksdb/rate_limiter.h"
|
2016-03-03 20:20:25 +01:00
|
|
|
#include "rocksdb/slice.h"
|
2014-01-25 01:15:05 +01:00
|
|
|
#include "rocksdb/slice_transform.h"
|
2017-01-26 00:54:09 +01:00
|
|
|
#include "rocksdb/utilities/object_registry.h"
|
2016-03-03 20:20:25 +01:00
|
|
|
#include "rocksdb/utilities/optimistic_transaction_db.h"
|
2016-06-03 01:24:14 +02:00
|
|
|
#include "rocksdb/utilities/options_util.h"
|
2016-05-24 08:35:23 +02:00
|
|
|
#include "rocksdb/utilities/sim_cache.h"
|
Pessimistic Transactions
Summary:
Initial implementation of Pessimistic Transactions. This diff contains the api changes discussed in D38913. This diff is pretty large, so let me know if people would prefer to meet up to discuss it.
MyRocks folks: please take a look at the API in include/rocksdb/utilities/transaction[_db].h and let me know if you have any issues.
Also, you'll notice a couple of TODOs in the implementation of RollbackToSavePoint(). After chatting with Siying, I'm going to send out a separate diff for an alternate implementation of this feature that implements the rollback inside of WriteBatch/WriteBatchWithIndex. We can then decide which route is preferable.
Next, I'm planning on doing some perf testing and then integrating this diff into MongoRocks for further testing.
Test Plan: Unit tests, db_bench parallel testing.
Reviewers: igor, rven, sdong, yhchiang, yoshinorim
Reviewed By: sdong
Subscribers: hermanlee4, maykov, spetrunia, leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D40869
2015-05-26 02:37:33 +02:00
|
|
|
#include "rocksdb/utilities/transaction.h"
|
|
|
|
#include "rocksdb/utilities/transaction_db.h"
|
2016-03-03 20:20:25 +01:00
|
|
|
#include "rocksdb/write_batch.h"
|
2017-07-29 01:23:50 +02:00
|
|
|
#include "util/cast_util.h"
|
2015-01-09 22:04:06 +01:00
|
|
|
#include "util/compression.h"
|
2016-03-03 20:20:25 +01:00
|
|
|
#include "util/crc32c.h"
|
2017-12-01 19:40:45 +01:00
|
|
|
#include "util/gflags_compat.h"
|
2011-08-22 23:08:51 +02:00
|
|
|
#include "util/mutexlock.h"
|
2011-03-18 23:37:00 +01:00
|
|
|
#include "util/random.h"
|
2016-12-10 01:36:43 +01:00
|
|
|
#include "util/stderr_logger.h"
|
2016-03-03 20:20:25 +01:00
|
|
|
#include "util/string_util.h"
|
2011-03-18 23:37:00 +01:00
|
|
|
#include "util/testutil.h"
|
2016-03-03 20:20:25 +01:00
|
|
|
#include "util/transaction_test_util.h"
|
2014-05-01 20:09:32 +02:00
|
|
|
#include "util/xxhash.h"
|
2016-08-09 19:16:32 +02:00
|
|
|
#include "utilities/blob_db/blob_db.h"
|
Benchmarking for Merge Operator
Summary:
Updated db_bench and utilities/merge_operators.h to allow for dynamic benchmarking
of merge operators in db_bench. Added a new test (--benchmarks=mergerandom), which performs
a bunch of random Merge() operations over random keys. Also added a "--merge_operator=" flag
so that the tester can easily benchmark different merge operators. Currently supports
the PutOperator and UInt64Add operator. Support for stringappend or list append may come later.
Test Plan:
1. make db_bench
2. Test the PutOperator (simulating Put) as follows:
./db_bench --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom --merge_operator=put
--threads=2
3. Test the UInt64AddOperator (simulating numeric addition) similarly:
./db_bench --value_size=8 --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom
--merge_operator=uint64add --threads=2
Reviewers: haobo, dhruba, zshao, MarkCallaghan
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11535
2013-08-16 02:13:07 +02:00
|
|
|
#include "utilities/merge_operators.h"
|
2018-03-06 19:20:52 +01:00
|
|
|
#include "utilities/merge_operators/bytesxor.h"
|
2017-02-27 21:59:34 +01:00
|
|
|
#include "utilities/persistent_cache/block_cache_tier.h"
|
2011-03-18 23:37:00 +01:00
|
|
|
|
2015-07-02 01:13:49 +02:00
|
|
|
#ifdef OS_WIN
|
2015-07-13 21:11:05 +02:00
|
|
|
#include <io.h> // open/close
|
2015-07-02 01:13:49 +02:00
|
|
|
#endif
|
|
|
|
|
2017-12-01 19:40:45 +01:00
|
|
|
using GFLAGS_NAMESPACE::ParseCommandLineFlags;
|
|
|
|
using GFLAGS_NAMESPACE::RegisterFlagValidator;
|
|
|
|
using GFLAGS_NAMESPACE::SetUsageMessage;
|
2013-08-23 23:49:57 +02:00
|
|
|
|
2016-10-19 01:30:57 +02:00
|
|
|
DEFINE_string(
|
|
|
|
benchmarks,
|
|
|
|
"fillseq,"
|
|
|
|
"fillseqdeterministic,"
|
|
|
|
"fillsync,"
|
|
|
|
"fillrandom,"
|
|
|
|
"filluniquerandomdeterministic,"
|
|
|
|
"overwrite,"
|
|
|
|
"readrandom,"
|
|
|
|
"newiterator,"
|
|
|
|
"newiteratorwhilewriting,"
|
|
|
|
"seekrandom,"
|
|
|
|
"seekrandomwhilewriting,"
|
|
|
|
"seekrandomwhilemerging,"
|
|
|
|
"readseq,"
|
|
|
|
"readreverse,"
|
|
|
|
"compact,"
|
2017-05-31 01:33:26 +02:00
|
|
|
"compactall,"
|
2016-10-19 01:30:57 +02:00
|
|
|
"multireadrandom,"
|
Generate mixed workload with Get, Put, Seek in db_bench (#4788)
Summary:
Based on the specific workload models (key access distribution, value size distribution, and iterator scan length distribution, the QPS variation), the MixGraph benchmark generate the synthetic workload according to these distributions which can reflect the real-world workload characteristics.
After user enable the tracing function, they will get the trace file. By analyzing the trace file with the trace_analyzer tool, user can generate a set of statistic data files including. The *_accessed_key_stats.txt, *-accessed_value_size_distribution.txt, *-iterator_length_distribution.txt, and *-qps_stats.txt are mainly used to fit the Matlab model fitting. After that, user can get the parameters of the workload distributions (the modeling details are described: [here](https://github.com/facebook/rocksdb/wiki/RocksDB-Trace%2C-Replay%2C-and-Analyzer))
The key access distribution follows the The two-term power model. The probability density function is: `f(x) = ax^{b}+c`. The corresponding parameters are key_dist_a, key_dist_b, and key_dist_c in db_bench
For the value size distribution and iterator scan length distribution, they both follow the Generalized Pareto Distribution. The probability density function is `f(x) = (1/sigma)(1+k*(x-theta)/sigma))^{-1-1/k)`. The parameters are: value_k, value_theta, value_sigma and iter_k, iter_theta, iter_sigma. For more information about the Generalized Pareto Distribution, users can find the [wiki](https://en.wikipedia.org/wiki/Generalized_Pareto_distribution) and [Matalb page](https://www.mathworks.com/help/stats/generalized-pareto-distribution.html)
As for the QPS, it follows the diurnal pattern. So Sine is a good model to fit it. `F(x) = sine_a*sin(sine_b*x + sine_c) + sine_d`. The trace_will tell you the average QPS in the print out resutls, which is sine_d. After user fit the "*-qps_stats.txt" to the Matlab model, user can get the sine_a, sine_b, and sine_c. By using the 4 parameters, user can control the QPS variation including the period, average, changes.
To use the bench mark, user can indicate the following parameters as examples:
```
-benchmarks="mixgraph" -key_dist_a=0.002312 -key_dist_b=0.3467 -value_k=0.9233 -value_sigma=226.4092 -iter_k=2.517 -iter_sigma=14.236 -mix_get_ratio=0.7 -mix_put_ratio=0.25 -mix_seek_ratio=0.05 -sine_mix_rate_interval_milliseconds=500 -sine_a=15000 -sine_b=1 -sine_d=20000
```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4788
Differential Revision: D13573940
Pulled By: sagar0
fbshipit-source-id: e184c27e07b4f1bc0b436c2be36c5090c1fb0222
2019-01-22 19:40:44 +01:00
|
|
|
"mixgraph,"
|
2016-10-19 01:30:57 +02:00
|
|
|
"readseq,"
|
|
|
|
"readtocache,"
|
|
|
|
"readreverse,"
|
|
|
|
"readwhilewriting,"
|
|
|
|
"readwhilemerging,"
|
2018-05-25 00:45:49 +02:00
|
|
|
"readwhilescanning,"
|
2016-10-19 01:30:57 +02:00
|
|
|
"readrandomwriterandom,"
|
|
|
|
"updaterandom,"
|
2018-03-06 19:20:52 +01:00
|
|
|
"xorupdaterandom,"
|
2016-10-19 01:30:57 +02:00
|
|
|
"randomwithverify,"
|
|
|
|
"fill100K,"
|
|
|
|
"crc32c,"
|
|
|
|
"xxhash,"
|
|
|
|
"compress,"
|
|
|
|
"uncompress,"
|
|
|
|
"acquireload,"
|
|
|
|
"fillseekseq,"
|
|
|
|
"randomtransaction,"
|
|
|
|
"randomreplacekeys,"
|
|
|
|
"timeseries",
|
|
|
|
|
|
|
|
"Comma-separated list of operations to run in the specified"
|
|
|
|
" order. Available benchmarks:\n"
|
|
|
|
"\tfillseq -- write N values in sequential key"
|
|
|
|
" order in async mode\n"
|
|
|
|
"\tfillseqdeterministic -- write N values in the specified"
|
|
|
|
" key order and keep the shape of the LSM tree\n"
|
|
|
|
"\tfillrandom -- write N values in random key order in async"
|
|
|
|
" mode\n"
|
|
|
|
"\tfilluniquerandomdeterministic -- write N values in a random"
|
|
|
|
" key order and keep the shape of the LSM tree\n"
|
|
|
|
"\toverwrite -- overwrite N values in random key order in"
|
|
|
|
" async mode\n"
|
|
|
|
"\tfillsync -- write N/100 values in random key order in "
|
|
|
|
"sync mode\n"
|
|
|
|
"\tfill100K -- write N/1000 100K values in random order in"
|
|
|
|
" async mode\n"
|
|
|
|
"\tdeleteseq -- delete N keys in sequential order\n"
|
|
|
|
"\tdeleterandom -- delete N keys in random order\n"
|
|
|
|
"\treadseq -- read N times sequentially\n"
|
|
|
|
"\treadtocache -- 1 thread reading database sequentially\n"
|
|
|
|
"\treadreverse -- read N times in reverse order\n"
|
|
|
|
"\treadrandom -- read N times in random order\n"
|
|
|
|
"\treadmissing -- read N missing keys in random order\n"
|
|
|
|
"\treadwhilewriting -- 1 writer, N threads doing random "
|
|
|
|
"reads\n"
|
|
|
|
"\treadwhilemerging -- 1 merger, N threads doing random "
|
|
|
|
"reads\n"
|
2018-05-25 00:45:49 +02:00
|
|
|
"\treadwhilescanning -- 1 thread doing full table scan, "
|
|
|
|
"N threads doing random reads\n"
|
2016-10-19 01:30:57 +02:00
|
|
|
"\treadrandomwriterandom -- N threads doing random-read, "
|
|
|
|
"random-write\n"
|
|
|
|
"\tupdaterandom -- N threads doing read-modify-write for random "
|
|
|
|
"keys\n"
|
2018-03-06 19:20:52 +01:00
|
|
|
"\txorupdaterandom -- N threads doing read-XOR-write for "
|
|
|
|
"random keys\n"
|
2016-10-19 01:30:57 +02:00
|
|
|
"\tappendrandom -- N threads doing read-modify-write with "
|
|
|
|
"growing values\n"
|
|
|
|
"\tmergerandom -- same as updaterandom/appendrandom using merge"
|
|
|
|
" operator. "
|
|
|
|
"Must be used with merge_operator\n"
|
|
|
|
"\treadrandommergerandom -- perform N random read-or-merge "
|
|
|
|
"operations. Must be used with merge_operator\n"
|
|
|
|
"\tnewiterator -- repeated iterator creation\n"
|
|
|
|
"\tseekrandom -- N random seeks, call Next seek_nexts times "
|
|
|
|
"per seek\n"
|
|
|
|
"\tseekrandomwhilewriting -- seekrandom and 1 thread doing "
|
|
|
|
"overwrite\n"
|
|
|
|
"\tseekrandomwhilemerging -- seekrandom and 1 thread doing "
|
|
|
|
"merge\n"
|
|
|
|
"\tcrc32c -- repeated crc32c of 4K of data\n"
|
|
|
|
"\txxhash -- repeated xxHash of 4K of data\n"
|
|
|
|
"\tacquireload -- load N*1000 times\n"
|
|
|
|
"\tfillseekseq -- write N values in sequential key, then read "
|
|
|
|
"them by seeking to each key\n"
|
|
|
|
"\trandomtransaction -- execute N random transactions and "
|
|
|
|
"verify correctness\n"
|
|
|
|
"\trandomreplacekeys -- randomly replaces N keys by deleting "
|
|
|
|
"the old version and putting the new version\n\n"
|
|
|
|
"\ttimeseries -- 1 writer generates time series data "
|
|
|
|
"and multiple readers doing random reads on id\n\n"
|
|
|
|
"Meta operations:\n"
|
2017-06-01 00:33:02 +02:00
|
|
|
"\tcompact -- Compact the entire DB; If multiple, randomly choose one\n"
|
|
|
|
"\tcompactall -- Compact the entire DB\n"
|
2016-10-19 01:30:57 +02:00
|
|
|
"\tstats -- Print DB stats\n"
|
2017-04-19 01:30:51 +02:00
|
|
|
"\tresetstats -- Reset DB stats\n"
|
2016-10-19 01:30:57 +02:00
|
|
|
"\tlevelstats -- Print the number of files and bytes per level\n"
|
|
|
|
"\tsstables -- Print sstable info\n"
|
2018-08-01 09:14:43 +02:00
|
|
|
"\theapprofile -- Dump a heap profile (if supported by this port)\n"
|
|
|
|
"\treplay -- replay the trace file specified with trace_file\n");
|
2013-10-24 16:43:14 +02:00
|
|
|
|
|
|
|
DEFINE_int64(num, 1000000, "Number of key/values to place in database");
|
|
|
|
|
|
|
|
DEFINE_int64(numdistinct, 1000,
|
|
|
|
"Number of distinct keys to use. Used in RandomWithVerify to "
|
|
|
|
"read/write on fewer keys so that gets are more likely to find the"
|
|
|
|
" key and puts are more likely to update the same key");
|
|
|
|
|
2014-01-11 02:33:56 +01:00
|
|
|
DEFINE_int64(merge_keys, -1,
|
|
|
|
"Number of distinct keys to use for MergeRandom and "
|
|
|
|
"ReadRandomMergeRandom. "
|
|
|
|
"If negative, there will be FLAGS_num keys.");
|
2014-08-19 03:15:01 +02:00
|
|
|
DEFINE_int32(num_column_families, 1, "Number of Column Families to use.");
|
2014-01-11 02:33:56 +01:00
|
|
|
|
2014-12-23 03:39:28 +01:00
|
|
|
DEFINE_int32(
|
2014-12-24 17:54:28 +01:00
|
|
|
num_hot_column_families, 0,
|
2014-12-23 03:39:28 +01:00
|
|
|
"Number of Hot Column Families. If more than 0, only write to this "
|
|
|
|
"number of column families. After finishing all the writes to them, "
|
|
|
|
"create new set of column families and insert to them. Only used "
|
|
|
|
"when num_column_families > 1.");
|
|
|
|
|
2017-08-11 22:46:41 +02:00
|
|
|
DEFINE_string(column_family_distribution, "",
|
|
|
|
"Comma-separated list of percentages, where the ith element "
|
|
|
|
"indicates the probability of an op using the ith column family. "
|
|
|
|
"The number of elements must be `num_hot_column_families` if "
|
|
|
|
"specified; otherwise, it must be `num_column_families`. The "
|
|
|
|
"sum of elements must be 100. E.g., if `num_column_families=4`, "
|
|
|
|
"and `num_hot_column_families=0`, a valid list could be "
|
|
|
|
"\"10,20,30,40\".");
|
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_int64(reads, -1, "Number of read operations to do. "
|
|
|
|
"If negative, do FLAGS_num reads.");
|
|
|
|
|
2016-06-28 23:10:24 +02:00
|
|
|
DEFINE_int64(deletes, -1, "Number of delete operations to do. "
|
|
|
|
"If negative, do FLAGS_num deletions.");
|
|
|
|
|
2014-03-28 17:21:20 +01:00
|
|
|
DEFINE_int32(bloom_locality, 0, "Control bloom filter probes locality");
|
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_int64(seed, 0, "Seed base for random number generators. "
|
|
|
|
"When 0 it is deterministic.");
|
|
|
|
|
|
|
|
DEFINE_int32(threads, 1, "Number of concurrent threads to run.");
|
|
|
|
|
|
|
|
DEFINE_int32(duration, 0, "Time in seconds for the random-ops tests to run."
|
|
|
|
" When 0 then num & reads determine the test duration");
|
|
|
|
|
|
|
|
DEFINE_int32(value_size, 100, "Size of each value");
|
2013-08-23 23:49:57 +02:00
|
|
|
|
2014-10-17 19:20:25 +02:00
|
|
|
DEFINE_int32(seek_nexts, 0,
|
|
|
|
"How many times to call Next() after Seek() in "
|
2015-03-30 20:28:25 +02:00
|
|
|
"fillseekseq, seekrandom, seekrandomwhilewriting and "
|
|
|
|
"seekrandomwhilemerging");
|
SkipListRep::LookaheadIterator
Summary:
This diff introduces the `lookahead` argument to `SkipListFactory()`. This is an
optimization for the tailing use case which includes many seeks. E.g. consider
the following operations on a skip list iterator:
Seek(x), Next(), Next(), Seek(x+2), Next(), Seek(x+3), Next(), Next(), ...
If `lookahead` is positive, `SkipListRep` will return an iterator which also
keeps track of the previously visited node. Seek() then first does a linear
search starting from that node (up to `lookahead` steps). As in the tailing
example above, this may require fewer than ~log(n) comparisons as with regular
skip list search.
Test Plan:
Added a new benchmark (`fillseekseq`) which simulates the usage pattern. It
first writes N records (with consecutive keys), then measures how much time it
takes to read them by calling `Seek()` and `Next()`.
$ time ./db_bench -num 10000000 -benchmarks fillseekseq -prefix_size 1 \
-key_size 8 -write_buffer_size $[1024*1024*1024] -value_size 50 \
-seekseq_next 2 -skip_list_lookahead=0
[...]
DB path: [/dev/shm/rocksdbtest/dbbench]
fillseekseq : 0.389 micros/op 2569047 ops/sec;
real 0m21.806s
user 0m12.106s
sys 0m9.672s
$ time ./db_bench [...] -skip_list_lookahead=2
[...]
DB path: [/dev/shm/rocksdbtest/dbbench]
fillseekseq : 0.153 micros/op 6540684 ops/sec;
real 0m19.469s
user 0m10.192s
sys 0m9.252s
Reviewers: ljin, sdong, igor
Reviewed By: igor
Subscribers: dhruba, leveldb, march, lovro
Differential Revision: https://reviews.facebook.net/D23997
2014-09-24 00:52:28 +02:00
|
|
|
|
2015-03-18 21:50:52 +01:00
|
|
|
DEFINE_bool(reverse_iterator, false,
|
|
|
|
"When true use Prev rather than Next for iterators that do "
|
|
|
|
"Seek and then Next");
|
|
|
|
|
2018-11-14 19:39:54 +01:00
|
|
|
DEFINE_int64(max_scan_distance, 0,
|
|
|
|
"Used to define iterate_upper_bound (or iterate_lower_bound "
|
|
|
|
"if FLAGS_reverse_iterator is set to true) when value is nonzero");
|
|
|
|
|
2014-08-27 19:39:31 +02:00
|
|
|
DEFINE_bool(use_uint64_comparator, false, "use Uint64 user comparator");
|
2013-01-03 21:11:50 +01:00
|
|
|
|
2014-09-25 22:34:51 +02:00
|
|
|
DEFINE_int64(batch_size, 1, "Batch size");
|
|
|
|
|
2018-03-05 22:08:17 +01:00
|
|
|
static bool ValidateKeySize(const char* /*flagname*/, int32_t /*value*/) {
|
2013-10-24 16:43:14 +02:00
|
|
|
return true;
|
|
|
|
}
|
2014-04-09 20:25:21 +02:00
|
|
|
|
2015-08-21 23:25:34 +02:00
|
|
|
static bool ValidateUint32Range(const char* flagname, uint64_t value) {
|
|
|
|
if (value > std::numeric_limits<uint32_t>::max()) {
|
Fixing race condition in DBTest.DynamicMemtableOptions
Summary:
This patch fixes a race condition in DBTEst.DynamicMemtableOptions. In rare cases,
it was possible that the main thread would fill up both memtables before the flush
job acquired its work. Then, the flush job was flushing both memtables together,
producing only one L0 file while the test expected two. Now, the test waits for
flushes to finish earlier, to make sure that the memtables are flushed in separate
flush jobs.
Test Plan:
Insert "usleep(10000);" after "IOSTATS_SET_THREAD_POOL_ID(Env::Priority::HIGH);" in BGWorkFlush()
to make the issue more likely. Then test with:
make db_test && time while ./db_test --gtest_filter=*DynamicMemtableOptions; do true; done
Reviewers: rven, sdong, yhchiang, anthony, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D45429
2015-08-25 02:04:18 +02:00
|
|
|
fprintf(stderr, "Invalid value for --%s: %lu, overflow\n", flagname,
|
2015-08-21 23:25:34 +02:00
|
|
|
(unsigned long)value);
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_int32(key_size, 16, "size of each key");
|
2013-03-01 23:10:09 +01:00
|
|
|
|
2014-04-11 21:15:09 +02:00
|
|
|
DEFINE_int32(num_multi_db, 0,
|
|
|
|
"Number of DBs used in the benchmark. 0 means single DB.");
|
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_double(compression_ratio, 0.5, "Arrange to generate values that shrink"
|
|
|
|
" to this fraction of their original size after compression");
|
2011-03-18 23:37:00 +01:00
|
|
|
|
2015-03-07 01:38:50 +01:00
|
|
|
DEFINE_double(read_random_exp_range, 0.0,
|
|
|
|
"Read random's key will be generated using distribution of "
|
2015-07-15 19:21:09 +02:00
|
|
|
"num * exp(-r) where r is uniform number from 0 to this value. "
|
2015-03-07 01:38:50 +01:00
|
|
|
"The larger the number is, the more skewed the reads are. "
|
|
|
|
"Only used in readrandom and multireadrandom benchmarks.");
|
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_bool(histogram, false, "Print histogram of operation timings");
|
2011-03-18 23:37:00 +01:00
|
|
|
|
2014-07-07 19:53:31 +02:00
|
|
|
DEFINE_bool(enable_numa, false,
|
|
|
|
"Make operations aware of NUMA architecture and bind memory "
|
|
|
|
"and cpus corresponding to nodes together. In NUMA, memory "
|
|
|
|
"in same node as CPUs are closer when compared to memory in "
|
|
|
|
"other nodes. Reads can be faster when the process is bound to "
|
|
|
|
"CPU and memory of same node. Use \"$numactl --hardware\" command "
|
|
|
|
"to see NUMA memory architecture.");
|
|
|
|
|
2014-12-02 21:09:20 +01:00
|
|
|
DEFINE_int64(db_write_buffer_size, rocksdb::Options().db_write_buffer_size,
|
|
|
|
"Number of bytes to buffer in all memtables before compacting");
|
|
|
|
|
2017-06-02 23:13:59 +02:00
|
|
|
DEFINE_bool(cost_write_buffer_to_cache, false,
|
|
|
|
"The usage of memtable is costed to the block cache");
|
|
|
|
|
Add monitoring for universal compaction and add counters for compaction IO
Summary:
Adds these counters
{ WAL_FILE_SYNCED, "rocksdb.wal.synced" }
number of writes that request a WAL sync
{ WAL_FILE_BYTES, "rocksdb.wal.bytes" },
number of bytes written to the WAL
{ WRITE_DONE_BY_SELF, "rocksdb.write.self" },
number of writes processed by the calling thread
{ WRITE_DONE_BY_OTHER, "rocksdb.write.other" },
number of writes not processed by the calling thread. Instead these were
processed by the current holder of the write lock
{ WRITE_WITH_WAL, "rocksdb.write.wal" },
number of writes that request WAL logging
{ COMPACT_READ_BYTES, "rocksdb.compact.read.bytes" },
number of bytes read during compaction
{ COMPACT_WRITE_BYTES, "rocksdb.compact.write.bytes" },
number of bytes written during compaction
Per-interval stats output was updated with WAL stats and correct stats for universal compaction
including a correct value for write-amplification. It now looks like:
Compactions
Level Files Size(MB) Score Time(sec) Read(MB) Write(MB) Rn(MB) Rnp1(MB) Wnew(MB) RW-Amplify Read(MB/s) Write(MB/s) Rn Rnp1 Wnp1 NewW Count Ln-stall Stall-cnt
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
0 7 464 46.4 281 3411 3875 3411 0 3875 2.1 12.1 13.8 621 0 240 240 628 0.0 0
Uptime(secs): 310.8 total, 2.0 interval
Writes cumulative: 9999999 total, 9999999 batches, 1.0 per batch, 1.22 ingest GB
WAL cumulative: 9999999 WAL writes, 9999999 WAL syncs, 1.00 writes per sync, 1.22 GB written
Compaction IO cumulative (GB): 1.22 new, 3.33 read, 3.78 write, 7.12 read+write
Compaction IO cumulative (MB/sec): 4.0 new, 11.0 read, 12.5 write, 23.4 read+write
Amplification cumulative: 4.1 write, 6.8 compaction
Writes interval: 100000 total, 100000 batches, 1.0 per batch, 12.5 ingest MB
WAL interval: 100000 WAL writes, 100000 WAL syncs, 1.00 writes per sync, 0.01 MB written
Compaction IO interval (MB): 12.49 new, 14.98 read, 21.50 write, 36.48 read+write
Compaction IO interval (MB/sec): 6.4 new, 7.6 read, 11.0 write, 18.6 read+write
Amplification interval: 101.7 write, 102.9 compaction
Stalls(secs): 142.924 level0_slowdown, 0.000 level0_numfiles, 0.805 memtable_compaction, 0.000 leveln_slowdown
Stalls(count): 132461 level0_slowdown, 0 level0_numfiles, 3 memtable_compaction, 0 leveln_slowdown
Task ID: #3329644, #3301695
Blame Rev:
Test Plan:
Revert Plan:
Database Impact:
Memcache Impact:
Other Notes:
EImportant:
- begin *PUBLIC* platform impact section -
Bugzilla: #
- end platform impact -
Reviewers: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D14583
2013-12-09 22:43:34 +01:00
|
|
|
DEFINE_int64(write_buffer_size, rocksdb::Options().write_buffer_size,
|
2013-10-24 16:43:14 +02:00
|
|
|
"Number of bytes to buffer in memtable before compacting");
|
2011-04-12 21:38:58 +02:00
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_int32(max_write_buffer_number,
|
|
|
|
rocksdb::Options().max_write_buffer_number,
|
|
|
|
"The number of in-memory memtables. Each memtable is of size"
|
2018-04-17 18:41:57 +02:00
|
|
|
" write_buffer_size bytes.");
|
2012-10-19 23:00:53 +02:00
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_int32(min_write_buffer_number_to_merge,
|
|
|
|
rocksdb::Options().min_write_buffer_number_to_merge,
|
|
|
|
"The minimum number of write buffers that will be merged together"
|
|
|
|
"before writing to storage. This is cheap because it is an"
|
|
|
|
"in-memory merge. If this feature is not enabled, then all these"
|
|
|
|
"write buffers are flushed to L0 as separate files and this "
|
|
|
|
"increases read amplification because a get request has to check"
|
|
|
|
" in all of these files. Also, an in-memory merge may result in"
|
|
|
|
" writing less data to storage if there are duplicate records "
|
|
|
|
" in each of these individual write buffers.");
|
2013-06-11 23:23:58 +02:00
|
|
|
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
DEFINE_int32(max_write_buffer_number_to_maintain,
|
|
|
|
rocksdb::Options().max_write_buffer_number_to_maintain,
|
|
|
|
"The total maximum number of write buffers to maintain in memory "
|
|
|
|
"including copies of buffers that have already been flushed. "
|
|
|
|
"Unlike max_write_buffer_number, this parameter does not affect "
|
|
|
|
"flushing. This controls the minimum amount of write history "
|
|
|
|
"that will be available in memory for conflict checking when "
|
|
|
|
"Transactions are used. If this value is too low, some "
|
|
|
|
"transactions may fail at commit time due to not being able to "
|
|
|
|
"determine whether there were any write conflicts. Setting this "
|
|
|
|
"value to 0 will cause write buffers to be freed immediately "
|
|
|
|
"after they are flushed. If this value is set to -1, "
|
|
|
|
"'max_write_buffer_number' will be used.");
|
|
|
|
|
2017-07-08 03:11:53 +02:00
|
|
|
DEFINE_int32(max_background_jobs,
|
|
|
|
rocksdb::Options().max_background_jobs,
|
|
|
|
"The maximum number of concurrent background jobs that can occur "
|
|
|
|
"in parallel.");
|
|
|
|
|
2017-08-04 00:36:28 +02:00
|
|
|
DEFINE_int32(num_bottom_pri_threads, 0,
|
|
|
|
"The number of threads in the bottom-priority thread pool (used "
|
|
|
|
"by universal compaction only).");
|
|
|
|
|
2017-08-04 06:30:37 +02:00
|
|
|
DEFINE_int32(num_high_pri_threads, 0,
|
|
|
|
"The maximum number of concurrent background compactions"
|
|
|
|
" that can occur in parallel.");
|
|
|
|
|
|
|
|
DEFINE_int32(num_low_pri_threads, 0,
|
|
|
|
"The maximum number of concurrent background compactions"
|
|
|
|
" that can occur in parallel.");
|
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_int32(max_background_compactions,
|
|
|
|
rocksdb::Options().max_background_compactions,
|
|
|
|
"The maximum number of concurrent background compactions"
|
|
|
|
" that can occur in parallel.");
|
2012-10-19 23:00:53 +02:00
|
|
|
|
2017-05-24 20:25:38 +02:00
|
|
|
DEFINE_int32(base_background_compactions, -1, "DEPRECATED");
|
2016-08-04 01:02:41 +02:00
|
|
|
|
2015-08-21 23:25:34 +02:00
|
|
|
DEFINE_uint64(subcompactions, 1,
|
Fixing race condition in DBTest.DynamicMemtableOptions
Summary:
This patch fixes a race condition in DBTEst.DynamicMemtableOptions. In rare cases,
it was possible that the main thread would fill up both memtables before the flush
job acquired its work. Then, the flush job was flushing both memtables together,
producing only one L0 file while the test expected two. Now, the test waits for
flushes to finish earlier, to make sure that the memtables are flushed in separate
flush jobs.
Test Plan:
Insert "usleep(10000);" after "IOSTATS_SET_THREAD_POOL_ID(Env::Priority::HIGH);" in BGWorkFlush()
to make the issue more likely. Then test with:
make db_test && time while ./db_test --gtest_filter=*DynamicMemtableOptions; do true; done
Reviewers: rven, sdong, yhchiang, anthony, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D45429
2015-08-25 02:04:18 +02:00
|
|
|
"Maximum number of subcompactions to divide L0-L1 compactions "
|
|
|
|
"into.");
|
|
|
|
static const bool FLAGS_subcompactions_dummy
|
2018-02-02 21:14:42 +01:00
|
|
|
__attribute__((__unused__)) = RegisterFlagValidator(&FLAGS_subcompactions,
|
Fixing race condition in DBTest.DynamicMemtableOptions
Summary:
This patch fixes a race condition in DBTEst.DynamicMemtableOptions. In rare cases,
it was possible that the main thread would fill up both memtables before the flush
job acquired its work. Then, the flush job was flushing both memtables together,
producing only one L0 file while the test expected two. Now, the test waits for
flushes to finish earlier, to make sure that the memtables are flushed in separate
flush jobs.
Test Plan:
Insert "usleep(10000);" after "IOSTATS_SET_THREAD_POOL_ID(Env::Priority::HIGH);" in BGWorkFlush()
to make the issue more likely. Then test with:
make db_test && time while ./db_test --gtest_filter=*DynamicMemtableOptions; do true; done
Reviewers: rven, sdong, yhchiang, anthony, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D45429
2015-08-25 02:04:18 +02:00
|
|
|
&ValidateUint32Range);
|
2015-08-21 23:25:34 +02:00
|
|
|
|
2014-03-05 19:27:17 +01:00
|
|
|
DEFINE_int32(max_background_flushes,
|
|
|
|
rocksdb::Options().max_background_flushes,
|
|
|
|
"The maximum number of concurrent background flushes"
|
|
|
|
" that can occur in parallel.");
|
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
static rocksdb::CompactionStyle FLAGS_compaction_style_e;
|
|
|
|
DEFINE_int32(compaction_style, (int32_t) rocksdb::Options().compaction_style,
|
2017-03-25 00:58:51 +01:00
|
|
|
"style of compaction: level-based, universal and fifo");
|
2013-06-14 07:09:08 +02:00
|
|
|
|
2015-09-22 02:16:31 +02:00
|
|
|
static rocksdb::CompactionPri FLAGS_compaction_pri_e;
|
2015-10-30 22:42:18 +01:00
|
|
|
DEFINE_int32(compaction_pri, (int32_t)rocksdb::Options().compaction_pri,
|
2015-09-22 02:16:31 +02:00
|
|
|
"priority of files to compaction: by size or by data age");
|
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_int32(universal_size_ratio, 0,
|
|
|
|
"Percentage flexibility while comparing file size"
|
|
|
|
" (for universal compaction only).");
|
2013-06-14 07:09:08 +02:00
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_int32(universal_min_merge_width, 0, "The minimum number of files in a"
|
|
|
|
" single compaction run (for universal compaction only).");
|
2013-09-10 01:06:10 +02:00
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_int32(universal_max_merge_width, 0, "The max number of files to compact"
|
|
|
|
" in universal style compaction");
|
2013-09-10 01:06:10 +02:00
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_int32(universal_max_size_amplification_percent, 0,
|
|
|
|
"The max size amplification for universal style compaction");
|
2013-06-14 07:09:08 +02:00
|
|
|
|
2013-12-03 21:32:07 +01:00
|
|
|
DEFINE_int32(universal_compression_size_percent, -1,
|
|
|
|
"The percentage of the database to compress for universal "
|
|
|
|
"compaction. -1 means compress everything.");
|
|
|
|
|
2015-07-20 20:52:39 +02:00
|
|
|
DEFINE_bool(universal_allow_trivial_move, false,
|
2015-08-11 20:46:15 +02:00
|
|
|
"Allow trivial move in universal compaction.");
|
2015-07-20 20:52:39 +02:00
|
|
|
|
2016-08-19 21:28:19 +02:00
|
|
|
DEFINE_int64(cache_size, 8 << 20, // 8MB
|
|
|
|
"Number of bytes to use as a cache of uncompressed data");
|
|
|
|
|
|
|
|
DEFINE_int32(cache_numshardbits, 6,
|
|
|
|
"Number of shards for the block cache"
|
|
|
|
" is 2 ** cache_numshardbits. Negative means use default settings."
|
|
|
|
" This is applied only if FLAGS_cache_size is non-negative.");
|
|
|
|
|
2017-01-04 03:24:15 +01:00
|
|
|
DEFINE_double(cache_high_pri_pool_ratio, 0.0,
|
|
|
|
"Ratio of block cache reserve for high pri blocks. "
|
|
|
|
"If > 0.0, we also enable "
|
|
|
|
"cache_index_and_filter_blocks_with_high_priority.");
|
|
|
|
|
2016-08-19 21:28:19 +02:00
|
|
|
DEFINE_bool(use_clock_cache, false,
|
|
|
|
"Replace default LRU block cache with clock cache.");
|
2016-05-24 08:35:23 +02:00
|
|
|
|
|
|
|
DEFINE_int64(simcache_size, -1,
|
|
|
|
"Number of bytes to use as a simcache of "
|
2016-08-19 21:28:19 +02:00
|
|
|
"uncompressed data. Nagative value disables simcache.");
|
2011-03-18 23:37:00 +01:00
|
|
|
|
2015-07-09 22:36:16 +02:00
|
|
|
DEFINE_bool(cache_index_and_filter_blocks, false,
|
|
|
|
"Cache index/filter blocks in block cache.");
|
|
|
|
|
2017-06-16 01:06:22 +02:00
|
|
|
DEFINE_bool(partition_index_and_filters, false,
|
|
|
|
"Partition index and filter blocks.");
|
|
|
|
|
2018-02-12 23:54:50 +01:00
|
|
|
DEFINE_bool(partition_index, false, "Partition index blocks");
|
|
|
|
|
2017-06-16 01:06:22 +02:00
|
|
|
DEFINE_int64(metadata_block_size,
|
|
|
|
rocksdb::BlockBasedTableOptions().metadata_block_size,
|
|
|
|
"Max partition size when partitioning index/filters");
|
|
|
|
|
|
|
|
// The default reduces the overhead of reading time with flash. With HDD, which
|
|
|
|
// offers much less throughput, however, this number better to be set to 1.
|
|
|
|
DEFINE_int32(ops_between_duration_checks, 1000,
|
|
|
|
"Check duration limit every x ops");
|
|
|
|
|
Adding pin_l0_filter_and_index_blocks_in_cache feature and related fixes.
Summary:
When a block based table file is opened, if prefetch_index_and_filter is true, it will prefetch the index and filter blocks, putting them into the block cache.
What this feature adds: when a L0 block based table file is opened, if pin_l0_filter_and_index_blocks_in_cache is true in the options (and prefetch_index_and_filter is true), then the filter and index blocks aren't released back to the block cache at the end of BlockBasedTableReader::Open(). Instead the table reader takes ownership of them, hence pinning them, ie. the LRU cache will never push them out. Meanwhile in the table reader, further accesses will not hit the block cache, thus avoiding lock contention.
Test Plan:
'export TEST_TMPDIR=/dev/shm/ && DISABLE_JEMALLOC=1 OPT=-g make all valgrind_check -j32' is OK.
I didn't run the Java tests, I don't have Java set up on my devserver.
Reviewers: sdong
Reviewed By: sdong
Subscribers: andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D56133
2016-04-01 19:42:39 +02:00
|
|
|
DEFINE_bool(pin_l0_filter_and_index_blocks_in_cache, false,
|
|
|
|
"Pin index/filter blocks of L0 files in block cache.");
|
|
|
|
|
2018-06-23 00:14:05 +02:00
|
|
|
DEFINE_bool(
|
|
|
|
pin_top_level_index_and_filter, false,
|
|
|
|
"Pin top-level index of partitioned index/filter blocks in block cache.");
|
|
|
|
|
2014-11-11 22:47:22 +01:00
|
|
|
DEFINE_int32(block_size,
|
|
|
|
static_cast<int32_t>(rocksdb::BlockBasedTableOptions().block_size),
|
2013-10-24 16:43:14 +02:00
|
|
|
"Number of bytes in a block.");
|
2012-09-13 18:23:18 +02:00
|
|
|
|
2018-06-05 04:59:44 +02:00
|
|
|
DEFINE_int32(
|
|
|
|
format_version,
|
|
|
|
static_cast<int32_t>(rocksdb::BlockBasedTableOptions().format_version),
|
|
|
|
"Format version of SST files.");
|
|
|
|
|
2014-08-25 23:22:05 +02:00
|
|
|
DEFINE_int32(block_restart_interval,
|
|
|
|
rocksdb::BlockBasedTableOptions().block_restart_interval,
|
2014-07-21 21:01:40 +02:00
|
|
|
"Number of keys between restart points "
|
2016-05-09 21:09:05 +02:00
|
|
|
"for delta encoding of keys in data block.");
|
|
|
|
|
|
|
|
DEFINE_int32(index_block_restart_interval,
|
|
|
|
rocksdb::BlockBasedTableOptions().index_block_restart_interval,
|
|
|
|
"Number of keys between restart points "
|
|
|
|
"for delta encoding of keys in index block.");
|
2014-07-21 21:01:40 +02:00
|
|
|
|
2016-08-27 03:55:58 +02:00
|
|
|
DEFINE_int32(read_amp_bytes_per_bit,
|
|
|
|
rocksdb::BlockBasedTableOptions().read_amp_bytes_per_bit,
|
|
|
|
"Number of bytes per bit to be used in block read-amp bitmap");
|
|
|
|
|
2018-01-11 00:06:29 +01:00
|
|
|
DEFINE_bool(enable_index_compression,
|
|
|
|
rocksdb::BlockBasedTableOptions().enable_index_compression,
|
|
|
|
"Compress the index block");
|
|
|
|
|
2018-03-27 05:14:24 +02:00
|
|
|
DEFINE_bool(block_align, rocksdb::BlockBasedTableOptions().block_align,
|
|
|
|
"Align data blocks on page size");
|
|
|
|
|
2018-08-17 03:29:13 +02:00
|
|
|
DEFINE_bool(use_data_block_hash_index, false,
|
|
|
|
"if use kDataBlockBinaryAndHash "
|
|
|
|
"instead of kDataBlockBinarySearch. "
|
|
|
|
"This is valid if only we use BlockTable");
|
|
|
|
|
|
|
|
DEFINE_double(data_block_hash_table_util_ratio, 0.75,
|
|
|
|
"util ratio for data block hash index table. "
|
|
|
|
"This is only valid if use_data_block_hash_index is "
|
|
|
|
"set to true");
|
|
|
|
|
2013-09-02 08:23:40 +02:00
|
|
|
DEFINE_int64(compressed_cache_size, -1,
|
|
|
|
"Number of bytes to use as a cache of compressed data.");
|
|
|
|
|
2015-06-23 19:25:45 +02:00
|
|
|
DEFINE_int64(row_cache_size, 0,
|
|
|
|
"Number of bytes to use as a cache of individual rows"
|
|
|
|
" (0 = disabled).");
|
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_int32(open_files, rocksdb::Options().max_open_files,
|
|
|
|
"Maximum number of files to keep open at the same time"
|
|
|
|
" (use default if == 0)");
|
2011-05-21 04:17:43 +02:00
|
|
|
|
2015-09-30 18:51:31 +02:00
|
|
|
DEFINE_int32(file_opening_threads, rocksdb::Options().max_file_opening_threads,
|
|
|
|
"If open_files is set to -1, this option set the number of "
|
|
|
|
"threads that will be used to open files during DB::Open()");
|
|
|
|
|
2017-07-27 21:10:41 +02:00
|
|
|
DEFINE_bool(new_table_reader_for_compaction_inputs, true,
|
2015-08-27 00:25:59 +02:00
|
|
|
"If true, uses a separate file handle for compaction inputs");
|
|
|
|
|
|
|
|
DEFINE_int32(compaction_readahead_size, 0, "Compaction readahead size");
|
|
|
|
|
2015-10-29 19:34:34 +01:00
|
|
|
DEFINE_int32(random_access_max_buffer_size, 1024 * 1024,
|
|
|
|
"Maximum windows randomaccess buffer size");
|
2015-10-27 22:44:16 +01:00
|
|
|
|
2015-11-16 21:56:21 +01:00
|
|
|
DEFINE_int32(writable_file_max_buffer_size, 1024 * 1024,
|
|
|
|
"Maximum write buffer for Writable File");
|
2015-10-30 06:10:25 +01:00
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_int32(bloom_bits, -1, "Bloom filter bits per key. Negative means"
|
|
|
|
" use default settings.");
|
2016-06-04 02:02:10 +02:00
|
|
|
DEFINE_double(memtable_bloom_size_ratio, 0,
|
|
|
|
"Ratio of memtable size used for bloom filter. 0 means no bloom "
|
|
|
|
"filter.");
|
2019-02-19 21:12:25 +01:00
|
|
|
DEFINE_bool(memtable_whole_key_filtering, false,
|
|
|
|
"Try to use whole key bloom filter in memtables.");
|
2016-07-27 03:05:30 +02:00
|
|
|
DEFINE_bool(memtable_use_huge_page, false,
|
|
|
|
"Try to use huge page in memtables.");
|
2012-04-17 17:36:46 +02:00
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_bool(use_existing_db, false, "If true, do not destroy the existing"
|
|
|
|
" database. If you set this flag and also specify a benchmark that"
|
|
|
|
" wants a fresh database, that benchmark will fail.");
|
2011-05-21 04:17:43 +02:00
|
|
|
|
2019-03-01 20:14:28 +01:00
|
|
|
DEFINE_bool(use_existing_keys, false,
|
|
|
|
"If true, uses existing keys in the DB, "
|
|
|
|
"rather than generating new ones. This involves some startup "
|
|
|
|
"latency to load all keys into memory. It is supported for the "
|
|
|
|
"same read/overwrite benchmarks as `-use_existing_db=true`, which "
|
|
|
|
"must also be set for this flag to be enabled. When this flag is "
|
|
|
|
"set, the value for `-num` will be ignored.");
|
|
|
|
|
Add argument --show_table_properties to db_bench
Summary:
Add argument --show_table_properties to db_bench
-show_table_properties (If true, then per-level table properties will be
printed on every stats-interval when stats_interval is set and
stats_per_interval is on.) type: bool default: false
Test Plan:
./db_bench --show_table_properties=1 --stats_interval=100000 --stats_per_interval=1
./db_bench --show_table_properties=1 --stats_interval=100000 --stats_per_interval=1 --num_column_families=2
Sample Output:
Compaction Stats [column_family_name_000001]
Level Files Size(MB) Score Read(GB) Rn(GB) Rnp1(GB) Write(GB) Wnew(GB) Moved(GB) W-Amp Rd(MB/s) Wr(MB/s) Comp(sec) Comp(cnt) Avg(sec) Stall(cnt) KeyIn KeyDrop
---------------------------------------------------------------------------------------------------------------------------------------------------------------------
L0 3/0 5 0.8 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 86.3 0 17 0.021 0 0 0
L1 5/0 9 0.9 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0 0 0.000 0 0 0
L2 9/0 16 0.2 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0 0 0.000 0 0 0
Sum 17/0 31 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 86.3 0 17 0.021 0 0 0
Int 0/0 0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 83.9 0 2 0.022 0 0 0
Flush(GB): cumulative 0.030, interval 0.004
Stalls(count): 0 level0_slowdown, 0 level0_numfiles, 0 memtable_compaction, 0 leveln_slowdown_soft, 0 leveln_slowdown_hard
Level[0]: # data blocks=2571; # entries=84813; raw key size=2035512; raw average key size=24.000000; raw value size=8481300; raw average value size=100.000000; data block size=5690119; index block size=82415; filter block size=0; (estimated) table size=5772534; filter policy name=N/A;
Level[1]: # data blocks=4285; # entries=141355; raw key size=3392520; raw average key size=24.000000; raw value size=14135500; raw average value size=100.000000; data block size=9487353; index block size=137377; filter block size=0; (estimated) table size=9624730; filter policy name=N/A;
Level[2]: # data blocks=7713; # entries=254439; raw key size=6106536; raw average key size=24.000000; raw value size=25443900; raw average value size=100.000000; data block size=17077893; index block size=247269; filter block size=0; (estimated) table size=17325162; filter policy name=N/A;
Level[3]: # data blocks=0; # entries=0; raw key size=0; raw average key size=0.000000; raw value size=0; raw average value size=0.000000; data block size=0; index block size=0; filter block size=0; (estimated) table size=0; filter policy name=N/A;
Level[4]: # data blocks=0; # entries=0; raw key size=0; raw average key size=0.000000; raw value size=0; raw average value size=0.000000; data block size=0; index block size=0; filter block size=0; (estimated) table size=0; filter policy name=N/A;
Level[5]: # data blocks=0; # entries=0; raw key size=0; raw average key size=0.000000; raw value size=0; raw average value size=0.000000; data block size=0; index block size=0; filter block size=0; (estimated) table size=0; filter policy name=N/A;
Level[6]: # data blocks=0; # entries=0; raw key size=0; raw average key size=0.000000; raw value size=0; raw average value size=0.000000; data block size=0; index block size=0; filter block size=0; (estimated) table size=0; filter policy name=N/A;
Reviewers: anthony, IslamAbdelRahman, MarkCallaghan, sdong, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D45651
2015-08-27 03:27:23 +02:00
|
|
|
DEFINE_bool(show_table_properties, false,
|
|
|
|
"If true, then per-level table"
|
|
|
|
" properties will be printed on every stats-interval when"
|
|
|
|
" stats_interval is set and stats_per_interval is on.");
|
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_string(db, "", "Use the db with the following name.");
|
2011-06-22 04:36:45 +02:00
|
|
|
|
2017-02-27 21:59:34 +01:00
|
|
|
// Read cache flags
|
|
|
|
|
|
|
|
DEFINE_string(read_cache_path, "",
|
|
|
|
"If not empty string, a read cache will be used in this path");
|
|
|
|
|
|
|
|
DEFINE_int64(read_cache_size, 4LL * 1024 * 1024 * 1024,
|
|
|
|
"Maximum size of the read cache");
|
|
|
|
|
|
|
|
DEFINE_bool(read_cache_direct_write, true,
|
|
|
|
"Whether to use Direct IO for writing to the read cache");
|
|
|
|
|
|
|
|
DEFINE_bool(read_cache_direct_read, true,
|
|
|
|
"Whether to use Direct IO for reading from read cache");
|
|
|
|
|
2018-07-14 02:18:39 +02:00
|
|
|
DEFINE_bool(use_keep_filter, false, "Whether to use a noop compaction filter");
|
2018-07-13 04:33:59 +02:00
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
static bool ValidateCacheNumshardbits(const char* flagname, int32_t value) {
|
|
|
|
if (value >= 20) {
|
|
|
|
fprintf(stderr, "Invalid value for --%s: %d, must be < 20\n",
|
|
|
|
flagname, value);
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
return true;
|
|
|
|
}
|
2012-05-17 02:22:33 +02:00
|
|
|
|
2017-07-14 01:25:53 +02:00
|
|
|
DEFINE_bool(verify_checksum, true,
|
|
|
|
"Verify checksum for every block read"
|
2013-10-24 16:43:14 +02:00
|
|
|
" from storage");
|
2012-05-19 09:05:48 +02:00
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_bool(statistics, false, "Database statistics");
|
2019-02-28 19:14:19 +01:00
|
|
|
DEFINE_int32(stats_level, rocksdb::StatsLevel::kExceptDetailedTimers,
|
|
|
|
"stats level for statistics");
|
2017-01-27 00:18:07 +01:00
|
|
|
DEFINE_string(statistics_string, "", "Serialized statistics string");
|
2013-10-04 06:49:15 +02:00
|
|
|
static class std::shared_ptr<rocksdb::Statistics> dbstats;
|
2012-05-30 08:18:16 +02:00
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_int64(writes, -1, "Number of write operations to do. If negative, do"
|
|
|
|
" --num reads.");
|
2012-08-16 00:44:23 +02:00
|
|
|
|
2017-04-07 20:25:42 +02:00
|
|
|
DEFINE_bool(finish_after_writes, false, "Write thread terminates after all writes are finished");
|
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_bool(sync, false, "Sync all writes to disk");
|
2012-08-20 01:05:02 +02:00
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_bool(use_fsync, false, "If true, issue fsync instead of fdatasync");
|
2013-03-01 20:25:13 +01:00
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_bool(disable_wal, false, "If true, do not write WAL for write.");
|
2012-10-15 19:10:56 +02:00
|
|
|
|
2014-02-25 19:43:46 +01:00
|
|
|
DEFINE_string(wal_dir, "", "If not empty, use the given dir for WAL");
|
|
|
|
|
2017-04-07 20:25:42 +02:00
|
|
|
DEFINE_string(truth_db, "/dev/shm/truth_db/dbbench",
|
|
|
|
"Truth key/values used when using verify");
|
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_int32(num_levels, 7, "The total number of levels");
|
2012-08-16 00:44:23 +02:00
|
|
|
|
2017-03-02 19:45:04 +01:00
|
|
|
DEFINE_int64(target_file_size_base, rocksdb::Options().target_file_size_base,
|
|
|
|
"Target file size at level-1");
|
2012-08-16 00:44:23 +02:00
|
|
|
|
2017-03-02 19:45:04 +01:00
|
|
|
DEFINE_int32(target_file_size_multiplier,
|
|
|
|
rocksdb::Options().target_file_size_multiplier,
|
2013-10-24 16:43:14 +02:00
|
|
|
"A multiplier to compute target level-N file size (N >= 2)");
|
2013-05-21 20:37:06 +02:00
|
|
|
|
2017-03-02 19:45:04 +01:00
|
|
|
DEFINE_uint64(max_bytes_for_level_base,
|
|
|
|
rocksdb::Options().max_bytes_for_level_base,
|
|
|
|
"Max bytes for level-1");
|
2012-08-20 01:05:02 +02:00
|
|
|
|
options.level_compaction_dynamic_level_bytes to allow RocksDB to pick size bases of levels dynamically.
Summary:
When having fixed max_bytes_for_level_base, the ratio of size of largest level and the second one can range from 0 to the multiplier. This makes LSM tree frequently irregular and unpredictable. It can also cause poor space amplification in some cases.
In this improvement (proposed by Igor Kabiljo), we introduce a parameter option.level_compaction_use_dynamic_max_bytes. When turning it on, RocksDB is free to pick a level base in the range of (options.max_bytes_for_level_base/options.max_bytes_for_level_multiplier, options.max_bytes_for_level_base] so that real level ratios are close to options.max_bytes_for_level_multiplier.
Test Plan: New unit tests and pass tests suites including valgrind.
Reviewers: MarkCallaghan, rven, yhchiang, igor, ikabiljo
Reviewed By: ikabiljo
Subscribers: yoshinorim, ikabiljo, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D31437
2015-02-05 20:44:17 +01:00
|
|
|
DEFINE_bool(level_compaction_dynamic_level_bytes, false,
|
|
|
|
"Whether level size base is dynamic");
|
|
|
|
|
2016-11-02 05:05:32 +01:00
|
|
|
DEFINE_double(max_bytes_for_level_multiplier, 10,
|
|
|
|
"A multiplier to compute max bytes for level-N (N >= 2)");
|
2012-08-20 01:05:02 +02:00
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
static std::vector<int> FLAGS_max_bytes_for_level_multiplier_additional_v;
|
|
|
|
DEFINE_string(max_bytes_for_level_multiplier_additional, "",
|
|
|
|
"A vector that specifies additional fanout per level");
|
2012-10-19 23:00:53 +02:00
|
|
|
|
Don't artificially inflate L0 score
Summary:
This turns out to be pretty bad because if we prioritize L0->L1 then L1 can grow artificially large, which makes L0->L1 more and more expensive. For example:
256MB @ L0 + 256MB @ L1 --> 512MB @ L1
256MB @ L0 + 512MB @ L1 --> 768MB @ L1
256MB @ L0 + 768MB @ L1 --> 1GB @ L1
....
256MB @ L0 + 10GB @ L1 --> 10.2GB @ L1
At some point we need to start compacting L1->L2 to speed up L0->L1.
Test Plan:
The performance improvement is massive for heavy write workload. This is the benchmark I ran: https://phabricator.fb.com/P19842671. Before this change, the benchmark took 47 minutes to complete. After, the benchmark finished in 2minutes. You can see full results here: https://phabricator.fb.com/P19842674
Also, we ran this diff on MongoDB on RocksDB on one replicaset. Before the change, our initial sync was so slow that it couldn't keep up with primary writes. After the change, the import finished without any issues
Reviewers: dynamike, MarkCallaghan, rven, yhchiang, sdong
Reviewed By: sdong
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D38637
2015-05-21 20:40:48 +02:00
|
|
|
DEFINE_int32(level0_stop_writes_trigger,
|
|
|
|
rocksdb::Options().level0_stop_writes_trigger,
|
|
|
|
"Number of files in level-0"
|
2013-10-24 16:43:14 +02:00
|
|
|
" that will trigger put stop.");
|
2012-08-31 08:06:17 +02:00
|
|
|
|
Don't artificially inflate L0 score
Summary:
This turns out to be pretty bad because if we prioritize L0->L1 then L1 can grow artificially large, which makes L0->L1 more and more expensive. For example:
256MB @ L0 + 256MB @ L1 --> 512MB @ L1
256MB @ L0 + 512MB @ L1 --> 768MB @ L1
256MB @ L0 + 768MB @ L1 --> 1GB @ L1
....
256MB @ L0 + 10GB @ L1 --> 10.2GB @ L1
At some point we need to start compacting L1->L2 to speed up L0->L1.
Test Plan:
The performance improvement is massive for heavy write workload. This is the benchmark I ran: https://phabricator.fb.com/P19842671. Before this change, the benchmark took 47 minutes to complete. After, the benchmark finished in 2minutes. You can see full results here: https://phabricator.fb.com/P19842674
Also, we ran this diff on MongoDB on RocksDB on one replicaset. Before the change, our initial sync was so slow that it couldn't keep up with primary writes. After the change, the import finished without any issues
Reviewers: dynamike, MarkCallaghan, rven, yhchiang, sdong
Reviewed By: sdong
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D38637
2015-05-21 20:40:48 +02:00
|
|
|
DEFINE_int32(level0_slowdown_writes_trigger,
|
|
|
|
rocksdb::Options().level0_slowdown_writes_trigger,
|
|
|
|
"Number of files in level-0"
|
2013-10-24 16:43:14 +02:00
|
|
|
" that will slow down writes.");
|
2013-02-20 02:51:06 +01:00
|
|
|
|
Don't artificially inflate L0 score
Summary:
This turns out to be pretty bad because if we prioritize L0->L1 then L1 can grow artificially large, which makes L0->L1 more and more expensive. For example:
256MB @ L0 + 256MB @ L1 --> 512MB @ L1
256MB @ L0 + 512MB @ L1 --> 768MB @ L1
256MB @ L0 + 768MB @ L1 --> 1GB @ L1
....
256MB @ L0 + 10GB @ L1 --> 10.2GB @ L1
At some point we need to start compacting L1->L2 to speed up L0->L1.
Test Plan:
The performance improvement is massive for heavy write workload. This is the benchmark I ran: https://phabricator.fb.com/P19842671. Before this change, the benchmark took 47 minutes to complete. After, the benchmark finished in 2minutes. You can see full results here: https://phabricator.fb.com/P19842674
Also, we ran this diff on MongoDB on RocksDB on one replicaset. Before the change, our initial sync was so slow that it couldn't keep up with primary writes. After the change, the import finished without any issues
Reviewers: dynamike, MarkCallaghan, rven, yhchiang, sdong
Reviewed By: sdong
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D38637
2015-05-21 20:40:48 +02:00
|
|
|
DEFINE_int32(level0_file_num_compaction_trigger,
|
|
|
|
rocksdb::Options().level0_file_num_compaction_trigger,
|
|
|
|
"Number of files in level-0"
|
2013-10-24 16:43:14 +02:00
|
|
|
" when compactions start");
|
2012-09-17 22:35:57 +02:00
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
static bool ValidateInt32Percent(const char* flagname, int32_t value) {
|
|
|
|
if (value <= 0 || value>=100) {
|
|
|
|
fprintf(stderr, "Invalid value for --%s: %d, 0< pct <100 \n",
|
|
|
|
flagname, value);
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
DEFINE_int32(readwritepercent, 90, "Ratio of reads to reads/writes (expressed"
|
|
|
|
" as percentage) for the ReadRandomWriteRandom workload. The "
|
|
|
|
"default value 90 means 90% operations out of all reads and writes"
|
|
|
|
" operations are reads. In other words, 9 gets for every 1 put.");
|
|
|
|
|
2014-01-11 02:33:56 +01:00
|
|
|
DEFINE_int32(mergereadpercent, 70, "Ratio of merges to merges&reads (expressed"
|
|
|
|
" as percentage) for the ReadRandomMergeRandom workload. The"
|
|
|
|
" default value 70 means 70% out of all read and merge operations"
|
|
|
|
" are merges. In other words, 7 merges for every 3 gets.");
|
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_int32(deletepercent, 2, "Percentage of deletes out of reads/writes/"
|
|
|
|
"deletes (used in RandomWithVerify only). RandomWithVerify "
|
|
|
|
"calculates writepercent as (100 - FLAGS_readwritepercent - "
|
|
|
|
"deletepercent), so deletepercent must be smaller than (100 - "
|
|
|
|
"FLAGS_readwritepercent)");
|
|
|
|
|
2016-05-05 16:32:10 +02:00
|
|
|
DEFINE_bool(optimize_filters_for_hits, false,
|
|
|
|
"Optimizes bloom filters for workloads for most lookups return "
|
|
|
|
"a value. For now this doesn't create bloom filters for the max "
|
|
|
|
"level of the LSM to reduce metadata that should fit in RAM. ");
|
|
|
|
|
2014-12-22 12:04:45 +01:00
|
|
|
DEFINE_uint64(delete_obsolete_files_period_micros, 0,
|
|
|
|
"Ignored. Left here for backward compatibility");
|
2013-10-24 16:43:14 +02:00
|
|
|
|
2018-10-19 02:16:36 +02:00
|
|
|
DEFINE_int64(writes_before_delete_range, 0,
|
|
|
|
"Number of writes before DeleteRange is called regularly.");
|
|
|
|
|
2016-11-16 02:33:27 +01:00
|
|
|
DEFINE_int64(writes_per_range_tombstone, 0,
|
2018-10-19 02:16:36 +02:00
|
|
|
"Number of writes between range tombstones");
|
2016-11-16 02:33:27 +01:00
|
|
|
|
|
|
|
DEFINE_int64(range_tombstone_width, 100, "Number of keys in tombstone's range");
|
|
|
|
|
|
|
|
DEFINE_int64(max_num_range_tombstones, 0,
|
|
|
|
"Maximum number of range tombstones "
|
|
|
|
"to insert.");
|
|
|
|
|
2016-12-07 03:02:43 +01:00
|
|
|
DEFINE_bool(expand_range_tombstones, false,
|
|
|
|
"Expand range tombstone into sequential regular tombstones.");
|
|
|
|
|
2015-10-14 21:43:00 +02:00
|
|
|
#ifndef ROCKSDB_LITE
|
2017-11-21 08:19:46 +01:00
|
|
|
// Transactions Options
|
Pessimistic Transactions
Summary:
Initial implementation of Pessimistic Transactions. This diff contains the api changes discussed in D38913. This diff is pretty large, so let me know if people would prefer to meet up to discuss it.
MyRocks folks: please take a look at the API in include/rocksdb/utilities/transaction[_db].h and let me know if you have any issues.
Also, you'll notice a couple of TODOs in the implementation of RollbackToSavePoint(). After chatting with Siying, I'm going to send out a separate diff for an alternate implementation of this feature that implements the rollback inside of WriteBatch/WriteBatchWithIndex. We can then decide which route is preferable.
Next, I'm planning on doing some perf testing and then integrating this diff into MongoRocks for further testing.
Test Plan: Unit tests, db_bench parallel testing.
Reviewers: igor, rven, sdong, yhchiang, yoshinorim
Reviewed By: sdong
Subscribers: hermanlee4, maykov, spetrunia, leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D40869
2015-05-26 02:37:33 +02:00
|
|
|
DEFINE_bool(optimistic_transaction_db, false,
|
2015-05-29 23:36:35 +02:00
|
|
|
"Open a OptimisticTransactionDB instance. "
|
|
|
|
"Required for randomtransaction benchmark.");
|
|
|
|
|
Pessimistic Transactions
Summary:
Initial implementation of Pessimistic Transactions. This diff contains the api changes discussed in D38913. This diff is pretty large, so let me know if people would prefer to meet up to discuss it.
MyRocks folks: please take a look at the API in include/rocksdb/utilities/transaction[_db].h and let me know if you have any issues.
Also, you'll notice a couple of TODOs in the implementation of RollbackToSavePoint(). After chatting with Siying, I'm going to send out a separate diff for an alternate implementation of this feature that implements the rollback inside of WriteBatch/WriteBatchWithIndex. We can then decide which route is preferable.
Next, I'm planning on doing some perf testing and then integrating this diff into MongoRocks for further testing.
Test Plan: Unit tests, db_bench parallel testing.
Reviewers: igor, rven, sdong, yhchiang, yoshinorim
Reviewed By: sdong
Subscribers: hermanlee4, maykov, spetrunia, leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D40869
2015-05-26 02:37:33 +02:00
|
|
|
DEFINE_bool(transaction_db, false,
|
|
|
|
"Open a TransactionDB instance. "
|
|
|
|
"Required for randomtransaction benchmark.");
|
|
|
|
|
2015-05-29 23:36:35 +02:00
|
|
|
DEFINE_uint64(transaction_sets, 2,
|
|
|
|
"Number of keys each transaction will "
|
|
|
|
"modify (use in RandomTransaction only). Max: 9999");
|
|
|
|
|
2015-08-29 01:09:34 +02:00
|
|
|
DEFINE_bool(transaction_set_snapshot, false,
|
|
|
|
"Setting to true will have each transaction call SetSnapshot()"
|
|
|
|
" upon creation.");
|
|
|
|
|
2015-05-29 23:36:35 +02:00
|
|
|
DEFINE_int32(transaction_sleep, 0,
|
|
|
|
"Max microseconds to sleep in between "
|
|
|
|
"reading and writing a value (used in RandomTransaction only). ");
|
|
|
|
|
2015-08-29 01:09:34 +02:00
|
|
|
DEFINE_uint64(transaction_lock_timeout, 100,
|
|
|
|
"If using a transaction_db, specifies the lock wait timeout in"
|
|
|
|
" milliseconds before failing a transaction waiting on a lock");
|
2016-06-03 01:24:14 +02:00
|
|
|
DEFINE_string(
|
|
|
|
options_file, "",
|
|
|
|
"The path to a RocksDB options file. If specified, then db_bench will "
|
|
|
|
"run with the RocksDB options in the default column family of the "
|
|
|
|
"specified options file. "
|
|
|
|
"Note that with this setting, db_bench will ONLY accept the following "
|
|
|
|
"RocksDB options related command-line arguments, all other arguments "
|
|
|
|
"that are related to RocksDB options will be ignored:\n"
|
|
|
|
"\t--use_existing_db\n"
|
2019-03-01 20:14:28 +01:00
|
|
|
"\t--use_existing_keys\n"
|
2016-06-03 01:24:14 +02:00
|
|
|
"\t--statistics\n"
|
|
|
|
"\t--row_cache_size\n"
|
|
|
|
"\t--row_cache_numshardbits\n"
|
|
|
|
"\t--enable_io_prio\n"
|
|
|
|
"\t--dump_malloc_stats\n"
|
|
|
|
"\t--num_multi_db\n");
|
2017-03-25 00:58:51 +01:00
|
|
|
|
2017-11-21 08:19:46 +01:00
|
|
|
// FIFO Compaction Options
|
2017-03-25 00:58:51 +01:00
|
|
|
DEFINE_uint64(fifo_compaction_max_table_files_size_mb, 0,
|
|
|
|
"The limit of total table file sizes to trigger FIFO compaction");
|
2017-11-21 08:19:46 +01:00
|
|
|
|
2017-05-05 03:14:29 +02:00
|
|
|
DEFINE_bool(fifo_compaction_allow_compaction, true,
|
|
|
|
"Allow compaction in FIFO compaction.");
|
2017-11-21 08:19:46 +01:00
|
|
|
|
2017-06-28 02:02:20 +02:00
|
|
|
DEFINE_uint64(fifo_compaction_ttl, 0, "TTL for the SST Files in seconds.");
|
2017-11-21 08:19:46 +01:00
|
|
|
|
|
|
|
// Blob DB Options
|
|
|
|
DEFINE_bool(use_blob_db, false,
|
|
|
|
"Open a BlobDB instance. "
|
2017-12-07 05:40:46 +01:00
|
|
|
"Required for large value benchmark.");
|
|
|
|
|
|
|
|
DEFINE_bool(blob_db_enable_gc, false, "Enable BlobDB garbage collection.");
|
|
|
|
|
|
|
|
DEFINE_bool(blob_db_is_fifo, false, "Enable FIFO eviction strategy in BlobDB.");
|
|
|
|
|
2018-03-06 20:46:20 +01:00
|
|
|
DEFINE_uint64(blob_db_max_db_size, 0,
|
2017-12-07 05:40:46 +01:00
|
|
|
"Max size limit of the directory where blob files are stored.");
|
|
|
|
|
|
|
|
DEFINE_uint64(blob_db_max_ttl_range, 86400,
|
|
|
|
"TTL range to generate BlobDB data (in seconds).");
|
|
|
|
|
|
|
|
DEFINE_uint64(blob_db_ttl_range_secs, 3600,
|
|
|
|
"TTL bucket size to use when creating blob files.");
|
|
|
|
|
|
|
|
DEFINE_uint64(blob_db_min_blob_size, 0,
|
|
|
|
"Smallest blob to store in a file. Blobs smaller than this "
|
|
|
|
"will be inlined with the key in the LSM tree.");
|
|
|
|
|
|
|
|
DEFINE_uint64(blob_db_bytes_per_sync, 0, "Bytes to sync blob file at.");
|
|
|
|
|
|
|
|
DEFINE_uint64(blob_db_file_size, 256 * 1024 * 1024,
|
|
|
|
"Target size of each blob file.");
|
2017-11-21 08:19:46 +01:00
|
|
|
|
2015-10-14 21:43:00 +02:00
|
|
|
#endif // ROCKSDB_LITE
|
2015-08-29 01:09:34 +02:00
|
|
|
|
2016-04-14 22:56:29 +02:00
|
|
|
DEFINE_bool(report_bg_io_stats, false,
|
Add options.compaction_measure_io_stats to print write I/O stats in compactions
Summary:
Add options.compaction_measure_io_stats to print out / pass to listener accumulated time spent on write calls. Example outputs in info logs:
2015/08/12-16:27:59.463944 7fd428bff700 (Original Log Time 2015/08/12-16:27:59.463922) EVENT_LOG_v1 {"time_micros": 1439422079463897, "job": 6, "event": "compaction_finished", "output_level": 1, "num_output_files": 4, "total_output_size": 6900525, "num_input_records": 111483, "num_output_records": 106877, "file_write_nanos": 15663206, "file_range_sync_nanos": 649588, "file_fsync_nanos": 349614797, "file_prepare_write_nanos": 1505812, "lsm_state": [2, 4, 0, 0, 0, 0, 0]}
Add two more counters in iostats_context.
Also add a parameter of db_bench.
Test Plan: Add a unit test. Also manually verify LOG outputs in db_bench
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D44115
2015-08-13 02:24:45 +02:00
|
|
|
"Measure times spents on I/Os while in compactions. ");
|
|
|
|
|
2016-12-10 01:36:43 +01:00
|
|
|
DEFINE_bool(use_stderr_info_logger, false,
|
|
|
|
"Write info logs to stderr instead of to LOG file. ");
|
|
|
|
|
2018-08-01 09:14:43 +02:00
|
|
|
DEFINE_string(trace_file, "", "Trace workload to a file. ");
|
|
|
|
|
2017-02-07 20:34:14 +01:00
|
|
|
static enum rocksdb::CompressionType StringToCompressionType(const char* ctype) {
|
2013-10-24 16:43:14 +02:00
|
|
|
assert(ctype);
|
|
|
|
|
|
|
|
if (!strcasecmp(ctype, "none"))
|
|
|
|
return rocksdb::kNoCompression;
|
|
|
|
else if (!strcasecmp(ctype, "snappy"))
|
|
|
|
return rocksdb::kSnappyCompression;
|
|
|
|
else if (!strcasecmp(ctype, "zlib"))
|
|
|
|
return rocksdb::kZlibCompression;
|
|
|
|
else if (!strcasecmp(ctype, "bzip2"))
|
|
|
|
return rocksdb::kBZip2Compression;
|
2014-02-08 03:12:30 +01:00
|
|
|
else if (!strcasecmp(ctype, "lz4"))
|
|
|
|
return rocksdb::kLZ4Compression;
|
|
|
|
else if (!strcasecmp(ctype, "lz4hc"))
|
|
|
|
return rocksdb::kLZ4HCCompression;
|
2016-04-20 07:54:24 +02:00
|
|
|
else if (!strcasecmp(ctype, "xpress"))
|
|
|
|
return rocksdb::kXpressCompression;
|
2015-08-28 00:40:42 +02:00
|
|
|
else if (!strcasecmp(ctype, "zstd"))
|
2016-09-02 00:28:40 +02:00
|
|
|
return rocksdb::kZSTD;
|
2013-10-24 16:43:14 +02:00
|
|
|
|
|
|
|
fprintf(stdout, "Cannot parse compression type '%s'\n", ctype);
|
2016-02-16 15:17:31 +01:00
|
|
|
return rocksdb::kSnappyCompression; // default value
|
2013-10-24 16:43:14 +02:00
|
|
|
}
|
2014-12-23 03:39:28 +01:00
|
|
|
|
2017-02-07 20:34:14 +01:00
|
|
|
static std::string ColumnFamilyName(size_t i) {
|
2014-12-23 03:39:28 +01:00
|
|
|
if (i == 0) {
|
|
|
|
return rocksdb::kDefaultColumnFamilyName;
|
|
|
|
} else {
|
|
|
|
char name[100];
|
2014-12-31 03:33:35 +01:00
|
|
|
snprintf(name, sizeof(name), "column_family_name_%06zu", i);
|
2014-12-23 03:39:28 +01:00
|
|
|
return std::string(name);
|
|
|
|
}
|
|
|
|
}
|
2014-04-10 06:17:14 +02:00
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_string(compression_type, "snappy",
|
|
|
|
"Algorithm to use to compress the database");
|
|
|
|
static enum rocksdb::CompressionType FLAGS_compression_type_e =
|
2013-10-04 06:49:15 +02:00
|
|
|
rocksdb::kSnappyCompression;
|
2012-09-14 20:21:54 +02:00
|
|
|
|
2019-03-18 20:07:35 +01:00
|
|
|
DEFINE_int64(sample_for_compression, 0, "Sample every N block for compression");
|
|
|
|
|
2018-08-09 19:17:56 +02:00
|
|
|
DEFINE_int32(compression_level, rocksdb::CompressionOptions().level,
|
|
|
|
"Compression level. The meaning of this value is library-"
|
|
|
|
"dependent. If unset, we try to use the default for the library "
|
|
|
|
"specified in `--compression_type`");
|
2013-12-03 21:32:07 +01:00
|
|
|
|
2018-08-09 19:17:56 +02:00
|
|
|
DEFINE_int32(compression_max_dict_bytes,
|
|
|
|
rocksdb::CompressionOptions().max_dict_bytes,
|
2016-04-28 02:39:18 +02:00
|
|
|
"Maximum size of dictionary used to prime the compression "
|
|
|
|
"library.");
|
|
|
|
|
2018-08-09 19:17:56 +02:00
|
|
|
DEFINE_int32(compression_zstd_max_train_bytes,
|
|
|
|
rocksdb::CompressionOptions().zstd_max_train_bytes,
|
2017-11-07 22:49:35 +01:00
|
|
|
"Maximum size of training data passed to zstd's dictionary "
|
|
|
|
"trainer.");
|
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_int32(min_level_to_compress, -1, "If non-negative, compression starts"
|
|
|
|
" from this level. Levels with number < min_level_to_compress are"
|
|
|
|
" not compressed. Otherwise, apply compression_type to "
|
|
|
|
"all levels.");
|
|
|
|
|
|
|
|
static bool ValidateTableCacheNumshardbits(const char* flagname,
|
|
|
|
int32_t value) {
|
|
|
|
if (0 >= value || value > 20) {
|
|
|
|
fprintf(stderr, "Invalid value for --%s: %d, must be 0 < val <= 20\n",
|
|
|
|
flagname, value);
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
DEFINE_int32(table_cache_numshardbits, 4, "");
|
2012-11-01 01:02:24 +01:00
|
|
|
|
2016-06-10 02:53:03 +02:00
|
|
|
#ifndef ROCKSDB_LITE
|
|
|
|
DEFINE_string(env_uri, "", "URI for registry Env lookup. Mutually exclusive"
|
|
|
|
" with --hdfs.");
|
|
|
|
#endif // ROCKSDB_LITE
|
|
|
|
DEFINE_string(hdfs, "", "Name of hdfs environment. Mutually exclusive with"
|
|
|
|
" --env_uri.");
|
2013-10-04 06:49:15 +02:00
|
|
|
static rocksdb::Env* FLAGS_env = rocksdb::Env::Default();
|
2012-06-08 10:11:14 +02:00
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_int64(stats_interval, 0, "Stats are reported every N operations when "
|
|
|
|
"this is greater than zero. When 0 the interval grows over time.");
|
2012-10-03 15:51:11 +02:00
|
|
|
|
2015-03-30 21:58:32 +02:00
|
|
|
DEFINE_int64(stats_interval_seconds, 0, "Report stats every N seconds. This "
|
|
|
|
"overrides stats_interval when both are > 0.");
|
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_int32(stats_per_interval, 0, "Reports additional stats per interval when"
|
|
|
|
" this is greater than 0.");
|
2012-10-26 22:37:21 +02:00
|
|
|
|
db_bench periodically writes QPS to CSV file
Summary:
This is part of an effort to better understand and optimize RocksDB stalls under high load. I added a feature to db_bench to periodically write QPS to CSV files. That way we can nicely see how our QPS changes in time (especially when DB is stalled) and can do a better job of evaluating our stall system (i.e. we want the QPS to be as constant as possible, as opposed to having bunch of stalls)
Cool part of CSV files is that we can easily graph them -- there are a bunch of tools available.
Test Plan:
Ran ./db_bench --report_interval_seconds=10 --benchmarks=fillrandom --num=10000000
and observed this in report.csv:
secs_elapsed,interval_qps
10,2725860
20,1980480
30,1863456
40,1454359
50,1460389
Reviewers: sdong, MarkCallaghan, rven, yhchiang
Reviewed By: yhchiang
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D40047
2015-06-12 23:31:53 +02:00
|
|
|
DEFINE_int64(report_interval_seconds, 0,
|
|
|
|
"If greater than zero, it will write simple stats in CVS format "
|
|
|
|
"to --report_file every N seconds");
|
|
|
|
|
|
|
|
DEFINE_string(report_file, "report.csv",
|
|
|
|
"Filename where some simple stats are reported to (if "
|
|
|
|
"--report_interval_seconds is bigger than 0)");
|
|
|
|
|
2015-03-06 20:22:06 +01:00
|
|
|
DEFINE_int32(thread_status_per_interval, 0,
|
|
|
|
"Takes and report a snapshot of the current status of each thread"
|
|
|
|
" when this is greater than 0.");
|
|
|
|
|
2016-06-30 03:44:22 +02:00
|
|
|
DEFINE_int32(perf_level, rocksdb::PerfLevel::kDisable, "Level of perf collection");
|
2013-11-18 20:32:54 +01:00
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
static bool ValidateRateLimit(const char* flagname, double value) {
|
2015-07-02 01:13:49 +02:00
|
|
|
const double EPSILON = 1e-10;
|
2013-10-24 16:43:14 +02:00
|
|
|
if ( value < -EPSILON ) {
|
|
|
|
fprintf(stderr, "Invalid value for --%s: %12.6f, must be >= 0.0\n",
|
|
|
|
flagname, value);
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
return true;
|
|
|
|
}
|
2015-11-19 03:10:20 +01:00
|
|
|
DEFINE_double(soft_rate_limit, 0.0, "DEPRECATED");
|
2013-08-06 00:43:49 +02:00
|
|
|
|
2015-09-11 23:31:23 +02:00
|
|
|
DEFINE_double(hard_rate_limit, 0.0, "DEPRECATED");
|
|
|
|
|
2015-11-19 03:10:20 +01:00
|
|
|
DEFINE_uint64(soft_pending_compaction_bytes_limit, 64ull * 1024 * 1024 * 1024,
|
|
|
|
"Slowdown writes if pending compaction bytes exceed this number");
|
|
|
|
|
2015-11-06 20:34:06 +01:00
|
|
|
DEFINE_uint64(hard_pending_compaction_bytes_limit, 128ull * 1024 * 1024 * 1024,
|
2015-09-11 23:31:23 +02:00
|
|
|
"Stop writes if pending compaction bytes exceed this number");
|
2012-10-26 22:37:21 +02:00
|
|
|
|
2015-12-23 22:49:53 +01:00
|
|
|
DEFINE_uint64(delayed_write_rate, 8388608u,
|
2015-05-16 00:52:51 +02:00
|
|
|
"Limited bytes allowed to DB when soft_rate_limit or "
|
|
|
|
"level0_slowdown_writes_trigger triggers");
|
|
|
|
|
2017-07-01 00:21:02 +02:00
|
|
|
DEFINE_bool(enable_pipelined_write, true,
|
|
|
|
"Allow WAL and memtable writes to be pipelined");
|
|
|
|
|
|
|
|
DEFINE_bool(allow_concurrent_memtable_write, true,
|
support for concurrent adds to memtable
Summary:
This diff adds support for concurrent adds to the skiplist memtable
implementations. Memory allocation is made thread-safe by the addition of
a spinlock, with small per-core buffers to avoid contention. Concurrent
memtable writes are made via an additional method and don't impose a
performance overhead on the non-concurrent case, so parallelism can be
selected on a per-batch basis.
Write thread synchronization is an increasing bottleneck for higher levels
of concurrency, so this diff adds --enable_write_thread_adaptive_yield
(default off). This feature causes threads joining a write batch
group to spin for a short time (default 100 usec) using sched_yield,
rather than going to sleep on a mutex. If the timing of the yield calls
indicates that another thread has actually run during the yield then
spinning is avoided. This option improves performance for concurrent
situations even without parallel adds, although it has the potential to
increase CPU usage (and the heuristic adaptation is not yet mature).
Parallel writes are not currently compatible with
inplace updates, update callbacks, or delete filtering.
Enable it with --allow_concurrent_memtable_write (and
--enable_write_thread_adaptive_yield). Parallel memtable writes
are performance neutral when there is no actual parallelism, and in
my experiments (SSD server-class Linux and varying contention and key
sizes for fillrandom) they are always a performance win when there is
more than one thread.
Statistics are updated earlier in the write path, dropping the number
of DB mutex acquisitions from 2 to 1 for almost all cases.
This diff was motivated and inspired by Yahoo's cLSM work. It is more
conservative than cLSM: RocksDB's write batch group leader role is
preserved (along with all of the existing flush and write throttling
logic) and concurrent writers are blocked until all memtable insertions
have completed and the sequence number has been advanced, to preserve
linearizability.
My test config is "db_bench -benchmarks=fillrandom -threads=$T
-batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T
-level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999
-disable_auto_compactions --max_write_buffer_number=8
-max_background_flushes=8 --disable_wal --write_buffer_size=160000000
--block_size=16384 --allow_concurrent_memtable_write" on a two-socket
Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1
thread I get ~440Kops/sec. Peak performance for 1 socket (numactl
-N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance
across both sockets happens at 30 threads, and is ~900Kops/sec, although
with fewer threads there is less performance loss when the system has
background work.
Test Plan:
1. concurrent stress tests for InlineSkipList and DynamicBloom
2. make clean; make check
3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench
4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench
5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench
6. make clean; OPT=-DROCKSDB_LITE make check
7. verify no perf regressions when disabled
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba
Differential Revision: https://reviews.facebook.net/D50589
2015-08-15 01:59:07 +02:00
|
|
|
"Allow multi-writers to update mem tables in parallel.");
|
|
|
|
|
2018-01-26 19:46:30 +01:00
|
|
|
DEFINE_bool(inplace_update_support, rocksdb::Options().inplace_update_support,
|
|
|
|
"Support in-place memtable update for smaller or same-size values");
|
|
|
|
|
|
|
|
DEFINE_uint64(inplace_update_num_locks,
|
|
|
|
rocksdb::Options().inplace_update_num_locks,
|
|
|
|
"Number of RW locks to protect in-place memtable updates");
|
|
|
|
|
2017-07-01 00:21:02 +02:00
|
|
|
DEFINE_bool(enable_write_thread_adaptive_yield, true,
|
support for concurrent adds to memtable
Summary:
This diff adds support for concurrent adds to the skiplist memtable
implementations. Memory allocation is made thread-safe by the addition of
a spinlock, with small per-core buffers to avoid contention. Concurrent
memtable writes are made via an additional method and don't impose a
performance overhead on the non-concurrent case, so parallelism can be
selected on a per-batch basis.
Write thread synchronization is an increasing bottleneck for higher levels
of concurrency, so this diff adds --enable_write_thread_adaptive_yield
(default off). This feature causes threads joining a write batch
group to spin for a short time (default 100 usec) using sched_yield,
rather than going to sleep on a mutex. If the timing of the yield calls
indicates that another thread has actually run during the yield then
spinning is avoided. This option improves performance for concurrent
situations even without parallel adds, although it has the potential to
increase CPU usage (and the heuristic adaptation is not yet mature).
Parallel writes are not currently compatible with
inplace updates, update callbacks, or delete filtering.
Enable it with --allow_concurrent_memtable_write (and
--enable_write_thread_adaptive_yield). Parallel memtable writes
are performance neutral when there is no actual parallelism, and in
my experiments (SSD server-class Linux and varying contention and key
sizes for fillrandom) they are always a performance win when there is
more than one thread.
Statistics are updated earlier in the write path, dropping the number
of DB mutex acquisitions from 2 to 1 for almost all cases.
This diff was motivated and inspired by Yahoo's cLSM work. It is more
conservative than cLSM: RocksDB's write batch group leader role is
preserved (along with all of the existing flush and write throttling
logic) and concurrent writers are blocked until all memtable insertions
have completed and the sequence number has been advanced, to preserve
linearizability.
My test config is "db_bench -benchmarks=fillrandom -threads=$T
-batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T
-level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999
-disable_auto_compactions --max_write_buffer_number=8
-max_background_flushes=8 --disable_wal --write_buffer_size=160000000
--block_size=16384 --allow_concurrent_memtable_write" on a two-socket
Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1
thread I get ~440Kops/sec. Peak performance for 1 socket (numactl
-N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance
across both sockets happens at 30 threads, and is ~900Kops/sec, although
with fewer threads there is less performance loss when the system has
background work.
Test Plan:
1. concurrent stress tests for InlineSkipList and DynamicBloom
2. make clean; make check
3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench
4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench
5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench
6. make clean; OPT=-DROCKSDB_LITE make check
7. verify no perf regressions when disabled
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba
Differential Revision: https://reviews.facebook.net/D50589
2015-08-15 01:59:07 +02:00
|
|
|
"Use a yielding spin loop for brief writer thread waits.");
|
|
|
|
|
|
|
|
DEFINE_uint64(
|
|
|
|
write_thread_max_yield_usec, 100,
|
|
|
|
"Maximum microseconds for enable_write_thread_adaptive_yield operation.");
|
|
|
|
|
|
|
|
DEFINE_uint64(write_thread_slow_yield_usec, 3,
|
|
|
|
"The threshold at which a slow yield is considered a signal that "
|
|
|
|
"other processes or threads want the core.");
|
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_int32(rate_limit_delay_max_milliseconds, 1000,
|
|
|
|
"When hard_rate_limit is set then this is the max time a put will"
|
|
|
|
" be stalled.");
|
2013-03-02 21:56:04 +01:00
|
|
|
|
2015-05-13 02:40:27 +02:00
|
|
|
DEFINE_uint64(rate_limiter_bytes_per_sec, 0, "Set options.rate_limiter value.");
|
|
|
|
|
2017-10-05 04:02:22 +02:00
|
|
|
DEFINE_bool(rate_limiter_auto_tuned, false,
|
|
|
|
"Enable dynamic adjustment of rate limit according to demand for "
|
|
|
|
"background I/O");
|
|
|
|
|
2018-06-15 21:07:38 +02:00
|
|
|
|
|
|
|
DEFINE_bool(sine_write_rate, false,
|
|
|
|
"Use a sine wave write_rate_limit");
|
|
|
|
|
|
|
|
DEFINE_uint64(sine_write_rate_interval_milliseconds, 10000,
|
|
|
|
"Interval of which the sine wave write_rate_limit is recalculated");
|
|
|
|
|
|
|
|
DEFINE_double(sine_a, 1,
|
|
|
|
"A in f(x) = A sin(bx + c) + d");
|
|
|
|
|
|
|
|
DEFINE_double(sine_b, 1,
|
|
|
|
"B in f(x) = A sin(bx + c) + d");
|
|
|
|
|
|
|
|
DEFINE_double(sine_c, 0,
|
|
|
|
"C in f(x) = A sin(bx + c) + d");
|
|
|
|
|
|
|
|
DEFINE_double(sine_d, 1,
|
|
|
|
"D in f(x) = A sin(bx + c) + d");
|
|
|
|
|
2017-06-13 23:51:22 +02:00
|
|
|
DEFINE_bool(rate_limit_bg_reads, false,
|
|
|
|
"Use options.rate_limiter on compaction reads");
|
|
|
|
|
2015-06-18 01:44:52 +02:00
|
|
|
DEFINE_uint64(
|
|
|
|
benchmark_write_rate_limit, 0,
|
2016-01-04 21:01:27 +01:00
|
|
|
"If non-zero, db_bench will rate-limit the writes going into RocksDB. This "
|
|
|
|
"is the global rate in bytes/second.");
|
2015-06-18 01:44:52 +02:00
|
|
|
|
Generate mixed workload with Get, Put, Seek in db_bench (#4788)
Summary:
Based on the specific workload models (key access distribution, value size distribution, and iterator scan length distribution, the QPS variation), the MixGraph benchmark generate the synthetic workload according to these distributions which can reflect the real-world workload characteristics.
After user enable the tracing function, they will get the trace file. By analyzing the trace file with the trace_analyzer tool, user can generate a set of statistic data files including. The *_accessed_key_stats.txt, *-accessed_value_size_distribution.txt, *-iterator_length_distribution.txt, and *-qps_stats.txt are mainly used to fit the Matlab model fitting. After that, user can get the parameters of the workload distributions (the modeling details are described: [here](https://github.com/facebook/rocksdb/wiki/RocksDB-Trace%2C-Replay%2C-and-Analyzer))
The key access distribution follows the The two-term power model. The probability density function is: `f(x) = ax^{b}+c`. The corresponding parameters are key_dist_a, key_dist_b, and key_dist_c in db_bench
For the value size distribution and iterator scan length distribution, they both follow the Generalized Pareto Distribution. The probability density function is `f(x) = (1/sigma)(1+k*(x-theta)/sigma))^{-1-1/k)`. The parameters are: value_k, value_theta, value_sigma and iter_k, iter_theta, iter_sigma. For more information about the Generalized Pareto Distribution, users can find the [wiki](https://en.wikipedia.org/wiki/Generalized_Pareto_distribution) and [Matalb page](https://www.mathworks.com/help/stats/generalized-pareto-distribution.html)
As for the QPS, it follows the diurnal pattern. So Sine is a good model to fit it. `F(x) = sine_a*sin(sine_b*x + sine_c) + sine_d`. The trace_will tell you the average QPS in the print out resutls, which is sine_d. After user fit the "*-qps_stats.txt" to the Matlab model, user can get the sine_a, sine_b, and sine_c. By using the 4 parameters, user can control the QPS variation including the period, average, changes.
To use the bench mark, user can indicate the following parameters as examples:
```
-benchmarks="mixgraph" -key_dist_a=0.002312 -key_dist_b=0.3467 -value_k=0.9233 -value_sigma=226.4092 -iter_k=2.517 -iter_sigma=14.236 -mix_get_ratio=0.7 -mix_put_ratio=0.25 -mix_seek_ratio=0.05 -sine_mix_rate_interval_milliseconds=500 -sine_a=15000 -sine_b=1 -sine_d=20000
```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4788
Differential Revision: D13573940
Pulled By: sagar0
fbshipit-source-id: e184c27e07b4f1bc0b436c2be36c5090c1fb0222
2019-01-22 19:40:44 +01:00
|
|
|
// the parameters of mix_graph
|
|
|
|
DEFINE_double(key_dist_a, 0.0,
|
|
|
|
"The parameter 'a' of key access distribution model "
|
|
|
|
"f(x)=a*x^b");
|
|
|
|
DEFINE_double(key_dist_b, 0.0,
|
|
|
|
"The parameter 'b' of key access distribution model "
|
|
|
|
"f(x)=a*x^b");
|
|
|
|
DEFINE_double(value_theta, 0.0,
|
|
|
|
"The parameter 'theta' of Generized Pareto Distribution "
|
|
|
|
"f(x)=(1/sigma)*(1+k*(x-theta)/sigma)^-(1/k+1)");
|
|
|
|
DEFINE_double(value_k, 0.0,
|
|
|
|
"The parameter 'k' of Generized Pareto Distribution "
|
|
|
|
"f(x)=(1/sigma)*(1+k*(x-theta)/sigma)^-(1/k+1)");
|
|
|
|
DEFINE_double(value_sigma, 0.0,
|
|
|
|
"The parameter 'theta' of Generized Pareto Distribution "
|
|
|
|
"f(x)=(1/sigma)*(1+k*(x-theta)/sigma)^-(1/k+1)");
|
|
|
|
DEFINE_double(iter_theta, 0.0,
|
|
|
|
"The parameter 'theta' of Generized Pareto Distribution "
|
|
|
|
"f(x)=(1/sigma)*(1+k*(x-theta)/sigma)^-(1/k+1)");
|
|
|
|
DEFINE_double(iter_k, 0.0,
|
|
|
|
"The parameter 'k' of Generized Pareto Distribution "
|
|
|
|
"f(x)=(1/sigma)*(1+k*(x-theta)/sigma)^-(1/k+1)");
|
|
|
|
DEFINE_double(iter_sigma, 0.0,
|
|
|
|
"The parameter 'sigma' of Generized Pareto Distribution "
|
|
|
|
"f(x)=(1/sigma)*(1+k*(x-theta)/sigma)^-(1/k+1)");
|
|
|
|
DEFINE_double(mix_get_ratio, 1.0,
|
|
|
|
"The ratio of Get queries of mix_graph workload");
|
|
|
|
DEFINE_double(mix_put_ratio, 0.0,
|
|
|
|
"The ratio of Put queries of mix_graph workload");
|
|
|
|
DEFINE_double(mix_seek_ratio, 0.0,
|
|
|
|
"The ratio of Seek queries of mix_graph workload");
|
|
|
|
DEFINE_int64(mix_max_scan_len, 10000, "The max scan length of Iterator");
|
|
|
|
DEFINE_int64(mix_ave_kv_size, 512,
|
|
|
|
"The average key-value size of this workload");
|
|
|
|
DEFINE_int64(mix_max_value_size, 1024, "The max value size of this workload");
|
|
|
|
DEFINE_double(
|
|
|
|
sine_mix_rate_noise, 0.0,
|
|
|
|
"Add the noise ratio to the sine rate, it is between 0.0 and 1.0");
|
|
|
|
DEFINE_bool(sine_mix_rate, false,
|
|
|
|
"Enable the sine QPS control on the mix workload");
|
|
|
|
DEFINE_uint64(
|
|
|
|
sine_mix_rate_interval_milliseconds, 10000,
|
|
|
|
"Interval of which the sine wave read_rate_limit is recalculated");
|
|
|
|
DEFINE_int64(mix_accesses, -1,
|
|
|
|
"The total query accesses of mix_graph workload");
|
|
|
|
|
2016-12-19 21:25:35 +01:00
|
|
|
DEFINE_uint64(
|
|
|
|
benchmark_read_rate_limit, 0,
|
|
|
|
"If non-zero, db_bench will rate-limit the reads from RocksDB. This "
|
|
|
|
"is the global rate in ops/second.");
|
|
|
|
|
2016-06-17 01:02:52 +02:00
|
|
|
DEFINE_uint64(max_compaction_bytes, rocksdb::Options().max_compaction_bytes,
|
|
|
|
"Max bytes allowed in one compaction");
|
2012-11-14 18:10:46 +01:00
|
|
|
|
2015-10-14 21:43:00 +02:00
|
|
|
#ifndef ROCKSDB_LITE
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_bool(readonly, false, "Run read only benchmarks.");
|
2018-10-25 05:13:06 +02:00
|
|
|
|
|
|
|
DEFINE_bool(print_malloc_stats, false,
|
|
|
|
"Print malloc stats to stdout after benchmarks finish.");
|
2015-10-14 21:43:00 +02:00
|
|
|
#endif // ROCKSDB_LITE
|
2012-11-06 04:18:49 +01:00
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_bool(disable_auto_compactions, false, "Do not auto trigger compactions");
|
2012-11-21 00:45:41 +01:00
|
|
|
|
2013-11-07 03:46:28 +01:00
|
|
|
DEFINE_uint64(wal_ttl_seconds, 0, "Set the TTL for the WAL Files in seconds.");
|
|
|
|
DEFINE_uint64(wal_size_limit_MB, 0, "Set the size limit for the WAL Files"
|
|
|
|
" in MB.");
|
2014-12-23 03:39:28 +01:00
|
|
|
DEFINE_uint64(max_total_wal_size, 0, "Set total max WAL size");
|
2012-11-26 22:56:45 +01:00
|
|
|
|
2017-04-13 22:07:33 +02:00
|
|
|
DEFINE_bool(mmap_read, rocksdb::Options().allow_mmap_reads,
|
2013-10-24 16:43:14 +02:00
|
|
|
"Allow reads to occur via mmap-ing files");
|
2013-03-15 01:00:04 +01:00
|
|
|
|
2017-04-13 22:07:33 +02:00
|
|
|
DEFINE_bool(mmap_write, rocksdb::Options().allow_mmap_writes,
|
2013-10-24 16:43:14 +02:00
|
|
|
"Allow writes to occur via mmap-ing files");
|
2013-03-15 01:00:04 +01:00
|
|
|
|
2017-04-13 22:07:33 +02:00
|
|
|
DEFINE_bool(use_direct_reads, rocksdb::Options().use_direct_reads,
|
2016-10-28 19:36:05 +02:00
|
|
|
"Use O_DIRECT for reading data");
|
|
|
|
|
2017-04-13 22:07:33 +02:00
|
|
|
DEFINE_bool(use_direct_io_for_flush_and_compaction,
|
|
|
|
rocksdb::Options().use_direct_io_for_flush_and_compaction,
|
2018-05-10 04:26:43 +02:00
|
|
|
"Use O_DIRECT for background flush and compaction writes");
|
2016-12-22 21:51:29 +01:00
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_bool(advise_random_on_open, rocksdb::Options().advise_random_on_open,
|
|
|
|
"Advise random access on table file open");
|
2013-05-18 00:53:01 +02:00
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_string(compaction_fadvice, "NORMAL",
|
|
|
|
"Access pattern advice when a file is compacted");
|
|
|
|
static auto FLAGS_compaction_fadvice_e =
|
2013-10-04 06:49:15 +02:00
|
|
|
rocksdb::Options().access_hint_on_compaction_start;
|
2013-05-18 00:53:01 +02:00
|
|
|
|
2014-02-07 18:47:47 +01:00
|
|
|
DEFINE_bool(use_tailing_iterator, false,
|
|
|
|
"Use tailing iterator to access a series of keys instead of get");
|
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_bool(use_adaptive_mutex, rocksdb::Options().use_adaptive_mutex,
|
|
|
|
"Use adaptive mutex");
|
|
|
|
|
|
|
|
DEFINE_uint64(bytes_per_sync, rocksdb::Options().bytes_per_sync,
|
2015-05-20 01:19:30 +02:00
|
|
|
"Allows OS to incrementally sync SST files to disk while they are"
|
2013-10-24 16:43:14 +02:00
|
|
|
" being written, in the background. Issue one request for every"
|
|
|
|
" bytes_per_sync written. 0 turns it off.");
|
2015-05-20 01:19:30 +02:00
|
|
|
|
|
|
|
DEFINE_uint64(wal_bytes_per_sync, rocksdb::Options().wal_bytes_per_sync,
|
|
|
|
"Allows OS to incrementally sync WAL files to disk while they are"
|
|
|
|
" being written, in the background. Issue one request for every"
|
|
|
|
" wal_bytes_per_sync written. 0 turns it off.");
|
|
|
|
|
Support for SingleDelete()
Summary:
This patch fixes #7460559. It introduces SingleDelete as a new database
operation. This operation can be used to delete keys that were never
overwritten (no put following another put of the same key). If an overwritten
key is single deleted the behavior is undefined. Single deletion of a
non-existent key has no effect but multiple consecutive single deletions are
not allowed (see limitations).
In contrast to the conventional Delete() operation, the deletion entry is
removed along with the value when the two are lined up in a compaction. Note:
The semantics are similar to @igor's prototype that allowed to have this
behavior on the granularity of a column family (
https://reviews.facebook.net/D42093 ). This new patch, however, is more
aggressive when it comes to removing tombstones: It removes the SingleDelete
together with the value whenever there is no snapshot between them while the
older patch only did this when the sequence number of the deletion was older
than the earliest snapshot.
Most of the complex additions are in the Compaction Iterator, all other changes
should be relatively straightforward. The patch also includes basic support for
single deletions in db_stress and db_bench.
Limitations:
- Not compatible with cuckoo hash tables
- Single deletions cannot be used in combination with merges and normal
deletions on the same key (other keys are not affected by this)
- Consecutive single deletions are currently not allowed (and older version of
this patch supported this so it could be resurrected if needed)
Test Plan: make all check
Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor
Reviewed By: igor
Subscribers: maykov, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D43179
2015-09-17 20:42:56 +02:00
|
|
|
DEFINE_bool(use_single_deletes, true,
|
|
|
|
"Use single deletes (used in RandomReplaceKeys only).");
|
|
|
|
|
|
|
|
DEFINE_double(stddev, 2000.0,
|
|
|
|
"Standard deviation of normal distribution used for picking keys"
|
|
|
|
" (used in RandomReplaceKeys only).");
|
|
|
|
|
2016-08-01 21:28:51 +02:00
|
|
|
DEFINE_int32(key_id_range, 100000,
|
|
|
|
"Range of possible value of key id (used in TimeSeries only).");
|
|
|
|
|
|
|
|
DEFINE_string(expire_style, "none",
|
|
|
|
"Style to remove expired time entries. Can be one of the options "
|
|
|
|
"below: none (do not expired data), compaction_filter (use a "
|
|
|
|
"compaction filter to remove expired data), delete (seek IDs and "
|
|
|
|
"remove expired data) (used in TimeSeries only).");
|
|
|
|
|
|
|
|
DEFINE_uint64(
|
|
|
|
time_range, 100000,
|
|
|
|
"Range of timestamp that store in the database (used in TimeSeries"
|
|
|
|
" only).");
|
|
|
|
|
|
|
|
DEFINE_int32(num_deletion_threads, 1,
|
|
|
|
"Number of threads to do deletion (used in TimeSeries and delete "
|
|
|
|
"expire_style only).");
|
|
|
|
|
2014-01-11 02:33:56 +01:00
|
|
|
DEFINE_int32(max_successive_merges, 0, "Maximum number of successive merge"
|
|
|
|
" operations on a key in the memtable");
|
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
static bool ValidatePrefixSize(const char* flagname, int32_t value) {
|
|
|
|
if (value < 0 || value>=2000000000) {
|
|
|
|
fprintf(stderr, "Invalid value for --%s: %d. 0<= PrefixSize <=2000000000\n",
|
|
|
|
flagname, value);
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
return true;
|
|
|
|
}
|
2014-02-12 19:41:55 +01:00
|
|
|
DEFINE_int32(prefix_size, 0, "control the prefix size for HashSkipList and "
|
|
|
|
"plain table");
|
2014-03-05 02:08:05 +01:00
|
|
|
DEFINE_int64(keys_per_prefix, 0, "control average number of keys generated "
|
|
|
|
"per prefix, 0 means no special handling of the prefix, "
|
|
|
|
"i.e. use the prefix comes with the generated random number.");
|
2016-12-02 01:50:59 +01:00
|
|
|
DEFINE_int32(memtable_insert_with_hint_prefix_size, 0,
|
|
|
|
"If non-zero, enable "
|
|
|
|
"memtable insert with hint with the given prefix size.");
|
2014-08-14 05:49:58 +02:00
|
|
|
DEFINE_bool(enable_io_prio, false, "Lower the background flush/compaction "
|
|
|
|
"threads' IO priority");
|
2018-04-24 17:38:01 +02:00
|
|
|
DEFINE_bool(enable_cpu_prio, false, "Lower the background flush/compaction "
|
|
|
|
"threads' CPU priority");
|
CuckooTable: add one option to allow identity function for the first hash function
Summary:
MurmurHash becomes expensive when we do millions Get() a second in one
thread. Add this option to allow the first hash function to use identity
function as hash function. It results in QPS increase from 3.7M/s to
~4.3M/s. I did not observe improvement for end to end RocksDB
performance. This may be caused by other bottlenecks that I will address
in a separate diff.
Test Plan:
```
[ljin@dev1964 rocksdb] ./cuckoo_table_reader_test --enable_perf --file_dir=/dev/shm --write --identity_as_first_hash=0
==== Test CuckooReaderTest.WhenKeyExists
==== Test CuckooReaderTest.WhenKeyExistsWithUint64Comparator
==== Test CuckooReaderTest.CheckIterator
==== Test CuckooReaderTest.CheckIteratorUint64
==== Test CuckooReaderTest.WhenKeyNotFound
==== Test CuckooReaderTest.TestReadPerformance
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.272us (3.7 Mqps) with batch size of 0, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.138us (7.2 Mqps) with batch size of 10, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.142us (7.1 Mqps) with batch size of 25, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.142us (7.0 Mqps) with batch size of 50, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.144us (6.9 Mqps) with batch size of 100, # of found keys 125829120
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.201us (5.0 Mqps) with batch size of 0, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.121us (8.3 Mqps) with batch size of 10, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.123us (8.1 Mqps) with batch size of 25, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.121us (8.3 Mqps) with batch size of 50, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.112us (8.9 Mqps) with batch size of 100, # of found keys 104857600
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.251us (4.0 Mqps) with batch size of 0, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.107us (9.4 Mqps) with batch size of 10, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.099us (10.1 Mqps) with batch size of 25, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.100us (10.0 Mqps) with batch size of 50, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.116us (8.6 Mqps) with batch size of 100, # of found keys 83886080
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.189us (5.3 Mqps) with batch size of 0, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.095us (10.5 Mqps) with batch size of 10, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.096us (10.4 Mqps) with batch size of 25, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.098us (10.2 Mqps) with batch size of 50, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.105us (9.5 Mqps) with batch size of 100, # of found keys 73400320
[ljin@dev1964 rocksdb] ./cuckoo_table_reader_test --enable_perf --file_dir=/dev/shm --write --identity_as_first_hash=1
==== Test CuckooReaderTest.WhenKeyExists
==== Test CuckooReaderTest.WhenKeyExistsWithUint64Comparator
==== Test CuckooReaderTest.CheckIterator
==== Test CuckooReaderTest.CheckIteratorUint64
==== Test CuckooReaderTest.WhenKeyNotFound
==== Test CuckooReaderTest.TestReadPerformance
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.230us (4.3 Mqps) with batch size of 0, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.086us (11.7 Mqps) with batch size of 10, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.088us (11.3 Mqps) with batch size of 25, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.083us (12.1 Mqps) with batch size of 50, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.083us (12.1 Mqps) with batch size of 100, # of found keys 125829120
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.159us (6.3 Mqps) with batch size of 0, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.078us (12.8 Mqps) with batch size of 10, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.080us (12.6 Mqps) with batch size of 25, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.080us (12.5 Mqps) with batch size of 50, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.082us (12.2 Mqps) with batch size of 100, # of found keys 104857600
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.154us (6.5 Mqps) with batch size of 0, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.077us (13.0 Mqps) with batch size of 10, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.077us (12.9 Mqps) with batch size of 25, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.078us (12.8 Mqps) with batch size of 50, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.079us (12.6 Mqps) with batch size of 100, # of found keys 83886080
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.218us (4.6 Mqps) with batch size of 0, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.083us (12.0 Mqps) with batch size of 10, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.085us (11.7 Mqps) with batch size of 25, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.086us (11.6 Mqps) with batch size of 50, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.078us (12.8 Mqps) with batch size of 100, # of found keys 73400320
```
Reviewers: sdong, igor, yhchiang
Reviewed By: igor
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D23451
2014-09-18 20:00:48 +02:00
|
|
|
DEFINE_bool(identity_as_first_hash, false, "the first hash function of cuckoo "
|
|
|
|
"table becomes an identity function. This is only valid when key "
|
|
|
|
"is 8 bytes");
|
2016-04-28 01:23:33 +02:00
|
|
|
DEFINE_bool(dump_malloc_stats, true, "Dump malloc stats in LOG ");
|
2018-07-13 03:30:37 +02:00
|
|
|
DEFINE_uint64(stats_dump_period_sec, rocksdb::Options().stats_dump_period_sec,
|
2018-07-13 00:55:07 +02:00
|
|
|
"Gap between printing stats to log in seconds");
|
2019-02-21 00:46:59 +01:00
|
|
|
DEFINE_uint64(stats_persist_period_sec,
|
|
|
|
rocksdb::Options().stats_persist_period_sec,
|
|
|
|
"Gap between persisting stats in seconds");
|
|
|
|
DEFINE_uint64(stats_history_buffer_size,
|
|
|
|
rocksdb::Options().stats_history_buffer_size,
|
|
|
|
"Max number of stats snapshots to keep in memory");
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
2019-04-11 23:24:09 +02:00
|
|
|
DEFINE_int64(multiread_stride, 0,
|
|
|
|
"Stride length for the keys in a MultiGet batch");
|
|
|
|
DEFINE_bool(multiread_batched, false, "Use the new MultiGet API");
|
2013-08-23 08:10:02 +02:00
|
|
|
|
|
|
|
enum RepFactory {
|
|
|
|
kSkipList,
|
|
|
|
kPrefixHash,
|
2014-03-05 19:27:17 +01:00
|
|
|
kVectorRep,
|
Add a new mem-table representation based on cuckoo hash.
Summary:
= Major Changes =
* Add a new mem-table representation, HashCuckooRep, which is based cuckoo hash.
Cuckoo hash uses multiple hash functions. This allows each key to have multiple
possible locations in the mem-table.
- Put: When insert a key, it will try to find whether one of its possible
locations is vacant and store the key. If none of its possible
locations are available, then it will kick out a victim key and
store at that location. The kicked-out victim key will then be
stored at a vacant space of its possible locations or kick-out
another victim. In this diff, the kick-out path (known as
cuckoo-path) is found using BFS, which guarantees to be the shortest.
- Get: Simply tries all possible locations of a key --- this guarantees
worst-case constant time complexity.
- Time complexity: O(1) for Get, and average O(1) for Put if the
fullness of the mem-table is below 80%.
- Default using two hash functions, the number of hash functions used
by the cuckoo-hash may dynamically increase if it fails to find a
short-enough kick-out path.
- Currently, HashCuckooRep does not support iteration and snapshots,
as our current main purpose of this is to optimize point access.
= Minor Changes =
* Add IsSnapshotSupported() to DB to indicate whether the current DB
supports snapshots. If it returns false, then DB::GetSnapshot() will
always return nullptr.
Test Plan:
Run existing tests. Will develop a test specifically for cuckoo hash in
the next diff.
Reviewers: sdong, haobo
Reviewed By: sdong
CC: leveldb, dhruba, igor
Differential Revision: https://reviews.facebook.net/D16155
2014-04-30 02:13:46 +02:00
|
|
|
kHashLinkedList,
|
2013-08-23 08:10:02 +02:00
|
|
|
};
|
2014-04-10 06:17:14 +02:00
|
|
|
|
2017-02-07 20:34:14 +01:00
|
|
|
static enum RepFactory StringToRepFactory(const char* ctype) {
|
2013-10-24 16:43:14 +02:00
|
|
|
assert(ctype);
|
|
|
|
|
|
|
|
if (!strcasecmp(ctype, "skip_list"))
|
|
|
|
return kSkipList;
|
|
|
|
else if (!strcasecmp(ctype, "prefix_hash"))
|
|
|
|
return kPrefixHash;
|
|
|
|
else if (!strcasecmp(ctype, "vector"))
|
|
|
|
return kVectorRep;
|
2014-03-05 19:27:17 +01:00
|
|
|
else if (!strcasecmp(ctype, "hash_linkedlist"))
|
|
|
|
return kHashLinkedList;
|
2013-10-24 16:43:14 +02:00
|
|
|
|
|
|
|
fprintf(stdout, "Cannot parse memreptable %s\n", ctype);
|
|
|
|
return kSkipList;
|
|
|
|
}
|
2014-04-10 06:17:14 +02:00
|
|
|
|
2013-08-23 08:10:02 +02:00
|
|
|
static enum RepFactory FLAGS_rep_factory;
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_string(memtablerep, "skip_list", "");
|
2014-03-05 19:27:17 +01:00
|
|
|
DEFINE_int64(hash_bucket_count, 1024 * 1024, "hash bucket count");
|
2014-02-12 19:41:55 +01:00
|
|
|
DEFINE_bool(use_plain_table, false, "if use plain table "
|
|
|
|
"instead of block-based table format");
|
2014-08-18 20:59:38 +02:00
|
|
|
DEFINE_bool(use_cuckoo_table, false, "if use cuckoo table format");
|
|
|
|
DEFINE_double(cuckoo_hash_ratio, 0.9, "Hash ratio for Cuckoo SST table.");
|
2014-07-17 02:32:30 +02:00
|
|
|
DEFINE_bool(use_hash_search, false, "if use kHashSearch "
|
|
|
|
"instead of kBinarySearch. "
|
|
|
|
"This is valid if only we use BlockTable");
|
2014-09-08 19:37:05 +02:00
|
|
|
DEFINE_bool(use_block_based_filter, false, "if use kBlockBasedFilter "
|
|
|
|
"instead of kFullFilter for filter block. "
|
|
|
|
"This is valid if only we use BlockTable");
|
2013-10-24 16:43:14 +02:00
|
|
|
DEFINE_string(merge_operator, "", "The merge operator to use with the database."
|
|
|
|
"If a new merge operator is specified, be sure to use fresh"
|
|
|
|
" database The possible merge operators are defined in"
|
|
|
|
" utilities/merge_operators.h");
|
SkipListRep::LookaheadIterator
Summary:
This diff introduces the `lookahead` argument to `SkipListFactory()`. This is an
optimization for the tailing use case which includes many seeks. E.g. consider
the following operations on a skip list iterator:
Seek(x), Next(), Next(), Seek(x+2), Next(), Seek(x+3), Next(), Next(), ...
If `lookahead` is positive, `SkipListRep` will return an iterator which also
keeps track of the previously visited node. Seek() then first does a linear
search starting from that node (up to `lookahead` steps). As in the tailing
example above, this may require fewer than ~log(n) comparisons as with regular
skip list search.
Test Plan:
Added a new benchmark (`fillseekseq`) which simulates the usage pattern. It
first writes N records (with consecutive keys), then measures how much time it
takes to read them by calling `Seek()` and `Next()`.
$ time ./db_bench -num 10000000 -benchmarks fillseekseq -prefix_size 1 \
-key_size 8 -write_buffer_size $[1024*1024*1024] -value_size 50 \
-seekseq_next 2 -skip_list_lookahead=0
[...]
DB path: [/dev/shm/rocksdbtest/dbbench]
fillseekseq : 0.389 micros/op 2569047 ops/sec;
real 0m21.806s
user 0m12.106s
sys 0m9.672s
$ time ./db_bench [...] -skip_list_lookahead=2
[...]
DB path: [/dev/shm/rocksdbtest/dbbench]
fillseekseq : 0.153 micros/op 6540684 ops/sec;
real 0m19.469s
user 0m10.192s
sys 0m9.252s
Reviewers: ljin, sdong, igor
Reviewed By: igor
Subscribers: dhruba, leveldb, march, lovro
Differential Revision: https://reviews.facebook.net/D23997
2014-09-24 00:52:28 +02:00
|
|
|
DEFINE_int32(skip_list_lookahead, 0, "Used with skip_list memtablerep; try "
|
|
|
|
"linear search first for this many steps from the previous "
|
|
|
|
"position");
|
2014-10-29 22:24:34 +01:00
|
|
|
DEFINE_bool(report_file_operations, false, "if report number of file "
|
|
|
|
"operations");
|
Benchmarking for Merge Operator
Summary:
Updated db_bench and utilities/merge_operators.h to allow for dynamic benchmarking
of merge operators in db_bench. Added a new test (--benchmarks=mergerandom), which performs
a bunch of random Merge() operations over random keys. Also added a "--merge_operator=" flag
so that the tester can easily benchmark different merge operators. Currently supports
the PutOperator and UInt64Add operator. Support for stringappend or list append may come later.
Test Plan:
1. make db_bench
2. Test the PutOperator (simulating Put) as follows:
./db_bench --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom --merge_operator=put
--threads=2
3. Test the UInt64AddOperator (simulating numeric addition) similarly:
./db_bench --value_size=8 --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom
--merge_operator=uint64add --threads=2
Reviewers: haobo, dhruba, zshao, MarkCallaghan
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11535
2013-08-16 02:13:07 +02:00
|
|
|
|
2018-02-02 21:14:42 +01:00
|
|
|
static const bool FLAGS_soft_rate_limit_dummy __attribute__((__unused__)) =
|
2014-05-09 02:25:13 +02:00
|
|
|
RegisterFlagValidator(&FLAGS_soft_rate_limit, &ValidateRateLimit);
|
2013-11-17 08:44:39 +01:00
|
|
|
|
2018-02-02 21:14:42 +01:00
|
|
|
static const bool FLAGS_hard_rate_limit_dummy __attribute__((__unused__)) =
|
2014-05-09 02:25:13 +02:00
|
|
|
RegisterFlagValidator(&FLAGS_hard_rate_limit, &ValidateRateLimit);
|
2013-11-17 08:44:39 +01:00
|
|
|
|
2018-02-02 21:14:42 +01:00
|
|
|
static const bool FLAGS_prefix_size_dummy __attribute__((__unused__)) =
|
2014-05-09 02:25:13 +02:00
|
|
|
RegisterFlagValidator(&FLAGS_prefix_size, &ValidatePrefixSize);
|
2013-11-17 08:44:39 +01:00
|
|
|
|
2018-02-02 21:14:42 +01:00
|
|
|
static const bool FLAGS_key_size_dummy __attribute__((__unused__)) =
|
2014-05-09 02:25:13 +02:00
|
|
|
RegisterFlagValidator(&FLAGS_key_size, &ValidateKeySize);
|
2013-11-17 08:44:39 +01:00
|
|
|
|
2018-02-02 21:14:42 +01:00
|
|
|
static const bool FLAGS_cache_numshardbits_dummy __attribute__((__unused__)) =
|
2014-05-09 02:25:13 +02:00
|
|
|
RegisterFlagValidator(&FLAGS_cache_numshardbits,
|
|
|
|
&ValidateCacheNumshardbits);
|
2013-11-17 08:44:39 +01:00
|
|
|
|
2018-02-02 21:14:42 +01:00
|
|
|
static const bool FLAGS_readwritepercent_dummy __attribute__((__unused__)) =
|
2014-05-09 02:25:13 +02:00
|
|
|
RegisterFlagValidator(&FLAGS_readwritepercent, &ValidateInt32Percent);
|
2013-11-17 08:44:39 +01:00
|
|
|
|
2014-06-20 10:23:02 +02:00
|
|
|
DEFINE_int32(disable_seek_compaction, false,
|
|
|
|
"Not used, left here for backwards compatibility");
|
|
|
|
|
2018-02-02 21:14:42 +01:00
|
|
|
static const bool FLAGS_deletepercent_dummy __attribute__((__unused__)) =
|
2014-05-09 02:25:13 +02:00
|
|
|
RegisterFlagValidator(&FLAGS_deletepercent, &ValidateInt32Percent);
|
2018-02-02 21:14:42 +01:00
|
|
|
static const bool FLAGS_table_cache_numshardbits_dummy __attribute__((__unused__)) =
|
2014-05-09 02:25:13 +02:00
|
|
|
RegisterFlagValidator(&FLAGS_table_cache_numshardbits,
|
|
|
|
&ValidateTableCacheNumshardbits);
|
2013-11-17 08:44:39 +01:00
|
|
|
|
2013-10-04 06:49:15 +02:00
|
|
|
namespace rocksdb {
|
2011-03-18 23:37:00 +01:00
|
|
|
|
2014-10-29 22:24:34 +01:00
|
|
|
namespace {
|
|
|
|
struct ReportFileOpCounters {
|
|
|
|
std::atomic<int> open_counter_;
|
|
|
|
std::atomic<int> read_counter_;
|
|
|
|
std::atomic<int> append_counter_;
|
|
|
|
std::atomic<uint64_t> bytes_read_;
|
|
|
|
std::atomic<uint64_t> bytes_written_;
|
|
|
|
};
|
|
|
|
|
|
|
|
// A special Env to records and report file operations in db_bench
|
|
|
|
class ReportFileOpEnv : public EnvWrapper {
|
|
|
|
public:
|
|
|
|
explicit ReportFileOpEnv(Env* base) : EnvWrapper(base) { reset(); }
|
|
|
|
|
|
|
|
void reset() {
|
|
|
|
counters_.open_counter_ = 0;
|
|
|
|
counters_.read_counter_ = 0;
|
|
|
|
counters_.append_counter_ = 0;
|
|
|
|
counters_.bytes_read_ = 0;
|
|
|
|
counters_.bytes_written_ = 0;
|
|
|
|
}
|
|
|
|
|
2018-11-09 20:17:34 +01:00
|
|
|
Status NewSequentialFile(const std::string& f,
|
|
|
|
std::unique_ptr<SequentialFile>* r,
|
2015-02-26 20:28:41 +01:00
|
|
|
const EnvOptions& soptions) override {
|
2014-10-29 22:24:34 +01:00
|
|
|
class CountingFile : public SequentialFile {
|
|
|
|
private:
|
2018-11-09 20:17:34 +01:00
|
|
|
std::unique_ptr<SequentialFile> target_;
|
2014-10-29 22:24:34 +01:00
|
|
|
ReportFileOpCounters* counters_;
|
|
|
|
|
|
|
|
public:
|
2019-03-27 18:18:56 +01:00
|
|
|
CountingFile(std::unique_ptr<SequentialFile>&& target,
|
2014-10-29 22:24:34 +01:00
|
|
|
ReportFileOpCounters* counters)
|
|
|
|
: target_(std::move(target)), counters_(counters) {}
|
|
|
|
|
2019-02-14 22:52:47 +01:00
|
|
|
Status Read(size_t n, Slice* result, char* scratch) override {
|
2014-10-29 22:24:34 +01:00
|
|
|
counters_->read_counter_.fetch_add(1, std::memory_order_relaxed);
|
|
|
|
Status rv = target_->Read(n, result, scratch);
|
|
|
|
counters_->bytes_read_.fetch_add(result->size(),
|
|
|
|
std::memory_order_relaxed);
|
|
|
|
return rv;
|
|
|
|
}
|
|
|
|
|
2019-02-14 22:52:47 +01:00
|
|
|
Status Skip(uint64_t n) override { return target_->Skip(n); }
|
2014-10-29 22:24:34 +01:00
|
|
|
};
|
|
|
|
|
|
|
|
Status s = target()->NewSequentialFile(f, r, soptions);
|
|
|
|
if (s.ok()) {
|
|
|
|
counters()->open_counter_.fetch_add(1, std::memory_order_relaxed);
|
|
|
|
r->reset(new CountingFile(std::move(*r), counters()));
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
Status NewRandomAccessFile(const std::string& f,
|
2018-11-09 20:17:34 +01:00
|
|
|
std::unique_ptr<RandomAccessFile>* r,
|
2015-02-26 20:28:41 +01:00
|
|
|
const EnvOptions& soptions) override {
|
2014-10-29 22:24:34 +01:00
|
|
|
class CountingFile : public RandomAccessFile {
|
|
|
|
private:
|
2018-11-09 20:17:34 +01:00
|
|
|
std::unique_ptr<RandomAccessFile> target_;
|
2014-10-29 22:24:34 +01:00
|
|
|
ReportFileOpCounters* counters_;
|
|
|
|
|
|
|
|
public:
|
2019-03-27 18:18:56 +01:00
|
|
|
CountingFile(std::unique_ptr<RandomAccessFile>&& target,
|
2014-10-29 22:24:34 +01:00
|
|
|
ReportFileOpCounters* counters)
|
|
|
|
: target_(std::move(target)), counters_(counters) {}
|
2019-02-14 22:52:47 +01:00
|
|
|
Status Read(uint64_t offset, size_t n, Slice* result,
|
|
|
|
char* scratch) const override {
|
2014-10-29 22:24:34 +01:00
|
|
|
counters_->read_counter_.fetch_add(1, std::memory_order_relaxed);
|
|
|
|
Status rv = target_->Read(offset, n, result, scratch);
|
|
|
|
counters_->bytes_read_.fetch_add(result->size(),
|
|
|
|
std::memory_order_relaxed);
|
|
|
|
return rv;
|
|
|
|
}
|
|
|
|
};
|
|
|
|
|
|
|
|
Status s = target()->NewRandomAccessFile(f, r, soptions);
|
|
|
|
if (s.ok()) {
|
|
|
|
counters()->open_counter_.fetch_add(1, std::memory_order_relaxed);
|
|
|
|
r->reset(new CountingFile(std::move(*r), counters()));
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
2018-11-09 20:17:34 +01:00
|
|
|
Status NewWritableFile(const std::string& f, std::unique_ptr<WritableFile>* r,
|
2015-02-26 20:28:41 +01:00
|
|
|
const EnvOptions& soptions) override {
|
2014-10-29 22:24:34 +01:00
|
|
|
class CountingFile : public WritableFile {
|
|
|
|
private:
|
2018-11-09 20:17:34 +01:00
|
|
|
std::unique_ptr<WritableFile> target_;
|
2014-10-29 22:24:34 +01:00
|
|
|
ReportFileOpCounters* counters_;
|
|
|
|
|
|
|
|
public:
|
2019-03-27 18:18:56 +01:00
|
|
|
CountingFile(std::unique_ptr<WritableFile>&& target,
|
2014-10-29 22:24:34 +01:00
|
|
|
ReportFileOpCounters* counters)
|
|
|
|
: target_(std::move(target)), counters_(counters) {}
|
|
|
|
|
2015-02-26 20:28:41 +01:00
|
|
|
Status Append(const Slice& data) override {
|
2014-10-29 22:24:34 +01:00
|
|
|
counters_->append_counter_.fetch_add(1, std::memory_order_relaxed);
|
|
|
|
Status rv = target_->Append(data);
|
|
|
|
counters_->bytes_written_.fetch_add(data.size(),
|
|
|
|
std::memory_order_relaxed);
|
|
|
|
return rv;
|
|
|
|
}
|
|
|
|
|
2015-09-11 18:57:02 +02:00
|
|
|
Status Truncate(uint64_t size) override { return target_->Truncate(size); }
|
2015-02-26 20:28:41 +01:00
|
|
|
Status Close() override { return target_->Close(); }
|
|
|
|
Status Flush() override { return target_->Flush(); }
|
|
|
|
Status Sync() override { return target_->Sync(); }
|
2014-10-29 22:24:34 +01:00
|
|
|
};
|
|
|
|
|
|
|
|
Status s = target()->NewWritableFile(f, r, soptions);
|
|
|
|
if (s.ok()) {
|
|
|
|
counters()->open_counter_.fetch_add(1, std::memory_order_relaxed);
|
|
|
|
r->reset(new CountingFile(std::move(*r), counters()));
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
// getter
|
|
|
|
ReportFileOpCounters* counters() { return &counters_; }
|
|
|
|
|
|
|
|
private:
|
|
|
|
ReportFileOpCounters counters_;
|
|
|
|
};
|
|
|
|
|
|
|
|
} // namespace
|
|
|
|
|
2011-08-22 23:08:51 +02:00
|
|
|
// Helper for quickly generating random data.
|
2011-03-18 23:37:00 +01:00
|
|
|
class RandomGenerator {
|
|
|
|
private:
|
|
|
|
std::string data_;
|
2012-11-06 21:02:18 +01:00
|
|
|
unsigned int pos_;
|
2011-03-18 23:37:00 +01:00
|
|
|
|
|
|
|
public:
|
|
|
|
RandomGenerator() {
|
|
|
|
// We use a limited amount of data over and over again and ensure
|
|
|
|
// that it is larger than the compression window (32KB), and also
|
|
|
|
// large enough to serve all typical value sizes we want to write.
|
|
|
|
Random rnd(301);
|
|
|
|
std::string piece;
|
2013-03-15 02:32:01 +01:00
|
|
|
while (data_.size() < (unsigned)std::max(1048576, FLAGS_value_size)) {
|
2011-03-18 23:37:00 +01:00
|
|
|
// Add a short fragment that is as compressible as specified
|
|
|
|
// by FLAGS_compression_ratio.
|
|
|
|
test::CompressibleString(&rnd, FLAGS_compression_ratio, 100, &piece);
|
|
|
|
data_.append(piece);
|
|
|
|
}
|
|
|
|
pos_ = 0;
|
|
|
|
}
|
|
|
|
|
2013-03-15 02:32:01 +01:00
|
|
|
Slice Generate(unsigned int len) {
|
2014-06-24 23:29:28 +02:00
|
|
|
assert(len <= data_.size());
|
2011-03-18 23:37:00 +01:00
|
|
|
if (pos_ + len > data_.size()) {
|
|
|
|
pos_ = 0;
|
|
|
|
}
|
|
|
|
pos_ += len;
|
|
|
|
return Slice(data_.data() + pos_ - len, len);
|
|
|
|
}
|
2017-05-10 23:54:35 +02:00
|
|
|
|
|
|
|
Slice GenerateWithTTL(unsigned int len) {
|
|
|
|
assert(len <= data_.size());
|
|
|
|
if (pos_ + len > data_.size()) {
|
|
|
|
pos_ = 0;
|
|
|
|
}
|
|
|
|
pos_ += len;
|
|
|
|
return Slice(data_.data() + pos_ - len, len);
|
|
|
|
}
|
2012-09-26 17:06:19 +02:00
|
|
|
};
|
2013-08-13 22:58:02 +02:00
|
|
|
|
2011-08-22 23:08:51 +02:00
|
|
|
static void AppendWithSpace(std::string* str, Slice msg) {
|
|
|
|
if (msg.empty()) return;
|
|
|
|
if (!str->empty()) {
|
|
|
|
str->push_back(' ');
|
|
|
|
}
|
|
|
|
str->append(msg.data(), msg.size());
|
|
|
|
}
|
|
|
|
|
2014-09-29 18:50:41 +02:00
|
|
|
struct DBWithColumnFamilies {
|
|
|
|
std::vector<ColumnFamilyHandle*> cfh;
|
|
|
|
DB* db;
|
2015-10-14 21:43:00 +02:00
|
|
|
#ifndef ROCKSDB_LITE
|
Pessimistic Transactions
Summary:
Initial implementation of Pessimistic Transactions. This diff contains the api changes discussed in D38913. This diff is pretty large, so let me know if people would prefer to meet up to discuss it.
MyRocks folks: please take a look at the API in include/rocksdb/utilities/transaction[_db].h and let me know if you have any issues.
Also, you'll notice a couple of TODOs in the implementation of RollbackToSavePoint(). After chatting with Siying, I'm going to send out a separate diff for an alternate implementation of this feature that implements the rollback inside of WriteBatch/WriteBatchWithIndex. We can then decide which route is preferable.
Next, I'm planning on doing some perf testing and then integrating this diff into MongoRocks for further testing.
Test Plan: Unit tests, db_bench parallel testing.
Reviewers: igor, rven, sdong, yhchiang, yoshinorim
Reviewed By: sdong
Subscribers: hermanlee4, maykov, spetrunia, leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D40869
2015-05-26 02:37:33 +02:00
|
|
|
OptimisticTransactionDB* opt_txn_db;
|
2015-10-14 21:43:00 +02:00
|
|
|
#endif // ROCKSDB_LITE
|
2014-12-23 03:39:28 +01:00
|
|
|
std::atomic<size_t> num_created; // Need to be updated after all the
|
|
|
|
// new entries in cfh are set.
|
|
|
|
size_t num_hot; // Number of column families to be queried at each moment.
|
|
|
|
// After each CreateNewCf(), another num_hot number of new
|
|
|
|
// Column families will be created and used to be queried.
|
|
|
|
port::Mutex create_cf_mutex; // Only one thread can execute CreateNewCf()
|
2017-08-11 22:46:41 +02:00
|
|
|
std::vector<int> cfh_idx_to_prob; // ith index holds probability of operating
|
|
|
|
// on cfh[i].
|
2014-12-23 03:39:28 +01:00
|
|
|
|
2015-10-14 21:43:00 +02:00
|
|
|
DBWithColumnFamilies()
|
|
|
|
: db(nullptr)
|
|
|
|
#ifndef ROCKSDB_LITE
|
|
|
|
, opt_txn_db(nullptr)
|
|
|
|
#endif // ROCKSDB_LITE
|
|
|
|
{
|
2014-09-29 18:50:41 +02:00
|
|
|
cfh.clear();
|
2015-03-12 01:57:16 +01:00
|
|
|
num_created = 0;
|
|
|
|
num_hot = 0;
|
2014-09-29 18:50:41 +02:00
|
|
|
}
|
2014-12-23 03:39:28 +01:00
|
|
|
|
|
|
|
DBWithColumnFamilies(const DBWithColumnFamilies& other)
|
|
|
|
: cfh(other.cfh),
|
|
|
|
db(other.db),
|
2015-10-14 21:43:00 +02:00
|
|
|
#ifndef ROCKSDB_LITE
|
Pessimistic Transactions
Summary:
Initial implementation of Pessimistic Transactions. This diff contains the api changes discussed in D38913. This diff is pretty large, so let me know if people would prefer to meet up to discuss it.
MyRocks folks: please take a look at the API in include/rocksdb/utilities/transaction[_db].h and let me know if you have any issues.
Also, you'll notice a couple of TODOs in the implementation of RollbackToSavePoint(). After chatting with Siying, I'm going to send out a separate diff for an alternate implementation of this feature that implements the rollback inside of WriteBatch/WriteBatchWithIndex. We can then decide which route is preferable.
Next, I'm planning on doing some perf testing and then integrating this diff into MongoRocks for further testing.
Test Plan: Unit tests, db_bench parallel testing.
Reviewers: igor, rven, sdong, yhchiang, yoshinorim
Reviewed By: sdong
Subscribers: hermanlee4, maykov, spetrunia, leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D40869
2015-05-26 02:37:33 +02:00
|
|
|
opt_txn_db(other.opt_txn_db),
|
2015-10-14 21:43:00 +02:00
|
|
|
#endif // ROCKSDB_LITE
|
2014-12-23 03:39:28 +01:00
|
|
|
num_created(other.num_created.load()),
|
2017-08-11 22:46:41 +02:00
|
|
|
num_hot(other.num_hot),
|
|
|
|
cfh_idx_to_prob(other.cfh_idx_to_prob) {
|
|
|
|
}
|
2014-12-23 03:39:28 +01:00
|
|
|
|
2015-05-29 23:36:35 +02:00
|
|
|
void DeleteDBs() {
|
|
|
|
std::for_each(cfh.begin(), cfh.end(),
|
|
|
|
[](ColumnFamilyHandle* cfhi) { delete cfhi; });
|
|
|
|
cfh.clear();
|
2015-10-14 21:43:00 +02:00
|
|
|
#ifndef ROCKSDB_LITE
|
Pessimistic Transactions
Summary:
Initial implementation of Pessimistic Transactions. This diff contains the api changes discussed in D38913. This diff is pretty large, so let me know if people would prefer to meet up to discuss it.
MyRocks folks: please take a look at the API in include/rocksdb/utilities/transaction[_db].h and let me know if you have any issues.
Also, you'll notice a couple of TODOs in the implementation of RollbackToSavePoint(). After chatting with Siying, I'm going to send out a separate diff for an alternate implementation of this feature that implements the rollback inside of WriteBatch/WriteBatchWithIndex. We can then decide which route is preferable.
Next, I'm planning on doing some perf testing and then integrating this diff into MongoRocks for further testing.
Test Plan: Unit tests, db_bench parallel testing.
Reviewers: igor, rven, sdong, yhchiang, yoshinorim
Reviewed By: sdong
Subscribers: hermanlee4, maykov, spetrunia, leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D40869
2015-05-26 02:37:33 +02:00
|
|
|
if (opt_txn_db) {
|
|
|
|
delete opt_txn_db;
|
|
|
|
opt_txn_db = nullptr;
|
2015-05-29 23:36:35 +02:00
|
|
|
} else {
|
|
|
|
delete db;
|
2015-10-14 21:43:00 +02:00
|
|
|
db = nullptr;
|
2015-05-29 23:36:35 +02:00
|
|
|
}
|
2015-10-14 21:43:00 +02:00
|
|
|
#else
|
|
|
|
delete db;
|
2015-05-29 23:36:35 +02:00
|
|
|
db = nullptr;
|
2015-10-14 21:43:00 +02:00
|
|
|
#endif // ROCKSDB_LITE
|
2015-05-29 23:36:35 +02:00
|
|
|
}
|
|
|
|
|
2014-12-23 03:39:28 +01:00
|
|
|
ColumnFamilyHandle* GetCfh(int64_t rand_num) {
|
|
|
|
assert(num_hot > 0);
|
2017-08-11 22:46:41 +02:00
|
|
|
size_t rand_offset = 0;
|
|
|
|
if (!cfh_idx_to_prob.empty()) {
|
|
|
|
assert(cfh_idx_to_prob.size() == num_hot);
|
|
|
|
int sum = 0;
|
|
|
|
while (sum + cfh_idx_to_prob[rand_offset] < rand_num % 100) {
|
|
|
|
sum += cfh_idx_to_prob[rand_offset];
|
|
|
|
++rand_offset;
|
|
|
|
}
|
|
|
|
assert(rand_offset < cfh_idx_to_prob.size());
|
|
|
|
} else {
|
|
|
|
rand_offset = rand_num % num_hot;
|
|
|
|
}
|
2014-12-23 03:39:28 +01:00
|
|
|
return cfh[num_created.load(std::memory_order_acquire) - num_hot +
|
2017-08-11 22:46:41 +02:00
|
|
|
rand_offset];
|
2014-12-23 03:39:28 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
// stage: assume CF from 0 to stage * num_hot has be created. Need to create
|
|
|
|
// stage * num_hot + 1 to stage * (num_hot + 1).
|
|
|
|
void CreateNewCf(ColumnFamilyOptions options, int64_t stage) {
|
|
|
|
MutexLock l(&create_cf_mutex);
|
|
|
|
if ((stage + 1) * num_hot <= num_created) {
|
|
|
|
// Already created.
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
auto new_num_created = num_created + num_hot;
|
|
|
|
assert(new_num_created <= cfh.size());
|
|
|
|
for (size_t i = num_created; i < new_num_created; i++) {
|
|
|
|
Status s =
|
|
|
|
db->CreateColumnFamily(options, ColumnFamilyName(i), &(cfh[i]));
|
|
|
|
if (!s.ok()) {
|
|
|
|
fprintf(stderr, "create column family error: %s\n",
|
|
|
|
s.ToString().c_str());
|
|
|
|
abort();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
num_created.store(new_num_created, std::memory_order_release);
|
|
|
|
}
|
2014-09-29 18:50:41 +02:00
|
|
|
};
|
|
|
|
|
db_bench periodically writes QPS to CSV file
Summary:
This is part of an effort to better understand and optimize RocksDB stalls under high load. I added a feature to db_bench to periodically write QPS to CSV files. That way we can nicely see how our QPS changes in time (especially when DB is stalled) and can do a better job of evaluating our stall system (i.e. we want the QPS to be as constant as possible, as opposed to having bunch of stalls)
Cool part of CSV files is that we can easily graph them -- there are a bunch of tools available.
Test Plan:
Ran ./db_bench --report_interval_seconds=10 --benchmarks=fillrandom --num=10000000
and observed this in report.csv:
secs_elapsed,interval_qps
10,2725860
20,1980480
30,1863456
40,1454359
50,1460389
Reviewers: sdong, MarkCallaghan, rven, yhchiang
Reviewed By: yhchiang
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D40047
2015-06-12 23:31:53 +02:00
|
|
|
// a class that reports stats to CSV file
|
|
|
|
class ReporterAgent {
|
|
|
|
public:
|
|
|
|
ReporterAgent(Env* env, const std::string& fname,
|
|
|
|
uint64_t report_interval_secs)
|
|
|
|
: env_(env),
|
|
|
|
total_ops_done_(0),
|
|
|
|
last_report_(0),
|
|
|
|
report_interval_secs_(report_interval_secs),
|
|
|
|
stop_(false) {
|
|
|
|
auto s = env_->NewWritableFile(fname, &report_file_, EnvOptions());
|
|
|
|
if (s.ok()) {
|
|
|
|
s = report_file_->Append(Header() + "\n");
|
|
|
|
}
|
|
|
|
if (s.ok()) {
|
|
|
|
s = report_file_->Flush();
|
|
|
|
}
|
|
|
|
if (!s.ok()) {
|
|
|
|
fprintf(stderr, "Can't open %s: %s\n", fname.c_str(),
|
|
|
|
s.ToString().c_str());
|
|
|
|
abort();
|
|
|
|
}
|
|
|
|
|
2017-02-06 23:43:55 +01:00
|
|
|
reporting_thread_ = port::Thread([&]() { SleepAndReport(); });
|
db_bench periodically writes QPS to CSV file
Summary:
This is part of an effort to better understand and optimize RocksDB stalls under high load. I added a feature to db_bench to periodically write QPS to CSV files. That way we can nicely see how our QPS changes in time (especially when DB is stalled) and can do a better job of evaluating our stall system (i.e. we want the QPS to be as constant as possible, as opposed to having bunch of stalls)
Cool part of CSV files is that we can easily graph them -- there are a bunch of tools available.
Test Plan:
Ran ./db_bench --report_interval_seconds=10 --benchmarks=fillrandom --num=10000000
and observed this in report.csv:
secs_elapsed,interval_qps
10,2725860
20,1980480
30,1863456
40,1454359
50,1460389
Reviewers: sdong, MarkCallaghan, rven, yhchiang
Reviewed By: yhchiang
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D40047
2015-06-12 23:31:53 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
~ReporterAgent() {
|
|
|
|
{
|
|
|
|
std::unique_lock<std::mutex> lk(mutex_);
|
|
|
|
stop_ = true;
|
|
|
|
stop_cv_.notify_all();
|
|
|
|
}
|
|
|
|
reporting_thread_.join();
|
|
|
|
}
|
|
|
|
|
|
|
|
// thread safe
|
|
|
|
void ReportFinishedOps(int64_t num_ops) {
|
|
|
|
total_ops_done_.fetch_add(num_ops);
|
|
|
|
}
|
|
|
|
|
|
|
|
private:
|
|
|
|
std::string Header() const { return "secs_elapsed,interval_qps"; }
|
|
|
|
void SleepAndReport() {
|
|
|
|
uint64_t kMicrosInSecond = 1000 * 1000;
|
|
|
|
auto time_started = env_->NowMicros();
|
|
|
|
while (true) {
|
|
|
|
{
|
|
|
|
std::unique_lock<std::mutex> lk(mutex_);
|
|
|
|
if (stop_ ||
|
|
|
|
stop_cv_.wait_for(lk, std::chrono::seconds(report_interval_secs_),
|
|
|
|
[&]() { return stop_; })) {
|
|
|
|
// stopping
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
// else -> timeout, which means time for a report!
|
|
|
|
}
|
|
|
|
auto total_ops_done_snapshot = total_ops_done_.load();
|
|
|
|
// round the seconds elapsed
|
|
|
|
auto secs_elapsed =
|
|
|
|
(env_->NowMicros() - time_started + kMicrosInSecond / 2) /
|
|
|
|
kMicrosInSecond;
|
|
|
|
std::string report = ToString(secs_elapsed) + "," +
|
|
|
|
ToString(total_ops_done_snapshot - last_report_) +
|
|
|
|
"\n";
|
|
|
|
auto s = report_file_->Append(report);
|
|
|
|
if (s.ok()) {
|
|
|
|
s = report_file_->Flush();
|
|
|
|
}
|
|
|
|
if (!s.ok()) {
|
|
|
|
fprintf(stderr,
|
|
|
|
"Can't write to report file (%s), stopping the reporting\n",
|
|
|
|
s.ToString().c_str());
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
last_report_ = total_ops_done_snapshot;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
Env* env_;
|
|
|
|
std::unique_ptr<WritableFile> report_file_;
|
|
|
|
std::atomic<int64_t> total_ops_done_;
|
|
|
|
int64_t last_report_;
|
|
|
|
const uint64_t report_interval_secs_;
|
2017-02-06 23:43:55 +01:00
|
|
|
rocksdb::port::Thread reporting_thread_;
|
db_bench periodically writes QPS to CSV file
Summary:
This is part of an effort to better understand and optimize RocksDB stalls under high load. I added a feature to db_bench to periodically write QPS to CSV files. That way we can nicely see how our QPS changes in time (especially when DB is stalled) and can do a better job of evaluating our stall system (i.e. we want the QPS to be as constant as possible, as opposed to having bunch of stalls)
Cool part of CSV files is that we can easily graph them -- there are a bunch of tools available.
Test Plan:
Ran ./db_bench --report_interval_seconds=10 --benchmarks=fillrandom --num=10000000
and observed this in report.csv:
secs_elapsed,interval_qps
10,2725860
20,1980480
30,1863456
40,1454359
50,1460389
Reviewers: sdong, MarkCallaghan, rven, yhchiang
Reviewed By: yhchiang
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D40047
2015-06-12 23:31:53 +02:00
|
|
|
std::mutex mutex_;
|
|
|
|
// will notify on stop
|
|
|
|
std::condition_variable stop_cv_;
|
|
|
|
bool stop_;
|
|
|
|
};
|
|
|
|
|
2015-12-08 02:33:18 +01:00
|
|
|
enum OperationType : unsigned char {
|
|
|
|
kRead = 0,
|
|
|
|
kWrite,
|
|
|
|
kDelete,
|
|
|
|
kSeek,
|
|
|
|
kMerge,
|
|
|
|
kUpdate,
|
|
|
|
kCompress,
|
|
|
|
kUncompress,
|
|
|
|
kCrc,
|
|
|
|
kHash,
|
|
|
|
kOthers
|
|
|
|
};
|
|
|
|
|
|
|
|
static std::unordered_map<OperationType, std::string, std::hash<unsigned char>>
|
|
|
|
OperationTypeString = {
|
|
|
|
{kRead, "read"},
|
|
|
|
{kWrite, "write"},
|
|
|
|
{kDelete, "delete"},
|
|
|
|
{kSeek, "seek"},
|
|
|
|
{kMerge, "merge"},
|
|
|
|
{kUpdate, "update"},
|
|
|
|
{kCompress, "compress"},
|
|
|
|
{kCompress, "uncompress"},
|
|
|
|
{kCrc, "crc"},
|
|
|
|
{kHash, "hash"},
|
|
|
|
{kOthers, "op"}
|
|
|
|
};
|
|
|
|
|
2016-08-25 21:57:35 +02:00
|
|
|
class CombinedStats;
|
2011-08-22 23:08:51 +02:00
|
|
|
class Stats {
|
|
|
|
private:
|
2012-10-03 15:51:11 +02:00
|
|
|
int id_;
|
2015-11-19 20:47:12 +01:00
|
|
|
uint64_t start_;
|
2018-06-15 21:07:38 +02:00
|
|
|
uint64_t sine_interval_;
|
2015-11-19 20:47:12 +01:00
|
|
|
uint64_t finish_;
|
2011-08-22 23:08:51 +02:00
|
|
|
double seconds_;
|
2015-11-19 20:47:12 +01:00
|
|
|
uint64_t done_;
|
|
|
|
uint64_t last_report_done_;
|
|
|
|
uint64_t next_report_;
|
|
|
|
uint64_t bytes_;
|
|
|
|
uint64_t last_op_finish_;
|
|
|
|
uint64_t last_report_finish_;
|
2016-03-12 01:54:25 +01:00
|
|
|
std::unordered_map<OperationType, std::shared_ptr<HistogramImpl>,
|
2015-12-08 02:33:18 +01:00
|
|
|
std::hash<unsigned char>> hist_;
|
2011-08-22 23:08:51 +02:00
|
|
|
std::string message_;
|
2013-03-19 14:36:03 +01:00
|
|
|
bool exclude_from_merge_;
|
db_bench periodically writes QPS to CSV file
Summary:
This is part of an effort to better understand and optimize RocksDB stalls under high load. I added a feature to db_bench to periodically write QPS to CSV files. That way we can nicely see how our QPS changes in time (especially when DB is stalled) and can do a better job of evaluating our stall system (i.e. we want the QPS to be as constant as possible, as opposed to having bunch of stalls)
Cool part of CSV files is that we can easily graph them -- there are a bunch of tools available.
Test Plan:
Ran ./db_bench --report_interval_seconds=10 --benchmarks=fillrandom --num=10000000
and observed this in report.csv:
secs_elapsed,interval_qps
10,2725860
20,1980480
30,1863456
40,1454359
50,1460389
Reviewers: sdong, MarkCallaghan, rven, yhchiang
Reviewed By: yhchiang
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D40047
2015-06-12 23:31:53 +02:00
|
|
|
ReporterAgent* reporter_agent_; // does not own
|
2016-08-25 21:57:35 +02:00
|
|
|
friend class CombinedStats;
|
2011-08-22 23:08:51 +02:00
|
|
|
|
|
|
|
public:
|
2012-10-03 15:51:11 +02:00
|
|
|
Stats() { Start(-1); }
|
2011-08-22 23:08:51 +02:00
|
|
|
|
db_bench periodically writes QPS to CSV file
Summary:
This is part of an effort to better understand and optimize RocksDB stalls under high load. I added a feature to db_bench to periodically write QPS to CSV files. That way we can nicely see how our QPS changes in time (especially when DB is stalled) and can do a better job of evaluating our stall system (i.e. we want the QPS to be as constant as possible, as opposed to having bunch of stalls)
Cool part of CSV files is that we can easily graph them -- there are a bunch of tools available.
Test Plan:
Ran ./db_bench --report_interval_seconds=10 --benchmarks=fillrandom --num=10000000
and observed this in report.csv:
secs_elapsed,interval_qps
10,2725860
20,1980480
30,1863456
40,1454359
50,1460389
Reviewers: sdong, MarkCallaghan, rven, yhchiang
Reviewed By: yhchiang
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D40047
2015-06-12 23:31:53 +02:00
|
|
|
void SetReporterAgent(ReporterAgent* reporter_agent) {
|
|
|
|
reporter_agent_ = reporter_agent;
|
|
|
|
}
|
|
|
|
|
2012-10-03 15:51:11 +02:00
|
|
|
void Start(int id) {
|
|
|
|
id_ = id;
|
|
|
|
next_report_ = FLAGS_stats_interval ? FLAGS_stats_interval : 100;
|
2011-08-22 23:08:51 +02:00
|
|
|
last_op_finish_ = start_;
|
2015-12-08 02:33:18 +01:00
|
|
|
hist_.clear();
|
2011-08-22 23:08:51 +02:00
|
|
|
done_ = 0;
|
2012-10-03 15:51:11 +02:00
|
|
|
last_report_done_ = 0;
|
2011-08-22 23:08:51 +02:00
|
|
|
bytes_ = 0;
|
|
|
|
seconds_ = 0;
|
2012-06-08 10:11:14 +02:00
|
|
|
start_ = FLAGS_env->NowMicros();
|
2018-06-15 21:07:38 +02:00
|
|
|
sine_interval_ = FLAGS_env->NowMicros();
|
2011-08-22 23:08:51 +02:00
|
|
|
finish_ = start_;
|
2012-10-03 15:51:11 +02:00
|
|
|
last_report_finish_ = start_;
|
2011-08-22 23:08:51 +02:00
|
|
|
message_.clear();
|
2013-03-19 14:36:03 +01:00
|
|
|
// When set, stats from this thread won't be merged with others.
|
|
|
|
exclude_from_merge_ = false;
|
2011-08-22 23:08:51 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
void Merge(const Stats& other) {
|
2013-03-19 14:36:03 +01:00
|
|
|
if (other.exclude_from_merge_)
|
|
|
|
return;
|
|
|
|
|
2015-12-08 02:33:18 +01:00
|
|
|
for (auto it = other.hist_.begin(); it != other.hist_.end(); ++it) {
|
|
|
|
auto this_it = hist_.find(it->first);
|
|
|
|
if (this_it != hist_.end()) {
|
2016-03-12 01:54:25 +01:00
|
|
|
this_it->second->Merge(*(other.hist_.at(it->first)));
|
2015-12-08 02:33:18 +01:00
|
|
|
} else {
|
|
|
|
hist_.insert({ it->first, it->second });
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2011-08-22 23:08:51 +02:00
|
|
|
done_ += other.done_;
|
|
|
|
bytes_ += other.bytes_;
|
|
|
|
seconds_ += other.seconds_;
|
|
|
|
if (other.start_ < start_) start_ = other.start_;
|
|
|
|
if (other.finish_ > finish_) finish_ = other.finish_;
|
|
|
|
|
|
|
|
// Just keep the messages from one thread
|
|
|
|
if (message_.empty()) message_ = other.message_;
|
|
|
|
}
|
|
|
|
|
|
|
|
void Stop() {
|
2012-06-08 10:11:14 +02:00
|
|
|
finish_ = FLAGS_env->NowMicros();
|
2011-08-22 23:08:51 +02:00
|
|
|
seconds_ = (finish_ - start_) * 1e-6;
|
|
|
|
}
|
|
|
|
|
|
|
|
void AddMessage(Slice msg) {
|
|
|
|
AppendWithSpace(&message_, msg);
|
|
|
|
}
|
|
|
|
|
2012-10-03 15:51:11 +02:00
|
|
|
void SetId(int id) { id_ = id; }
|
2013-03-19 14:36:03 +01:00
|
|
|
void SetExcludeFromMerge() { exclude_from_merge_ = true; }
|
2012-10-03 15:51:11 +02:00
|
|
|
|
2015-03-06 20:22:06 +01:00
|
|
|
void PrintThreadStatus() {
|
|
|
|
std::vector<ThreadStatus> thread_list;
|
|
|
|
FLAGS_env->GetThreadList(&thread_list);
|
|
|
|
|
2015-05-07 07:50:35 +02:00
|
|
|
fprintf(stderr, "\n%18s %10s %12s %20s %13s %45s %12s %s\n",
|
2015-03-10 22:51:28 +01:00
|
|
|
"ThreadID", "ThreadType", "cfName", "Operation",
|
2015-05-07 07:50:35 +02:00
|
|
|
"ElapsedTime", "Stage", "State", "OperationProperties");
|
2015-03-06 20:22:06 +01:00
|
|
|
|
2015-03-10 22:51:28 +01:00
|
|
|
int64_t current_time = 0;
|
|
|
|
Env::Default()->GetCurrentTime(¤t_time);
|
2015-03-06 20:22:06 +01:00
|
|
|
for (auto ts : thread_list) {
|
2015-05-07 07:50:35 +02:00
|
|
|
fprintf(stderr, "%18" PRIu64 " %10s %12s %20s %13s %45s %12s",
|
2015-03-06 20:22:06 +01:00
|
|
|
ts.thread_id,
|
|
|
|
ThreadStatus::GetThreadTypeName(ts.thread_type).c_str(),
|
|
|
|
ts.cf_name.c_str(),
|
|
|
|
ThreadStatus::GetOperationName(ts.operation_type).c_str(),
|
2015-03-24 00:35:04 +01:00
|
|
|
ThreadStatus::MicrosToString(ts.op_elapsed_micros).c_str(),
|
2015-03-13 18:45:40 +01:00
|
|
|
ThreadStatus::GetOperationStageName(ts.operation_stage).c_str(),
|
2015-03-06 20:22:06 +01:00
|
|
|
ThreadStatus::GetStateName(ts.state_type).c_str());
|
2015-05-07 07:50:35 +02:00
|
|
|
|
|
|
|
auto op_properties = ThreadStatus::InterpretOperationProperties(
|
|
|
|
ts.operation_type, ts.op_properties);
|
|
|
|
for (const auto& op_prop : op_properties) {
|
|
|
|
fprintf(stderr, " %s %" PRIu64" |",
|
|
|
|
op_prop.first.c_str(), op_prop.second);
|
|
|
|
}
|
|
|
|
fprintf(stderr, "\n");
|
2015-03-06 20:22:06 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2018-06-15 21:07:38 +02:00
|
|
|
void ResetSineInterval() {
|
|
|
|
sine_interval_ = FLAGS_env->NowMicros();
|
|
|
|
}
|
|
|
|
|
|
|
|
uint64_t GetSineInterval() {
|
|
|
|
return sine_interval_;
|
|
|
|
}
|
|
|
|
|
|
|
|
uint64_t GetStart() {
|
|
|
|
return start_;
|
|
|
|
}
|
|
|
|
|
2016-01-04 21:01:27 +01:00
|
|
|
void ResetLastOpTime() {
|
|
|
|
// Set to now to avoid latency from calls to SleepForMicroseconds
|
|
|
|
last_op_finish_ = FLAGS_env->NowMicros();
|
|
|
|
}
|
|
|
|
|
2015-12-08 02:33:18 +01:00
|
|
|
void FinishedOps(DBWithColumnFamilies* db_with_cfh, DB* db, int64_t num_ops,
|
|
|
|
enum OperationType op_type = kOthers) {
|
db_bench periodically writes QPS to CSV file
Summary:
This is part of an effort to better understand and optimize RocksDB stalls under high load. I added a feature to db_bench to periodically write QPS to CSV files. That way we can nicely see how our QPS changes in time (especially when DB is stalled) and can do a better job of evaluating our stall system (i.e. we want the QPS to be as constant as possible, as opposed to having bunch of stalls)
Cool part of CSV files is that we can easily graph them -- there are a bunch of tools available.
Test Plan:
Ran ./db_bench --report_interval_seconds=10 --benchmarks=fillrandom --num=10000000
and observed this in report.csv:
secs_elapsed,interval_qps
10,2725860
20,1980480
30,1863456
40,1454359
50,1460389
Reviewers: sdong, MarkCallaghan, rven, yhchiang
Reviewed By: yhchiang
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D40047
2015-06-12 23:31:53 +02:00
|
|
|
if (reporter_agent_) {
|
|
|
|
reporter_agent_->ReportFinishedOps(num_ops);
|
|
|
|
}
|
2011-08-22 23:08:51 +02:00
|
|
|
if (FLAGS_histogram) {
|
2015-11-19 20:47:12 +01:00
|
|
|
uint64_t now = FLAGS_env->NowMicros();
|
|
|
|
uint64_t micros = now - last_op_finish_;
|
2015-12-08 02:33:18 +01:00
|
|
|
|
|
|
|
if (hist_.find(op_type) == hist_.end())
|
|
|
|
{
|
2016-03-12 01:54:25 +01:00
|
|
|
auto hist_temp = std::make_shared<HistogramImpl>();
|
|
|
|
hist_.insert({op_type, std::move(hist_temp)});
|
2015-12-08 02:33:18 +01:00
|
|
|
}
|
2016-03-12 01:54:25 +01:00
|
|
|
hist_[op_type]->Add(micros);
|
2015-12-08 02:33:18 +01:00
|
|
|
|
2012-10-03 15:51:11 +02:00
|
|
|
if (micros > 20000 && !FLAGS_stats_interval) {
|
2015-11-19 20:47:12 +01:00
|
|
|
fprintf(stderr, "long op: %" PRIu64 " micros%30s\r", micros, "");
|
2011-08-22 23:08:51 +02:00
|
|
|
fflush(stderr);
|
|
|
|
}
|
|
|
|
last_op_finish_ = now;
|
|
|
|
}
|
|
|
|
|
2014-06-23 22:23:02 +02:00
|
|
|
done_ += num_ops;
|
2011-08-22 23:08:51 +02:00
|
|
|
if (done_ >= next_report_) {
|
2012-10-03 15:51:11 +02:00
|
|
|
if (!FLAGS_stats_interval) {
|
|
|
|
if (next_report_ < 1000) next_report_ += 100;
|
|
|
|
else if (next_report_ < 5000) next_report_ += 500;
|
|
|
|
else if (next_report_ < 10000) next_report_ += 1000;
|
|
|
|
else if (next_report_ < 50000) next_report_ += 5000;
|
|
|
|
else if (next_report_ < 100000) next_report_ += 10000;
|
|
|
|
else if (next_report_ < 500000) next_report_ += 50000;
|
|
|
|
else next_report_ += 100000;
|
2014-03-05 02:08:05 +01:00
|
|
|
fprintf(stderr, "... finished %" PRIu64 " ops%30s\r", done_, "");
|
2012-10-03 15:51:11 +02:00
|
|
|
} else {
|
2015-11-19 20:47:12 +01:00
|
|
|
uint64_t now = FLAGS_env->NowMicros();
|
2015-03-30 21:58:32 +02:00
|
|
|
int64_t usecs_since_last = now - last_report_finish_;
|
|
|
|
|
|
|
|
// Determine whether to print status where interval is either
|
|
|
|
// each N operations or each N seconds.
|
|
|
|
|
|
|
|
if (FLAGS_stats_interval_seconds &&
|
|
|
|
usecs_since_last < (FLAGS_stats_interval_seconds * 1000000)) {
|
|
|
|
// Don't check again for this many operations
|
|
|
|
next_report_ += FLAGS_stats_interval;
|
|
|
|
|
|
|
|
} else {
|
2014-09-29 18:50:41 +02:00
|
|
|
|
2015-03-30 21:58:32 +02:00
|
|
|
fprintf(stderr,
|
|
|
|
"%s ... thread %d: (%" PRIu64 ",%" PRIu64 ") ops and "
|
|
|
|
"(%.1f,%.1f) ops/second in (%.6f,%.6f) seconds\n",
|
2015-11-19 20:47:12 +01:00
|
|
|
FLAGS_env->TimeToString(now/1000000).c_str(),
|
2015-03-30 21:58:32 +02:00
|
|
|
id_,
|
|
|
|
done_ - last_report_done_, done_,
|
|
|
|
(done_ - last_report_done_) /
|
|
|
|
(usecs_since_last / 1000000.0),
|
|
|
|
done_ / ((now - start_) / 1000000.0),
|
|
|
|
(now - last_report_finish_) / 1000000.0,
|
|
|
|
(now - start_) / 1000000.0);
|
|
|
|
|
2016-08-15 20:07:05 +02:00
|
|
|
if (id_ == 0 && FLAGS_stats_per_interval) {
|
2015-03-30 21:58:32 +02:00
|
|
|
std::string stats;
|
|
|
|
|
|
|
|
if (db_with_cfh && db_with_cfh->num_created.load()) {
|
|
|
|
for (size_t i = 0; i < db_with_cfh->num_created.load(); ++i) {
|
|
|
|
if (db->GetProperty(db_with_cfh->cfh[i], "rocksdb.cfstats",
|
|
|
|
&stats))
|
|
|
|
fprintf(stderr, "%s\n", stats.c_str());
|
Add argument --show_table_properties to db_bench
Summary:
Add argument --show_table_properties to db_bench
-show_table_properties (If true, then per-level table properties will be
printed on every stats-interval when stats_interval is set and
stats_per_interval is on.) type: bool default: false
Test Plan:
./db_bench --show_table_properties=1 --stats_interval=100000 --stats_per_interval=1
./db_bench --show_table_properties=1 --stats_interval=100000 --stats_per_interval=1 --num_column_families=2
Sample Output:
Compaction Stats [column_family_name_000001]
Level Files Size(MB) Score Read(GB) Rn(GB) Rnp1(GB) Write(GB) Wnew(GB) Moved(GB) W-Amp Rd(MB/s) Wr(MB/s) Comp(sec) Comp(cnt) Avg(sec) Stall(cnt) KeyIn KeyDrop
---------------------------------------------------------------------------------------------------------------------------------------------------------------------
L0 3/0 5 0.8 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 86.3 0 17 0.021 0 0 0
L1 5/0 9 0.9 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0 0 0.000 0 0 0
L2 9/0 16 0.2 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0 0 0.000 0 0 0
Sum 17/0 31 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 86.3 0 17 0.021 0 0 0
Int 0/0 0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 83.9 0 2 0.022 0 0 0
Flush(GB): cumulative 0.030, interval 0.004
Stalls(count): 0 level0_slowdown, 0 level0_numfiles, 0 memtable_compaction, 0 leveln_slowdown_soft, 0 leveln_slowdown_hard
Level[0]: # data blocks=2571; # entries=84813; raw key size=2035512; raw average key size=24.000000; raw value size=8481300; raw average value size=100.000000; data block size=5690119; index block size=82415; filter block size=0; (estimated) table size=5772534; filter policy name=N/A;
Level[1]: # data blocks=4285; # entries=141355; raw key size=3392520; raw average key size=24.000000; raw value size=14135500; raw average value size=100.000000; data block size=9487353; index block size=137377; filter block size=0; (estimated) table size=9624730; filter policy name=N/A;
Level[2]: # data blocks=7713; # entries=254439; raw key size=6106536; raw average key size=24.000000; raw value size=25443900; raw average value size=100.000000; data block size=17077893; index block size=247269; filter block size=0; (estimated) table size=17325162; filter policy name=N/A;
Level[3]: # data blocks=0; # entries=0; raw key size=0; raw average key size=0.000000; raw value size=0; raw average value size=0.000000; data block size=0; index block size=0; filter block size=0; (estimated) table size=0; filter policy name=N/A;
Level[4]: # data blocks=0; # entries=0; raw key size=0; raw average key size=0.000000; raw value size=0; raw average value size=0.000000; data block size=0; index block size=0; filter block size=0; (estimated) table size=0; filter policy name=N/A;
Level[5]: # data blocks=0; # entries=0; raw key size=0; raw average key size=0.000000; raw value size=0; raw average value size=0.000000; data block size=0; index block size=0; filter block size=0; (estimated) table size=0; filter policy name=N/A;
Level[6]: # data blocks=0; # entries=0; raw key size=0; raw average key size=0.000000; raw value size=0; raw average value size=0.000000; data block size=0; index block size=0; filter block size=0; (estimated) table size=0; filter policy name=N/A;
Reviewers: anthony, IslamAbdelRahman, MarkCallaghan, sdong, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D45651
2015-08-27 03:27:23 +02:00
|
|
|
if (FLAGS_show_table_properties) {
|
|
|
|
for (int level = 0; level < FLAGS_num_levels; ++level) {
|
|
|
|
if (db->GetProperty(
|
|
|
|
db_with_cfh->cfh[i],
|
|
|
|
"rocksdb.aggregated-table-properties-at-level" +
|
|
|
|
ToString(level),
|
|
|
|
&stats)) {
|
|
|
|
if (stats.find("# entries=0") == std::string::npos) {
|
|
|
|
fprintf(stderr, "Level[%d]: %s\n", level,
|
|
|
|
stats.c_str());
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
} else if (db) {
|
|
|
|
if (db->GetProperty("rocksdb.stats", &stats)) {
|
|
|
|
fprintf(stderr, "%s\n", stats.c_str());
|
|
|
|
}
|
|
|
|
if (FLAGS_show_table_properties) {
|
|
|
|
for (int level = 0; level < FLAGS_num_levels; ++level) {
|
|
|
|
if (db->GetProperty(
|
|
|
|
"rocksdb.aggregated-table-properties-at-level" +
|
|
|
|
ToString(level),
|
|
|
|
&stats)) {
|
|
|
|
if (stats.find("# entries=0") == std::string::npos) {
|
|
|
|
fprintf(stderr, "Level[%d]: %s\n", level, stats.c_str());
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2015-03-30 21:58:32 +02:00
|
|
|
}
|
|
|
|
}
|
2014-09-29 18:50:41 +02:00
|
|
|
}
|
2012-10-23 19:34:09 +02:00
|
|
|
|
2015-03-30 21:58:32 +02:00
|
|
|
next_report_ += FLAGS_stats_interval;
|
|
|
|
last_report_finish_ = now;
|
|
|
|
last_report_done_ = done_;
|
|
|
|
}
|
2012-10-03 15:51:11 +02:00
|
|
|
}
|
2015-03-06 20:22:06 +01:00
|
|
|
if (id_ == 0 && FLAGS_thread_status_per_interval) {
|
|
|
|
PrintThreadStatus();
|
|
|
|
}
|
|
|
|
fflush(stderr);
|
2011-08-22 23:08:51 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void AddBytes(int64_t n) {
|
|
|
|
bytes_ += n;
|
|
|
|
}
|
|
|
|
|
|
|
|
void Report(const Slice& name) {
|
|
|
|
// Pretend at least one op was done in case we are running a benchmark
|
2014-06-23 22:23:02 +02:00
|
|
|
// that does not call FinishedOps().
|
2011-08-22 23:08:51 +02:00
|
|
|
if (done_ < 1) done_ = 1;
|
|
|
|
|
|
|
|
std::string extra;
|
|
|
|
if (bytes_ > 0) {
|
|
|
|
// Rate is computed on actual elapsed time, not the sum of per-thread
|
|
|
|
// elapsed times.
|
|
|
|
double elapsed = (finish_ - start_) * 1e-6;
|
|
|
|
char rate[100];
|
|
|
|
snprintf(rate, sizeof(rate), "%6.1f MB/s",
|
|
|
|
(bytes_ / 1048576.0) / elapsed);
|
|
|
|
extra = rate;
|
|
|
|
}
|
|
|
|
AppendWithSpace(&extra, message_);
|
2012-05-11 20:32:37 +02:00
|
|
|
double elapsed = (finish_ - start_) * 1e-6;
|
|
|
|
double throughput = (double)done_/elapsed;
|
2011-08-22 23:08:51 +02:00
|
|
|
|
2012-05-10 08:32:20 +02:00
|
|
|
fprintf(stdout, "%-12s : %11.3f micros/op %ld ops/sec;%s%s\n",
|
2011-08-22 23:08:51 +02:00
|
|
|
name.ToString().c_str(),
|
2019-02-06 02:15:42 +01:00
|
|
|
seconds_ * 1e6 / done_,
|
2012-05-10 08:32:20 +02:00
|
|
|
(long)throughput,
|
2011-08-22 23:08:51 +02:00
|
|
|
(extra.empty() ? "" : " "),
|
|
|
|
extra.c_str());
|
|
|
|
if (FLAGS_histogram) {
|
2015-12-08 02:33:18 +01:00
|
|
|
for (auto it = hist_.begin(); it != hist_.end(); ++it) {
|
|
|
|
fprintf(stdout, "Microseconds per %s:\n%s\n",
|
|
|
|
OperationTypeString[it->first].c_str(),
|
2016-03-12 01:54:25 +01:00
|
|
|
it->second->ToString().c_str());
|
2015-12-08 02:33:18 +01:00
|
|
|
}
|
2011-08-22 23:08:51 +02:00
|
|
|
}
|
2014-10-29 22:24:34 +01:00
|
|
|
if (FLAGS_report_file_operations) {
|
|
|
|
ReportFileOpEnv* env = static_cast<ReportFileOpEnv*>(FLAGS_env);
|
|
|
|
ReportFileOpCounters* counters = env->counters();
|
|
|
|
fprintf(stdout, "Num files opened: %d\n",
|
|
|
|
counters->open_counter_.load(std::memory_order_relaxed));
|
|
|
|
fprintf(stdout, "Num Read(): %d\n",
|
|
|
|
counters->read_counter_.load(std::memory_order_relaxed));
|
|
|
|
fprintf(stdout, "Num Append(): %d\n",
|
|
|
|
counters->append_counter_.load(std::memory_order_relaxed));
|
|
|
|
fprintf(stdout, "Num bytes read: %" PRIu64 "\n",
|
|
|
|
counters->bytes_read_.load(std::memory_order_relaxed));
|
|
|
|
fprintf(stdout, "Num bytes written: %" PRIu64 "\n",
|
|
|
|
counters->bytes_written_.load(std::memory_order_relaxed));
|
|
|
|
env->reset();
|
|
|
|
}
|
2011-08-22 23:08:51 +02:00
|
|
|
fflush(stdout);
|
|
|
|
}
|
|
|
|
};
|
|
|
|
|
2016-08-25 21:57:35 +02:00
|
|
|
class CombinedStats {
|
|
|
|
public:
|
|
|
|
void AddStats(const Stats& stat) {
|
|
|
|
uint64_t total_ops = stat.done_;
|
|
|
|
uint64_t total_bytes_ = stat.bytes_;
|
|
|
|
double elapsed;
|
|
|
|
|
|
|
|
if (total_ops < 1) {
|
|
|
|
total_ops = 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
elapsed = (stat.finish_ - stat.start_) * 1e-6;
|
|
|
|
throughput_ops_.emplace_back(total_ops / elapsed);
|
|
|
|
|
|
|
|
if (total_bytes_ > 0) {
|
|
|
|
double mbs = (total_bytes_ / 1048576.0);
|
|
|
|
throughput_mbs_.emplace_back(mbs / elapsed);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void Report(const std::string& bench_name) {
|
|
|
|
const char* name = bench_name.c_str();
|
|
|
|
int num_runs = static_cast<int>(throughput_ops_.size());
|
|
|
|
|
|
|
|
if (throughput_mbs_.size() == throughput_ops_.size()) {
|
|
|
|
fprintf(stdout,
|
|
|
|
"%s [AVG %d runs] : %d ops/sec; %6.1f MB/sec\n"
|
|
|
|
"%s [MEDIAN %d runs] : %d ops/sec; %6.1f MB/sec\n",
|
|
|
|
name, num_runs, static_cast<int>(CalcAvg(throughput_ops_)),
|
|
|
|
CalcAvg(throughput_mbs_), name, num_runs,
|
|
|
|
static_cast<int>(CalcMedian(throughput_ops_)),
|
|
|
|
CalcMedian(throughput_mbs_));
|
|
|
|
} else {
|
|
|
|
fprintf(stdout,
|
|
|
|
"%s [AVG %d runs] : %d ops/sec\n"
|
|
|
|
"%s [MEDIAN %d runs] : %d ops/sec\n",
|
|
|
|
name, num_runs, static_cast<int>(CalcAvg(throughput_ops_)), name,
|
|
|
|
num_runs, static_cast<int>(CalcMedian(throughput_ops_)));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
private:
|
|
|
|
double CalcAvg(std::vector<double> data) {
|
|
|
|
double avg = 0;
|
|
|
|
for (double x : data) {
|
|
|
|
avg += x;
|
|
|
|
}
|
|
|
|
avg = avg / data.size();
|
|
|
|
return avg;
|
|
|
|
}
|
|
|
|
|
|
|
|
double CalcMedian(std::vector<double> data) {
|
|
|
|
assert(data.size() > 0);
|
|
|
|
std::sort(data.begin(), data.end());
|
|
|
|
|
|
|
|
size_t mid = data.size() / 2;
|
|
|
|
if (data.size() % 2 == 1) {
|
|
|
|
// Odd number of entries
|
|
|
|
return data[mid];
|
|
|
|
} else {
|
|
|
|
// Even number of entries
|
|
|
|
return (data[mid] + data[mid - 1]) / 2;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
std::vector<double> throughput_ops_;
|
|
|
|
std::vector<double> throughput_mbs_;
|
|
|
|
};
|
|
|
|
|
2016-08-01 21:28:51 +02:00
|
|
|
class TimestampEmulator {
|
|
|
|
private:
|
|
|
|
std::atomic<uint64_t> timestamp_;
|
|
|
|
|
|
|
|
public:
|
|
|
|
TimestampEmulator() : timestamp_(0) {}
|
|
|
|
uint64_t Get() const { return timestamp_.load(); }
|
|
|
|
void Inc() { timestamp_++; }
|
|
|
|
};
|
|
|
|
|
2011-08-22 23:08:51 +02:00
|
|
|
// State shared by all concurrent executions of the same benchmark.
|
|
|
|
struct SharedState {
|
|
|
|
port::Mutex mu;
|
|
|
|
port::CondVar cv;
|
|
|
|
int total;
|
2013-11-18 20:32:54 +01:00
|
|
|
int perf_level;
|
2015-06-18 01:44:52 +02:00
|
|
|
std::shared_ptr<RateLimiter> write_rate_limiter;
|
2016-12-19 21:25:35 +01:00
|
|
|
std::shared_ptr<RateLimiter> read_rate_limiter;
|
2011-08-22 23:08:51 +02:00
|
|
|
|
|
|
|
// Each thread goes through the following states:
|
|
|
|
// (1) initializing
|
|
|
|
// (2) waiting for others to be initialized
|
|
|
|
// (3) running
|
|
|
|
// (4) done
|
|
|
|
|
2012-05-22 21:18:23 +02:00
|
|
|
long num_initialized;
|
|
|
|
long num_done;
|
2011-08-22 23:08:51 +02:00
|
|
|
bool start;
|
|
|
|
|
2013-11-18 20:32:54 +01:00
|
|
|
SharedState() : cv(&mu), perf_level(FLAGS_perf_level) { }
|
2011-08-22 23:08:51 +02:00
|
|
|
};
|
|
|
|
|
|
|
|
// Per-thread state for concurrent executions of the same benchmark.
|
|
|
|
struct ThreadState {
|
|
|
|
int tid; // 0..n-1 when running in n threads
|
2013-08-23 07:37:13 +02:00
|
|
|
Random64 rand; // Has different seeds for different threads
|
2011-08-22 23:08:51 +02:00
|
|
|
Stats stats;
|
2011-09-01 21:08:02 +02:00
|
|
|
SharedState* shared;
|
2011-08-22 23:08:51 +02:00
|
|
|
|
2012-11-29 01:42:36 +01:00
|
|
|
/* implicit */ ThreadState(int index)
|
2011-08-22 23:08:51 +02:00
|
|
|
: tid(index),
|
2013-01-03 21:11:50 +01:00
|
|
|
rand((FLAGS_seed ? FLAGS_seed : 1000) + index) {
|
2011-08-22 23:08:51 +02:00
|
|
|
}
|
|
|
|
};
|
|
|
|
|
2013-03-01 20:25:13 +01:00
|
|
|
class Duration {
|
|
|
|
public:
|
2015-11-19 20:47:12 +01:00
|
|
|
Duration(uint64_t max_seconds, int64_t max_ops, int64_t ops_per_stage = 0) {
|
2013-03-01 20:25:13 +01:00
|
|
|
max_seconds_ = max_seconds;
|
|
|
|
max_ops_= max_ops;
|
2014-12-23 03:39:28 +01:00
|
|
|
ops_per_stage_ = (ops_per_stage > 0) ? ops_per_stage : max_ops;
|
2013-03-01 20:25:13 +01:00
|
|
|
ops_ = 0;
|
|
|
|
start_at_ = FLAGS_env->NowMicros();
|
|
|
|
}
|
|
|
|
|
2014-12-23 03:39:28 +01:00
|
|
|
int64_t GetStage() { return std::min(ops_, max_ops_ - 1) / ops_per_stage_; }
|
|
|
|
|
2014-04-08 20:21:09 +02:00
|
|
|
bool Done(int64_t increment) {
|
2013-06-12 21:42:21 +02:00
|
|
|
if (increment <= 0) increment = 1; // avoid Done(0) and infinite loops
|
2013-03-01 20:25:13 +01:00
|
|
|
ops_ += increment;
|
|
|
|
|
|
|
|
if (max_seconds_) {
|
2013-06-12 21:42:21 +02:00
|
|
|
// Recheck every appx 1000 ops (exact iff increment is factor of 1000)
|
2017-06-16 01:06:22 +02:00
|
|
|
auto granularity = FLAGS_ops_between_duration_checks;
|
|
|
|
if ((ops_ / granularity) != ((ops_ - increment) / granularity)) {
|
2015-11-19 20:47:12 +01:00
|
|
|
uint64_t now = FLAGS_env->NowMicros();
|
|
|
|
return ((now - start_at_) / 1000000) >= max_seconds_;
|
2013-03-01 20:25:13 +01:00
|
|
|
} else {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
return ops_ > max_ops_;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
private:
|
2015-11-19 20:47:12 +01:00
|
|
|
uint64_t max_seconds_;
|
2014-03-05 02:08:05 +01:00
|
|
|
int64_t max_ops_;
|
2014-12-23 03:39:28 +01:00
|
|
|
int64_t ops_per_stage_;
|
2014-03-05 02:08:05 +01:00
|
|
|
int64_t ops_;
|
2015-11-19 20:47:12 +01:00
|
|
|
uint64_t start_at_;
|
2013-03-01 20:25:13 +01:00
|
|
|
};
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
class Benchmark {
|
|
|
|
private:
|
2014-08-25 23:22:05 +02:00
|
|
|
std::shared_ptr<Cache> cache_;
|
|
|
|
std::shared_ptr<Cache> compressed_cache_;
|
|
|
|
std::shared_ptr<const FilterPolicy> filter_policy_;
|
2013-08-23 23:49:57 +02:00
|
|
|
const SliceTransform* prefix_extractor_;
|
2014-08-19 03:15:01 +02:00
|
|
|
DBWithColumnFamilies db_;
|
|
|
|
std::vector<DBWithColumnFamilies> multi_dbs_;
|
2014-03-05 02:08:05 +01:00
|
|
|
int64_t num_;
|
2011-08-22 23:08:51 +02:00
|
|
|
int value_size_;
|
2013-03-01 23:10:09 +01:00
|
|
|
int key_size_;
|
2014-03-05 02:08:05 +01:00
|
|
|
int prefix_size_;
|
|
|
|
int64_t keys_per_prefix_;
|
2014-04-08 20:21:09 +02:00
|
|
|
int64_t entries_per_batch_;
|
2018-10-19 02:16:36 +02:00
|
|
|
int64_t writes_before_delete_range_;
|
2016-11-16 02:33:27 +01:00
|
|
|
int64_t writes_per_range_tombstone_;
|
|
|
|
int64_t range_tombstone_width_;
|
|
|
|
int64_t max_num_range_tombstones_;
|
2011-08-22 23:08:51 +02:00
|
|
|
WriteOptions write_options_;
|
2014-12-23 03:39:28 +01:00
|
|
|
Options open_options_; // keep options around to properly destroy db later
|
2018-11-27 06:32:02 +01:00
|
|
|
#ifndef ROCKSDB_LITE
|
2018-08-01 09:14:43 +02:00
|
|
|
TraceOptions trace_options_;
|
2018-11-27 06:32:02 +01:00
|
|
|
#endif
|
2014-03-05 02:08:05 +01:00
|
|
|
int64_t reads_;
|
2016-06-28 23:10:24 +02:00
|
|
|
int64_t deletes_;
|
2015-03-07 01:38:50 +01:00
|
|
|
double read_random_exp_range_;
|
2014-03-05 02:08:05 +01:00
|
|
|
int64_t writes_;
|
|
|
|
int64_t readwrites_;
|
|
|
|
int64_t merge_keys_;
|
2014-10-29 22:24:34 +01:00
|
|
|
bool report_file_operations_;
|
2017-05-25 23:08:16 +02:00
|
|
|
bool use_blob_db_;
|
2019-03-01 20:14:28 +01:00
|
|
|
std::vector<std::string> keys_;
|
2014-06-23 19:46:16 +02:00
|
|
|
|
Auto recovery from out of space errors (#4164)
Summary:
This commit implements automatic recovery from a Status::NoSpace() error
during background operations such as write callback, flush and
compaction. The broad design is as follows -
1. Compaction errors are treated as soft errors and don't put the
database in read-only mode. A compaction is delayed until enough free
disk space is available to accomodate the compaction outputs, which is
estimated based on the input size. This means that users can continue to
write, and we rely on the WriteController to delay or stop writes if the
compaction debt becomes too high due to persistent low disk space
condition
2. Errors during write callback and flush are treated as hard errors,
i.e the database is put in read-only mode and goes back to read-write
only fater certain recovery actions are taken.
3. Both types of recovery rely on the SstFileManagerImpl to poll for
sufficient disk space. We assume that there is a 1-1 mapping between an
SFM and the underlying OS storage container. For cases where multiple
DBs are hosted on a single storage container, the user is expected to
allocate a single SFM instance and use the same one for all the DBs. If
no SFM is specified by the user, DBImpl::Open() will allocate one, but
this will be one per DB and each DB will recover independently. The
recovery implemented by SFM is as follows -
a) On the first occurance of an out of space error during compaction,
subsequent
compactions will be delayed until the disk free space check indicates
enough available space. The required space is computed as the sum of
input sizes.
b) The free space check requirement will be removed once the amount of
free space is greater than the size reserved by in progress
compactions when the first error occured
c) If the out of space error is a hard error, a background thread in
SFM will poll for sufficient headroom before triggering the recovery
of the database and putting it in write-only mode. The headroom is
calculated as the sum of the write_buffer_size of all the DB instances
associated with the SFM
4. EventListener callbacks will be called at the start and completion of
automatic recovery. Users can disable the auto recov ery in the start
callback, and later initiate it manually by calling DB::Resume()
Todo:
1. More extensive testing
2. Add disk full condition to db_stress (follow-on PR)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4164
Differential Revision: D9846378
Pulled By: anand1976
fbshipit-source-id: 80ea875dbd7f00205e19c82215ff6e37da10da4a
2018-09-15 22:36:19 +02:00
|
|
|
class ErrorHandlerListener : public EventListener {
|
|
|
|
public:
|
2019-02-16 01:10:37 +01:00
|
|
|
#ifndef ROCKSDB_LITE
|
Auto recovery from out of space errors (#4164)
Summary:
This commit implements automatic recovery from a Status::NoSpace() error
during background operations such as write callback, flush and
compaction. The broad design is as follows -
1. Compaction errors are treated as soft errors and don't put the
database in read-only mode. A compaction is delayed until enough free
disk space is available to accomodate the compaction outputs, which is
estimated based on the input size. This means that users can continue to
write, and we rely on the WriteController to delay or stop writes if the
compaction debt becomes too high due to persistent low disk space
condition
2. Errors during write callback and flush are treated as hard errors,
i.e the database is put in read-only mode and goes back to read-write
only fater certain recovery actions are taken.
3. Both types of recovery rely on the SstFileManagerImpl to poll for
sufficient disk space. We assume that there is a 1-1 mapping between an
SFM and the underlying OS storage container. For cases where multiple
DBs are hosted on a single storage container, the user is expected to
allocate a single SFM instance and use the same one for all the DBs. If
no SFM is specified by the user, DBImpl::Open() will allocate one, but
this will be one per DB and each DB will recover independently. The
recovery implemented by SFM is as follows -
a) On the first occurance of an out of space error during compaction,
subsequent
compactions will be delayed until the disk free space check indicates
enough available space. The required space is computed as the sum of
input sizes.
b) The free space check requirement will be removed once the amount of
free space is greater than the size reserved by in progress
compactions when the first error occured
c) If the out of space error is a hard error, a background thread in
SFM will poll for sufficient headroom before triggering the recovery
of the database and putting it in write-only mode. The headroom is
calculated as the sum of the write_buffer_size of all the DB instances
associated with the SFM
4. EventListener callbacks will be called at the start and completion of
automatic recovery. Users can disable the auto recov ery in the start
callback, and later initiate it manually by calling DB::Resume()
Todo:
1. More extensive testing
2. Add disk full condition to db_stress (follow-on PR)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4164
Differential Revision: D9846378
Pulled By: anand1976
fbshipit-source-id: 80ea875dbd7f00205e19c82215ff6e37da10da4a
2018-09-15 22:36:19 +02:00
|
|
|
ErrorHandlerListener()
|
|
|
|
: mutex_(),
|
|
|
|
cv_(&mutex_),
|
|
|
|
no_auto_recovery_(false),
|
|
|
|
recovery_complete_(false) {}
|
|
|
|
|
2019-02-14 22:52:47 +01:00
|
|
|
~ErrorHandlerListener() override {}
|
Auto recovery from out of space errors (#4164)
Summary:
This commit implements automatic recovery from a Status::NoSpace() error
during background operations such as write callback, flush and
compaction. The broad design is as follows -
1. Compaction errors are treated as soft errors and don't put the
database in read-only mode. A compaction is delayed until enough free
disk space is available to accomodate the compaction outputs, which is
estimated based on the input size. This means that users can continue to
write, and we rely on the WriteController to delay or stop writes if the
compaction debt becomes too high due to persistent low disk space
condition
2. Errors during write callback and flush are treated as hard errors,
i.e the database is put in read-only mode and goes back to read-write
only fater certain recovery actions are taken.
3. Both types of recovery rely on the SstFileManagerImpl to poll for
sufficient disk space. We assume that there is a 1-1 mapping between an
SFM and the underlying OS storage container. For cases where multiple
DBs are hosted on a single storage container, the user is expected to
allocate a single SFM instance and use the same one for all the DBs. If
no SFM is specified by the user, DBImpl::Open() will allocate one, but
this will be one per DB and each DB will recover independently. The
recovery implemented by SFM is as follows -
a) On the first occurance of an out of space error during compaction,
subsequent
compactions will be delayed until the disk free space check indicates
enough available space. The required space is computed as the sum of
input sizes.
b) The free space check requirement will be removed once the amount of
free space is greater than the size reserved by in progress
compactions when the first error occured
c) If the out of space error is a hard error, a background thread in
SFM will poll for sufficient headroom before triggering the recovery
of the database and putting it in write-only mode. The headroom is
calculated as the sum of the write_buffer_size of all the DB instances
associated with the SFM
4. EventListener callbacks will be called at the start and completion of
automatic recovery. Users can disable the auto recov ery in the start
callback, and later initiate it manually by calling DB::Resume()
Todo:
1. More extensive testing
2. Add disk full condition to db_stress (follow-on PR)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4164
Differential Revision: D9846378
Pulled By: anand1976
fbshipit-source-id: 80ea875dbd7f00205e19c82215ff6e37da10da4a
2018-09-15 22:36:19 +02:00
|
|
|
|
|
|
|
void OnErrorRecoveryBegin(BackgroundErrorReason /*reason*/,
|
2019-02-14 22:52:47 +01:00
|
|
|
Status /*bg_error*/,
|
|
|
|
bool* auto_recovery) override {
|
Auto recovery from out of space errors (#4164)
Summary:
This commit implements automatic recovery from a Status::NoSpace() error
during background operations such as write callback, flush and
compaction. The broad design is as follows -
1. Compaction errors are treated as soft errors and don't put the
database in read-only mode. A compaction is delayed until enough free
disk space is available to accomodate the compaction outputs, which is
estimated based on the input size. This means that users can continue to
write, and we rely on the WriteController to delay or stop writes if the
compaction debt becomes too high due to persistent low disk space
condition
2. Errors during write callback and flush are treated as hard errors,
i.e the database is put in read-only mode and goes back to read-write
only fater certain recovery actions are taken.
3. Both types of recovery rely on the SstFileManagerImpl to poll for
sufficient disk space. We assume that there is a 1-1 mapping between an
SFM and the underlying OS storage container. For cases where multiple
DBs are hosted on a single storage container, the user is expected to
allocate a single SFM instance and use the same one for all the DBs. If
no SFM is specified by the user, DBImpl::Open() will allocate one, but
this will be one per DB and each DB will recover independently. The
recovery implemented by SFM is as follows -
a) On the first occurance of an out of space error during compaction,
subsequent
compactions will be delayed until the disk free space check indicates
enough available space. The required space is computed as the sum of
input sizes.
b) The free space check requirement will be removed once the amount of
free space is greater than the size reserved by in progress
compactions when the first error occured
c) If the out of space error is a hard error, a background thread in
SFM will poll for sufficient headroom before triggering the recovery
of the database and putting it in write-only mode. The headroom is
calculated as the sum of the write_buffer_size of all the DB instances
associated with the SFM
4. EventListener callbacks will be called at the start and completion of
automatic recovery. Users can disable the auto recov ery in the start
callback, and later initiate it manually by calling DB::Resume()
Todo:
1. More extensive testing
2. Add disk full condition to db_stress (follow-on PR)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4164
Differential Revision: D9846378
Pulled By: anand1976
fbshipit-source-id: 80ea875dbd7f00205e19c82215ff6e37da10da4a
2018-09-15 22:36:19 +02:00
|
|
|
if (*auto_recovery && no_auto_recovery_) {
|
|
|
|
*auto_recovery = false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2019-02-14 22:52:47 +01:00
|
|
|
void OnErrorRecoveryCompleted(Status /*old_bg_error*/) override {
|
Auto recovery from out of space errors (#4164)
Summary:
This commit implements automatic recovery from a Status::NoSpace() error
during background operations such as write callback, flush and
compaction. The broad design is as follows -
1. Compaction errors are treated as soft errors and don't put the
database in read-only mode. A compaction is delayed until enough free
disk space is available to accomodate the compaction outputs, which is
estimated based on the input size. This means that users can continue to
write, and we rely on the WriteController to delay or stop writes if the
compaction debt becomes too high due to persistent low disk space
condition
2. Errors during write callback and flush are treated as hard errors,
i.e the database is put in read-only mode and goes back to read-write
only fater certain recovery actions are taken.
3. Both types of recovery rely on the SstFileManagerImpl to poll for
sufficient disk space. We assume that there is a 1-1 mapping between an
SFM and the underlying OS storage container. For cases where multiple
DBs are hosted on a single storage container, the user is expected to
allocate a single SFM instance and use the same one for all the DBs. If
no SFM is specified by the user, DBImpl::Open() will allocate one, but
this will be one per DB and each DB will recover independently. The
recovery implemented by SFM is as follows -
a) On the first occurance of an out of space error during compaction,
subsequent
compactions will be delayed until the disk free space check indicates
enough available space. The required space is computed as the sum of
input sizes.
b) The free space check requirement will be removed once the amount of
free space is greater than the size reserved by in progress
compactions when the first error occured
c) If the out of space error is a hard error, a background thread in
SFM will poll for sufficient headroom before triggering the recovery
of the database and putting it in write-only mode. The headroom is
calculated as the sum of the write_buffer_size of all the DB instances
associated with the SFM
4. EventListener callbacks will be called at the start and completion of
automatic recovery. Users can disable the auto recov ery in the start
callback, and later initiate it manually by calling DB::Resume()
Todo:
1. More extensive testing
2. Add disk full condition to db_stress (follow-on PR)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4164
Differential Revision: D9846378
Pulled By: anand1976
fbshipit-source-id: 80ea875dbd7f00205e19c82215ff6e37da10da4a
2018-09-15 22:36:19 +02:00
|
|
|
InstrumentedMutexLock l(&mutex_);
|
|
|
|
recovery_complete_ = true;
|
|
|
|
cv_.SignalAll();
|
|
|
|
}
|
|
|
|
|
|
|
|
bool WaitForRecovery(uint64_t /*abs_time_us*/) {
|
|
|
|
InstrumentedMutexLock l(&mutex_);
|
|
|
|
if (!recovery_complete_) {
|
|
|
|
cv_.Wait(/*abs_time_us*/);
|
|
|
|
}
|
|
|
|
if (recovery_complete_) {
|
|
|
|
recovery_complete_ = false;
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
void EnableAutoRecovery(bool enable = true) { no_auto_recovery_ = !enable; }
|
|
|
|
|
|
|
|
private:
|
|
|
|
InstrumentedMutex mutex_;
|
|
|
|
InstrumentedCondVar cv_;
|
|
|
|
bool no_auto_recovery_;
|
|
|
|
bool recovery_complete_;
|
2019-02-16 01:10:37 +01:00
|
|
|
#else // ROCKSDB_LITE
|
|
|
|
bool WaitForRecovery(uint64_t /*abs_time_us*/) { return true; }
|
|
|
|
void EnableAutoRecovery(bool /*enable*/) {}
|
|
|
|
#endif // ROCKSDB_LITE
|
Auto recovery from out of space errors (#4164)
Summary:
This commit implements automatic recovery from a Status::NoSpace() error
during background operations such as write callback, flush and
compaction. The broad design is as follows -
1. Compaction errors are treated as soft errors and don't put the
database in read-only mode. A compaction is delayed until enough free
disk space is available to accomodate the compaction outputs, which is
estimated based on the input size. This means that users can continue to
write, and we rely on the WriteController to delay or stop writes if the
compaction debt becomes too high due to persistent low disk space
condition
2. Errors during write callback and flush are treated as hard errors,
i.e the database is put in read-only mode and goes back to read-write
only fater certain recovery actions are taken.
3. Both types of recovery rely on the SstFileManagerImpl to poll for
sufficient disk space. We assume that there is a 1-1 mapping between an
SFM and the underlying OS storage container. For cases where multiple
DBs are hosted on a single storage container, the user is expected to
allocate a single SFM instance and use the same one for all the DBs. If
no SFM is specified by the user, DBImpl::Open() will allocate one, but
this will be one per DB and each DB will recover independently. The
recovery implemented by SFM is as follows -
a) On the first occurance of an out of space error during compaction,
subsequent
compactions will be delayed until the disk free space check indicates
enough available space. The required space is computed as the sum of
input sizes.
b) The free space check requirement will be removed once the amount of
free space is greater than the size reserved by in progress
compactions when the first error occured
c) If the out of space error is a hard error, a background thread in
SFM will poll for sufficient headroom before triggering the recovery
of the database and putting it in write-only mode. The headroom is
calculated as the sum of the write_buffer_size of all the DB instances
associated with the SFM
4. EventListener callbacks will be called at the start and completion of
automatic recovery. Users can disable the auto recov ery in the start
callback, and later initiate it manually by calling DB::Resume()
Todo:
1. More extensive testing
2. Add disk full condition to db_stress (follow-on PR)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4164
Differential Revision: D9846378
Pulled By: anand1976
fbshipit-source-id: 80ea875dbd7f00205e19c82215ff6e37da10da4a
2018-09-15 22:36:19 +02:00
|
|
|
};
|
|
|
|
|
|
|
|
std::shared_ptr<ErrorHandlerListener> listener_;
|
|
|
|
|
2014-06-23 19:46:16 +02:00
|
|
|
bool SanityCheck() {
|
|
|
|
if (FLAGS_compression_ratio > 1) {
|
|
|
|
fprintf(stderr, "compression_ratio should be between 0 and 1\n");
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2019-01-19 04:10:17 +01:00
|
|
|
inline bool CompressSlice(const CompressionInfo& compression_info,
|
2018-06-05 21:51:05 +02:00
|
|
|
const Slice& input, std::string* compressed) {
|
2015-08-11 20:46:15 +02:00
|
|
|
bool ok = true;
|
|
|
|
switch (FLAGS_compression_type_e) {
|
|
|
|
case rocksdb::kSnappyCompression:
|
2019-01-19 04:10:17 +01:00
|
|
|
ok = Snappy_Compress(compression_info, input.data(), input.size(),
|
2018-06-05 21:51:05 +02:00
|
|
|
compressed);
|
2015-08-11 20:46:15 +02:00
|
|
|
break;
|
|
|
|
case rocksdb::kZlibCompression:
|
2019-01-19 04:10:17 +01:00
|
|
|
ok = Zlib_Compress(compression_info, 2, input.data(), input.size(),
|
2018-06-05 21:51:05 +02:00
|
|
|
compressed);
|
2015-08-11 20:46:15 +02:00
|
|
|
break;
|
|
|
|
case rocksdb::kBZip2Compression:
|
2019-01-19 04:10:17 +01:00
|
|
|
ok = BZip2_Compress(compression_info, 2, input.data(), input.size(),
|
2018-06-05 21:51:05 +02:00
|
|
|
compressed);
|
2015-08-11 20:46:15 +02:00
|
|
|
break;
|
|
|
|
case rocksdb::kLZ4Compression:
|
2019-01-19 04:10:17 +01:00
|
|
|
ok = LZ4_Compress(compression_info, 2, input.data(), input.size(),
|
2018-06-05 21:51:05 +02:00
|
|
|
compressed);
|
2015-08-11 20:46:15 +02:00
|
|
|
break;
|
|
|
|
case rocksdb::kLZ4HCCompression:
|
2019-01-19 04:10:17 +01:00
|
|
|
ok = LZ4HC_Compress(compression_info, 2, input.data(), input.size(),
|
2018-06-05 21:51:05 +02:00
|
|
|
compressed);
|
2015-08-11 20:46:15 +02:00
|
|
|
break;
|
2016-04-20 07:54:24 +02:00
|
|
|
case rocksdb::kXpressCompression:
|
|
|
|
ok = XPRESS_Compress(input.data(),
|
|
|
|
input.size(), compressed);
|
|
|
|
break;
|
2016-09-02 00:28:40 +02:00
|
|
|
case rocksdb::kZSTD:
|
2019-01-19 04:10:17 +01:00
|
|
|
ok = ZSTD_Compress(compression_info, input.data(), input.size(),
|
2018-06-05 21:51:05 +02:00
|
|
|
compressed);
|
2015-08-28 00:40:42 +02:00
|
|
|
break;
|
2015-08-11 20:46:15 +02:00
|
|
|
default:
|
|
|
|
ok = false;
|
|
|
|
}
|
|
|
|
return ok;
|
|
|
|
}
|
|
|
|
|
2011-03-22 19:32:49 +01:00
|
|
|
void PrintHeader() {
|
|
|
|
PrintEnvironment();
|
2013-03-01 23:10:09 +01:00
|
|
|
fprintf(stdout, "Keys: %d bytes each\n", FLAGS_key_size);
|
2011-03-22 19:32:49 +01:00
|
|
|
fprintf(stdout, "Values: %d bytes each (%d bytes after compression)\n",
|
|
|
|
FLAGS_value_size,
|
|
|
|
static_cast<int>(FLAGS_value_size * FLAGS_compression_ratio + 0.5));
|
2014-03-05 02:08:05 +01:00
|
|
|
fprintf(stdout, "Entries: %" PRIu64 "\n", num_);
|
|
|
|
fprintf(stdout, "Prefix: %d bytes\n", FLAGS_prefix_size);
|
|
|
|
fprintf(stdout, "Keys per prefix: %" PRIu64 "\n", keys_per_prefix_);
|
2011-03-22 19:32:49 +01:00
|
|
|
fprintf(stdout, "RawSize: %.1f MB (estimated)\n",
|
2013-03-01 23:10:09 +01:00
|
|
|
((static_cast<int64_t>(FLAGS_key_size + FLAGS_value_size) * num_)
|
2011-04-12 21:38:58 +02:00
|
|
|
/ 1048576.0));
|
2011-03-22 19:32:49 +01:00
|
|
|
fprintf(stdout, "FileSize: %.1f MB (estimated)\n",
|
2013-10-24 16:43:14 +02:00
|
|
|
(((FLAGS_key_size + FLAGS_value_size * FLAGS_compression_ratio)
|
|
|
|
* num_)
|
2011-03-22 19:32:49 +01:00
|
|
|
/ 1048576.0));
|
2016-01-04 21:01:27 +01:00
|
|
|
fprintf(stdout, "Write rate: %" PRIu64 " bytes/second\n",
|
|
|
|
FLAGS_benchmark_write_rate_limit);
|
2016-12-19 21:25:35 +01:00
|
|
|
fprintf(stdout, "Read rate: %" PRIu64 " ops/second\n",
|
|
|
|
FLAGS_benchmark_read_rate_limit);
|
2014-07-07 19:53:31 +02:00
|
|
|
if (FLAGS_enable_numa) {
|
|
|
|
fprintf(stderr, "Running in NUMA enabled mode.\n");
|
|
|
|
#ifndef NUMA
|
|
|
|
fprintf(stderr, "NUMA is not defined in the system.\n");
|
|
|
|
exit(1);
|
|
|
|
#else
|
|
|
|
if (numa_available() == -1) {
|
|
|
|
fprintf(stderr, "NUMA is not supported by the system.\n");
|
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
}
|
2015-08-11 20:46:15 +02:00
|
|
|
|
2015-12-17 21:22:36 +01:00
|
|
|
auto compression = CompressionTypeToString(FLAGS_compression_type_e);
|
|
|
|
fprintf(stdout, "Compression: %s\n", compression.c_str());
|
2019-03-18 20:07:35 +01:00
|
|
|
fprintf(stdout, "Compression sampling rate: %" PRId64 "\n",
|
|
|
|
FLAGS_sample_for_compression);
|
2012-09-14 20:21:54 +02:00
|
|
|
|
2013-08-23 08:10:02 +02:00
|
|
|
switch (FLAGS_rep_factory) {
|
|
|
|
case kPrefixHash:
|
|
|
|
fprintf(stdout, "Memtablerep: prefix_hash\n");
|
|
|
|
break;
|
|
|
|
case kSkipList:
|
|
|
|
fprintf(stdout, "Memtablerep: skip_list\n");
|
|
|
|
break;
|
|
|
|
case kVectorRep:
|
|
|
|
fprintf(stdout, "Memtablerep: vector\n");
|
|
|
|
break;
|
2014-03-05 19:27:17 +01:00
|
|
|
case kHashLinkedList:
|
|
|
|
fprintf(stdout, "Memtablerep: hash_linkedlist\n");
|
|
|
|
break;
|
2013-08-23 08:10:02 +02:00
|
|
|
}
|
2013-11-18 20:32:54 +01:00
|
|
|
fprintf(stdout, "Perf Level: %d\n", FLAGS_perf_level);
|
2013-08-23 08:10:02 +02:00
|
|
|
|
2015-12-17 21:22:36 +01:00
|
|
|
PrintWarnings(compression.c_str());
|
2011-03-22 19:32:49 +01:00
|
|
|
fprintf(stdout, "------------------------------------------------\n");
|
|
|
|
}
|
|
|
|
|
2015-08-11 20:46:15 +02:00
|
|
|
void PrintWarnings(const char* compression) {
|
2011-03-22 19:32:49 +01:00
|
|
|
#if defined(__GNUC__) && !defined(__OPTIMIZE__)
|
|
|
|
fprintf(stdout,
|
|
|
|
"WARNING: Optimization is disabled: benchmarks unnecessarily slow\n"
|
|
|
|
);
|
|
|
|
#endif
|
|
|
|
#ifndef NDEBUG
|
|
|
|
fprintf(stdout,
|
|
|
|
"WARNING: Assertions are enabled; benchmarks unnecessarily slow\n");
|
|
|
|
#endif
|
2013-10-24 16:43:14 +02:00
|
|
|
if (FLAGS_compression_type_e != rocksdb::kNoCompression) {
|
2012-09-14 20:21:54 +02:00
|
|
|
// The test string should not be too small.
|
|
|
|
const int len = FLAGS_block_size;
|
2015-08-11 20:46:15 +02:00
|
|
|
std::string input_str(len, 'y');
|
2012-09-14 20:21:54 +02:00
|
|
|
std::string compressed;
|
2019-01-19 04:10:17 +01:00
|
|
|
CompressionOptions opts;
|
|
|
|
CompressionContext context(FLAGS_compression_type_e);
|
|
|
|
CompressionInfo info(opts, context, CompressionDict::GetEmptyDict(),
|
2019-03-18 20:07:35 +01:00
|
|
|
FLAGS_compression_type_e,
|
|
|
|
FLAGS_sample_for_compression);
|
2019-01-19 04:10:17 +01:00
|
|
|
bool result = CompressSlice(info, Slice(input_str), &compressed);
|
2012-09-14 20:21:54 +02:00
|
|
|
|
|
|
|
if (!result) {
|
2015-08-11 20:46:15 +02:00
|
|
|
fprintf(stdout, "WARNING: %s compression is not enabled\n",
|
|
|
|
compression);
|
|
|
|
} else if (compressed.size() >= input_str.size()) {
|
|
|
|
fprintf(stdout, "WARNING: %s compression is not effective\n",
|
|
|
|
compression);
|
2012-09-14 20:21:54 +02:00
|
|
|
}
|
2011-04-12 21:38:58 +02:00
|
|
|
}
|
2011-03-22 19:32:49 +01:00
|
|
|
}
|
|
|
|
|
2013-11-17 08:44:39 +01:00
|
|
|
// Current the following isn't equivalent to OS_LINUX.
|
|
|
|
#if defined(__linux)
|
|
|
|
static Slice TrimSpace(Slice s) {
|
|
|
|
unsigned int start = 0;
|
|
|
|
while (start < s.size() && isspace(s[start])) {
|
|
|
|
start++;
|
|
|
|
}
|
2014-11-21 20:20:42 +01:00
|
|
|
unsigned int limit = static_cast<unsigned int>(s.size());
|
2013-11-17 08:44:39 +01:00
|
|
|
while (limit > start && isspace(s[limit-1])) {
|
|
|
|
limit--;
|
|
|
|
}
|
|
|
|
return Slice(s.data() + start, limit - start);
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
2011-03-22 19:32:49 +01:00
|
|
|
void PrintEnvironment() {
|
2016-04-12 23:05:09 +02:00
|
|
|
fprintf(stderr, "RocksDB: version %d.%d\n",
|
2011-03-22 19:32:49 +01:00
|
|
|
kMajorVersion, kMinorVersion);
|
|
|
|
|
|
|
|
#if defined(__linux)
|
2013-02-15 20:53:17 +01:00
|
|
|
time_t now = time(nullptr);
|
2015-12-16 00:26:20 +01:00
|
|
|
char buf[52];
|
|
|
|
// Lint complains about ctime() usage, so replace it with ctime_r(). The
|
|
|
|
// requirement is to provide a buffer which is at least 26 bytes.
|
|
|
|
fprintf(stderr, "Date: %s",
|
|
|
|
ctime_r(&now, buf)); // ctime_r() adds newline
|
2011-03-22 19:32:49 +01:00
|
|
|
|
|
|
|
FILE* cpuinfo = fopen("/proc/cpuinfo", "r");
|
2013-02-15 20:53:17 +01:00
|
|
|
if (cpuinfo != nullptr) {
|
2011-03-22 19:32:49 +01:00
|
|
|
char line[1000];
|
|
|
|
int num_cpus = 0;
|
|
|
|
std::string cpu_type;
|
|
|
|
std::string cache_size;
|
2013-02-15 20:53:17 +01:00
|
|
|
while (fgets(line, sizeof(line), cpuinfo) != nullptr) {
|
2011-03-22 19:32:49 +01:00
|
|
|
const char* sep = strchr(line, ':');
|
2013-02-15 20:53:17 +01:00
|
|
|
if (sep == nullptr) {
|
2011-03-22 19:32:49 +01:00
|
|
|
continue;
|
|
|
|
}
|
|
|
|
Slice key = TrimSpace(Slice(line, sep - 1 - line));
|
|
|
|
Slice val = TrimSpace(Slice(sep + 1));
|
|
|
|
if (key == "model name") {
|
|
|
|
++num_cpus;
|
|
|
|
cpu_type = val.ToString();
|
|
|
|
} else if (key == "cache size") {
|
|
|
|
cache_size = val.ToString();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
fclose(cpuinfo);
|
|
|
|
fprintf(stderr, "CPU: %d * %s\n", num_cpus, cpu_type.c_str());
|
|
|
|
fprintf(stderr, "CPUCache: %s\n", cache_size.c_str());
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
}
|
|
|
|
|
2016-08-01 21:28:51 +02:00
|
|
|
static bool KeyExpired(const TimestampEmulator* timestamp_emulator,
|
|
|
|
const Slice& key) {
|
|
|
|
const char* pos = key.data();
|
|
|
|
pos += 8;
|
|
|
|
uint64_t timestamp = 0;
|
|
|
|
if (port::kLittleEndian) {
|
|
|
|
int bytes_to_fill = 8;
|
|
|
|
for (int i = 0; i < bytes_to_fill; ++i) {
|
|
|
|
timestamp |= (static_cast<uint64_t>(static_cast<unsigned char>(pos[i]))
|
|
|
|
<< ((bytes_to_fill - i - 1) << 3));
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
memcpy(×tamp, pos, sizeof(timestamp));
|
|
|
|
}
|
|
|
|
return timestamp_emulator->Get() - timestamp > FLAGS_time_range;
|
|
|
|
}
|
|
|
|
|
|
|
|
class ExpiredTimeFilter : public CompactionFilter {
|
|
|
|
public:
|
|
|
|
explicit ExpiredTimeFilter(
|
|
|
|
const std::shared_ptr<TimestampEmulator>& timestamp_emulator)
|
|
|
|
: timestamp_emulator_(timestamp_emulator) {}
|
2018-03-05 22:08:17 +01:00
|
|
|
bool Filter(int /*level*/, const Slice& key,
|
|
|
|
const Slice& /*existing_value*/, std::string* /*new_value*/,
|
|
|
|
bool* /*value_changed*/) const override {
|
2016-08-01 21:28:51 +02:00
|
|
|
return KeyExpired(timestamp_emulator_.get(), key);
|
|
|
|
}
|
|
|
|
const char* Name() const override { return "ExpiredTimeFilter"; }
|
|
|
|
|
|
|
|
private:
|
|
|
|
std::shared_ptr<TimestampEmulator> timestamp_emulator_;
|
|
|
|
};
|
|
|
|
|
2018-07-13 04:33:59 +02:00
|
|
|
class KeepFilter : public CompactionFilter {
|
|
|
|
public:
|
2019-02-14 22:52:47 +01:00
|
|
|
bool Filter(int /*level*/, const Slice& /*key*/, const Slice& /*value*/,
|
|
|
|
std::string* /*new_value*/,
|
|
|
|
bool* /*value_changed*/) const override {
|
2018-07-13 04:33:59 +02:00
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2019-02-14 22:52:47 +01:00
|
|
|
const char* Name() const override { return "KeepFilter"; }
|
2018-07-13 04:33:59 +02:00
|
|
|
};
|
|
|
|
|
2016-08-19 21:28:19 +02:00
|
|
|
std::shared_ptr<Cache> NewCache(int64_t capacity) {
|
|
|
|
if (capacity <= 0) {
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
if (FLAGS_use_clock_cache) {
|
|
|
|
auto cache = NewClockCache((size_t)capacity, FLAGS_cache_numshardbits);
|
|
|
|
if (!cache) {
|
|
|
|
fprintf(stderr, "Clock cache not supported.");
|
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
return cache;
|
|
|
|
} else {
|
2017-01-04 03:24:15 +01:00
|
|
|
return NewLRUCache((size_t)capacity, FLAGS_cache_numshardbits,
|
|
|
|
false /*strict_capacity_limit*/,
|
|
|
|
FLAGS_cache_high_pri_pool_ratio);
|
2016-08-19 21:28:19 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
public:
|
2011-04-12 21:38:58 +02:00
|
|
|
Benchmark()
|
2016-08-19 21:28:19 +02:00
|
|
|
: cache_(NewCache(FLAGS_cache_size)),
|
|
|
|
compressed_cache_(NewCache(FLAGS_compressed_cache_size)),
|
2015-03-07 01:38:50 +01:00
|
|
|
filter_policy_(FLAGS_bloom_bits >= 0
|
|
|
|
? NewBloomFilterPolicy(FLAGS_bloom_bits,
|
|
|
|
FLAGS_use_block_based_filter)
|
|
|
|
: nullptr),
|
|
|
|
prefix_extractor_(NewFixedPrefixTransform(FLAGS_prefix_size)),
|
|
|
|
num_(FLAGS_num),
|
|
|
|
value_size_(FLAGS_value_size),
|
|
|
|
key_size_(FLAGS_key_size),
|
|
|
|
prefix_size_(FLAGS_prefix_size),
|
|
|
|
keys_per_prefix_(FLAGS_keys_per_prefix),
|
|
|
|
entries_per_batch_(1),
|
|
|
|
reads_(FLAGS_reads < 0 ? FLAGS_num : FLAGS_reads),
|
|
|
|
read_random_exp_range_(0.0),
|
|
|
|
writes_(FLAGS_writes < 0 ? FLAGS_num : FLAGS_writes),
|
|
|
|
readwrites_(
|
|
|
|
(FLAGS_writes < 0 && FLAGS_reads < 0)
|
|
|
|
? FLAGS_num
|
|
|
|
: ((FLAGS_writes > FLAGS_reads) ? FLAGS_writes : FLAGS_reads)),
|
|
|
|
merge_keys_(FLAGS_merge_keys < 0 ? FLAGS_num : FLAGS_merge_keys),
|
2017-05-25 23:08:16 +02:00
|
|
|
report_file_operations_(FLAGS_report_file_operations),
|
|
|
|
#ifndef ROCKSDB_LITE
|
2018-06-04 21:04:52 +02:00
|
|
|
use_blob_db_(FLAGS_use_blob_db)
|
2017-05-25 23:08:16 +02:00
|
|
|
#else
|
2018-06-04 21:04:52 +02:00
|
|
|
use_blob_db_(false)
|
2017-05-25 23:08:16 +02:00
|
|
|
#endif // !ROCKSDB_LITE
|
2018-06-04 21:04:52 +02:00
|
|
|
{
|
2016-05-24 08:35:23 +02:00
|
|
|
// use simcache instead of cache
|
|
|
|
if (FLAGS_simcache_size >= 0) {
|
|
|
|
if (FLAGS_cache_numshardbits >= 1) {
|
|
|
|
cache_ =
|
|
|
|
NewSimCache(cache_, FLAGS_simcache_size, FLAGS_cache_numshardbits);
|
|
|
|
} else {
|
|
|
|
cache_ = NewSimCache(cache_, FLAGS_simcache_size, 0);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2014-10-29 22:24:34 +01:00
|
|
|
if (report_file_operations_) {
|
|
|
|
if (!FLAGS_hdfs.empty()) {
|
|
|
|
fprintf(stderr,
|
|
|
|
"--hdfs and --report_file_operations cannot be enabled "
|
|
|
|
"at the same time");
|
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
FLAGS_env = new ReportFileOpEnv(rocksdb::Env::Default());
|
|
|
|
}
|
|
|
|
|
2014-03-05 02:08:05 +01:00
|
|
|
if (FLAGS_prefix_size > FLAGS_key_size) {
|
|
|
|
fprintf(stderr, "prefix size is larger than key size");
|
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
std::vector<std::string> files;
|
2012-06-08 10:11:14 +02:00
|
|
|
FLAGS_env->GetChildren(FLAGS_db, &files);
|
2015-12-16 00:26:20 +01:00
|
|
|
for (size_t i = 0; i < files.size(); i++) {
|
2011-03-18 23:37:00 +01:00
|
|
|
if (Slice(files[i]).starts_with("heap-")) {
|
2013-10-24 16:43:14 +02:00
|
|
|
FLAGS_env->DeleteFile(FLAGS_db + "/" + files[i]);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
}
|
2011-05-21 04:17:43 +02:00
|
|
|
if (!FLAGS_use_existing_db) {
|
benchmark.sh won't run through all tests properly if one specifies wal_dir to be different than db directory.
Summary:
A command line like this to run all the tests:
source benchmark.config.sh && nohup ./benchmark.sh 'bulkload,fillseq,overwrite,filluniquerandom,readrandom,readwhilewriting'
where
benchmark.config.sh is:
export DB_DIR=/data/mysql/rocksdata
export WAL_DIR=/txlogs/rockswal
export OUTPUT_DIR=/root/rocks_benchmarking/output
Will fail for the tests that need a new DB .
Also 1) set disable_data_sync=0 and 2) add debug mode to run through all the tests more quickly
Test Plan: run ./benchmark.sh 'debug,bulkload,fillseq,overwrite,filluniquerandom,readrandom,readwhilewriting' and verify that there are no complaints about WAL dir not being empty.
Reviewers: sdong, yhchiang, rven, igor
Reviewed By: igor
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D30909
2015-01-06 00:36:47 +01:00
|
|
|
Options options;
|
|
|
|
if (!FLAGS_wal_dir.empty()) {
|
|
|
|
options.wal_dir = FLAGS_wal_dir;
|
|
|
|
}
|
2017-07-11 23:05:39 +02:00
|
|
|
#ifndef ROCKSDB_LITE
|
|
|
|
if (use_blob_db_) {
|
|
|
|
blob_db::DestroyBlobDB(FLAGS_db, options, blob_db::BlobDBOptions());
|
|
|
|
}
|
|
|
|
#endif // !ROCKSDB_LITE
|
benchmark.sh won't run through all tests properly if one specifies wal_dir to be different than db directory.
Summary:
A command line like this to run all the tests:
source benchmark.config.sh && nohup ./benchmark.sh 'bulkload,fillseq,overwrite,filluniquerandom,readrandom,readwhilewriting'
where
benchmark.config.sh is:
export DB_DIR=/data/mysql/rocksdata
export WAL_DIR=/txlogs/rockswal
export OUTPUT_DIR=/root/rocks_benchmarking/output
Will fail for the tests that need a new DB .
Also 1) set disable_data_sync=0 and 2) add debug mode to run through all the tests more quickly
Test Plan: run ./benchmark.sh 'debug,bulkload,fillseq,overwrite,filluniquerandom,readrandom,readwhilewriting' and verify that there are no complaints about WAL dir not being empty.
Reviewers: sdong, yhchiang, rven, igor
Reviewed By: igor
Subscribers: dhruba
Differential Revision: https://reviews.facebook.net/D30909
2015-01-06 00:36:47 +01:00
|
|
|
DestroyDB(FLAGS_db, options);
|
2016-10-19 01:30:57 +02:00
|
|
|
if (!FLAGS_wal_dir.empty()) {
|
|
|
|
FLAGS_env->DeleteDir(FLAGS_wal_dir);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (FLAGS_num_multi_db > 1) {
|
|
|
|
FLAGS_env->CreateDir(FLAGS_db);
|
|
|
|
if (!FLAGS_wal_dir.empty()) {
|
|
|
|
FLAGS_env->CreateDir(FLAGS_wal_dir);
|
|
|
|
}
|
|
|
|
}
|
2011-05-21 04:17:43 +02:00
|
|
|
}
|
Auto recovery from out of space errors (#4164)
Summary:
This commit implements automatic recovery from a Status::NoSpace() error
during background operations such as write callback, flush and
compaction. The broad design is as follows -
1. Compaction errors are treated as soft errors and don't put the
database in read-only mode. A compaction is delayed until enough free
disk space is available to accomodate the compaction outputs, which is
estimated based on the input size. This means that users can continue to
write, and we rely on the WriteController to delay or stop writes if the
compaction debt becomes too high due to persistent low disk space
condition
2. Errors during write callback and flush are treated as hard errors,
i.e the database is put in read-only mode and goes back to read-write
only fater certain recovery actions are taken.
3. Both types of recovery rely on the SstFileManagerImpl to poll for
sufficient disk space. We assume that there is a 1-1 mapping between an
SFM and the underlying OS storage container. For cases where multiple
DBs are hosted on a single storage container, the user is expected to
allocate a single SFM instance and use the same one for all the DBs. If
no SFM is specified by the user, DBImpl::Open() will allocate one, but
this will be one per DB and each DB will recover independently. The
recovery implemented by SFM is as follows -
a) On the first occurance of an out of space error during compaction,
subsequent
compactions will be delayed until the disk free space check indicates
enough available space. The required space is computed as the sum of
input sizes.
b) The free space check requirement will be removed once the amount of
free space is greater than the size reserved by in progress
compactions when the first error occured
c) If the out of space error is a hard error, a background thread in
SFM will poll for sufficient headroom before triggering the recovery
of the database and putting it in write-only mode. The headroom is
calculated as the sum of the write_buffer_size of all the DB instances
associated with the SFM
4. EventListener callbacks will be called at the start and completion of
automatic recovery. Users can disable the auto recov ery in the start
callback, and later initiate it manually by calling DB::Resume()
Todo:
1. More extensive testing
2. Add disk full condition to db_stress (follow-on PR)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4164
Differential Revision: D9846378
Pulled By: anand1976
fbshipit-source-id: 80ea875dbd7f00205e19c82215ff6e37da10da4a
2018-09-15 22:36:19 +02:00
|
|
|
|
|
|
|
listener_.reset(new ErrorHandlerListener());
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
~Benchmark() {
|
2015-05-29 23:36:35 +02:00
|
|
|
db_.DeleteDBs();
|
2013-08-23 23:49:57 +02:00
|
|
|
delete prefix_extractor_;
|
2015-03-14 00:41:00 +01:00
|
|
|
if (cache_.get() != nullptr) {
|
|
|
|
// this will leak, but we're shutting down so nobody cares
|
|
|
|
cache_->DisownData();
|
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
2015-02-19 23:27:48 +01:00
|
|
|
Slice AllocateKey(std::unique_ptr<const char[]>* key_guard) {
|
2015-03-26 09:53:42 +01:00
|
|
|
char* data = new char[key_size_];
|
|
|
|
const char* const_data = data;
|
|
|
|
key_guard->reset(const_data);
|
2015-02-19 23:27:48 +01:00
|
|
|
return Slice(key_guard->get(), key_size_);
|
2014-04-08 20:21:09 +02:00
|
|
|
}
|
|
|
|
|
2014-03-05 02:08:05 +01:00
|
|
|
// Generate key according to the given specification and random number.
|
|
|
|
// The resulting key will have the following format (if keys_per_prefix_
|
2016-08-01 21:28:51 +02:00
|
|
|
// is positive), extra trailing bytes are either cut off or padded with '0'.
|
2014-03-05 02:08:05 +01:00
|
|
|
// The prefix value is derived from key value.
|
|
|
|
// ----------------------------
|
|
|
|
// | prefix 00000 | key 00000 |
|
|
|
|
// ----------------------------
|
|
|
|
// If keys_per_prefix_ is 0, the key is simply a binary representation of
|
|
|
|
// random number followed by trailing '0's
|
|
|
|
// ----------------------------
|
|
|
|
// | key 00000 |
|
|
|
|
// ----------------------------
|
2014-04-08 20:21:09 +02:00
|
|
|
void GenerateKeyFromInt(uint64_t v, int64_t num_keys, Slice* key) {
|
2019-03-01 20:14:28 +01:00
|
|
|
if (!keys_.empty()) {
|
|
|
|
assert(FLAGS_use_existing_keys);
|
|
|
|
assert(keys_.size() == static_cast<size_t>(num_keys));
|
|
|
|
assert(v < static_cast<uint64_t>(num_keys));
|
|
|
|
*key = keys_[v];
|
|
|
|
return;
|
|
|
|
}
|
2014-04-08 20:21:09 +02:00
|
|
|
char* start = const_cast<char*>(key->data());
|
2014-03-05 02:08:05 +01:00
|
|
|
char* pos = start;
|
|
|
|
if (keys_per_prefix_ > 0) {
|
|
|
|
int64_t num_prefix = num_keys / keys_per_prefix_;
|
|
|
|
int64_t prefix = v % num_prefix;
|
|
|
|
int bytes_to_fill = std::min(prefix_size_, 8);
|
|
|
|
if (port::kLittleEndian) {
|
|
|
|
for (int i = 0; i < bytes_to_fill; ++i) {
|
|
|
|
pos[i] = (prefix >> ((bytes_to_fill - i - 1) << 3)) & 0xFF;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
memcpy(pos, static_cast<void*>(&prefix), bytes_to_fill);
|
|
|
|
}
|
|
|
|
if (prefix_size_ > 8) {
|
|
|
|
// fill the rest with 0s
|
|
|
|
memset(pos + 8, '0', prefix_size_ - 8);
|
|
|
|
}
|
|
|
|
pos += prefix_size_;
|
|
|
|
}
|
|
|
|
|
|
|
|
int bytes_to_fill = std::min(key_size_ - static_cast<int>(pos - start), 8);
|
|
|
|
if (port::kLittleEndian) {
|
|
|
|
for (int i = 0; i < bytes_to_fill; ++i) {
|
|
|
|
pos[i] = (v >> ((bytes_to_fill - i - 1) << 3)) & 0xFF;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
memcpy(pos, static_cast<void*>(&v), bytes_to_fill);
|
|
|
|
}
|
|
|
|
pos += bytes_to_fill;
|
|
|
|
if (key_size_ > pos - start) {
|
|
|
|
memset(pos, '0', key_size_ - (pos - start));
|
|
|
|
}
|
2013-08-13 22:58:02 +02:00
|
|
|
}
|
|
|
|
|
2016-10-19 01:30:57 +02:00
|
|
|
std::string GetPathForMultiple(std::string base_name, size_t id) {
|
|
|
|
if (!base_name.empty()) {
|
|
|
|
#ifndef OS_WIN
|
|
|
|
if (base_name.back() != '/') {
|
|
|
|
base_name += '/';
|
|
|
|
}
|
|
|
|
#else
|
|
|
|
if (base_name.back() != '\\') {
|
|
|
|
base_name += '\\';
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
}
|
2014-11-25 05:44:49 +01:00
|
|
|
return base_name + ToString(id);
|
2014-04-11 21:15:09 +02:00
|
|
|
}
|
|
|
|
|
2017-04-07 20:25:42 +02:00
|
|
|
void VerifyDBFromDB(std::string& truth_db_name) {
|
|
|
|
DBWithColumnFamilies truth_db;
|
|
|
|
auto s = DB::OpenForReadOnly(open_options_, truth_db_name, &truth_db.db);
|
|
|
|
if (!s.ok()) {
|
|
|
|
fprintf(stderr, "open error: %s\n", s.ToString().c_str());
|
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
ReadOptions ro;
|
|
|
|
ro.total_order_seek = true;
|
|
|
|
std::unique_ptr<Iterator> truth_iter(truth_db.db->NewIterator(ro));
|
|
|
|
std::unique_ptr<Iterator> db_iter(db_.db->NewIterator(ro));
|
|
|
|
// Verify that all the key/values in truth_db are retrivable in db with ::Get
|
|
|
|
fprintf(stderr, "Verifying db >= truth_db with ::Get...\n");
|
|
|
|
for (truth_iter->SeekToFirst(); truth_iter->Valid(); truth_iter->Next()) {
|
|
|
|
std::string value;
|
|
|
|
s = db_.db->Get(ro, truth_iter->key(), &value);
|
|
|
|
assert(s.ok());
|
|
|
|
// TODO(myabandeh): provide debugging hints
|
|
|
|
assert(Slice(value) == truth_iter->value());
|
|
|
|
}
|
|
|
|
// Verify that the db iterator does not give any extra key/value
|
|
|
|
fprintf(stderr, "Verifying db == truth_db...\n");
|
|
|
|
for (db_iter->SeekToFirst(), truth_iter->SeekToFirst(); db_iter->Valid(); db_iter->Next(), truth_iter->Next()) {
|
|
|
|
assert(truth_iter->Valid());
|
|
|
|
assert(truth_iter->value() == db_iter->value());
|
|
|
|
}
|
|
|
|
// No more key should be left unchecked in truth_db
|
|
|
|
assert(!truth_iter->Valid());
|
|
|
|
fprintf(stderr, "...Verified\n");
|
|
|
|
}
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
void Run() {
|
2014-06-23 19:46:16 +02:00
|
|
|
if (!SanityCheck()) {
|
|
|
|
exit(1);
|
|
|
|
}
|
2014-12-23 03:39:28 +01:00
|
|
|
Open(&open_options_);
|
Initial script for the new regression test
Summary:
This diff includes an initial script running a set of benchmarks for
regression test. The script does the following things:
checkout the specified rocksdb commit (or origin/master as default)
make clean && DEBUG_LEVEL=0 make db_bench
setup test directories
run set of benchmarks and store results
Currently, the script will run couple benchmarks, store all the benchmark
output, extract micros per op and percentile information for each benchmark
and store them in a single SUMMARY.csv file. The SUMMARY.csv will make the
follow-up regression detection easier.
In addition, the current script only takes env arguments to set important
attributes of db_bench. Will follow-up with a patch that allows db_bench
to construct options from an options file.
Test Plan:
NUM_KEYS=100 ./tools/regression_test.sh
Sample SUMMARY.csv file:
commit id, benchmark, ms-per-op, p50, p75, p99, p99.9, p99.99
7e23ddf575890510e7d2fc7a79b31a1bbf317917, fillseq, 15.28, 54.66, 77.14, 5000.00, 17900.00, 18483.00
7e23ddf575890510e7d2fc7a79b31a1bbf317917, overwrite, 13.54, 57.69, 86.39, 3000.00, 15600.00, 17013.00
7e23ddf575890510e7d2fc7a79b31a1bbf317917, readrandom, 1.04, 0.80, 1.67, 293.33, 395.00, 504.00
7e23ddf575890510e7d2fc7a79b31a1bbf317917, readwhilewriting, 2.75, 1.01, 1.87, 200.00, 460.00, 485.00
7e23ddf575890510e7d2fc7a79b31a1bbf317917, deleterandom, 3.64, 48.12, 70.09, 200.00, 336.67, 347.00
7e23ddf575890510e7d2fc7a79b31a1bbf317917, seekrandom, 24.31, 391.87, 513.69, 872.73, 990.00, 1048.00
7e23ddf575890510e7d2fc7a79b31a1bbf317917, seekrandomwhilewriting, 14.02, 185.14, 294.15, 700.00, 1440.00, 1527.00
Reviewers: sdong, IslamAbdelRahman, kradhakrishnan, yiwu, andrewkr, gunnarku
Reviewed By: gunnarku
Subscribers: gunnarku, MarkCallaghan, andrewkr, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D57597
2016-05-09 22:32:57 +02:00
|
|
|
PrintHeader();
|
2015-08-11 20:46:15 +02:00
|
|
|
std::stringstream benchmark_stream(FLAGS_benchmarks);
|
|
|
|
std::string name;
|
2016-08-01 21:28:51 +02:00
|
|
|
std::unique_ptr<ExpiredTimeFilter> filter;
|
2015-08-11 20:46:15 +02:00
|
|
|
while (std::getline(benchmark_stream, name, ',')) {
|
2013-08-13 22:58:02 +02:00
|
|
|
// Sanitize parameters
|
2011-08-22 23:08:51 +02:00
|
|
|
num_ = FLAGS_num;
|
2011-09-01 21:08:02 +02:00
|
|
|
reads_ = (FLAGS_reads < 0 ? FLAGS_num : FLAGS_reads);
|
2012-05-31 02:04:35 +02:00
|
|
|
writes_ = (FLAGS_writes < 0 ? FLAGS_num : FLAGS_writes);
|
2016-06-28 23:10:24 +02:00
|
|
|
deletes_ = (FLAGS_deletes < 0 ? FLAGS_num : FLAGS_deletes);
|
2011-08-22 23:08:51 +02:00
|
|
|
value_size_ = FLAGS_value_size;
|
2013-03-01 23:10:09 +01:00
|
|
|
key_size_ = FLAGS_key_size;
|
2014-10-23 03:43:27 +02:00
|
|
|
entries_per_batch_ = FLAGS_batch_size;
|
2018-10-19 02:16:36 +02:00
|
|
|
writes_before_delete_range_ = FLAGS_writes_before_delete_range;
|
2016-11-16 02:33:27 +01:00
|
|
|
writes_per_range_tombstone_ = FLAGS_writes_per_range_tombstone;
|
|
|
|
range_tombstone_width_ = FLAGS_range_tombstone_width;
|
|
|
|
max_num_range_tombstones_ = FLAGS_max_num_range_tombstones;
|
2011-08-22 23:08:51 +02:00
|
|
|
write_options_ = WriteOptions();
|
2015-03-07 01:38:50 +01:00
|
|
|
read_random_exp_range_ = FLAGS_read_random_exp_range;
|
2012-05-31 02:04:35 +02:00
|
|
|
if (FLAGS_sync) {
|
|
|
|
write_options_.sync = true;
|
|
|
|
}
|
2012-08-20 01:05:02 +02:00
|
|
|
write_options_.disableWAL = FLAGS_disable_wal;
|
|
|
|
|
2013-02-15 20:53:17 +01:00
|
|
|
void (Benchmark::*method)(ThreadState*) = nullptr;
|
2015-05-29 23:36:35 +02:00
|
|
|
void (Benchmark::*post_process_method)() = nullptr;
|
|
|
|
|
2011-08-22 23:08:51 +02:00
|
|
|
bool fresh_db = false;
|
2011-09-01 21:08:02 +02:00
|
|
|
int num_threads = FLAGS_threads;
|
2011-03-22 19:32:49 +01:00
|
|
|
|
2016-08-25 21:57:35 +02:00
|
|
|
int num_repeat = 1;
|
|
|
|
int num_warmup = 0;
|
|
|
|
if (!name.empty() && *name.rbegin() == ']') {
|
|
|
|
auto it = name.find('[');
|
|
|
|
if (it == std::string::npos) {
|
|
|
|
fprintf(stderr, "unknown benchmark arguments '%s'\n", name.c_str());
|
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
std::string args = name.substr(it + 1);
|
|
|
|
args.resize(args.size() - 1);
|
|
|
|
name.resize(it);
|
|
|
|
|
|
|
|
std::string bench_arg;
|
|
|
|
std::stringstream args_stream(args);
|
|
|
|
while (std::getline(args_stream, bench_arg, '-')) {
|
|
|
|
if (bench_arg.empty()) {
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
if (bench_arg[0] == 'X') {
|
|
|
|
// Repeat the benchmark n times
|
|
|
|
std::string num_str = bench_arg.substr(1);
|
|
|
|
num_repeat = std::stoi(num_str);
|
|
|
|
} else if (bench_arg[0] == 'W') {
|
|
|
|
// Warm up the benchmark for n times
|
|
|
|
std::string num_str = bench_arg.substr(1);
|
|
|
|
num_warmup = std::stoi(num_str);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2016-10-19 01:30:57 +02:00
|
|
|
// Both fillseqdeterministic and filluniquerandomdeterministic
|
|
|
|
// fill the levels except the max level with UNIQUE_RANDOM
|
|
|
|
// and fill the max level with fillseq and filluniquerandom, respectively
|
|
|
|
if (name == "fillseqdeterministic" ||
|
|
|
|
name == "filluniquerandomdeterministic") {
|
|
|
|
if (!FLAGS_disable_auto_compactions) {
|
|
|
|
fprintf(stderr,
|
|
|
|
"Please disable_auto_compactions in FillDeterministic "
|
|
|
|
"benchmark\n");
|
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
if (num_threads > 1) {
|
|
|
|
fprintf(stderr,
|
|
|
|
"filldeterministic multithreaded not supported"
|
|
|
|
", use 1 thread\n");
|
|
|
|
num_threads = 1;
|
|
|
|
}
|
|
|
|
fresh_db = true;
|
|
|
|
if (name == "fillseqdeterministic") {
|
|
|
|
method = &Benchmark::WriteSeqDeterministic;
|
|
|
|
} else {
|
|
|
|
method = &Benchmark::WriteUniqueRandomDeterministic;
|
|
|
|
}
|
|
|
|
} else if (name == "fillseq") {
|
2011-08-22 23:08:51 +02:00
|
|
|
fresh_db = true;
|
|
|
|
method = &Benchmark::WriteSeq;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "fillbatch") {
|
2011-08-22 23:08:51 +02:00
|
|
|
fresh_db = true;
|
|
|
|
entries_per_batch_ = 1000;
|
|
|
|
method = &Benchmark::WriteSeq;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "fillrandom") {
|
2011-08-22 23:08:51 +02:00
|
|
|
fresh_db = true;
|
|
|
|
method = &Benchmark::WriteRandom;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "filluniquerandom") {
|
2013-06-15 01:17:56 +02:00
|
|
|
fresh_db = true;
|
|
|
|
if (num_threads > 1) {
|
2015-08-11 20:46:15 +02:00
|
|
|
fprintf(stderr,
|
|
|
|
"filluniquerandom multithreaded not supported"
|
|
|
|
", use 1 thread");
|
2014-04-05 00:37:28 +02:00
|
|
|
num_threads = 1;
|
2013-06-15 01:17:56 +02:00
|
|
|
}
|
|
|
|
method = &Benchmark::WriteUniqueRandom;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "overwrite") {
|
2011-08-22 23:08:51 +02:00
|
|
|
method = &Benchmark::WriteRandom;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "fillsync") {
|
2011-08-22 23:08:51 +02:00
|
|
|
fresh_db = true;
|
|
|
|
num_ /= 1000;
|
|
|
|
write_options_.sync = true;
|
|
|
|
method = &Benchmark::WriteRandom;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "fill100K") {
|
2011-08-22 23:08:51 +02:00
|
|
|
fresh_db = true;
|
|
|
|
num_ /= 1000;
|
|
|
|
value_size_ = 100 * 1000;
|
|
|
|
method = &Benchmark::WriteRandom;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "readseq") {
|
2011-08-22 23:08:51 +02:00
|
|
|
method = &Benchmark::ReadSequential;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "readtocache") {
|
2013-12-19 01:50:48 +01:00
|
|
|
method = &Benchmark::ReadSequential;
|
|
|
|
num_threads = 1;
|
|
|
|
reads_ = num_;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "readreverse") {
|
2011-08-22 23:08:51 +02:00
|
|
|
method = &Benchmark::ReadReverse;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "readrandom") {
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
2019-04-11 23:24:09 +02:00
|
|
|
if (FLAGS_multiread_stride) {
|
|
|
|
fprintf(stderr, "entries_per_batch = %" PRIi64 "\n",
|
|
|
|
entries_per_batch_);
|
|
|
|
}
|
2011-08-22 23:08:51 +02:00
|
|
|
method = &Benchmark::ReadRandom;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "readrandomfast") {
|
2014-09-25 20:14:01 +02:00
|
|
|
method = &Benchmark::ReadRandomFast;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "multireadrandom") {
|
2014-09-30 01:05:25 +02:00
|
|
|
fprintf(stderr, "entries_per_batch = %" PRIi64 "\n",
|
|
|
|
entries_per_batch_);
|
2014-04-08 20:21:09 +02:00
|
|
|
method = &Benchmark::MultiReadRandom;
|
Generate mixed workload with Get, Put, Seek in db_bench (#4788)
Summary:
Based on the specific workload models (key access distribution, value size distribution, and iterator scan length distribution, the QPS variation), the MixGraph benchmark generate the synthetic workload according to these distributions which can reflect the real-world workload characteristics.
After user enable the tracing function, they will get the trace file. By analyzing the trace file with the trace_analyzer tool, user can generate a set of statistic data files including. The *_accessed_key_stats.txt, *-accessed_value_size_distribution.txt, *-iterator_length_distribution.txt, and *-qps_stats.txt are mainly used to fit the Matlab model fitting. After that, user can get the parameters of the workload distributions (the modeling details are described: [here](https://github.com/facebook/rocksdb/wiki/RocksDB-Trace%2C-Replay%2C-and-Analyzer))
The key access distribution follows the The two-term power model. The probability density function is: `f(x) = ax^{b}+c`. The corresponding parameters are key_dist_a, key_dist_b, and key_dist_c in db_bench
For the value size distribution and iterator scan length distribution, they both follow the Generalized Pareto Distribution. The probability density function is `f(x) = (1/sigma)(1+k*(x-theta)/sigma))^{-1-1/k)`. The parameters are: value_k, value_theta, value_sigma and iter_k, iter_theta, iter_sigma. For more information about the Generalized Pareto Distribution, users can find the [wiki](https://en.wikipedia.org/wiki/Generalized_Pareto_distribution) and [Matalb page](https://www.mathworks.com/help/stats/generalized-pareto-distribution.html)
As for the QPS, it follows the diurnal pattern. So Sine is a good model to fit it. `F(x) = sine_a*sin(sine_b*x + sine_c) + sine_d`. The trace_will tell you the average QPS in the print out resutls, which is sine_d. After user fit the "*-qps_stats.txt" to the Matlab model, user can get the sine_a, sine_b, and sine_c. By using the 4 parameters, user can control the QPS variation including the period, average, changes.
To use the bench mark, user can indicate the following parameters as examples:
```
-benchmarks="mixgraph" -key_dist_a=0.002312 -key_dist_b=0.3467 -value_k=0.9233 -value_sigma=226.4092 -iter_k=2.517 -iter_sigma=14.236 -mix_get_ratio=0.7 -mix_put_ratio=0.25 -mix_seek_ratio=0.05 -sine_mix_rate_interval_milliseconds=500 -sine_a=15000 -sine_b=1 -sine_d=20000
```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4788
Differential Revision: D13573940
Pulled By: sagar0
fbshipit-source-id: e184c27e07b4f1bc0b436c2be36c5090c1fb0222
2019-01-22 19:40:44 +01:00
|
|
|
} else if (name == "mixgraph") {
|
|
|
|
method = &Benchmark::MixGraph;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "readmissing") {
|
2014-04-08 20:21:09 +02:00
|
|
|
++key_size_;
|
|
|
|
method = &Benchmark::ReadRandom;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "newiterator") {
|
2014-02-03 22:13:36 +01:00
|
|
|
method = &Benchmark::IteratorCreation;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "newiteratorwhilewriting") {
|
2014-04-10 19:15:59 +02:00
|
|
|
num_threads++; // Add extra thread for writing
|
|
|
|
method = &Benchmark::IteratorCreationWhileWriting;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "seekrandom") {
|
2012-04-17 17:36:46 +02:00
|
|
|
method = &Benchmark::SeekRandom;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "seekrandomwhilewriting") {
|
2014-04-11 18:47:20 +02:00
|
|
|
num_threads++; // Add extra thread for writing
|
|
|
|
method = &Benchmark::SeekRandomWhileWriting;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "seekrandomwhilemerging") {
|
2015-03-30 20:28:25 +02:00
|
|
|
num_threads++; // Add extra thread for merging
|
|
|
|
method = &Benchmark::SeekRandomWhileMerging;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "readrandomsmall") {
|
2011-05-21 04:17:43 +02:00
|
|
|
reads_ /= 1000;
|
2011-08-22 23:08:51 +02:00
|
|
|
method = &Benchmark::ReadRandom;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "deleteseq") {
|
2012-04-17 17:36:46 +02:00
|
|
|
method = &Benchmark::DeleteSeq;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "deleterandom") {
|
2012-04-17 17:36:46 +02:00
|
|
|
method = &Benchmark::DeleteRandom;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "readwhilewriting") {
|
2011-09-01 21:08:02 +02:00
|
|
|
num_threads++; // Add extra thread for writing
|
|
|
|
method = &Benchmark::ReadWhileWriting;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "readwhilemerging") {
|
2015-03-18 21:50:52 +01:00
|
|
|
num_threads++; // Add extra thread for writing
|
|
|
|
method = &Benchmark::ReadWhileMerging;
|
2018-05-25 00:45:49 +02:00
|
|
|
} else if (name == "readwhilescanning") {
|
|
|
|
num_threads++; // Add extra thread for scaning
|
|
|
|
method = &Benchmark::ReadWhileScanning;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "readrandomwriterandom") {
|
2012-08-31 08:06:17 +02:00
|
|
|
method = &Benchmark::ReadRandomWriteRandom;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "readrandommergerandom") {
|
2014-01-11 02:33:56 +01:00
|
|
|
if (FLAGS_merge_operator.empty()) {
|
|
|
|
fprintf(stdout, "%-12s : skipped (--merge_operator is unknown)\n",
|
2015-08-11 20:46:15 +02:00
|
|
|
name.c_str());
|
2014-04-08 20:21:09 +02:00
|
|
|
exit(1);
|
2014-01-11 02:33:56 +01:00
|
|
|
}
|
2014-04-08 20:21:09 +02:00
|
|
|
method = &Benchmark::ReadRandomMergeRandom;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "updaterandom") {
|
2013-03-01 20:25:13 +01:00
|
|
|
method = &Benchmark::UpdateRandom;
|
2018-03-06 19:20:52 +01:00
|
|
|
} else if (name == "xorupdaterandom") {
|
|
|
|
method = &Benchmark::XORUpdateRandom;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "appendrandom") {
|
Benchmarking for Merge Operator
Summary:
Updated db_bench and utilities/merge_operators.h to allow for dynamic benchmarking
of merge operators in db_bench. Added a new test (--benchmarks=mergerandom), which performs
a bunch of random Merge() operations over random keys. Also added a "--merge_operator=" flag
so that the tester can easily benchmark different merge operators. Currently supports
the PutOperator and UInt64Add operator. Support for stringappend or list append may come later.
Test Plan:
1. make db_bench
2. Test the PutOperator (simulating Put) as follows:
./db_bench --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom --merge_operator=put
--threads=2
3. Test the UInt64AddOperator (simulating numeric addition) similarly:
./db_bench --value_size=8 --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom
--merge_operator=uint64add --threads=2
Reviewers: haobo, dhruba, zshao, MarkCallaghan
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11535
2013-08-16 02:13:07 +02:00
|
|
|
method = &Benchmark::AppendRandom;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "mergerandom") {
|
Benchmarking for Merge Operator
Summary:
Updated db_bench and utilities/merge_operators.h to allow for dynamic benchmarking
of merge operators in db_bench. Added a new test (--benchmarks=mergerandom), which performs
a bunch of random Merge() operations over random keys. Also added a "--merge_operator=" flag
so that the tester can easily benchmark different merge operators. Currently supports
the PutOperator and UInt64Add operator. Support for stringappend or list append may come later.
Test Plan:
1. make db_bench
2. Test the PutOperator (simulating Put) as follows:
./db_bench --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom --merge_operator=put
--threads=2
3. Test the UInt64AddOperator (simulating numeric addition) similarly:
./db_bench --value_size=8 --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom
--merge_operator=uint64add --threads=2
Reviewers: haobo, dhruba, zshao, MarkCallaghan
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11535
2013-08-16 02:13:07 +02:00
|
|
|
if (FLAGS_merge_operator.empty()) {
|
|
|
|
fprintf(stdout, "%-12s : skipped (--merge_operator is unknown)\n",
|
2015-08-11 20:46:15 +02:00
|
|
|
name.c_str());
|
2014-04-08 20:21:09 +02:00
|
|
|
exit(1);
|
Benchmarking for Merge Operator
Summary:
Updated db_bench and utilities/merge_operators.h to allow for dynamic benchmarking
of merge operators in db_bench. Added a new test (--benchmarks=mergerandom), which performs
a bunch of random Merge() operations over random keys. Also added a "--merge_operator=" flag
so that the tester can easily benchmark different merge operators. Currently supports
the PutOperator and UInt64Add operator. Support for stringappend or list append may come later.
Test Plan:
1. make db_bench
2. Test the PutOperator (simulating Put) as follows:
./db_bench --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom --merge_operator=put
--threads=2
3. Test the UInt64AddOperator (simulating numeric addition) similarly:
./db_bench --value_size=8 --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom
--merge_operator=uint64add --threads=2
Reviewers: haobo, dhruba, zshao, MarkCallaghan
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11535
2013-08-16 02:13:07 +02:00
|
|
|
}
|
2014-04-08 20:21:09 +02:00
|
|
|
method = &Benchmark::MergeRandom;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "randomwithverify") {
|
2013-02-20 02:51:06 +01:00
|
|
|
method = &Benchmark::RandomWithVerify;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "fillseekseq") {
|
SkipListRep::LookaheadIterator
Summary:
This diff introduces the `lookahead` argument to `SkipListFactory()`. This is an
optimization for the tailing use case which includes many seeks. E.g. consider
the following operations on a skip list iterator:
Seek(x), Next(), Next(), Seek(x+2), Next(), Seek(x+3), Next(), Next(), ...
If `lookahead` is positive, `SkipListRep` will return an iterator which also
keeps track of the previously visited node. Seek() then first does a linear
search starting from that node (up to `lookahead` steps). As in the tailing
example above, this may require fewer than ~log(n) comparisons as with regular
skip list search.
Test Plan:
Added a new benchmark (`fillseekseq`) which simulates the usage pattern. It
first writes N records (with consecutive keys), then measures how much time it
takes to read them by calling `Seek()` and `Next()`.
$ time ./db_bench -num 10000000 -benchmarks fillseekseq -prefix_size 1 \
-key_size 8 -write_buffer_size $[1024*1024*1024] -value_size 50 \
-seekseq_next 2 -skip_list_lookahead=0
[...]
DB path: [/dev/shm/rocksdbtest/dbbench]
fillseekseq : 0.389 micros/op 2569047 ops/sec;
real 0m21.806s
user 0m12.106s
sys 0m9.672s
$ time ./db_bench [...] -skip_list_lookahead=2
[...]
DB path: [/dev/shm/rocksdbtest/dbbench]
fillseekseq : 0.153 micros/op 6540684 ops/sec;
real 0m19.469s
user 0m10.192s
sys 0m9.252s
Reviewers: ljin, sdong, igor
Reviewed By: igor
Subscribers: dhruba, leveldb, march, lovro
Differential Revision: https://reviews.facebook.net/D23997
2014-09-24 00:52:28 +02:00
|
|
|
method = &Benchmark::WriteSeqSeekSeq;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "compact") {
|
2011-08-22 23:08:51 +02:00
|
|
|
method = &Benchmark::Compact;
|
2017-05-31 01:33:26 +02:00
|
|
|
} else if (name == "compactall") {
|
|
|
|
CompactAll();
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "crc32c") {
|
2011-08-22 23:08:51 +02:00
|
|
|
method = &Benchmark::Crc32c;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "xxhash") {
|
2014-05-01 20:09:32 +02:00
|
|
|
method = &Benchmark::xxHash;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "acquireload") {
|
2011-08-22 23:08:51 +02:00
|
|
|
method = &Benchmark::AcquireLoad;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "compress") {
|
2014-02-08 03:12:30 +01:00
|
|
|
method = &Benchmark::Compress;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "uncompress") {
|
2014-02-08 03:12:30 +01:00
|
|
|
method = &Benchmark::Uncompress;
|
2015-10-14 21:43:00 +02:00
|
|
|
#ifndef ROCKSDB_LITE
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "randomtransaction") {
|
2015-05-29 23:36:35 +02:00
|
|
|
method = &Benchmark::RandomTransaction;
|
|
|
|
post_process_method = &Benchmark::RandomTransactionVerify;
|
2015-10-14 21:43:00 +02:00
|
|
|
#endif // ROCKSDB_LITE
|
Support for SingleDelete()
Summary:
This patch fixes #7460559. It introduces SingleDelete as a new database
operation. This operation can be used to delete keys that were never
overwritten (no put following another put of the same key). If an overwritten
key is single deleted the behavior is undefined. Single deletion of a
non-existent key has no effect but multiple consecutive single deletions are
not allowed (see limitations).
In contrast to the conventional Delete() operation, the deletion entry is
removed along with the value when the two are lined up in a compaction. Note:
The semantics are similar to @igor's prototype that allowed to have this
behavior on the granularity of a column family (
https://reviews.facebook.net/D42093 ). This new patch, however, is more
aggressive when it comes to removing tombstones: It removes the SingleDelete
together with the value whenever there is no snapshot between them while the
older patch only did this when the sequence number of the deletion was older
than the earliest snapshot.
Most of the complex additions are in the Compaction Iterator, all other changes
should be relatively straightforward. The patch also includes basic support for
single deletions in db_stress and db_bench.
Limitations:
- Not compatible with cuckoo hash tables
- Single deletions cannot be used in combination with merges and normal
deletions on the same key (other keys are not affected by this)
- Consecutive single deletions are currently not allowed (and older version of
this patch supported this so it could be resurrected if needed)
Test Plan: make all check
Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor
Reviewed By: igor
Subscribers: maykov, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D43179
2015-09-17 20:42:56 +02:00
|
|
|
} else if (name == "randomreplacekeys") {
|
|
|
|
fresh_db = true;
|
|
|
|
method = &Benchmark::RandomReplaceKeys;
|
2016-08-01 21:28:51 +02:00
|
|
|
} else if (name == "timeseries") {
|
|
|
|
timestamp_emulator_.reset(new TimestampEmulator());
|
|
|
|
if (FLAGS_expire_style == "compaction_filter") {
|
|
|
|
filter.reset(new ExpiredTimeFilter(timestamp_emulator_));
|
|
|
|
fprintf(stdout, "Compaction filter is used to remove expired data");
|
|
|
|
open_options_.compaction_filter = filter.get();
|
|
|
|
}
|
|
|
|
fresh_db = true;
|
|
|
|
method = &Benchmark::TimeSeries;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "stats") {
|
2013-10-05 07:32:05 +02:00
|
|
|
PrintStats("rocksdb.stats");
|
2017-04-19 01:30:51 +02:00
|
|
|
} else if (name == "resetstats") {
|
|
|
|
ResetStats();
|
2017-04-07 20:25:42 +02:00
|
|
|
} else if (name == "verify") {
|
|
|
|
VerifyDBFromDB(FLAGS_truth_db);
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "levelstats") {
|
2013-10-05 07:32:05 +02:00
|
|
|
PrintStats("rocksdb.levelstats");
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (name == "sstables") {
|
2013-10-05 07:32:05 +02:00
|
|
|
PrintStats("rocksdb.sstables");
|
2018-08-01 09:14:43 +02:00
|
|
|
} else if (name == "replay") {
|
|
|
|
if (num_threads > 1) {
|
|
|
|
fprintf(stderr, "Multi-threaded replay is not yet supported\n");
|
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
if (FLAGS_trace_file == "") {
|
|
|
|
fprintf(stderr, "Please set --trace_file to be replayed from\n");
|
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
method = &Benchmark::Replay;
|
2015-08-11 20:46:15 +02:00
|
|
|
} else if (!name.empty()) { // No error message for empty name
|
|
|
|
fprintf(stderr, "unknown benchmark '%s'\n", name.c_str());
|
|
|
|
exit(1);
|
2011-04-12 21:38:58 +02:00
|
|
|
}
|
2011-08-22 23:08:51 +02:00
|
|
|
|
|
|
|
if (fresh_db) {
|
|
|
|
if (FLAGS_use_existing_db) {
|
|
|
|
fprintf(stdout, "%-12s : skipped (--use_existing_db is true)\n",
|
2015-08-11 20:46:15 +02:00
|
|
|
name.c_str());
|
2013-02-15 20:53:17 +01:00
|
|
|
method = nullptr;
|
2011-08-22 23:08:51 +02:00
|
|
|
} else {
|
2014-08-19 03:15:01 +02:00
|
|
|
if (db_.db != nullptr) {
|
2015-05-29 23:36:35 +02:00
|
|
|
db_.DeleteDBs();
|
2014-12-23 03:39:28 +01:00
|
|
|
DestroyDB(FLAGS_db, open_options_);
|
2014-04-11 21:15:09 +02:00
|
|
|
}
|
2016-10-19 01:30:57 +02:00
|
|
|
Options options = open_options_;
|
2014-04-11 21:15:09 +02:00
|
|
|
for (size_t i = 0; i < multi_dbs_.size(); i++) {
|
2014-08-19 03:15:01 +02:00
|
|
|
delete multi_dbs_[i].db;
|
2016-10-19 01:30:57 +02:00
|
|
|
if (!open_options_.wal_dir.empty()) {
|
|
|
|
options.wal_dir = GetPathForMultiple(open_options_.wal_dir, i);
|
|
|
|
}
|
|
|
|
DestroyDB(GetPathForMultiple(FLAGS_db, i), options);
|
2014-04-11 21:15:09 +02:00
|
|
|
}
|
|
|
|
multi_dbs_.clear();
|
2011-08-22 23:08:51 +02:00
|
|
|
}
|
2014-12-23 03:39:28 +01:00
|
|
|
Open(&open_options_); // use open_options for the last accessed
|
2011-08-22 23:08:51 +02:00
|
|
|
}
|
|
|
|
|
2013-02-15 20:53:17 +01:00
|
|
|
if (method != nullptr) {
|
2013-10-24 16:43:14 +02:00
|
|
|
fprintf(stdout, "DB path: [%s]\n", FLAGS_db.c_str());
|
2018-08-01 09:14:43 +02:00
|
|
|
|
2018-08-03 20:56:23 +02:00
|
|
|
#ifndef ROCKSDB_LITE
|
2018-08-01 09:14:43 +02:00
|
|
|
// A trace_file option can be provided both for trace and replay
|
|
|
|
// operations. But db_bench does not support tracing and replaying at
|
|
|
|
// the same time, for now. So, start tracing only when it is not a
|
|
|
|
// replay.
|
|
|
|
if (FLAGS_trace_file != "" && name != "replay") {
|
|
|
|
std::unique_ptr<TraceWriter> trace_writer;
|
|
|
|
Status s = NewFileTraceWriter(FLAGS_env, EnvOptions(),
|
|
|
|
FLAGS_trace_file, &trace_writer);
|
|
|
|
if (!s.ok()) {
|
|
|
|
fprintf(stderr, "Encountered an error starting a trace, %s\n",
|
|
|
|
s.ToString().c_str());
|
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
s = db_.db->StartTrace(trace_options_, std::move(trace_writer));
|
|
|
|
if (!s.ok()) {
|
|
|
|
fprintf(stderr, "Encountered an error starting a trace, %s\n",
|
|
|
|
s.ToString().c_str());
|
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
fprintf(stdout, "Tracing the workload to: [%s]\n",
|
|
|
|
FLAGS_trace_file.c_str());
|
|
|
|
}
|
2018-08-03 20:56:23 +02:00
|
|
|
#endif // ROCKSDB_LITE
|
2018-08-01 09:14:43 +02:00
|
|
|
|
2016-08-25 21:57:35 +02:00
|
|
|
if (num_warmup > 0) {
|
|
|
|
printf("Warming up benchmark by running %d times\n", num_warmup);
|
|
|
|
}
|
|
|
|
|
|
|
|
for (int i = 0; i < num_warmup; i++) {
|
|
|
|
RunBenchmark(num_threads, name, method);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (num_repeat > 1) {
|
|
|
|
printf("Running benchmark for %d times\n", num_repeat);
|
|
|
|
}
|
|
|
|
|
|
|
|
CombinedStats combined_stats;
|
|
|
|
for (int i = 0; i < num_repeat; i++) {
|
|
|
|
Stats stats = RunBenchmark(num_threads, name, method);
|
|
|
|
combined_stats.AddStats(stats);
|
|
|
|
}
|
|
|
|
if (num_repeat > 1) {
|
|
|
|
combined_stats.Report(name);
|
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
2015-05-29 23:36:35 +02:00
|
|
|
if (post_process_method != nullptr) {
|
|
|
|
(this->*post_process_method)();
|
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
2018-08-01 09:14:43 +02:00
|
|
|
|
2018-08-03 20:56:23 +02:00
|
|
|
#ifndef ROCKSDB_LITE
|
2018-08-01 09:14:43 +02:00
|
|
|
if (name != "replay" && FLAGS_trace_file != "") {
|
|
|
|
Status s = db_.db->EndTrace();
|
|
|
|
if (!s.ok()) {
|
|
|
|
fprintf(stderr, "Encountered an error ending the trace, %s\n",
|
|
|
|
s.ToString().c_str());
|
|
|
|
}
|
|
|
|
}
|
2018-08-03 20:56:23 +02:00
|
|
|
#endif // ROCKSDB_LITE
|
2018-08-01 09:14:43 +02:00
|
|
|
|
2013-06-19 05:28:41 +02:00
|
|
|
if (FLAGS_statistics) {
|
2015-12-16 03:20:10 +01:00
|
|
|
fprintf(stdout, "STATISTICS:\n%s\n", dbstats->ToString().c_str());
|
2013-06-19 05:28:41 +02:00
|
|
|
}
|
2016-05-25 02:27:44 +02:00
|
|
|
if (FLAGS_simcache_size >= 0) {
|
2016-05-24 08:35:23 +02:00
|
|
|
fprintf(stdout, "SIMULATOR CACHE STATISTICS:\n%s\n",
|
2017-07-29 01:23:50 +02:00
|
|
|
static_cast_with_check<SimCache, Cache>(cache_.get())
|
|
|
|
->ToString()
|
|
|
|
.c_str());
|
2016-05-24 08:35:23 +02:00
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
2011-03-22 19:32:49 +01:00
|
|
|
private:
|
2016-08-01 21:28:51 +02:00
|
|
|
std::shared_ptr<TimestampEmulator> timestamp_emulator_;
|
2015-03-30 18:51:11 +02:00
|
|
|
|
2011-08-22 23:08:51 +02:00
|
|
|
struct ThreadArg {
|
|
|
|
Benchmark* bm;
|
|
|
|
SharedState* shared;
|
|
|
|
ThreadState* thread;
|
|
|
|
void (Benchmark::*method)(ThreadState*);
|
|
|
|
};
|
|
|
|
|
|
|
|
static void ThreadBody(void* v) {
|
|
|
|
ThreadArg* arg = reinterpret_cast<ThreadArg*>(v);
|
|
|
|
SharedState* shared = arg->shared;
|
|
|
|
ThreadState* thread = arg->thread;
|
|
|
|
{
|
|
|
|
MutexLock l(&shared->mu);
|
|
|
|
shared->num_initialized++;
|
|
|
|
if (shared->num_initialized >= shared->total) {
|
|
|
|
shared->cv.SignalAll();
|
|
|
|
}
|
|
|
|
while (!shared->start) {
|
|
|
|
shared->cv.Wait();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-11-18 20:32:54 +01:00
|
|
|
SetPerfLevel(static_cast<PerfLevel> (shared->perf_level));
|
2018-10-17 20:18:00 +02:00
|
|
|
perf_context.EnablePerLevelPerfContext();
|
2012-10-03 15:51:11 +02:00
|
|
|
thread->stats.Start(thread->tid);
|
2011-08-22 23:08:51 +02:00
|
|
|
(arg->bm->*(arg->method))(thread);
|
|
|
|
thread->stats.Stop();
|
|
|
|
|
|
|
|
{
|
|
|
|
MutexLock l(&shared->mu);
|
|
|
|
shared->num_done++;
|
|
|
|
if (shared->num_done >= shared->total) {
|
|
|
|
shared->cv.SignalAll();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2016-08-25 21:57:35 +02:00
|
|
|
Stats RunBenchmark(int n, Slice name,
|
|
|
|
void (Benchmark::*method)(ThreadState*)) {
|
2011-08-22 23:08:51 +02:00
|
|
|
SharedState shared;
|
|
|
|
shared.total = n;
|
|
|
|
shared.num_initialized = 0;
|
|
|
|
shared.num_done = 0;
|
|
|
|
shared.start = false;
|
2015-06-18 01:44:52 +02:00
|
|
|
if (FLAGS_benchmark_write_rate_limit > 0) {
|
|
|
|
shared.write_rate_limiter.reset(
|
|
|
|
NewGenericRateLimiter(FLAGS_benchmark_write_rate_limit));
|
|
|
|
}
|
2016-12-19 21:25:35 +01:00
|
|
|
if (FLAGS_benchmark_read_rate_limit > 0) {
|
2017-06-13 23:51:22 +02:00
|
|
|
shared.read_rate_limiter.reset(NewGenericRateLimiter(
|
|
|
|
FLAGS_benchmark_read_rate_limit, 100000 /* refill_period_us */,
|
|
|
|
10 /* fairness */, RateLimiter::Mode::kReadsOnly));
|
2016-12-19 21:25:35 +01:00
|
|
|
}
|
2011-08-22 23:08:51 +02:00
|
|
|
|
db_bench periodically writes QPS to CSV file
Summary:
This is part of an effort to better understand and optimize RocksDB stalls under high load. I added a feature to db_bench to periodically write QPS to CSV files. That way we can nicely see how our QPS changes in time (especially when DB is stalled) and can do a better job of evaluating our stall system (i.e. we want the QPS to be as constant as possible, as opposed to having bunch of stalls)
Cool part of CSV files is that we can easily graph them -- there are a bunch of tools available.
Test Plan:
Ran ./db_bench --report_interval_seconds=10 --benchmarks=fillrandom --num=10000000
and observed this in report.csv:
secs_elapsed,interval_qps
10,2725860
20,1980480
30,1863456
40,1454359
50,1460389
Reviewers: sdong, MarkCallaghan, rven, yhchiang
Reviewed By: yhchiang
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D40047
2015-06-12 23:31:53 +02:00
|
|
|
std::unique_ptr<ReporterAgent> reporter_agent;
|
|
|
|
if (FLAGS_report_interval_seconds > 0) {
|
|
|
|
reporter_agent.reset(new ReporterAgent(FLAGS_env, FLAGS_report_file,
|
|
|
|
FLAGS_report_interval_seconds));
|
|
|
|
}
|
|
|
|
|
2011-08-22 23:08:51 +02:00
|
|
|
ThreadArg* arg = new ThreadArg[n];
|
2014-07-07 19:53:31 +02:00
|
|
|
|
2011-08-22 23:08:51 +02:00
|
|
|
for (int i = 0; i < n; i++) {
|
2014-07-07 19:53:31 +02:00
|
|
|
#ifdef NUMA
|
|
|
|
if (FLAGS_enable_numa) {
|
|
|
|
// Performs a local allocation of memory to threads in numa node.
|
|
|
|
int n_nodes = numa_num_task_nodes(); // Number of nodes in NUMA.
|
|
|
|
numa_exit_on_error = 1;
|
|
|
|
int numa_node = i % n_nodes;
|
|
|
|
bitmask* nodes = numa_allocate_nodemask();
|
|
|
|
numa_bitmask_clearall(nodes);
|
|
|
|
numa_bitmask_setbit(nodes, numa_node);
|
|
|
|
// numa_bind() call binds the process to the node and these
|
|
|
|
// properties are passed on to the thread that is created in
|
|
|
|
// StartThread method called later in the loop.
|
|
|
|
numa_bind(nodes);
|
|
|
|
numa_set_strict(1);
|
|
|
|
numa_free_nodemask(nodes);
|
|
|
|
}
|
|
|
|
#endif
|
2011-08-22 23:08:51 +02:00
|
|
|
arg[i].bm = this;
|
|
|
|
arg[i].method = method;
|
|
|
|
arg[i].shared = &shared;
|
|
|
|
arg[i].thread = new ThreadState(i);
|
db_bench periodically writes QPS to CSV file
Summary:
This is part of an effort to better understand and optimize RocksDB stalls under high load. I added a feature to db_bench to periodically write QPS to CSV files. That way we can nicely see how our QPS changes in time (especially when DB is stalled) and can do a better job of evaluating our stall system (i.e. we want the QPS to be as constant as possible, as opposed to having bunch of stalls)
Cool part of CSV files is that we can easily graph them -- there are a bunch of tools available.
Test Plan:
Ran ./db_bench --report_interval_seconds=10 --benchmarks=fillrandom --num=10000000
and observed this in report.csv:
secs_elapsed,interval_qps
10,2725860
20,1980480
30,1863456
40,1454359
50,1460389
Reviewers: sdong, MarkCallaghan, rven, yhchiang
Reviewed By: yhchiang
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D40047
2015-06-12 23:31:53 +02:00
|
|
|
arg[i].thread->stats.SetReporterAgent(reporter_agent.get());
|
2011-09-01 21:08:02 +02:00
|
|
|
arg[i].thread->shared = &shared;
|
2012-06-08 10:11:14 +02:00
|
|
|
FLAGS_env->StartThread(ThreadBody, &arg[i]);
|
2011-08-22 23:08:51 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
shared.mu.Lock();
|
|
|
|
while (shared.num_initialized < n) {
|
|
|
|
shared.cv.Wait();
|
|
|
|
}
|
|
|
|
|
|
|
|
shared.start = true;
|
|
|
|
shared.cv.SignalAll();
|
|
|
|
while (shared.num_done < n) {
|
|
|
|
shared.cv.Wait();
|
|
|
|
}
|
|
|
|
shared.mu.Unlock();
|
|
|
|
|
2013-03-19 14:36:03 +01:00
|
|
|
// Stats for some threads can be excluded.
|
|
|
|
Stats merge_stats;
|
|
|
|
for (int i = 0; i < n; i++) {
|
|
|
|
merge_stats.Merge(arg[i].thread->stats);
|
2011-08-22 23:08:51 +02:00
|
|
|
}
|
2013-03-19 14:36:03 +01:00
|
|
|
merge_stats.Report(name);
|
2011-08-22 23:08:51 +02:00
|
|
|
|
2016-09-02 19:58:08 +02:00
|
|
|
for (int i = 0; i < n; i++) {
|
|
|
|
delete arg[i].thread;
|
|
|
|
}
|
|
|
|
delete[] arg;
|
|
|
|
|
2016-08-25 21:57:35 +02:00
|
|
|
return merge_stats;
|
2011-08-22 23:08:51 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
void Crc32c(ThreadState* thread) {
|
2011-03-25 21:27:43 +01:00
|
|
|
// Checksum about 500MB of data total
|
Port 3 way SSE4.2 crc32c implementation from Folly
Summary:
**# Summary**
RocksDB uses SSE crc32 intrinsics to calculate the crc32 values but it does it in single way fashion (not pipelined on single CPU core). Intel's whitepaper () published an algorithm that uses 3-way pipelining for the crc32 intrinsics, then use pclmulqdq intrinsic to combine the values. Because pclmulqdq has overhead on its own, this algorithm will show perf gains on buffers larger than 216 bytes, which makes RocksDB a perfect user, since most of the buffers RocksDB call crc32c on is over 4KB. Initial db_bench show tremendous CPU gain.
This change uses the 3-way SSE algorithm by default. The old SSE algorithm is now behind a compiler tag NO_THREEWAY_CRC32C. If user compiles the code with NO_THREEWAY_CRC32C=1 then the old SSE Crc32c algorithm would be used. If the server does not have SSE4.2 at the run time the slow way (Non SSE) will be used.
**# Performance Test Results**
We ran the FillRandom and ReadRandom benchmarks in db_bench. ReadRandom is the point of interest here since it calculates the CRC32 for the in-mem buffers. We did 3 runs for each algorithm.
Before this change the CRC32 value computation takes about 11.5% of total CPU cost, and with the new 3-way algorithm it reduced to around 4.5%. The overall throughput also improved from 25.53MB/s to 27.63MB/s.
1) ReadRandom in db_bench overall metrics
PER RUN
Algorithm | run | micros/op | ops/sec |Throughput (MB/s)
3-way | 1 | 4.143 | 241387 | 26.7
3-way | 2 | 3.775 | 264872 | 29.3
3-way | 3 | 4.116 | 242929 | 26.9
FastCrc32c|1 | 4.037 | 247727 | 27.4
FastCrc32c|2 | 4.648 | 215166 | 23.8
FastCrc32c|3 | 4.352 | 229799 | 25.4
AVG
Algorithm | Average of micros/op | Average of ops/sec | Average of Throughput (MB/s)
3-way | 4.01 | 249,729 | 27.63
FastCrc32c | 4.35 | 230,897 | 25.53
2) Crc32c computation CPU cost (inclusive samples percentage)
PER RUN
Implementation | run | TotalSamples | Crc32c percentage
3-way | 1 | 4,572,250,000 | 4.37%
3-way | 2 | 3,779,250,000 | 4.62%
3-way | 3 | 4,129,500,000 | 4.48%
FastCrc32c | 1 | 4,663,500,000 | 11.24%
FastCrc32c | 2 | 4,047,500,000 | 12.34%
FastCrc32c | 3 | 4,366,750,000 | 11.68%
**# Test Plan**
make -j64 corruption_test && ./corruption_test
By default it uses 3-way SSE algorithm
NO_THREEWAY_CRC32C=1 make -j64 corruption_test && ./corruption_test
make clean && DEBUG_LEVEL=0 make -j64 db_bench
make clean && DEBUG_LEVEL=0 NO_THREEWAY_CRC32C=1 make -j64 db_bench
Closes https://github.com/facebook/rocksdb/pull/3173
Differential Revision: D6330882
Pulled By: yingsu00
fbshipit-source-id: 8ec3d89719533b63b536a736663ca6f0dd4482e9
2017-12-20 03:20:50 +01:00
|
|
|
const int size = FLAGS_block_size; // use --block_size option for db_bench
|
|
|
|
std::string labels = "(" + ToString(FLAGS_block_size) + " per op)";
|
|
|
|
const char* label = labels.c_str();
|
|
|
|
|
2011-03-28 22:43:44 +02:00
|
|
|
std::string data(size, 'x');
|
2011-03-25 21:27:43 +01:00
|
|
|
int64_t bytes = 0;
|
|
|
|
uint32_t crc = 0;
|
|
|
|
while (bytes < 500 * 1048576) {
|
|
|
|
crc = crc32c::Value(data.data(), size);
|
2015-12-08 02:33:18 +01:00
|
|
|
thread->stats.FinishedOps(nullptr, nullptr, 1, kCrc);
|
2011-03-25 21:27:43 +01:00
|
|
|
bytes += size;
|
|
|
|
}
|
|
|
|
// Print so result is not dead
|
|
|
|
fprintf(stderr, "... crc=0x%x\r", static_cast<unsigned int>(crc));
|
|
|
|
|
2011-08-22 23:08:51 +02:00
|
|
|
thread->stats.AddBytes(bytes);
|
|
|
|
thread->stats.AddMessage(label);
|
2011-03-25 21:27:43 +01:00
|
|
|
}
|
|
|
|
|
2014-05-01 20:09:32 +02:00
|
|
|
void xxHash(ThreadState* thread) {
|
|
|
|
// Checksum about 500MB of data total
|
|
|
|
const int size = 4096;
|
|
|
|
const char* label = "(4K per op)";
|
|
|
|
std::string data(size, 'x');
|
|
|
|
int64_t bytes = 0;
|
|
|
|
unsigned int xxh32 = 0;
|
|
|
|
while (bytes < 500 * 1048576) {
|
|
|
|
xxh32 = XXH32(data.data(), size, 0);
|
2015-12-08 02:33:18 +01:00
|
|
|
thread->stats.FinishedOps(nullptr, nullptr, 1, kHash);
|
2014-05-01 20:09:32 +02:00
|
|
|
bytes += size;
|
|
|
|
}
|
|
|
|
// Print so result is not dead
|
|
|
|
fprintf(stderr, "... xxh32=0x%x\r", static_cast<unsigned int>(xxh32));
|
|
|
|
|
|
|
|
thread->stats.AddBytes(bytes);
|
|
|
|
thread->stats.AddMessage(label);
|
|
|
|
}
|
|
|
|
|
2011-08-22 23:08:51 +02:00
|
|
|
void AcquireLoad(ThreadState* thread) {
|
2011-05-28 02:53:58 +02:00
|
|
|
int dummy;
|
2014-10-27 23:41:05 +01:00
|
|
|
std::atomic<void*> ap(&dummy);
|
2011-05-28 02:53:58 +02:00
|
|
|
int count = 0;
|
2013-02-15 20:53:17 +01:00
|
|
|
void *ptr = nullptr;
|
2011-08-22 23:08:51 +02:00
|
|
|
thread->stats.AddMessage("(each op is 1000 loads)");
|
2011-05-28 02:53:58 +02:00
|
|
|
while (count < 100000) {
|
|
|
|
for (int i = 0; i < 1000; i++) {
|
2014-10-27 22:50:21 +01:00
|
|
|
ptr = ap.load(std::memory_order_acquire);
|
2011-05-28 02:53:58 +02:00
|
|
|
}
|
|
|
|
count++;
|
2015-12-08 02:33:18 +01:00
|
|
|
thread->stats.FinishedOps(nullptr, nullptr, 1, kOthers);
|
2011-05-28 02:53:58 +02:00
|
|
|
}
|
2016-02-16 15:17:31 +01:00
|
|
|
if (ptr == nullptr) exit(1); // Disable unused variable warning.
|
2011-05-28 02:53:58 +02:00
|
|
|
}
|
|
|
|
|
2014-02-08 03:12:30 +01:00
|
|
|
void Compress(ThreadState *thread) {
|
2011-08-22 23:08:51 +02:00
|
|
|
RandomGenerator gen;
|
2014-08-25 23:22:05 +02:00
|
|
|
Slice input = gen.Generate(FLAGS_block_size);
|
2011-04-12 21:38:58 +02:00
|
|
|
int64_t bytes = 0;
|
|
|
|
int64_t produced = 0;
|
|
|
|
bool ok = true;
|
|
|
|
std::string compressed;
|
2019-01-19 04:10:17 +01:00
|
|
|
CompressionOptions opts;
|
|
|
|
CompressionContext context(FLAGS_compression_type_e);
|
|
|
|
CompressionInfo info(opts, context, CompressionDict::GetEmptyDict(),
|
2019-03-18 20:07:35 +01:00
|
|
|
FLAGS_compression_type_e,
|
|
|
|
FLAGS_sample_for_compression);
|
2014-02-08 03:12:30 +01:00
|
|
|
// Compress 1G
|
|
|
|
while (ok && bytes < int64_t(1) << 30) {
|
2016-01-15 23:13:08 +01:00
|
|
|
compressed.clear();
|
2019-01-19 04:10:17 +01:00
|
|
|
ok = CompressSlice(info, input, &compressed);
|
2011-04-12 21:38:58 +02:00
|
|
|
produced += compressed.size();
|
|
|
|
bytes += input.size();
|
2015-12-08 02:33:18 +01:00
|
|
|
thread->stats.FinishedOps(nullptr, nullptr, 1, kCompress);
|
2011-04-12 21:38:58 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
if (!ok) {
|
2014-02-08 03:12:30 +01:00
|
|
|
thread->stats.AddMessage("(compression failure)");
|
2011-04-12 21:38:58 +02:00
|
|
|
} else {
|
2016-12-15 04:20:42 +01:00
|
|
|
char buf[340];
|
2011-04-12 21:38:58 +02:00
|
|
|
snprintf(buf, sizeof(buf), "(output: %.1f%%)",
|
|
|
|
(produced * 100.0) / bytes);
|
2011-08-22 23:08:51 +02:00
|
|
|
thread->stats.AddMessage(buf);
|
|
|
|
thread->stats.AddBytes(bytes);
|
2011-04-12 21:38:58 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2014-02-08 03:12:30 +01:00
|
|
|
void Uncompress(ThreadState *thread) {
|
2011-08-22 23:08:51 +02:00
|
|
|
RandomGenerator gen;
|
2014-08-25 23:22:05 +02:00
|
|
|
Slice input = gen.Generate(FLAGS_block_size);
|
2011-04-12 21:38:58 +02:00
|
|
|
std::string compressed;
|
2014-02-08 03:12:30 +01:00
|
|
|
|
2019-01-19 04:10:17 +01:00
|
|
|
CompressionContext compression_ctx(FLAGS_compression_type_e);
|
|
|
|
CompressionOptions compression_opts;
|
2019-03-18 20:07:35 +01:00
|
|
|
CompressionInfo compression_info(
|
|
|
|
compression_opts, compression_ctx, CompressionDict::GetEmptyDict(),
|
|
|
|
FLAGS_compression_type_e, FLAGS_sample_for_compression);
|
2018-06-04 21:04:52 +02:00
|
|
|
UncompressionContext uncompression_ctx(FLAGS_compression_type_e);
|
2019-01-19 04:10:17 +01:00
|
|
|
UncompressionInfo uncompression_info(uncompression_ctx,
|
|
|
|
UncompressionDict::GetEmptyDict(),
|
|
|
|
FLAGS_compression_type_e);
|
2018-06-04 21:04:52 +02:00
|
|
|
|
2019-01-19 04:10:17 +01:00
|
|
|
bool ok = CompressSlice(compression_info, input, &compressed);
|
2011-04-12 21:38:58 +02:00
|
|
|
int64_t bytes = 0;
|
2014-02-08 03:12:30 +01:00
|
|
|
int decompress_size;
|
|
|
|
while (ok && bytes < 1024 * 1048576) {
|
2018-10-03 02:21:54 +02:00
|
|
|
CacheAllocationPtr uncompressed;
|
2014-02-08 03:12:30 +01:00
|
|
|
switch (FLAGS_compression_type_e) {
|
2015-12-08 02:12:43 +01:00
|
|
|
case rocksdb::kSnappyCompression: {
|
|
|
|
// get size and allocate here to make comparison fair
|
|
|
|
size_t ulength = 0;
|
|
|
|
if (!Snappy_GetUncompressedLength(compressed.data(),
|
|
|
|
compressed.size(), &ulength)) {
|
|
|
|
ok = false;
|
|
|
|
break;
|
|
|
|
}
|
2018-10-03 02:21:54 +02:00
|
|
|
uncompressed = AllocateBlock(ulength, nullptr);
|
2015-12-08 02:12:43 +01:00
|
|
|
ok = Snappy_Uncompress(compressed.data(), compressed.size(),
|
2018-10-03 02:21:54 +02:00
|
|
|
uncompressed.get());
|
2015-12-08 02:12:43 +01:00
|
|
|
break;
|
|
|
|
}
|
2014-02-08 03:12:30 +01:00
|
|
|
case rocksdb::kZlibCompression:
|
2019-01-19 04:10:17 +01:00
|
|
|
uncompressed = Zlib_Uncompress(uncompression_info, compressed.data(),
|
2018-06-05 21:51:05 +02:00
|
|
|
compressed.size(), &decompress_size, 2);
|
2018-10-03 02:21:54 +02:00
|
|
|
ok = uncompressed.get() != nullptr;
|
2014-02-08 03:12:30 +01:00
|
|
|
break;
|
|
|
|
case rocksdb::kBZip2Compression:
|
2015-01-09 22:04:06 +01:00
|
|
|
uncompressed = BZip2_Uncompress(compressed.data(), compressed.size(),
|
2015-01-15 01:24:24 +01:00
|
|
|
&decompress_size, 2);
|
2018-10-03 02:21:54 +02:00
|
|
|
ok = uncompressed.get() != nullptr;
|
2014-02-08 03:12:30 +01:00
|
|
|
break;
|
|
|
|
case rocksdb::kLZ4Compression:
|
2019-01-19 04:10:17 +01:00
|
|
|
uncompressed = LZ4_Uncompress(uncompression_info, compressed.data(),
|
2018-06-05 21:51:05 +02:00
|
|
|
compressed.size(), &decompress_size, 2);
|
2018-10-03 02:21:54 +02:00
|
|
|
ok = uncompressed.get() != nullptr;
|
2014-02-08 03:12:30 +01:00
|
|
|
break;
|
|
|
|
case rocksdb::kLZ4HCCompression:
|
2019-01-19 04:10:17 +01:00
|
|
|
uncompressed = LZ4_Uncompress(uncompression_info, compressed.data(),
|
2018-06-05 21:51:05 +02:00
|
|
|
compressed.size(), &decompress_size, 2);
|
2018-10-03 02:21:54 +02:00
|
|
|
ok = uncompressed.get() != nullptr;
|
2014-02-08 03:12:30 +01:00
|
|
|
break;
|
2016-04-20 07:54:24 +02:00
|
|
|
case rocksdb::kXpressCompression:
|
2018-10-03 02:21:54 +02:00
|
|
|
uncompressed.reset(XPRESS_Uncompress(
|
|
|
|
compressed.data(), compressed.size(), &decompress_size));
|
|
|
|
ok = uncompressed.get() != nullptr;
|
2016-04-20 07:54:24 +02:00
|
|
|
break;
|
2016-09-02 00:28:40 +02:00
|
|
|
case rocksdb::kZSTD:
|
2019-01-19 04:10:17 +01:00
|
|
|
uncompressed = ZSTD_Uncompress(uncompression_info, compressed.data(),
|
2018-06-05 21:51:05 +02:00
|
|
|
compressed.size(), &decompress_size);
|
2018-10-03 02:21:54 +02:00
|
|
|
ok = uncompressed.get() != nullptr;
|
2015-08-28 00:40:42 +02:00
|
|
|
break;
|
2014-02-08 03:12:30 +01:00
|
|
|
default:
|
|
|
|
ok = false;
|
|
|
|
}
|
2011-07-21 04:40:18 +02:00
|
|
|
bytes += input.size();
|
2015-12-08 02:33:18 +01:00
|
|
|
thread->stats.FinishedOps(nullptr, nullptr, 1, kUncompress);
|
2011-04-12 21:38:58 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
if (!ok) {
|
2014-02-08 03:12:30 +01:00
|
|
|
thread->stats.AddMessage("(compression failure)");
|
2011-04-12 21:38:58 +02:00
|
|
|
} else {
|
2011-08-22 23:08:51 +02:00
|
|
|
thread->stats.AddBytes(bytes);
|
2011-04-12 21:38:58 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2016-06-03 01:24:14 +02:00
|
|
|
// Returns true if the options is initialized from the specified
|
|
|
|
// options file.
|
|
|
|
bool InitializeOptionsFromFile(Options* opts) {
|
|
|
|
#ifndef ROCKSDB_LITE
|
|
|
|
printf("Initializing RocksDB Options from the specified file\n");
|
|
|
|
DBOptions db_opts;
|
|
|
|
std::vector<ColumnFamilyDescriptor> cf_descs;
|
|
|
|
if (FLAGS_options_file != "") {
|
|
|
|
auto s = LoadOptionsFromFile(FLAGS_options_file, Env::Default(), &db_opts,
|
|
|
|
&cf_descs);
|
|
|
|
if (s.ok()) {
|
|
|
|
*opts = Options(db_opts, cf_descs[0].options);
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
fprintf(stderr, "Unable to load options file %s --- %s\n",
|
|
|
|
FLAGS_options_file.c_str(), s.ToString().c_str());
|
|
|
|
exit(1);
|
|
|
|
}
|
2018-04-16 02:19:57 +02:00
|
|
|
#else
|
|
|
|
(void)opts;
|
2016-06-03 01:24:14 +02:00
|
|
|
#endif
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
void InitializeOptionsFromFlags(Options* opts) {
|
|
|
|
printf("Initializing RocksDB Options from command-line flags\n");
|
2014-12-08 21:53:24 +01:00
|
|
|
Options& options = *opts;
|
|
|
|
|
2014-08-19 03:15:01 +02:00
|
|
|
assert(db_.db == nullptr);
|
2014-12-08 21:53:24 +01:00
|
|
|
|
2016-06-03 01:24:14 +02:00
|
|
|
options.max_open_files = FLAGS_open_files;
|
2017-06-02 23:13:59 +02:00
|
|
|
if (FLAGS_cost_write_buffer_to_cache || FLAGS_db_write_buffer_size != 0) {
|
|
|
|
options.write_buffer_manager.reset(
|
|
|
|
new WriteBufferManager(FLAGS_db_write_buffer_size, cache_));
|
|
|
|
}
|
2011-03-22 19:32:49 +01:00
|
|
|
options.write_buffer_size = FLAGS_write_buffer_size;
|
2012-10-19 23:00:53 +02:00
|
|
|
options.max_write_buffer_number = FLAGS_max_write_buffer_number;
|
2013-06-11 23:23:58 +02:00
|
|
|
options.min_write_buffer_number_to_merge =
|
|
|
|
FLAGS_min_write_buffer_number_to_merge;
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
options.max_write_buffer_number_to_maintain =
|
|
|
|
FLAGS_max_write_buffer_number_to_maintain;
|
2017-07-08 03:11:53 +02:00
|
|
|
options.max_background_jobs = FLAGS_max_background_jobs;
|
2012-10-19 23:00:53 +02:00
|
|
|
options.max_background_compactions = FLAGS_max_background_compactions;
|
2015-08-21 23:25:34 +02:00
|
|
|
options.max_subcompactions = static_cast<uint32_t>(FLAGS_subcompactions);
|
2014-03-05 19:27:17 +01:00
|
|
|
options.max_background_flushes = FLAGS_max_background_flushes;
|
2013-10-24 16:43:14 +02:00
|
|
|
options.compaction_style = FLAGS_compaction_style_e;
|
2015-09-22 02:16:31 +02:00
|
|
|
options.compaction_pri = FLAGS_compaction_pri_e;
|
2016-10-28 19:36:05 +02:00
|
|
|
options.allow_mmap_reads = FLAGS_mmap_read;
|
|
|
|
options.allow_mmap_writes = FLAGS_mmap_write;
|
|
|
|
options.use_direct_reads = FLAGS_use_direct_reads;
|
2017-04-13 22:07:33 +02:00
|
|
|
options.use_direct_io_for_flush_and_compaction =
|
|
|
|
FLAGS_use_direct_io_for_flush_and_compaction;
|
2017-03-28 22:26:53 +02:00
|
|
|
#ifndef ROCKSDB_LITE
|
2019-02-15 18:48:44 +01:00
|
|
|
options.ttl = FLAGS_fifo_compaction_ttl;
|
2017-03-25 00:58:51 +01:00
|
|
|
options.compaction_options_fifo = CompactionOptionsFIFO(
|
2017-05-05 03:14:29 +02:00
|
|
|
FLAGS_fifo_compaction_max_table_files_size_mb * 1024 * 1024,
|
2019-02-15 18:48:44 +01:00
|
|
|
FLAGS_fifo_compaction_allow_compaction);
|
2017-03-28 22:26:53 +02:00
|
|
|
#endif // ROCKSDB_LITE
|
2014-07-17 02:32:30 +02:00
|
|
|
if (FLAGS_prefix_size != 0) {
|
2014-03-10 20:56:46 +01:00
|
|
|
options.prefix_extractor.reset(
|
|
|
|
NewFixedPrefixTransform(FLAGS_prefix_size));
|
|
|
|
}
|
2014-08-27 19:39:31 +02:00
|
|
|
if (FLAGS_use_uint64_comparator) {
|
|
|
|
options.comparator = test::Uint64Comparator();
|
|
|
|
if (FLAGS_key_size != 8) {
|
|
|
|
fprintf(stderr, "Using Uint64 comparator but key size is not 8.\n");
|
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
}
|
2016-12-10 01:36:43 +01:00
|
|
|
if (FLAGS_use_stderr_info_logger) {
|
|
|
|
options.info_log.reset(new StderrLogger());
|
|
|
|
}
|
2016-07-27 03:05:30 +02:00
|
|
|
options.memtable_huge_page_size = FLAGS_memtable_use_huge_page ? 2048 : 0;
|
2016-06-04 02:02:10 +02:00
|
|
|
options.memtable_prefix_bloom_size_ratio = FLAGS_memtable_bloom_size_ratio;
|
2019-02-19 21:12:25 +01:00
|
|
|
options.memtable_whole_key_filtering = FLAGS_memtable_whole_key_filtering;
|
2016-12-02 01:50:59 +01:00
|
|
|
if (FLAGS_memtable_insert_with_hint_prefix_size > 0) {
|
|
|
|
options.memtable_insert_with_hint_prefix_extractor.reset(
|
|
|
|
NewCappedPrefixTransform(
|
|
|
|
FLAGS_memtable_insert_with_hint_prefix_size));
|
|
|
|
}
|
2014-03-28 17:21:20 +01:00
|
|
|
options.bloom_locality = FLAGS_bloom_locality;
|
2015-09-30 18:51:31 +02:00
|
|
|
options.max_file_opening_threads = FLAGS_file_opening_threads;
|
2015-08-27 00:25:59 +02:00
|
|
|
options.new_table_reader_for_compaction_inputs =
|
|
|
|
FLAGS_new_table_reader_for_compaction_inputs;
|
|
|
|
options.compaction_readahead_size = FLAGS_compaction_readahead_size;
|
2015-10-27 22:44:16 +01:00
|
|
|
options.random_access_max_buffer_size = FLAGS_random_access_max_buffer_size;
|
2015-10-30 06:10:25 +01:00
|
|
|
options.writable_file_max_buffer_size = FLAGS_writable_file_max_buffer_size;
|
2012-08-27 21:10:26 +02:00
|
|
|
options.use_fsync = FLAGS_use_fsync;
|
2012-10-15 19:10:56 +02:00
|
|
|
options.num_levels = FLAGS_num_levels;
|
2012-08-16 00:44:23 +02:00
|
|
|
options.target_file_size_base = FLAGS_target_file_size_base;
|
|
|
|
options.target_file_size_multiplier = FLAGS_target_file_size_multiplier;
|
|
|
|
options.max_bytes_for_level_base = FLAGS_max_bytes_for_level_base;
|
options.level_compaction_dynamic_level_bytes to allow RocksDB to pick size bases of levels dynamically.
Summary:
When having fixed max_bytes_for_level_base, the ratio of size of largest level and the second one can range from 0 to the multiplier. This makes LSM tree frequently irregular and unpredictable. It can also cause poor space amplification in some cases.
In this improvement (proposed by Igor Kabiljo), we introduce a parameter option.level_compaction_use_dynamic_max_bytes. When turning it on, RocksDB is free to pick a level base in the range of (options.max_bytes_for_level_base/options.max_bytes_for_level_multiplier, options.max_bytes_for_level_base] so that real level ratios are close to options.max_bytes_for_level_multiplier.
Test Plan: New unit tests and pass tests suites including valgrind.
Reviewers: MarkCallaghan, rven, yhchiang, igor, ikabiljo
Reviewed By: ikabiljo
Subscribers: yoshinorim, ikabiljo, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D31437
2015-02-05 20:44:17 +01:00
|
|
|
options.level_compaction_dynamic_level_bytes =
|
|
|
|
FLAGS_level_compaction_dynamic_level_bytes;
|
2012-08-16 00:44:23 +02:00
|
|
|
options.max_bytes_for_level_multiplier =
|
|
|
|
FLAGS_max_bytes_for_level_multiplier;
|
2014-03-05 19:27:17 +01:00
|
|
|
if ((FLAGS_prefix_size == 0) && (FLAGS_rep_factory == kPrefixHash ||
|
|
|
|
FLAGS_rep_factory == kHashLinkedList)) {
|
|
|
|
fprintf(stderr, "prefix_size should be non-zero if PrefixHash or "
|
|
|
|
"HashLinkedList memtablerep is used\n");
|
2013-08-23 08:10:02 +02:00
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
switch (FLAGS_rep_factory) {
|
|
|
|
case kSkipList:
|
SkipListRep::LookaheadIterator
Summary:
This diff introduces the `lookahead` argument to `SkipListFactory()`. This is an
optimization for the tailing use case which includes many seeks. E.g. consider
the following operations on a skip list iterator:
Seek(x), Next(), Next(), Seek(x+2), Next(), Seek(x+3), Next(), Next(), ...
If `lookahead` is positive, `SkipListRep` will return an iterator which also
keeps track of the previously visited node. Seek() then first does a linear
search starting from that node (up to `lookahead` steps). As in the tailing
example above, this may require fewer than ~log(n) comparisons as with regular
skip list search.
Test Plan:
Added a new benchmark (`fillseekseq`) which simulates the usage pattern. It
first writes N records (with consecutive keys), then measures how much time it
takes to read them by calling `Seek()` and `Next()`.
$ time ./db_bench -num 10000000 -benchmarks fillseekseq -prefix_size 1 \
-key_size 8 -write_buffer_size $[1024*1024*1024] -value_size 50 \
-seekseq_next 2 -skip_list_lookahead=0
[...]
DB path: [/dev/shm/rocksdbtest/dbbench]
fillseekseq : 0.389 micros/op 2569047 ops/sec;
real 0m21.806s
user 0m12.106s
sys 0m9.672s
$ time ./db_bench [...] -skip_list_lookahead=2
[...]
DB path: [/dev/shm/rocksdbtest/dbbench]
fillseekseq : 0.153 micros/op 6540684 ops/sec;
real 0m19.469s
user 0m10.192s
sys 0m9.252s
Reviewers: ljin, sdong, igor
Reviewed By: igor
Subscribers: dhruba, leveldb, march, lovro
Differential Revision: https://reviews.facebook.net/D23997
2014-09-24 00:52:28 +02:00
|
|
|
options.memtable_factory.reset(new SkipListFactory(
|
|
|
|
FLAGS_skip_list_lookahead));
|
2013-08-23 08:10:02 +02:00
|
|
|
break;
|
2014-11-12 22:05:12 +01:00
|
|
|
#ifndef ROCKSDB_LITE
|
|
|
|
case kPrefixHash:
|
|
|
|
options.memtable_factory.reset(
|
|
|
|
NewHashSkipListRepFactory(FLAGS_hash_bucket_count));
|
|
|
|
break;
|
2014-03-05 19:27:17 +01:00
|
|
|
case kHashLinkedList:
|
|
|
|
options.memtable_factory.reset(NewHashLinkListRepFactory(
|
|
|
|
FLAGS_hash_bucket_count));
|
|
|
|
break;
|
2013-08-23 08:10:02 +02:00
|
|
|
case kVectorRep:
|
|
|
|
options.memtable_factory.reset(
|
|
|
|
new VectorRepFactory
|
|
|
|
);
|
|
|
|
break;
|
2014-11-12 22:05:12 +01:00
|
|
|
#else
|
|
|
|
default:
|
|
|
|
fprintf(stderr, "Only skip list is supported in lite mode\n");
|
|
|
|
exit(1);
|
|
|
|
#endif // ROCKSDB_LITE
|
2013-08-23 08:10:02 +02:00
|
|
|
}
|
2014-02-12 19:41:55 +01:00
|
|
|
if (FLAGS_use_plain_table) {
|
2014-11-12 22:05:12 +01:00
|
|
|
#ifndef ROCKSDB_LITE
|
2014-03-05 19:27:17 +01:00
|
|
|
if (FLAGS_rep_factory != kPrefixHash &&
|
|
|
|
FLAGS_rep_factory != kHashLinkedList) {
|
2014-02-12 19:41:55 +01:00
|
|
|
fprintf(stderr, "Waring: plain table is used with skipList\n");
|
|
|
|
}
|
|
|
|
|
|
|
|
int bloom_bits_per_key = FLAGS_bloom_bits;
|
|
|
|
if (bloom_bits_per_key < 0) {
|
|
|
|
bloom_bits_per_key = 0;
|
|
|
|
}
|
2014-07-18 09:08:38 +02:00
|
|
|
|
|
|
|
PlainTableOptions plain_table_options;
|
|
|
|
plain_table_options.user_key_len = FLAGS_key_size;
|
|
|
|
plain_table_options.bloom_bits_per_key = bloom_bits_per_key;
|
|
|
|
plain_table_options.hash_table_ratio = 0.75;
|
|
|
|
options.table_factory = std::shared_ptr<TableFactory>(
|
|
|
|
NewPlainTableFactory(plain_table_options));
|
2014-11-12 22:05:12 +01:00
|
|
|
#else
|
|
|
|
fprintf(stderr, "Plain table is not supported in lite mode\n");
|
|
|
|
exit(1);
|
|
|
|
#endif // ROCKSDB_LITE
|
2014-08-18 20:59:38 +02:00
|
|
|
} else if (FLAGS_use_cuckoo_table) {
|
2014-11-12 22:05:12 +01:00
|
|
|
#ifndef ROCKSDB_LITE
|
2014-08-18 20:59:38 +02:00
|
|
|
if (FLAGS_cuckoo_hash_ratio > 1 || FLAGS_cuckoo_hash_ratio < 0) {
|
|
|
|
fprintf(stderr, "Invalid cuckoo_hash_ratio\n");
|
|
|
|
exit(1);
|
|
|
|
}
|
2018-01-29 23:13:05 +01:00
|
|
|
|
|
|
|
if (!FLAGS_mmap_read) {
|
|
|
|
fprintf(stderr, "cuckoo table format requires mmap read to operate\n");
|
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
|
CuckooTable: add one option to allow identity function for the first hash function
Summary:
MurmurHash becomes expensive when we do millions Get() a second in one
thread. Add this option to allow the first hash function to use identity
function as hash function. It results in QPS increase from 3.7M/s to
~4.3M/s. I did not observe improvement for end to end RocksDB
performance. This may be caused by other bottlenecks that I will address
in a separate diff.
Test Plan:
```
[ljin@dev1964 rocksdb] ./cuckoo_table_reader_test --enable_perf --file_dir=/dev/shm --write --identity_as_first_hash=0
==== Test CuckooReaderTest.WhenKeyExists
==== Test CuckooReaderTest.WhenKeyExistsWithUint64Comparator
==== Test CuckooReaderTest.CheckIterator
==== Test CuckooReaderTest.CheckIteratorUint64
==== Test CuckooReaderTest.WhenKeyNotFound
==== Test CuckooReaderTest.TestReadPerformance
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.272us (3.7 Mqps) with batch size of 0, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.138us (7.2 Mqps) with batch size of 10, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.142us (7.1 Mqps) with batch size of 25, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.142us (7.0 Mqps) with batch size of 50, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.144us (6.9 Mqps) with batch size of 100, # of found keys 125829120
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.201us (5.0 Mqps) with batch size of 0, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.121us (8.3 Mqps) with batch size of 10, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.123us (8.1 Mqps) with batch size of 25, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.121us (8.3 Mqps) with batch size of 50, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.112us (8.9 Mqps) with batch size of 100, # of found keys 104857600
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.251us (4.0 Mqps) with batch size of 0, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.107us (9.4 Mqps) with batch size of 10, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.099us (10.1 Mqps) with batch size of 25, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.100us (10.0 Mqps) with batch size of 50, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.116us (8.6 Mqps) with batch size of 100, # of found keys 83886080
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.189us (5.3 Mqps) with batch size of 0, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.095us (10.5 Mqps) with batch size of 10, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.096us (10.4 Mqps) with batch size of 25, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.098us (10.2 Mqps) with batch size of 50, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.105us (9.5 Mqps) with batch size of 100, # of found keys 73400320
[ljin@dev1964 rocksdb] ./cuckoo_table_reader_test --enable_perf --file_dir=/dev/shm --write --identity_as_first_hash=1
==== Test CuckooReaderTest.WhenKeyExists
==== Test CuckooReaderTest.WhenKeyExistsWithUint64Comparator
==== Test CuckooReaderTest.CheckIterator
==== Test CuckooReaderTest.CheckIteratorUint64
==== Test CuckooReaderTest.WhenKeyNotFound
==== Test CuckooReaderTest.TestReadPerformance
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.230us (4.3 Mqps) with batch size of 0, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.086us (11.7 Mqps) with batch size of 10, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.088us (11.3 Mqps) with batch size of 25, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.083us (12.1 Mqps) with batch size of 50, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.083us (12.1 Mqps) with batch size of 100, # of found keys 125829120
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.159us (6.3 Mqps) with batch size of 0, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.078us (12.8 Mqps) with batch size of 10, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.080us (12.6 Mqps) with batch size of 25, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.080us (12.5 Mqps) with batch size of 50, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.082us (12.2 Mqps) with batch size of 100, # of found keys 104857600
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.154us (6.5 Mqps) with batch size of 0, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.077us (13.0 Mqps) with batch size of 10, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.077us (12.9 Mqps) with batch size of 25, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.078us (12.8 Mqps) with batch size of 50, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.079us (12.6 Mqps) with batch size of 100, # of found keys 83886080
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.218us (4.6 Mqps) with batch size of 0, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.083us (12.0 Mqps) with batch size of 10, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.085us (11.7 Mqps) with batch size of 25, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.086us (11.6 Mqps) with batch size of 50, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.078us (12.8 Mqps) with batch size of 100, # of found keys 73400320
```
Reviewers: sdong, igor, yhchiang
Reviewed By: igor
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D23451
2014-09-18 20:00:48 +02:00
|
|
|
rocksdb::CuckooTableOptions table_options;
|
|
|
|
table_options.hash_table_ratio = FLAGS_cuckoo_hash_ratio;
|
|
|
|
table_options.identity_as_first_hash = FLAGS_identity_as_first_hash;
|
2014-08-18 20:59:38 +02:00
|
|
|
options.table_factory = std::shared_ptr<TableFactory>(
|
CuckooTable: add one option to allow identity function for the first hash function
Summary:
MurmurHash becomes expensive when we do millions Get() a second in one
thread. Add this option to allow the first hash function to use identity
function as hash function. It results in QPS increase from 3.7M/s to
~4.3M/s. I did not observe improvement for end to end RocksDB
performance. This may be caused by other bottlenecks that I will address
in a separate diff.
Test Plan:
```
[ljin@dev1964 rocksdb] ./cuckoo_table_reader_test --enable_perf --file_dir=/dev/shm --write --identity_as_first_hash=0
==== Test CuckooReaderTest.WhenKeyExists
==== Test CuckooReaderTest.WhenKeyExistsWithUint64Comparator
==== Test CuckooReaderTest.CheckIterator
==== Test CuckooReaderTest.CheckIteratorUint64
==== Test CuckooReaderTest.WhenKeyNotFound
==== Test CuckooReaderTest.TestReadPerformance
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.272us (3.7 Mqps) with batch size of 0, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.138us (7.2 Mqps) with batch size of 10, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.142us (7.1 Mqps) with batch size of 25, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.142us (7.0 Mqps) with batch size of 50, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.144us (6.9 Mqps) with batch size of 100, # of found keys 125829120
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.201us (5.0 Mqps) with batch size of 0, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.121us (8.3 Mqps) with batch size of 10, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.123us (8.1 Mqps) with batch size of 25, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.121us (8.3 Mqps) with batch size of 50, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.112us (8.9 Mqps) with batch size of 100, # of found keys 104857600
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.251us (4.0 Mqps) with batch size of 0, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.107us (9.4 Mqps) with batch size of 10, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.099us (10.1 Mqps) with batch size of 25, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.100us (10.0 Mqps) with batch size of 50, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.116us (8.6 Mqps) with batch size of 100, # of found keys 83886080
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.189us (5.3 Mqps) with batch size of 0, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.095us (10.5 Mqps) with batch size of 10, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.096us (10.4 Mqps) with batch size of 25, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.098us (10.2 Mqps) with batch size of 50, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.105us (9.5 Mqps) with batch size of 100, # of found keys 73400320
[ljin@dev1964 rocksdb] ./cuckoo_table_reader_test --enable_perf --file_dir=/dev/shm --write --identity_as_first_hash=1
==== Test CuckooReaderTest.WhenKeyExists
==== Test CuckooReaderTest.WhenKeyExistsWithUint64Comparator
==== Test CuckooReaderTest.CheckIterator
==== Test CuckooReaderTest.CheckIteratorUint64
==== Test CuckooReaderTest.WhenKeyNotFound
==== Test CuckooReaderTest.TestReadPerformance
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.230us (4.3 Mqps) with batch size of 0, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.086us (11.7 Mqps) with batch size of 10, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.088us (11.3 Mqps) with batch size of 25, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.083us (12.1 Mqps) with batch size of 50, # of found keys 125829120
With 125829120 items, utilization is 93.75%, number of hash functions: 2.
Time taken per op is 0.083us (12.1 Mqps) with batch size of 100, # of found keys 125829120
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.159us (6.3 Mqps) with batch size of 0, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.078us (12.8 Mqps) with batch size of 10, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.080us (12.6 Mqps) with batch size of 25, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.080us (12.5 Mqps) with batch size of 50, # of found keys 104857600
With 104857600 items, utilization is 78.12%, number of hash functions: 2.
Time taken per op is 0.082us (12.2 Mqps) with batch size of 100, # of found keys 104857600
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.154us (6.5 Mqps) with batch size of 0, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.077us (13.0 Mqps) with batch size of 10, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.077us (12.9 Mqps) with batch size of 25, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.078us (12.8 Mqps) with batch size of 50, # of found keys 83886080
With 83886080 items, utilization is 62.50%, number of hash functions: 2.
Time taken per op is 0.079us (12.6 Mqps) with batch size of 100, # of found keys 83886080
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.218us (4.6 Mqps) with batch size of 0, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.083us (12.0 Mqps) with batch size of 10, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.085us (11.7 Mqps) with batch size of 25, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.086us (11.6 Mqps) with batch size of 50, # of found keys 73400320
With 73400320 items, utilization is 54.69%, number of hash functions: 2.
Time taken per op is 0.078us (12.8 Mqps) with batch size of 100, # of found keys 73400320
```
Reviewers: sdong, igor, yhchiang
Reviewed By: igor
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D23451
2014-09-18 20:00:48 +02:00
|
|
|
NewCuckooTableFactory(table_options));
|
2014-11-12 22:05:12 +01:00
|
|
|
#else
|
|
|
|
fprintf(stderr, "Cuckoo table is not supported in lite mode\n");
|
|
|
|
exit(1);
|
|
|
|
#endif // ROCKSDB_LITE
|
2014-07-17 02:32:30 +02:00
|
|
|
} else {
|
|
|
|
BlockBasedTableOptions block_based_options;
|
|
|
|
if (FLAGS_use_hash_search) {
|
2014-08-11 19:47:52 +02:00
|
|
|
if (FLAGS_prefix_size == 0) {
|
|
|
|
fprintf(stderr,
|
|
|
|
"prefix_size not assigned when enable use_hash_search \n");
|
|
|
|
exit(1);
|
|
|
|
}
|
2014-07-17 02:32:30 +02:00
|
|
|
block_based_options.index_type = BlockBasedTableOptions::kHashSearch;
|
|
|
|
} else {
|
|
|
|
block_based_options.index_type = BlockBasedTableOptions::kBinarySearch;
|
|
|
|
}
|
2018-02-12 23:54:50 +01:00
|
|
|
if (FLAGS_partition_index_and_filters || FLAGS_partition_index) {
|
2017-06-16 01:06:22 +02:00
|
|
|
if (FLAGS_use_hash_search) {
|
|
|
|
fprintf(stderr,
|
|
|
|
"use_hash_search is incompatible with "
|
2018-02-12 23:54:50 +01:00
|
|
|
"partition index and is ignored");
|
2017-06-16 01:06:22 +02:00
|
|
|
}
|
|
|
|
block_based_options.index_type =
|
|
|
|
BlockBasedTableOptions::kTwoLevelIndexSearch;
|
|
|
|
block_based_options.metadata_block_size = FLAGS_metadata_block_size;
|
2018-02-12 23:54:50 +01:00
|
|
|
if (FLAGS_partition_index_and_filters) {
|
|
|
|
block_based_options.partition_filters = true;
|
|
|
|
}
|
2017-06-16 01:06:22 +02:00
|
|
|
}
|
2014-08-25 23:22:05 +02:00
|
|
|
if (cache_ == nullptr) {
|
|
|
|
block_based_options.no_block_cache = true;
|
|
|
|
}
|
2015-07-09 22:36:16 +02:00
|
|
|
block_based_options.cache_index_and_filter_blocks =
|
|
|
|
FLAGS_cache_index_and_filter_blocks;
|
Adding pin_l0_filter_and_index_blocks_in_cache feature and related fixes.
Summary:
When a block based table file is opened, if prefetch_index_and_filter is true, it will prefetch the index and filter blocks, putting them into the block cache.
What this feature adds: when a L0 block based table file is opened, if pin_l0_filter_and_index_blocks_in_cache is true in the options (and prefetch_index_and_filter is true), then the filter and index blocks aren't released back to the block cache at the end of BlockBasedTableReader::Open(). Instead the table reader takes ownership of them, hence pinning them, ie. the LRU cache will never push them out. Meanwhile in the table reader, further accesses will not hit the block cache, thus avoiding lock contention.
Test Plan:
'export TEST_TMPDIR=/dev/shm/ && DISABLE_JEMALLOC=1 OPT=-g make all valgrind_check -j32' is OK.
I didn't run the Java tests, I don't have Java set up on my devserver.
Reviewers: sdong
Reviewed By: sdong
Subscribers: andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D56133
2016-04-01 19:42:39 +02:00
|
|
|
block_based_options.pin_l0_filter_and_index_blocks_in_cache =
|
|
|
|
FLAGS_pin_l0_filter_and_index_blocks_in_cache;
|
2018-06-23 00:14:05 +02:00
|
|
|
block_based_options.pin_top_level_index_and_filter =
|
|
|
|
FLAGS_pin_top_level_index_and_filter;
|
2017-01-04 03:24:15 +01:00
|
|
|
if (FLAGS_cache_high_pri_pool_ratio > 1e-6) { // > 0.0 + eps
|
|
|
|
block_based_options.cache_index_and_filter_blocks_with_high_priority =
|
|
|
|
true;
|
|
|
|
}
|
2014-08-25 23:22:05 +02:00
|
|
|
block_based_options.block_cache = cache_;
|
|
|
|
block_based_options.block_cache_compressed = compressed_cache_;
|
|
|
|
block_based_options.block_size = FLAGS_block_size;
|
|
|
|
block_based_options.block_restart_interval = FLAGS_block_restart_interval;
|
2016-05-09 21:09:05 +02:00
|
|
|
block_based_options.index_block_restart_interval =
|
|
|
|
FLAGS_index_block_restart_interval;
|
2014-08-25 23:22:05 +02:00
|
|
|
block_based_options.filter_policy = filter_policy_;
|
2018-06-05 04:59:44 +02:00
|
|
|
block_based_options.format_version =
|
|
|
|
static_cast<uint32_t>(FLAGS_format_version);
|
2016-08-27 03:55:58 +02:00
|
|
|
block_based_options.read_amp_bytes_per_bit = FLAGS_read_amp_bytes_per_bit;
|
2018-01-11 00:06:29 +01:00
|
|
|
block_based_options.enable_index_compression =
|
|
|
|
FLAGS_enable_index_compression;
|
2018-03-27 05:14:24 +02:00
|
|
|
block_based_options.block_align = FLAGS_block_align;
|
2018-08-17 03:29:13 +02:00
|
|
|
if (FLAGS_use_data_block_hash_index) {
|
|
|
|
block_based_options.data_block_index_type =
|
|
|
|
rocksdb::BlockBasedTableOptions::kDataBlockBinaryAndHash;
|
|
|
|
} else {
|
|
|
|
block_based_options.data_block_index_type =
|
|
|
|
rocksdb::BlockBasedTableOptions::kDataBlockBinarySearch;
|
|
|
|
}
|
|
|
|
block_based_options.data_block_hash_table_util_ratio =
|
|
|
|
FLAGS_data_block_hash_table_util_ratio;
|
2017-02-27 21:59:34 +01:00
|
|
|
if (FLAGS_read_cache_path != "") {
|
|
|
|
#ifndef ROCKSDB_LITE
|
|
|
|
Status rc_status;
|
|
|
|
|
|
|
|
// Read cache need to be provided with a the Logger, we will put all
|
|
|
|
// reac cache logs in the read cache path in a file named rc_LOG
|
|
|
|
rc_status = FLAGS_env->CreateDirIfMissing(FLAGS_read_cache_path);
|
|
|
|
std::shared_ptr<Logger> read_cache_logger;
|
|
|
|
if (rc_status.ok()) {
|
|
|
|
rc_status = FLAGS_env->NewLogger(FLAGS_read_cache_path + "/rc_LOG",
|
|
|
|
&read_cache_logger);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (rc_status.ok()) {
|
|
|
|
PersistentCacheConfig rc_cfg(FLAGS_env, FLAGS_read_cache_path,
|
|
|
|
FLAGS_read_cache_size,
|
|
|
|
read_cache_logger);
|
|
|
|
|
|
|
|
rc_cfg.enable_direct_reads = FLAGS_read_cache_direct_read;
|
|
|
|
rc_cfg.enable_direct_writes = FLAGS_read_cache_direct_write;
|
|
|
|
rc_cfg.writer_qdepth = 4;
|
|
|
|
rc_cfg.writer_dispatch_size = 4 * 1024;
|
|
|
|
|
|
|
|
auto pcache = std::make_shared<BlockCacheTier>(rc_cfg);
|
|
|
|
block_based_options.persistent_cache = pcache;
|
|
|
|
rc_status = pcache->Open();
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!rc_status.ok()) {
|
|
|
|
fprintf(stderr, "Error initializing read cache, %s\n",
|
|
|
|
rc_status.ToString().c_str());
|
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
#else
|
|
|
|
fprintf(stderr, "Read cache is not supported in LITE\n");
|
|
|
|
exit(1);
|
|
|
|
|
|
|
|
#endif
|
|
|
|
}
|
2014-07-17 02:32:30 +02:00
|
|
|
options.table_factory.reset(
|
|
|
|
NewBlockBasedTableFactory(block_based_options));
|
2014-02-12 19:41:55 +01:00
|
|
|
}
|
2013-10-24 16:43:14 +02:00
|
|
|
if (FLAGS_max_bytes_for_level_multiplier_additional_v.size() > 0) {
|
|
|
|
if (FLAGS_max_bytes_for_level_multiplier_additional_v.size() !=
|
2013-05-21 20:37:06 +02:00
|
|
|
(unsigned int)FLAGS_num_levels) {
|
|
|
|
fprintf(stderr, "Insufficient number of fanouts specified %d\n",
|
2013-10-24 16:43:14 +02:00
|
|
|
(int)FLAGS_max_bytes_for_level_multiplier_additional_v.size());
|
2013-05-21 20:37:06 +02:00
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
options.max_bytes_for_level_multiplier_additional =
|
2013-10-24 16:43:14 +02:00
|
|
|
FLAGS_max_bytes_for_level_multiplier_additional_v;
|
2013-05-21 20:37:06 +02:00
|
|
|
}
|
2012-08-20 01:05:02 +02:00
|
|
|
options.level0_stop_writes_trigger = FLAGS_level0_stop_writes_trigger;
|
2012-10-23 19:34:09 +02:00
|
|
|
options.level0_file_num_compaction_trigger =
|
2012-10-19 23:00:53 +02:00
|
|
|
FLAGS_level0_file_num_compaction_trigger;
|
2012-08-20 01:05:02 +02:00
|
|
|
options.level0_slowdown_writes_trigger =
|
|
|
|
FLAGS_level0_slowdown_writes_trigger;
|
2013-10-24 16:43:14 +02:00
|
|
|
options.compression = FLAGS_compression_type_e;
|
2019-03-18 20:07:35 +01:00
|
|
|
options.sample_for_compression = FLAGS_sample_for_compression;
|
2013-11-07 03:46:28 +01:00
|
|
|
options.WAL_ttl_seconds = FLAGS_wal_ttl_seconds;
|
|
|
|
options.WAL_size_limit_MB = FLAGS_wal_size_limit_MB;
|
2014-12-23 03:39:28 +01:00
|
|
|
options.max_total_wal_size = FLAGS_max_total_wal_size;
|
|
|
|
|
2012-10-28 07:13:17 +01:00
|
|
|
if (FLAGS_min_level_to_compress >= 0) {
|
|
|
|
assert(FLAGS_min_level_to_compress <= FLAGS_num_levels);
|
2013-01-24 19:54:26 +01:00
|
|
|
options.compression_per_level.resize(FLAGS_num_levels);
|
2012-11-12 22:50:55 +01:00
|
|
|
for (int i = 0; i < FLAGS_min_level_to_compress; i++) {
|
2012-10-28 07:13:17 +01:00
|
|
|
options.compression_per_level[i] = kNoCompression;
|
|
|
|
}
|
2013-03-15 02:32:01 +01:00
|
|
|
for (int i = FLAGS_min_level_to_compress;
|
2012-10-28 07:13:17 +01:00
|
|
|
i < FLAGS_num_levels; i++) {
|
2013-10-24 16:43:14 +02:00
|
|
|
options.compression_per_level[i] = FLAGS_compression_type_e;
|
2012-10-28 07:13:17 +01:00
|
|
|
}
|
|
|
|
}
|
2013-08-06 00:43:49 +02:00
|
|
|
options.soft_rate_limit = FLAGS_soft_rate_limit;
|
|
|
|
options.hard_rate_limit = FLAGS_hard_rate_limit;
|
2015-12-18 03:10:48 +01:00
|
|
|
options.soft_pending_compaction_bytes_limit =
|
|
|
|
FLAGS_soft_pending_compaction_bytes_limit;
|
2015-09-11 23:31:23 +02:00
|
|
|
options.hard_pending_compaction_bytes_limit =
|
|
|
|
FLAGS_hard_pending_compaction_bytes_limit;
|
2015-05-16 00:52:51 +02:00
|
|
|
options.delayed_write_rate = FLAGS_delayed_write_rate;
|
support for concurrent adds to memtable
Summary:
This diff adds support for concurrent adds to the skiplist memtable
implementations. Memory allocation is made thread-safe by the addition of
a spinlock, with small per-core buffers to avoid contention. Concurrent
memtable writes are made via an additional method and don't impose a
performance overhead on the non-concurrent case, so parallelism can be
selected on a per-batch basis.
Write thread synchronization is an increasing bottleneck for higher levels
of concurrency, so this diff adds --enable_write_thread_adaptive_yield
(default off). This feature causes threads joining a write batch
group to spin for a short time (default 100 usec) using sched_yield,
rather than going to sleep on a mutex. If the timing of the yield calls
indicates that another thread has actually run during the yield then
spinning is avoided. This option improves performance for concurrent
situations even without parallel adds, although it has the potential to
increase CPU usage (and the heuristic adaptation is not yet mature).
Parallel writes are not currently compatible with
inplace updates, update callbacks, or delete filtering.
Enable it with --allow_concurrent_memtable_write (and
--enable_write_thread_adaptive_yield). Parallel memtable writes
are performance neutral when there is no actual parallelism, and in
my experiments (SSD server-class Linux and varying contention and key
sizes for fillrandom) they are always a performance win when there is
more than one thread.
Statistics are updated earlier in the write path, dropping the number
of DB mutex acquisitions from 2 to 1 for almost all cases.
This diff was motivated and inspired by Yahoo's cLSM work. It is more
conservative than cLSM: RocksDB's write batch group leader role is
preserved (along with all of the existing flush and write throttling
logic) and concurrent writers are blocked until all memtable insertions
have completed and the sequence number has been advanced, to preserve
linearizability.
My test config is "db_bench -benchmarks=fillrandom -threads=$T
-batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T
-level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999
-disable_auto_compactions --max_write_buffer_number=8
-max_background_flushes=8 --disable_wal --write_buffer_size=160000000
--block_size=16384 --allow_concurrent_memtable_write" on a two-socket
Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1
thread I get ~440Kops/sec. Peak performance for 1 socket (numactl
-N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance
across both sockets happens at 30 threads, and is ~900Kops/sec, although
with fewer threads there is less performance loss when the system has
background work.
Test Plan:
1. concurrent stress tests for InlineSkipList and DynamicBloom
2. make clean; make check
3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench
4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench
5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench
6. make clean; OPT=-DROCKSDB_LITE make check
7. verify no perf regressions when disabled
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba
Differential Revision: https://reviews.facebook.net/D50589
2015-08-15 01:59:07 +02:00
|
|
|
options.allow_concurrent_memtable_write =
|
|
|
|
FLAGS_allow_concurrent_memtable_write;
|
2018-01-26 19:46:30 +01:00
|
|
|
options.inplace_update_support = FLAGS_inplace_update_support;
|
|
|
|
options.inplace_update_num_locks = FLAGS_inplace_update_num_locks;
|
support for concurrent adds to memtable
Summary:
This diff adds support for concurrent adds to the skiplist memtable
implementations. Memory allocation is made thread-safe by the addition of
a spinlock, with small per-core buffers to avoid contention. Concurrent
memtable writes are made via an additional method and don't impose a
performance overhead on the non-concurrent case, so parallelism can be
selected on a per-batch basis.
Write thread synchronization is an increasing bottleneck for higher levels
of concurrency, so this diff adds --enable_write_thread_adaptive_yield
(default off). This feature causes threads joining a write batch
group to spin for a short time (default 100 usec) using sched_yield,
rather than going to sleep on a mutex. If the timing of the yield calls
indicates that another thread has actually run during the yield then
spinning is avoided. This option improves performance for concurrent
situations even without parallel adds, although it has the potential to
increase CPU usage (and the heuristic adaptation is not yet mature).
Parallel writes are not currently compatible with
inplace updates, update callbacks, or delete filtering.
Enable it with --allow_concurrent_memtable_write (and
--enable_write_thread_adaptive_yield). Parallel memtable writes
are performance neutral when there is no actual parallelism, and in
my experiments (SSD server-class Linux and varying contention and key
sizes for fillrandom) they are always a performance win when there is
more than one thread.
Statistics are updated earlier in the write path, dropping the number
of DB mutex acquisitions from 2 to 1 for almost all cases.
This diff was motivated and inspired by Yahoo's cLSM work. It is more
conservative than cLSM: RocksDB's write batch group leader role is
preserved (along with all of the existing flush and write throttling
logic) and concurrent writers are blocked until all memtable insertions
have completed and the sequence number has been advanced, to preserve
linearizability.
My test config is "db_bench -benchmarks=fillrandom -threads=$T
-batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T
-level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999
-disable_auto_compactions --max_write_buffer_number=8
-max_background_flushes=8 --disable_wal --write_buffer_size=160000000
--block_size=16384 --allow_concurrent_memtable_write" on a two-socket
Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1
thread I get ~440Kops/sec. Peak performance for 1 socket (numactl
-N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance
across both sockets happens at 30 threads, and is ~900Kops/sec, although
with fewer threads there is less performance loss when the system has
background work.
Test Plan:
1. concurrent stress tests for InlineSkipList and DynamicBloom
2. make clean; make check
3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench
4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench
5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench
6. make clean; OPT=-DROCKSDB_LITE make check
7. verify no perf regressions when disabled
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba
Differential Revision: https://reviews.facebook.net/D50589
2015-08-15 01:59:07 +02:00
|
|
|
options.enable_write_thread_adaptive_yield =
|
|
|
|
FLAGS_enable_write_thread_adaptive_yield;
|
2017-07-01 00:21:02 +02:00
|
|
|
options.enable_pipelined_write = FLAGS_enable_pipelined_write;
|
support for concurrent adds to memtable
Summary:
This diff adds support for concurrent adds to the skiplist memtable
implementations. Memory allocation is made thread-safe by the addition of
a spinlock, with small per-core buffers to avoid contention. Concurrent
memtable writes are made via an additional method and don't impose a
performance overhead on the non-concurrent case, so parallelism can be
selected on a per-batch basis.
Write thread synchronization is an increasing bottleneck for higher levels
of concurrency, so this diff adds --enable_write_thread_adaptive_yield
(default off). This feature causes threads joining a write batch
group to spin for a short time (default 100 usec) using sched_yield,
rather than going to sleep on a mutex. If the timing of the yield calls
indicates that another thread has actually run during the yield then
spinning is avoided. This option improves performance for concurrent
situations even without parallel adds, although it has the potential to
increase CPU usage (and the heuristic adaptation is not yet mature).
Parallel writes are not currently compatible with
inplace updates, update callbacks, or delete filtering.
Enable it with --allow_concurrent_memtable_write (and
--enable_write_thread_adaptive_yield). Parallel memtable writes
are performance neutral when there is no actual parallelism, and in
my experiments (SSD server-class Linux and varying contention and key
sizes for fillrandom) they are always a performance win when there is
more than one thread.
Statistics are updated earlier in the write path, dropping the number
of DB mutex acquisitions from 2 to 1 for almost all cases.
This diff was motivated and inspired by Yahoo's cLSM work. It is more
conservative than cLSM: RocksDB's write batch group leader role is
preserved (along with all of the existing flush and write throttling
logic) and concurrent writers are blocked until all memtable insertions
have completed and the sequence number has been advanced, to preserve
linearizability.
My test config is "db_bench -benchmarks=fillrandom -threads=$T
-batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T
-level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999
-disable_auto_compactions --max_write_buffer_number=8
-max_background_flushes=8 --disable_wal --write_buffer_size=160000000
--block_size=16384 --allow_concurrent_memtable_write" on a two-socket
Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1
thread I get ~440Kops/sec. Peak performance for 1 socket (numactl
-N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance
across both sockets happens at 30 threads, and is ~900Kops/sec, although
with fewer threads there is less performance loss when the system has
background work.
Test Plan:
1. concurrent stress tests for InlineSkipList and DynamicBloom
2. make clean; make check
3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench
4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench
5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench
6. make clean; OPT=-DROCKSDB_LITE make check
7. verify no perf regressions when disabled
Reviewers: igor, sdong
Reviewed By: sdong
Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba
Differential Revision: https://reviews.facebook.net/D50589
2015-08-15 01:59:07 +02:00
|
|
|
options.write_thread_max_yield_usec = FLAGS_write_thread_max_yield_usec;
|
|
|
|
options.write_thread_slow_yield_usec = FLAGS_write_thread_slow_yield_usec;
|
2013-08-06 00:43:49 +02:00
|
|
|
options.rate_limit_delay_max_milliseconds =
|
|
|
|
FLAGS_rate_limit_delay_max_milliseconds;
|
2012-11-01 01:02:24 +01:00
|
|
|
options.table_cache_numshardbits = FLAGS_table_cache_numshardbits;
|
2016-06-17 01:02:52 +02:00
|
|
|
options.max_compaction_bytes = FLAGS_max_compaction_bytes;
|
2012-11-21 00:45:41 +01:00
|
|
|
options.disable_auto_compactions = FLAGS_disable_auto_compactions;
|
2016-05-05 16:32:10 +02:00
|
|
|
options.optimize_filters_for_hits = FLAGS_optimize_filters_for_hits;
|
2013-03-15 01:00:04 +01:00
|
|
|
|
|
|
|
// fill storage options
|
2013-05-18 00:53:01 +02:00
|
|
|
options.advise_random_on_open = FLAGS_advise_random_on_open;
|
2013-10-24 16:43:14 +02:00
|
|
|
options.access_hint_on_compaction_start = FLAGS_compaction_fadvice_e;
|
2013-06-01 01:30:17 +02:00
|
|
|
options.use_adaptive_mutex = FLAGS_use_adaptive_mutex;
|
2013-06-14 07:49:46 +02:00
|
|
|
options.bytes_per_sync = FLAGS_bytes_per_sync;
|
2015-05-20 01:19:30 +02:00
|
|
|
options.wal_bytes_per_sync = FLAGS_wal_bytes_per_sync;
|
2013-06-01 01:30:17 +02:00
|
|
|
|
Benchmarking for Merge Operator
Summary:
Updated db_bench and utilities/merge_operators.h to allow for dynamic benchmarking
of merge operators in db_bench. Added a new test (--benchmarks=mergerandom), which performs
a bunch of random Merge() operations over random keys. Also added a "--merge_operator=" flag
so that the tester can easily benchmark different merge operators. Currently supports
the PutOperator and UInt64Add operator. Support for stringappend or list append may come later.
Test Plan:
1. make db_bench
2. Test the PutOperator (simulating Put) as follows:
./db_bench --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom --merge_operator=put
--threads=2
3. Test the UInt64AddOperator (simulating numeric addition) similarly:
./db_bench --value_size=8 --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom
--merge_operator=uint64add --threads=2
Reviewers: haobo, dhruba, zshao, MarkCallaghan
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11535
2013-08-16 02:13:07 +02:00
|
|
|
// merge operator options
|
2013-08-20 22:35:28 +02:00
|
|
|
options.merge_operator = MergeOperators::CreateFromStringId(
|
|
|
|
FLAGS_merge_operator);
|
|
|
|
if (options.merge_operator == nullptr && !FLAGS_merge_operator.empty()) {
|
Benchmarking for Merge Operator
Summary:
Updated db_bench and utilities/merge_operators.h to allow for dynamic benchmarking
of merge operators in db_bench. Added a new test (--benchmarks=mergerandom), which performs
a bunch of random Merge() operations over random keys. Also added a "--merge_operator=" flag
so that the tester can easily benchmark different merge operators. Currently supports
the PutOperator and UInt64Add operator. Support for stringappend or list append may come later.
Test Plan:
1. make db_bench
2. Test the PutOperator (simulating Put) as follows:
./db_bench --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom --merge_operator=put
--threads=2
3. Test the UInt64AddOperator (simulating numeric addition) similarly:
./db_bench --value_size=8 --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom
--merge_operator=uint64add --threads=2
Reviewers: haobo, dhruba, zshao, MarkCallaghan
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11535
2013-08-16 02:13:07 +02:00
|
|
|
fprintf(stderr, "invalid merge operator: %s\n",
|
|
|
|
FLAGS_merge_operator.c_str());
|
|
|
|
exit(1);
|
|
|
|
}
|
2014-01-11 02:33:56 +01:00
|
|
|
options.max_successive_merges = FLAGS_max_successive_merges;
|
2016-04-14 22:56:29 +02:00
|
|
|
options.report_bg_io_stats = FLAGS_report_bg_io_stats;
|
Benchmarking for Merge Operator
Summary:
Updated db_bench and utilities/merge_operators.h to allow for dynamic benchmarking
of merge operators in db_bench. Added a new test (--benchmarks=mergerandom), which performs
a bunch of random Merge() operations over random keys. Also added a "--merge_operator=" flag
so that the tester can easily benchmark different merge operators. Currently supports
the PutOperator and UInt64Add operator. Support for stringappend or list append may come later.
Test Plan:
1. make db_bench
2. Test the PutOperator (simulating Put) as follows:
./db_bench --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom --merge_operator=put
--threads=2
3. Test the UInt64AddOperator (simulating numeric addition) similarly:
./db_bench --value_size=8 --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom
--merge_operator=uint64add --threads=2
Reviewers: haobo, dhruba, zshao, MarkCallaghan
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11535
2013-08-16 02:13:07 +02:00
|
|
|
|
2013-09-10 01:06:10 +02:00
|
|
|
// set universal style compaction configurations, if applicable
|
|
|
|
if (FLAGS_universal_size_ratio != 0) {
|
|
|
|
options.compaction_options_universal.size_ratio =
|
|
|
|
FLAGS_universal_size_ratio;
|
|
|
|
}
|
|
|
|
if (FLAGS_universal_min_merge_width != 0) {
|
|
|
|
options.compaction_options_universal.min_merge_width =
|
|
|
|
FLAGS_universal_min_merge_width;
|
|
|
|
}
|
|
|
|
if (FLAGS_universal_max_merge_width != 0) {
|
|
|
|
options.compaction_options_universal.max_merge_width =
|
|
|
|
FLAGS_universal_max_merge_width;
|
|
|
|
}
|
|
|
|
if (FLAGS_universal_max_size_amplification_percent != 0) {
|
|
|
|
options.compaction_options_universal.max_size_amplification_percent =
|
|
|
|
FLAGS_universal_max_size_amplification_percent;
|
|
|
|
}
|
2013-12-03 21:32:07 +01:00
|
|
|
if (FLAGS_universal_compression_size_percent != -1) {
|
|
|
|
options.compaction_options_universal.compression_size_percent =
|
|
|
|
FLAGS_universal_compression_size_percent;
|
|
|
|
}
|
2015-07-20 20:52:39 +02:00
|
|
|
options.compaction_options_universal.allow_trivial_move =
|
|
|
|
FLAGS_universal_allow_trivial_move;
|
2015-03-06 20:22:06 +01:00
|
|
|
if (FLAGS_thread_status_per_interval > 0) {
|
|
|
|
options.enable_thread_tracking = true;
|
|
|
|
}
|
2013-09-10 01:06:10 +02:00
|
|
|
|
2015-10-14 21:43:00 +02:00
|
|
|
#ifndef ROCKSDB_LITE
|
2015-05-29 23:36:35 +02:00
|
|
|
if (FLAGS_readonly && FLAGS_transaction_db) {
|
|
|
|
fprintf(stderr, "Cannot use readonly flag with transaction_db\n");
|
|
|
|
exit(1);
|
|
|
|
}
|
2015-10-14 21:43:00 +02:00
|
|
|
#endif // ROCKSDB_LITE
|
2015-05-29 23:36:35 +02:00
|
|
|
|
2016-06-03 01:24:14 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
void InitializeOptionsGeneral(Options* opts) {
|
|
|
|
Options& options = *opts;
|
|
|
|
|
2017-08-03 01:13:08 +02:00
|
|
|
options.create_missing_column_families = FLAGS_num_column_families > 1;
|
2016-06-03 01:24:14 +02:00
|
|
|
options.statistics = dbstats;
|
|
|
|
options.wal_dir = FLAGS_wal_dir;
|
|
|
|
options.create_if_missing = !FLAGS_use_existing_db;
|
2016-10-31 22:50:40 +01:00
|
|
|
options.dump_malloc_stats = FLAGS_dump_malloc_stats;
|
2018-07-13 03:30:37 +02:00
|
|
|
options.stats_dump_period_sec =
|
|
|
|
static_cast<unsigned int>(FLAGS_stats_dump_period_sec);
|
2019-02-21 00:46:59 +01:00
|
|
|
options.stats_persist_period_sec =
|
|
|
|
static_cast<unsigned int>(FLAGS_stats_persist_period_sec);
|
|
|
|
options.stats_history_buffer_size =
|
|
|
|
static_cast<size_t>(FLAGS_stats_history_buffer_size);
|
2016-06-03 01:24:14 +02:00
|
|
|
|
2017-11-07 22:49:35 +01:00
|
|
|
options.compression_opts.level = FLAGS_compression_level;
|
|
|
|
options.compression_opts.max_dict_bytes = FLAGS_compression_max_dict_bytes;
|
|
|
|
options.compression_opts.zstd_max_train_bytes =
|
|
|
|
FLAGS_compression_zstd_max_train_bytes;
|
2018-07-19 01:14:02 +02:00
|
|
|
// If this is a block based table, set some related options
|
|
|
|
if (options.table_factory->Name() == BlockBasedTableFactory::kName &&
|
|
|
|
options.table_factory->GetOptions() != nullptr) {
|
|
|
|
BlockBasedTableOptions* table_options =
|
|
|
|
reinterpret_cast<BlockBasedTableOptions*>(
|
|
|
|
options.table_factory->GetOptions());
|
|
|
|
if (FLAGS_cache_size) {
|
2018-07-14 01:27:37 +02:00
|
|
|
table_options->block_cache = cache_;
|
|
|
|
}
|
2018-07-19 01:14:02 +02:00
|
|
|
if (FLAGS_bloom_bits >= 0) {
|
|
|
|
table_options->filter_policy.reset(NewBloomFilterPolicy(
|
|
|
|
FLAGS_bloom_bits, FLAGS_use_block_based_filter));
|
|
|
|
}
|
2018-07-14 01:27:37 +02:00
|
|
|
}
|
2016-06-03 01:24:14 +02:00
|
|
|
if (FLAGS_row_cache_size) {
|
|
|
|
if (FLAGS_cache_numshardbits >= 1) {
|
|
|
|
options.row_cache =
|
|
|
|
NewLRUCache(FLAGS_row_cache_size, FLAGS_cache_numshardbits);
|
|
|
|
} else {
|
|
|
|
options.row_cache = NewLRUCache(FLAGS_row_cache_size);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (FLAGS_enable_io_prio) {
|
|
|
|
FLAGS_env->LowerThreadPoolIOPriority(Env::LOW);
|
|
|
|
FLAGS_env->LowerThreadPoolIOPriority(Env::HIGH);
|
|
|
|
}
|
2018-04-24 17:38:01 +02:00
|
|
|
if (FLAGS_enable_cpu_prio) {
|
|
|
|
FLAGS_env->LowerThreadPoolCPUPriority(Env::LOW);
|
|
|
|
FLAGS_env->LowerThreadPoolCPUPriority(Env::HIGH);
|
|
|
|
}
|
2016-12-01 19:00:56 +01:00
|
|
|
options.env = FLAGS_env;
|
2018-06-15 21:07:38 +02:00
|
|
|
if (FLAGS_sine_write_rate) {
|
|
|
|
FLAGS_benchmark_write_rate_limit = static_cast<uint64_t>(SineRate(0));
|
|
|
|
}
|
2016-06-03 01:24:14 +02:00
|
|
|
|
2017-09-21 20:03:53 +02:00
|
|
|
if (FLAGS_rate_limiter_bytes_per_sec > 0) {
|
|
|
|
if (FLAGS_rate_limit_bg_reads &&
|
|
|
|
!FLAGS_new_table_reader_for_compaction_inputs) {
|
|
|
|
fprintf(stderr,
|
|
|
|
"rate limit compaction reads must have "
|
|
|
|
"new_table_reader_for_compaction_inputs set\n");
|
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
options.rate_limiter.reset(NewGenericRateLimiter(
|
|
|
|
FLAGS_rate_limiter_bytes_per_sec, 100 * 1000 /* refill_period_us */,
|
|
|
|
10 /* fairness */,
|
|
|
|
FLAGS_rate_limit_bg_reads ? RateLimiter::Mode::kReadsOnly
|
2017-10-05 04:02:22 +02:00
|
|
|
: RateLimiter::Mode::kWritesOnly,
|
|
|
|
FLAGS_rate_limiter_auto_tuned));
|
2017-09-21 20:03:53 +02:00
|
|
|
}
|
|
|
|
|
Auto recovery from out of space errors (#4164)
Summary:
This commit implements automatic recovery from a Status::NoSpace() error
during background operations such as write callback, flush and
compaction. The broad design is as follows -
1. Compaction errors are treated as soft errors and don't put the
database in read-only mode. A compaction is delayed until enough free
disk space is available to accomodate the compaction outputs, which is
estimated based on the input size. This means that users can continue to
write, and we rely on the WriteController to delay or stop writes if the
compaction debt becomes too high due to persistent low disk space
condition
2. Errors during write callback and flush are treated as hard errors,
i.e the database is put in read-only mode and goes back to read-write
only fater certain recovery actions are taken.
3. Both types of recovery rely on the SstFileManagerImpl to poll for
sufficient disk space. We assume that there is a 1-1 mapping between an
SFM and the underlying OS storage container. For cases where multiple
DBs are hosted on a single storage container, the user is expected to
allocate a single SFM instance and use the same one for all the DBs. If
no SFM is specified by the user, DBImpl::Open() will allocate one, but
this will be one per DB and each DB will recover independently. The
recovery implemented by SFM is as follows -
a) On the first occurance of an out of space error during compaction,
subsequent
compactions will be delayed until the disk free space check indicates
enough available space. The required space is computed as the sum of
input sizes.
b) The free space check requirement will be removed once the amount of
free space is greater than the size reserved by in progress
compactions when the first error occured
c) If the out of space error is a hard error, a background thread in
SFM will poll for sufficient headroom before triggering the recovery
of the database and putting it in write-only mode. The headroom is
calculated as the sum of the write_buffer_size of all the DB instances
associated with the SFM
4. EventListener callbacks will be called at the start and completion of
automatic recovery. Users can disable the auto recov ery in the start
callback, and later initiate it manually by calling DB::Resume()
Todo:
1. More extensive testing
2. Add disk full condition to db_stress (follow-on PR)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4164
Differential Revision: D9846378
Pulled By: anand1976
fbshipit-source-id: 80ea875dbd7f00205e19c82215ff6e37da10da4a
2018-09-15 22:36:19 +02:00
|
|
|
options.listeners.emplace_back(listener_);
|
2014-04-11 21:15:09 +02:00
|
|
|
if (FLAGS_num_multi_db <= 1) {
|
|
|
|
OpenDb(options, FLAGS_db, &db_);
|
|
|
|
} else {
|
|
|
|
multi_dbs_.clear();
|
2014-08-19 03:15:01 +02:00
|
|
|
multi_dbs_.resize(FLAGS_num_multi_db);
|
2016-10-19 01:30:57 +02:00
|
|
|
auto wal_dir = options.wal_dir;
|
2014-04-29 21:33:57 +02:00
|
|
|
for (int i = 0; i < FLAGS_num_multi_db; i++) {
|
2016-10-19 01:30:57 +02:00
|
|
|
if (!wal_dir.empty()) {
|
|
|
|
options.wal_dir = GetPathForMultiple(wal_dir, i);
|
|
|
|
}
|
|
|
|
OpenDb(options, GetPathForMultiple(FLAGS_db, i), &multi_dbs_[i]);
|
2014-04-11 21:15:09 +02:00
|
|
|
}
|
2016-10-19 01:30:57 +02:00
|
|
|
options.wal_dir = wal_dir;
|
2014-04-11 21:15:09 +02:00
|
|
|
}
|
2018-07-13 04:33:59 +02:00
|
|
|
|
|
|
|
// KeepFilter is a noop filter, this can be used to test compaction filter
|
|
|
|
if (FLAGS_use_keep_filter) {
|
|
|
|
options.compaction_filter = new KeepFilter();
|
|
|
|
fprintf(stdout, "A noop compaction filter is used\n");
|
|
|
|
}
|
2019-03-01 20:14:28 +01:00
|
|
|
|
|
|
|
if (FLAGS_use_existing_keys) {
|
|
|
|
// Only work on single database
|
|
|
|
assert(db_.db != nullptr);
|
|
|
|
ReadOptions read_opts;
|
|
|
|
read_opts.total_order_seek = true;
|
|
|
|
Iterator* iter = db_.db->NewIterator(read_opts);
|
|
|
|
for (iter->SeekToFirst(); iter->Valid(); iter->Next()) {
|
|
|
|
keys_.emplace_back(iter->key().ToString());
|
|
|
|
}
|
|
|
|
delete iter;
|
|
|
|
FLAGS_num = keys_.size();
|
|
|
|
}
|
2016-06-03 01:24:14 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
void Open(Options* opts) {
|
|
|
|
if (!InitializeOptionsFromFile(opts)) {
|
|
|
|
InitializeOptionsFromFlags(opts);
|
2014-04-11 21:15:09 +02:00
|
|
|
}
|
2016-04-28 01:23:33 +02:00
|
|
|
|
2016-06-03 01:24:14 +02:00
|
|
|
InitializeOptionsGeneral(opts);
|
2014-04-11 21:15:09 +02:00
|
|
|
}
|
|
|
|
|
2017-07-11 23:05:39 +02:00
|
|
|
void OpenDb(Options options, const std::string& db_name,
|
2014-08-19 03:15:01 +02:00
|
|
|
DBWithColumnFamilies* db) {
|
2012-11-06 04:18:49 +01:00
|
|
|
Status s;
|
2014-08-19 03:15:01 +02:00
|
|
|
// Open with column families if necessary.
|
|
|
|
if (FLAGS_num_column_families > 1) {
|
2014-12-23 03:39:28 +01:00
|
|
|
size_t num_hot = FLAGS_num_column_families;
|
|
|
|
if (FLAGS_num_hot_column_families > 0 &&
|
|
|
|
FLAGS_num_hot_column_families < FLAGS_num_column_families) {
|
|
|
|
num_hot = FLAGS_num_hot_column_families;
|
|
|
|
} else {
|
|
|
|
FLAGS_num_hot_column_families = FLAGS_num_column_families;
|
|
|
|
}
|
2014-08-19 03:15:01 +02:00
|
|
|
std::vector<ColumnFamilyDescriptor> column_families;
|
2014-12-23 03:39:28 +01:00
|
|
|
for (size_t i = 0; i < num_hot; i++) {
|
2014-08-19 03:15:01 +02:00
|
|
|
column_families.push_back(ColumnFamilyDescriptor(
|
|
|
|
ColumnFamilyName(i), ColumnFamilyOptions(options)));
|
|
|
|
}
|
2017-08-11 22:46:41 +02:00
|
|
|
std::vector<int> cfh_idx_to_prob;
|
|
|
|
if (!FLAGS_column_family_distribution.empty()) {
|
|
|
|
std::stringstream cf_prob_stream(FLAGS_column_family_distribution);
|
|
|
|
std::string cf_prob;
|
|
|
|
int sum = 0;
|
|
|
|
while (std::getline(cf_prob_stream, cf_prob, ',')) {
|
|
|
|
cfh_idx_to_prob.push_back(std::stoi(cf_prob));
|
|
|
|
sum += cfh_idx_to_prob.back();
|
|
|
|
}
|
|
|
|
if (sum != 100) {
|
|
|
|
fprintf(stderr, "column_family_distribution items must sum to 100\n");
|
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
if (cfh_idx_to_prob.size() != num_hot) {
|
|
|
|
fprintf(stderr,
|
|
|
|
"got %" ROCKSDB_PRIszt
|
|
|
|
" column_family_distribution items; expected "
|
|
|
|
"%" ROCKSDB_PRIszt "\n",
|
|
|
|
cfh_idx_to_prob.size(), num_hot);
|
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
}
|
2015-10-14 21:43:00 +02:00
|
|
|
#ifndef ROCKSDB_LITE
|
2014-08-19 03:15:01 +02:00
|
|
|
if (FLAGS_readonly) {
|
|
|
|
s = DB::OpenForReadOnly(options, db_name, column_families,
|
|
|
|
&db->cfh, &db->db);
|
Pessimistic Transactions
Summary:
Initial implementation of Pessimistic Transactions. This diff contains the api changes discussed in D38913. This diff is pretty large, so let me know if people would prefer to meet up to discuss it.
MyRocks folks: please take a look at the API in include/rocksdb/utilities/transaction[_db].h and let me know if you have any issues.
Also, you'll notice a couple of TODOs in the implementation of RollbackToSavePoint(). After chatting with Siying, I'm going to send out a separate diff for an alternate implementation of this feature that implements the rollback inside of WriteBatch/WriteBatchWithIndex. We can then decide which route is preferable.
Next, I'm planning on doing some perf testing and then integrating this diff into MongoRocks for further testing.
Test Plan: Unit tests, db_bench parallel testing.
Reviewers: igor, rven, sdong, yhchiang, yoshinorim
Reviewed By: sdong
Subscribers: hermanlee4, maykov, spetrunia, leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D40869
2015-05-26 02:37:33 +02:00
|
|
|
} else if (FLAGS_optimistic_transaction_db) {
|
2015-05-29 23:36:35 +02:00
|
|
|
s = OptimisticTransactionDB::Open(options, db_name, column_families,
|
Pessimistic Transactions
Summary:
Initial implementation of Pessimistic Transactions. This diff contains the api changes discussed in D38913. This diff is pretty large, so let me know if people would prefer to meet up to discuss it.
MyRocks folks: please take a look at the API in include/rocksdb/utilities/transaction[_db].h and let me know if you have any issues.
Also, you'll notice a couple of TODOs in the implementation of RollbackToSavePoint(). After chatting with Siying, I'm going to send out a separate diff for an alternate implementation of this feature that implements the rollback inside of WriteBatch/WriteBatchWithIndex. We can then decide which route is preferable.
Next, I'm planning on doing some perf testing and then integrating this diff into MongoRocks for further testing.
Test Plan: Unit tests, db_bench parallel testing.
Reviewers: igor, rven, sdong, yhchiang, yoshinorim
Reviewed By: sdong
Subscribers: hermanlee4, maykov, spetrunia, leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D40869
2015-05-26 02:37:33 +02:00
|
|
|
&db->cfh, &db->opt_txn_db);
|
|
|
|
if (s.ok()) {
|
|
|
|
db->db = db->opt_txn_db->GetBaseDB();
|
|
|
|
}
|
|
|
|
} else if (FLAGS_transaction_db) {
|
|
|
|
TransactionDB* ptr;
|
|
|
|
TransactionDBOptions txn_db_options;
|
|
|
|
s = TransactionDB::Open(options, txn_db_options, db_name,
|
|
|
|
column_families, &db->cfh, &ptr);
|
2015-05-29 23:36:35 +02:00
|
|
|
if (s.ok()) {
|
Pessimistic Transactions
Summary:
Initial implementation of Pessimistic Transactions. This diff contains the api changes discussed in D38913. This diff is pretty large, so let me know if people would prefer to meet up to discuss it.
MyRocks folks: please take a look at the API in include/rocksdb/utilities/transaction[_db].h and let me know if you have any issues.
Also, you'll notice a couple of TODOs in the implementation of RollbackToSavePoint(). After chatting with Siying, I'm going to send out a separate diff for an alternate implementation of this feature that implements the rollback inside of WriteBatch/WriteBatchWithIndex. We can then decide which route is preferable.
Next, I'm planning on doing some perf testing and then integrating this diff into MongoRocks for further testing.
Test Plan: Unit tests, db_bench parallel testing.
Reviewers: igor, rven, sdong, yhchiang, yoshinorim
Reviewed By: sdong
Subscribers: hermanlee4, maykov, spetrunia, leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D40869
2015-05-26 02:37:33 +02:00
|
|
|
db->db = ptr;
|
2015-05-29 23:36:35 +02:00
|
|
|
}
|
2014-08-19 03:15:01 +02:00
|
|
|
} else {
|
|
|
|
s = DB::Open(options, db_name, column_families, &db->cfh, &db->db);
|
|
|
|
}
|
2015-10-14 21:43:00 +02:00
|
|
|
#else
|
|
|
|
s = DB::Open(options, db_name, column_families, &db->cfh, &db->db);
|
|
|
|
#endif // ROCKSDB_LITE
|
2014-12-23 03:39:28 +01:00
|
|
|
db->cfh.resize(FLAGS_num_column_families);
|
|
|
|
db->num_created = num_hot;
|
|
|
|
db->num_hot = num_hot;
|
2017-08-11 22:46:41 +02:00
|
|
|
db->cfh_idx_to_prob = std::move(cfh_idx_to_prob);
|
2015-10-14 21:43:00 +02:00
|
|
|
#ifndef ROCKSDB_LITE
|
2014-08-19 03:15:01 +02:00
|
|
|
} else if (FLAGS_readonly) {
|
|
|
|
s = DB::OpenForReadOnly(options, db_name, &db->db);
|
Pessimistic Transactions
Summary:
Initial implementation of Pessimistic Transactions. This diff contains the api changes discussed in D38913. This diff is pretty large, so let me know if people would prefer to meet up to discuss it.
MyRocks folks: please take a look at the API in include/rocksdb/utilities/transaction[_db].h and let me know if you have any issues.
Also, you'll notice a couple of TODOs in the implementation of RollbackToSavePoint(). After chatting with Siying, I'm going to send out a separate diff for an alternate implementation of this feature that implements the rollback inside of WriteBatch/WriteBatchWithIndex. We can then decide which route is preferable.
Next, I'm planning on doing some perf testing and then integrating this diff into MongoRocks for further testing.
Test Plan: Unit tests, db_bench parallel testing.
Reviewers: igor, rven, sdong, yhchiang, yoshinorim
Reviewed By: sdong
Subscribers: hermanlee4, maykov, spetrunia, leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D40869
2015-05-26 02:37:33 +02:00
|
|
|
} else if (FLAGS_optimistic_transaction_db) {
|
|
|
|
s = OptimisticTransactionDB::Open(options, db_name, &db->opt_txn_db);
|
|
|
|
if (s.ok()) {
|
|
|
|
db->db = db->opt_txn_db->GetBaseDB();
|
|
|
|
}
|
2015-05-29 23:36:35 +02:00
|
|
|
} else if (FLAGS_transaction_db) {
|
2017-10-19 19:48:47 +02:00
|
|
|
TransactionDB* ptr = nullptr;
|
Pessimistic Transactions
Summary:
Initial implementation of Pessimistic Transactions. This diff contains the api changes discussed in D38913. This diff is pretty large, so let me know if people would prefer to meet up to discuss it.
MyRocks folks: please take a look at the API in include/rocksdb/utilities/transaction[_db].h and let me know if you have any issues.
Also, you'll notice a couple of TODOs in the implementation of RollbackToSavePoint(). After chatting with Siying, I'm going to send out a separate diff for an alternate implementation of this feature that implements the rollback inside of WriteBatch/WriteBatchWithIndex. We can then decide which route is preferable.
Next, I'm planning on doing some perf testing and then integrating this diff into MongoRocks for further testing.
Test Plan: Unit tests, db_bench parallel testing.
Reviewers: igor, rven, sdong, yhchiang, yoshinorim
Reviewed By: sdong
Subscribers: hermanlee4, maykov, spetrunia, leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D40869
2015-05-26 02:37:33 +02:00
|
|
|
TransactionDBOptions txn_db_options;
|
2017-07-11 23:05:39 +02:00
|
|
|
s = CreateLoggerFromOptions(db_name, options, &options.info_log);
|
|
|
|
if (s.ok()) {
|
|
|
|
s = TransactionDB::Open(options, txn_db_options, db_name, &ptr);
|
|
|
|
}
|
2015-05-29 23:36:35 +02:00
|
|
|
if (s.ok()) {
|
Pessimistic Transactions
Summary:
Initial implementation of Pessimistic Transactions. This diff contains the api changes discussed in D38913. This diff is pretty large, so let me know if people would prefer to meet up to discuss it.
MyRocks folks: please take a look at the API in include/rocksdb/utilities/transaction[_db].h and let me know if you have any issues.
Also, you'll notice a couple of TODOs in the implementation of RollbackToSavePoint(). After chatting with Siying, I'm going to send out a separate diff for an alternate implementation of this feature that implements the rollback inside of WriteBatch/WriteBatchWithIndex. We can then decide which route is preferable.
Next, I'm planning on doing some perf testing and then integrating this diff into MongoRocks for further testing.
Test Plan: Unit tests, db_bench parallel testing.
Reviewers: igor, rven, sdong, yhchiang, yoshinorim
Reviewed By: sdong
Subscribers: hermanlee4, maykov, spetrunia, leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D40869
2015-05-26 02:37:33 +02:00
|
|
|
db->db = ptr;
|
2015-05-29 23:36:35 +02:00
|
|
|
}
|
2016-08-09 19:16:32 +02:00
|
|
|
} else if (FLAGS_use_blob_db) {
|
2017-05-10 23:54:35 +02:00
|
|
|
blob_db::BlobDBOptions blob_db_options;
|
2017-11-21 08:19:46 +01:00
|
|
|
blob_db_options.enable_garbage_collection = FLAGS_blob_db_enable_gc;
|
2017-12-07 05:40:46 +01:00
|
|
|
blob_db_options.is_fifo = FLAGS_blob_db_is_fifo;
|
2018-03-06 20:46:20 +01:00
|
|
|
blob_db_options.max_db_size = FLAGS_blob_db_max_db_size;
|
2017-12-07 05:40:46 +01:00
|
|
|
blob_db_options.ttl_range_secs = FLAGS_blob_db_ttl_range_secs;
|
|
|
|
blob_db_options.min_blob_size = FLAGS_blob_db_min_blob_size;
|
|
|
|
blob_db_options.bytes_per_sync = FLAGS_blob_db_bytes_per_sync;
|
|
|
|
blob_db_options.blob_file_size = FLAGS_blob_db_file_size;
|
2017-10-19 19:48:47 +02:00
|
|
|
blob_db::BlobDB* ptr = nullptr;
|
2017-08-01 21:48:22 +02:00
|
|
|
s = blob_db::BlobDB::Open(options, blob_db_options, db_name, &ptr);
|
2017-05-10 23:54:35 +02:00
|
|
|
if (s.ok()) {
|
|
|
|
db->db = ptr;
|
|
|
|
}
|
|
|
|
#endif // ROCKSDB_LITE
|
2012-11-06 04:18:49 +01:00
|
|
|
} else {
|
2014-08-19 03:15:01 +02:00
|
|
|
s = DB::Open(options, db_name, &db->db);
|
2012-11-06 04:18:49 +01:00
|
|
|
}
|
2011-03-22 19:32:49 +01:00
|
|
|
if (!s.ok()) {
|
|
|
|
fprintf(stderr, "open error: %s\n", s.ToString().c_str());
|
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-06-15 01:17:56 +02:00
|
|
|
enum WriteMode {
|
|
|
|
RANDOM, SEQUENTIAL, UNIQUE_RANDOM
|
|
|
|
};
|
|
|
|
|
2016-10-19 01:30:57 +02:00
|
|
|
void WriteSeqDeterministic(ThreadState* thread) {
|
|
|
|
DoDeterministicCompact(thread, open_options_.compaction_style, SEQUENTIAL);
|
|
|
|
}
|
|
|
|
|
|
|
|
void WriteUniqueRandomDeterministic(ThreadState* thread) {
|
|
|
|
DoDeterministicCompact(thread, open_options_.compaction_style,
|
|
|
|
UNIQUE_RANDOM);
|
|
|
|
}
|
|
|
|
|
2011-08-22 23:08:51 +02:00
|
|
|
void WriteSeq(ThreadState* thread) {
|
2013-06-15 01:17:56 +02:00
|
|
|
DoWrite(thread, SEQUENTIAL);
|
2011-08-22 23:08:51 +02:00
|
|
|
}
|
2011-03-22 19:32:49 +01:00
|
|
|
|
2011-08-22 23:08:51 +02:00
|
|
|
void WriteRandom(ThreadState* thread) {
|
2013-06-15 01:17:56 +02:00
|
|
|
DoWrite(thread, RANDOM);
|
2011-08-22 23:08:51 +02:00
|
|
|
}
|
|
|
|
|
2013-06-15 01:17:56 +02:00
|
|
|
void WriteUniqueRandom(ThreadState* thread) {
|
|
|
|
DoWrite(thread, UNIQUE_RANDOM);
|
|
|
|
}
|
|
|
|
|
2014-04-09 20:25:21 +02:00
|
|
|
class KeyGenerator {
|
|
|
|
public:
|
2018-03-05 22:08:17 +01:00
|
|
|
KeyGenerator(Random64* rand, WriteMode mode, uint64_t num,
|
|
|
|
uint64_t /*num_per_set*/ = 64 * 1024)
|
|
|
|
: rand_(rand), mode_(mode), num_(num), next_(0) {
|
2014-04-09 20:25:21 +02:00
|
|
|
if (mode_ == UNIQUE_RANDOM) {
|
|
|
|
// NOTE: if memory consumption of this approach becomes a concern,
|
|
|
|
// we can either break it into pieces and only random shuffle a section
|
|
|
|
// each time. Alternatively, use a bit map implementation
|
|
|
|
// (https://reviews.facebook.net/differential/diff/54627/)
|
|
|
|
values_.resize(num_);
|
|
|
|
for (uint64_t i = 0; i < num_; ++i) {
|
|
|
|
values_[i] = i;
|
|
|
|
}
|
2014-11-11 22:47:22 +01:00
|
|
|
std::shuffle(
|
|
|
|
values_.begin(), values_.end(),
|
|
|
|
std::default_random_engine(static_cast<unsigned int>(FLAGS_seed)));
|
2014-04-09 20:25:21 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
uint64_t Next() {
|
|
|
|
switch (mode_) {
|
|
|
|
case SEQUENTIAL:
|
|
|
|
return next_++;
|
|
|
|
case RANDOM:
|
|
|
|
return rand_->Next() % num_;
|
|
|
|
case UNIQUE_RANDOM:
|
2018-01-29 20:30:37 +01:00
|
|
|
assert(next_ < num_);
|
2014-04-09 20:25:21 +02:00
|
|
|
return values_[next_++];
|
|
|
|
}
|
|
|
|
assert(false);
|
|
|
|
return std::numeric_limits<uint64_t>::max();
|
|
|
|
}
|
|
|
|
|
|
|
|
private:
|
|
|
|
Random64* rand_;
|
|
|
|
WriteMode mode_;
|
|
|
|
const uint64_t num_;
|
|
|
|
uint64_t next_;
|
|
|
|
std::vector<uint64_t> values_;
|
|
|
|
};
|
|
|
|
|
2014-04-11 21:15:09 +02:00
|
|
|
DB* SelectDB(ThreadState* thread) {
|
2014-08-19 03:15:01 +02:00
|
|
|
return SelectDBWithCfh(thread)->db;
|
|
|
|
}
|
|
|
|
|
|
|
|
DBWithColumnFamilies* SelectDBWithCfh(ThreadState* thread) {
|
|
|
|
return SelectDBWithCfh(thread->rand.Next());
|
|
|
|
}
|
|
|
|
|
|
|
|
DBWithColumnFamilies* SelectDBWithCfh(uint64_t rand_int) {
|
|
|
|
if (db_.db != nullptr) {
|
|
|
|
return &db_;
|
|
|
|
} else {
|
|
|
|
return &multi_dbs_[rand_int % multi_dbs_.size()];
|
2014-04-11 21:15:09 +02:00
|
|
|
}
|
|
|
|
}
|
2014-04-09 20:25:21 +02:00
|
|
|
|
2018-06-15 21:07:38 +02:00
|
|
|
double SineRate(double x) {
|
|
|
|
return FLAGS_sine_a*sin((FLAGS_sine_b*x) + FLAGS_sine_c) + FLAGS_sine_d;
|
|
|
|
}
|
|
|
|
|
2013-06-15 01:17:56 +02:00
|
|
|
void DoWrite(ThreadState* thread, WriteMode write_mode) {
|
|
|
|
const int test_duration = write_mode == RANDOM ? FLAGS_duration : 0;
|
2014-03-05 19:27:17 +01:00
|
|
|
const int64_t num_ops = writes_ == 0 ? num_ : writes_;
|
2013-06-15 01:17:56 +02:00
|
|
|
|
2014-04-11 21:15:09 +02:00
|
|
|
size_t num_key_gens = 1;
|
2014-08-19 03:15:01 +02:00
|
|
|
if (db_.db == nullptr) {
|
2014-04-11 21:15:09 +02:00
|
|
|
num_key_gens = multi_dbs_.size();
|
|
|
|
}
|
|
|
|
std::vector<std::unique_ptr<KeyGenerator>> key_gens(num_key_gens);
|
2014-12-23 03:39:28 +01:00
|
|
|
int64_t max_ops = num_ops * num_key_gens;
|
|
|
|
int64_t ops_per_stage = max_ops;
|
|
|
|
if (FLAGS_num_column_families > 1 && FLAGS_num_hot_column_families > 0) {
|
|
|
|
ops_per_stage = (max_ops - 1) / (FLAGS_num_column_families /
|
|
|
|
FLAGS_num_hot_column_families) +
|
|
|
|
1;
|
|
|
|
}
|
|
|
|
|
|
|
|
Duration duration(test_duration, max_ops, ops_per_stage);
|
2014-04-11 21:15:09 +02:00
|
|
|
for (size_t i = 0; i < num_key_gens; i++) {
|
2018-09-20 01:26:50 +02:00
|
|
|
key_gens[i].reset(new KeyGenerator(&(thread->rand), write_mode,
|
|
|
|
num_ + max_num_range_tombstones_,
|
2014-12-23 03:39:28 +01:00
|
|
|
ops_per_stage));
|
2014-04-11 21:15:09 +02:00
|
|
|
}
|
2013-03-01 20:25:13 +01:00
|
|
|
|
2011-08-22 23:08:51 +02:00
|
|
|
if (num_ != FLAGS_num) {
|
2011-03-22 19:32:49 +01:00
|
|
|
char msg[100];
|
2014-03-05 02:08:05 +01:00
|
|
|
snprintf(msg, sizeof(msg), "(%" PRIu64 " ops)", num_);
|
2011-08-22 23:08:51 +02:00
|
|
|
thread->stats.AddMessage(msg);
|
2011-03-22 19:32:49 +01:00
|
|
|
}
|
|
|
|
|
2011-08-22 23:08:51 +02:00
|
|
|
RandomGenerator gen;
|
2011-03-18 23:37:00 +01:00
|
|
|
WriteBatch batch;
|
|
|
|
Status s;
|
2011-08-22 23:08:51 +02:00
|
|
|
int64_t bytes = 0;
|
2014-04-08 20:21:09 +02:00
|
|
|
|
2015-02-19 23:27:48 +01:00
|
|
|
std::unique_ptr<const char[]> key_guard;
|
|
|
|
Slice key = AllocateKey(&key_guard);
|
2016-11-16 02:33:27 +01:00
|
|
|
std::unique_ptr<const char[]> begin_key_guard;
|
|
|
|
Slice begin_key = AllocateKey(&begin_key_guard);
|
|
|
|
std::unique_ptr<const char[]> end_key_guard;
|
|
|
|
Slice end_key = AllocateKey(&end_key_guard);
|
2016-12-07 03:02:43 +01:00
|
|
|
std::vector<std::unique_ptr<const char[]>> expanded_key_guards;
|
|
|
|
std::vector<Slice> expanded_keys;
|
|
|
|
if (FLAGS_expand_range_tombstones) {
|
|
|
|
expanded_key_guards.resize(range_tombstone_width_);
|
|
|
|
for (auto& expanded_key_guard : expanded_key_guards) {
|
|
|
|
expanded_keys.emplace_back(AllocateKey(&expanded_key_guard));
|
|
|
|
}
|
|
|
|
}
|
2016-11-16 02:33:27 +01:00
|
|
|
|
2014-12-23 03:39:28 +01:00
|
|
|
int64_t stage = 0;
|
2016-11-16 02:33:27 +01:00
|
|
|
int64_t num_written = 0;
|
2013-03-01 20:25:13 +01:00
|
|
|
while (!duration.Done(entries_per_batch_)) {
|
2014-12-23 03:39:28 +01:00
|
|
|
if (duration.GetStage() != stage) {
|
|
|
|
stage = duration.GetStage();
|
|
|
|
if (db_.db != nullptr) {
|
|
|
|
db_.CreateNewCf(open_options_, stage);
|
|
|
|
} else {
|
|
|
|
for (auto& db : multi_dbs_) {
|
|
|
|
db.CreateNewCf(open_options_, stage);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2016-01-04 21:01:27 +01:00
|
|
|
|
2014-08-19 03:15:01 +02:00
|
|
|
size_t id = thread->rand.Next() % num_key_gens;
|
|
|
|
DBWithColumnFamilies* db_with_cfh = SelectDBWithCfh(id);
|
2011-03-18 23:37:00 +01:00
|
|
|
batch.Clear();
|
2016-01-04 21:01:27 +01:00
|
|
|
|
|
|
|
if (thread->shared->write_rate_limiter.get() != nullptr) {
|
|
|
|
thread->shared->write_rate_limiter->Request(
|
2017-03-17 01:45:28 +01:00
|
|
|
entries_per_batch_ * (value_size_ + key_size_), Env::IO_HIGH,
|
2017-06-13 23:51:22 +02:00
|
|
|
nullptr /* stats */, RateLimiter::OpType::kWrite);
|
2016-01-04 21:01:27 +01:00
|
|
|
// Set time at which last op finished to Now() to hide latency and
|
|
|
|
// sleep from rate limiter. Also, do the check once per batch, not
|
|
|
|
// once per write.
|
|
|
|
thread->stats.ResetLastOpTime();
|
|
|
|
}
|
|
|
|
|
2014-04-08 20:21:09 +02:00
|
|
|
for (int64_t j = 0; j < entries_per_batch_; j++) {
|
2014-08-19 03:15:01 +02:00
|
|
|
int64_t rand_num = key_gens[id]->Next();
|
|
|
|
GenerateKeyFromInt(rand_num, FLAGS_num, &key);
|
2017-05-25 23:08:16 +02:00
|
|
|
if (use_blob_db_) {
|
|
|
|
#ifndef ROCKSDB_LITE
|
2017-05-10 23:54:35 +02:00
|
|
|
Slice val = gen.Generate(value_size_);
|
2017-12-07 05:40:46 +01:00
|
|
|
int ttl = rand() % FLAGS_blob_db_max_ttl_range;
|
2017-05-10 23:54:35 +02:00
|
|
|
blob_db::BlobDB* blobdb =
|
|
|
|
static_cast<blob_db::BlobDB*>(db_with_cfh->db);
|
|
|
|
s = blobdb->PutWithTTL(write_options_, key, val, ttl);
|
2017-05-25 23:08:16 +02:00
|
|
|
#endif // ROCKSDB_LITE
|
2016-08-09 19:16:32 +02:00
|
|
|
} else if (FLAGS_num_column_families <= 1) {
|
2014-08-19 03:15:01 +02:00
|
|
|
batch.Put(key, gen.Generate(value_size_));
|
|
|
|
} else {
|
|
|
|
// We use same rand_num as seed for key and column family so that we
|
|
|
|
// can deterministically find the cfh corresponding to a particular
|
|
|
|
// key while reading the key.
|
2014-12-23 03:39:28 +01:00
|
|
|
batch.Put(db_with_cfh->GetCfh(rand_num), key,
|
|
|
|
gen.Generate(value_size_));
|
2014-08-19 03:15:01 +02:00
|
|
|
}
|
2014-04-08 20:21:09 +02:00
|
|
|
bytes += value_size_ + key_size_;
|
2016-11-16 02:33:27 +01:00
|
|
|
++num_written;
|
|
|
|
if (writes_per_range_tombstone_ > 0 &&
|
2018-10-19 02:16:36 +02:00
|
|
|
num_written > writes_before_delete_range_ &&
|
|
|
|
(num_written - writes_before_delete_range_) /
|
|
|
|
writes_per_range_tombstone_ <=
|
2016-11-16 02:33:27 +01:00
|
|
|
max_num_range_tombstones_ &&
|
2018-10-19 02:16:36 +02:00
|
|
|
(num_written - writes_before_delete_range_) %
|
|
|
|
writes_per_range_tombstone_ ==
|
|
|
|
0) {
|
2016-11-16 02:33:27 +01:00
|
|
|
int64_t begin_num = key_gens[id]->Next();
|
2016-12-07 03:02:43 +01:00
|
|
|
if (FLAGS_expand_range_tombstones) {
|
|
|
|
for (int64_t offset = 0; offset < range_tombstone_width_;
|
|
|
|
++offset) {
|
|
|
|
GenerateKeyFromInt(begin_num + offset, FLAGS_num,
|
|
|
|
&expanded_keys[offset]);
|
2017-05-25 23:08:16 +02:00
|
|
|
if (use_blob_db_) {
|
|
|
|
#ifndef ROCKSDB_LITE
|
2016-12-07 03:02:43 +01:00
|
|
|
s = db_with_cfh->db->Delete(write_options_,
|
|
|
|
expanded_keys[offset]);
|
2017-05-25 23:08:16 +02:00
|
|
|
#endif // ROCKSDB_LITE
|
2016-12-07 03:02:43 +01:00
|
|
|
} else if (FLAGS_num_column_families <= 1) {
|
|
|
|
batch.Delete(expanded_keys[offset]);
|
|
|
|
} else {
|
|
|
|
batch.Delete(db_with_cfh->GetCfh(rand_num),
|
|
|
|
expanded_keys[offset]);
|
|
|
|
}
|
|
|
|
}
|
2016-11-16 02:33:27 +01:00
|
|
|
} else {
|
2016-12-07 03:02:43 +01:00
|
|
|
GenerateKeyFromInt(begin_num, FLAGS_num, &begin_key);
|
|
|
|
GenerateKeyFromInt(begin_num + range_tombstone_width_, FLAGS_num,
|
|
|
|
&end_key);
|
2017-05-25 23:08:16 +02:00
|
|
|
if (use_blob_db_) {
|
|
|
|
#ifndef ROCKSDB_LITE
|
2016-12-07 03:02:43 +01:00
|
|
|
s = db_with_cfh->db->DeleteRange(
|
|
|
|
write_options_, db_with_cfh->db->DefaultColumnFamily(),
|
|
|
|
begin_key, end_key);
|
2017-05-25 23:08:16 +02:00
|
|
|
#endif // ROCKSDB_LITE
|
2016-12-07 03:02:43 +01:00
|
|
|
} else if (FLAGS_num_column_families <= 1) {
|
|
|
|
batch.DeleteRange(begin_key, end_key);
|
|
|
|
} else {
|
|
|
|
batch.DeleteRange(db_with_cfh->GetCfh(rand_num), begin_key,
|
|
|
|
end_key);
|
|
|
|
}
|
2016-11-16 02:33:27 +01:00
|
|
|
}
|
|
|
|
}
|
2011-04-12 21:38:58 +02:00
|
|
|
}
|
2017-05-25 23:08:16 +02:00
|
|
|
if (!use_blob_db_) {
|
2016-08-09 19:16:32 +02:00
|
|
|
s = db_with_cfh->db->Write(write_options_, &batch);
|
|
|
|
}
|
2014-09-29 18:50:41 +02:00
|
|
|
thread->stats.FinishedOps(db_with_cfh, db_with_cfh->db,
|
2015-12-08 02:33:18 +01:00
|
|
|
entries_per_batch_, kWrite);
|
2018-06-15 21:07:38 +02:00
|
|
|
if (FLAGS_sine_write_rate) {
|
|
|
|
uint64_t now = FLAGS_env->NowMicros();
|
|
|
|
|
|
|
|
uint64_t usecs_since_last;
|
|
|
|
if (now > thread->stats.GetSineInterval()) {
|
|
|
|
usecs_since_last = now - thread->stats.GetSineInterval();
|
|
|
|
} else {
|
|
|
|
usecs_since_last = 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (usecs_since_last >
|
|
|
|
(FLAGS_sine_write_rate_interval_milliseconds * uint64_t{1000})) {
|
|
|
|
double usecs_since_start =
|
|
|
|
static_cast<double>(now - thread->stats.GetStart());
|
|
|
|
thread->stats.ResetSineInterval();
|
|
|
|
uint64_t write_rate =
|
|
|
|
static_cast<uint64_t>(SineRate(usecs_since_start / 1000000.0));
|
|
|
|
thread->shared->write_rate_limiter.reset(
|
|
|
|
NewGenericRateLimiter(write_rate));
|
|
|
|
}
|
|
|
|
}
|
Auto recovery from out of space errors (#4164)
Summary:
This commit implements automatic recovery from a Status::NoSpace() error
during background operations such as write callback, flush and
compaction. The broad design is as follows -
1. Compaction errors are treated as soft errors and don't put the
database in read-only mode. A compaction is delayed until enough free
disk space is available to accomodate the compaction outputs, which is
estimated based on the input size. This means that users can continue to
write, and we rely on the WriteController to delay or stop writes if the
compaction debt becomes too high due to persistent low disk space
condition
2. Errors during write callback and flush are treated as hard errors,
i.e the database is put in read-only mode and goes back to read-write
only fater certain recovery actions are taken.
3. Both types of recovery rely on the SstFileManagerImpl to poll for
sufficient disk space. We assume that there is a 1-1 mapping between an
SFM and the underlying OS storage container. For cases where multiple
DBs are hosted on a single storage container, the user is expected to
allocate a single SFM instance and use the same one for all the DBs. If
no SFM is specified by the user, DBImpl::Open() will allocate one, but
this will be one per DB and each DB will recover independently. The
recovery implemented by SFM is as follows -
a) On the first occurance of an out of space error during compaction,
subsequent
compactions will be delayed until the disk free space check indicates
enough available space. The required space is computed as the sum of
input sizes.
b) The free space check requirement will be removed once the amount of
free space is greater than the size reserved by in progress
compactions when the first error occured
c) If the out of space error is a hard error, a background thread in
SFM will poll for sufficient headroom before triggering the recovery
of the database and putting it in write-only mode. The headroom is
calculated as the sum of the write_buffer_size of all the DB instances
associated with the SFM
4. EventListener callbacks will be called at the start and completion of
automatic recovery. Users can disable the auto recov ery in the start
callback, and later initiate it manually by calling DB::Resume()
Todo:
1. More extensive testing
2. Add disk full condition to db_stress (follow-on PR)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4164
Differential Revision: D9846378
Pulled By: anand1976
fbshipit-source-id: 80ea875dbd7f00205e19c82215ff6e37da10da4a
2018-09-15 22:36:19 +02:00
|
|
|
if (!s.ok()) {
|
|
|
|
s = listener_->WaitForRecovery(600000000) ? Status::OK() : s;
|
|
|
|
}
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
if (!s.ok()) {
|
|
|
|
fprintf(stderr, "put error: %s\n", s.ToString().c_str());
|
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
}
|
2011-08-22 23:08:51 +02:00
|
|
|
thread->stats.AddBytes(bytes);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
2016-10-19 01:30:57 +02:00
|
|
|
Status DoDeterministicCompact(ThreadState* thread,
|
|
|
|
CompactionStyle compaction_style,
|
|
|
|
WriteMode write_mode) {
|
|
|
|
#ifndef ROCKSDB_LITE
|
|
|
|
ColumnFamilyMetaData meta;
|
|
|
|
std::vector<DB*> db_list;
|
|
|
|
if (db_.db != nullptr) {
|
|
|
|
db_list.push_back(db_.db);
|
|
|
|
} else {
|
|
|
|
for (auto& db : multi_dbs_) {
|
|
|
|
db_list.push_back(db.db);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
std::vector<Options> options_list;
|
|
|
|
for (auto db : db_list) {
|
|
|
|
options_list.push_back(db->GetOptions());
|
2017-03-25 00:58:51 +01:00
|
|
|
if (compaction_style != kCompactionStyleFIFO) {
|
|
|
|
db->SetOptions({{"disable_auto_compactions", "1"},
|
|
|
|
{"level0_slowdown_writes_trigger", "400000000"},
|
|
|
|
{"level0_stop_writes_trigger", "400000000"}});
|
|
|
|
} else {
|
|
|
|
db->SetOptions({{"disable_auto_compactions", "1"}});
|
|
|
|
}
|
2016-10-19 01:30:57 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
assert(!db_list.empty());
|
|
|
|
auto num_db = db_list.size();
|
|
|
|
size_t num_levels = static_cast<size_t>(open_options_.num_levels);
|
|
|
|
size_t output_level = open_options_.num_levels - 1;
|
|
|
|
std::vector<std::vector<std::vector<SstFileMetaData>>> sorted_runs(num_db);
|
|
|
|
std::vector<size_t> num_files_at_level0(num_db, 0);
|
|
|
|
if (compaction_style == kCompactionStyleLevel) {
|
|
|
|
if (num_levels == 0) {
|
|
|
|
return Status::InvalidArgument("num_levels should be larger than 1");
|
|
|
|
}
|
|
|
|
bool should_stop = false;
|
|
|
|
while (!should_stop) {
|
|
|
|
if (sorted_runs[0].empty()) {
|
|
|
|
DoWrite(thread, write_mode);
|
|
|
|
} else {
|
|
|
|
DoWrite(thread, UNIQUE_RANDOM);
|
|
|
|
}
|
|
|
|
for (size_t i = 0; i < num_db; i++) {
|
|
|
|
auto db = db_list[i];
|
|
|
|
db->Flush(FlushOptions());
|
|
|
|
db->GetColumnFamilyMetaData(&meta);
|
|
|
|
if (num_files_at_level0[i] == meta.levels[0].files.size() ||
|
|
|
|
writes_ == 0) {
|
|
|
|
should_stop = true;
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
sorted_runs[i].emplace_back(
|
|
|
|
meta.levels[0].files.begin(),
|
|
|
|
meta.levels[0].files.end() - num_files_at_level0[i]);
|
|
|
|
num_files_at_level0[i] = meta.levels[0].files.size();
|
|
|
|
if (sorted_runs[i].back().size() == 1) {
|
|
|
|
should_stop = true;
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
if (sorted_runs[i].size() == output_level) {
|
|
|
|
auto& L1 = sorted_runs[i].back();
|
|
|
|
L1.erase(L1.begin(), L1.begin() + L1.size() / 3);
|
|
|
|
should_stop = true;
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
}
|
2017-01-06 23:42:18 +01:00
|
|
|
writes_ /= static_cast<int64_t>(open_options_.max_bytes_for_level_multiplier);
|
2016-10-19 01:30:57 +02:00
|
|
|
}
|
|
|
|
for (size_t i = 0; i < num_db; i++) {
|
|
|
|
if (sorted_runs[i].size() < num_levels - 1) {
|
2017-01-06 23:42:18 +01:00
|
|
|
fprintf(stderr, "n is too small to fill %" ROCKSDB_PRIszt " levels\n", num_levels);
|
2016-10-19 01:30:57 +02:00
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
for (size_t i = 0; i < num_db; i++) {
|
|
|
|
auto db = db_list[i];
|
|
|
|
auto compactionOptions = CompactionOptions();
|
2018-04-17 02:56:59 +02:00
|
|
|
compactionOptions.compression = FLAGS_compression_type_e;
|
2016-10-19 01:30:57 +02:00
|
|
|
auto options = db->GetOptions();
|
|
|
|
MutableCFOptions mutable_cf_options(options);
|
|
|
|
for (size_t j = 0; j < sorted_runs[i].size(); j++) {
|
|
|
|
compactionOptions.output_file_size_limit =
|
2018-05-04 01:35:46 +02:00
|
|
|
MaxFileSizeForLevel(mutable_cf_options,
|
|
|
|
static_cast<int>(output_level), compaction_style);
|
2016-10-19 01:30:57 +02:00
|
|
|
std::cout << sorted_runs[i][j].size() << std::endl;
|
|
|
|
db->CompactFiles(compactionOptions, {sorted_runs[i][j].back().name,
|
|
|
|
sorted_runs[i][j].front().name},
|
|
|
|
static_cast<int>(output_level - j) /*level*/);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
} else if (compaction_style == kCompactionStyleUniversal) {
|
|
|
|
auto ratio = open_options_.compaction_options_universal.size_ratio;
|
|
|
|
bool should_stop = false;
|
|
|
|
while (!should_stop) {
|
|
|
|
if (sorted_runs[0].empty()) {
|
|
|
|
DoWrite(thread, write_mode);
|
|
|
|
} else {
|
|
|
|
DoWrite(thread, UNIQUE_RANDOM);
|
|
|
|
}
|
|
|
|
for (size_t i = 0; i < num_db; i++) {
|
|
|
|
auto db = db_list[i];
|
|
|
|
db->Flush(FlushOptions());
|
|
|
|
db->GetColumnFamilyMetaData(&meta);
|
|
|
|
if (num_files_at_level0[i] == meta.levels[0].files.size() ||
|
|
|
|
writes_ == 0) {
|
|
|
|
should_stop = true;
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
sorted_runs[i].emplace_back(
|
|
|
|
meta.levels[0].files.begin(),
|
|
|
|
meta.levels[0].files.end() - num_files_at_level0[i]);
|
|
|
|
num_files_at_level0[i] = meta.levels[0].files.size();
|
|
|
|
if (sorted_runs[i].back().size() == 1) {
|
|
|
|
should_stop = true;
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
num_files_at_level0[i] = meta.levels[0].files.size();
|
|
|
|
}
|
2017-01-06 23:42:18 +01:00
|
|
|
writes_ = static_cast<int64_t>(writes_* static_cast<double>(100) / (ratio + 200));
|
2016-10-19 01:30:57 +02:00
|
|
|
}
|
|
|
|
for (size_t i = 0; i < num_db; i++) {
|
|
|
|
if (sorted_runs[i].size() < num_levels) {
|
2017-01-06 23:42:18 +01:00
|
|
|
fprintf(stderr, "n is too small to fill %" ROCKSDB_PRIszt " levels\n", num_levels);
|
2016-10-19 01:30:57 +02:00
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
for (size_t i = 0; i < num_db; i++) {
|
|
|
|
auto db = db_list[i];
|
|
|
|
auto compactionOptions = CompactionOptions();
|
2018-04-17 02:56:59 +02:00
|
|
|
compactionOptions.compression = FLAGS_compression_type_e;
|
2016-10-19 01:30:57 +02:00
|
|
|
auto options = db->GetOptions();
|
|
|
|
MutableCFOptions mutable_cf_options(options);
|
|
|
|
for (size_t j = 0; j < sorted_runs[i].size(); j++) {
|
|
|
|
compactionOptions.output_file_size_limit =
|
2018-05-04 01:35:46 +02:00
|
|
|
MaxFileSizeForLevel(mutable_cf_options,
|
|
|
|
static_cast<int>(output_level), compaction_style);
|
2016-10-19 01:30:57 +02:00
|
|
|
db->CompactFiles(
|
|
|
|
compactionOptions,
|
|
|
|
{sorted_runs[i][j].back().name, sorted_runs[i][j].front().name},
|
|
|
|
(output_level > j ? static_cast<int>(output_level - j)
|
|
|
|
: 0) /*level*/);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
} else if (compaction_style == kCompactionStyleFIFO) {
|
2017-03-25 00:58:51 +01:00
|
|
|
if (num_levels != 1) {
|
|
|
|
return Status::InvalidArgument(
|
|
|
|
"num_levels should be 1 for FIFO compaction");
|
|
|
|
}
|
|
|
|
if (FLAGS_num_multi_db != 0) {
|
|
|
|
return Status::InvalidArgument("Doesn't support multiDB");
|
|
|
|
}
|
|
|
|
auto db = db_list[0];
|
|
|
|
std::vector<std::string> file_names;
|
|
|
|
while (true) {
|
|
|
|
if (sorted_runs[0].empty()) {
|
|
|
|
DoWrite(thread, write_mode);
|
|
|
|
} else {
|
|
|
|
DoWrite(thread, UNIQUE_RANDOM);
|
|
|
|
}
|
|
|
|
db->Flush(FlushOptions());
|
|
|
|
db->GetColumnFamilyMetaData(&meta);
|
|
|
|
auto total_size = meta.levels[0].size;
|
|
|
|
if (total_size >=
|
|
|
|
db->GetOptions().compaction_options_fifo.max_table_files_size) {
|
|
|
|
for (auto file_meta : meta.levels[0].files) {
|
|
|
|
file_names.emplace_back(file_meta.name);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
// TODO(shuzhang1989): Investigate why CompactFiles not working
|
|
|
|
// auto compactionOptions = CompactionOptions();
|
|
|
|
// db->CompactFiles(compactionOptions, file_names, 0);
|
|
|
|
auto compactionOptions = CompactRangeOptions();
|
|
|
|
db->CompactRange(compactionOptions, nullptr, nullptr);
|
2016-10-19 01:30:57 +02:00
|
|
|
} else {
|
|
|
|
fprintf(stdout,
|
|
|
|
"%-12s : skipped (-compaction_stype=kCompactionStyleNone)\n",
|
|
|
|
"filldeterministic");
|
|
|
|
return Status::InvalidArgument("None compaction is not supported");
|
|
|
|
}
|
|
|
|
|
|
|
|
// Verify seqno and key range
|
|
|
|
// Note: the seqno get changed at the max level by implementation
|
|
|
|
// optimization, so skip the check of the max level.
|
|
|
|
#ifndef NDEBUG
|
|
|
|
for (size_t k = 0; k < num_db; k++) {
|
|
|
|
auto db = db_list[k];
|
|
|
|
db->GetColumnFamilyMetaData(&meta);
|
|
|
|
// verify the number of sorted runs
|
|
|
|
if (compaction_style == kCompactionStyleLevel) {
|
|
|
|
assert(num_levels - 1 == sorted_runs[k].size());
|
|
|
|
} else if (compaction_style == kCompactionStyleUniversal) {
|
|
|
|
assert(meta.levels[0].files.size() + num_levels - 1 ==
|
|
|
|
sorted_runs[k].size());
|
|
|
|
} else if (compaction_style == kCompactionStyleFIFO) {
|
|
|
|
// TODO(gzh): FIFO compaction
|
2017-03-25 00:58:51 +01:00
|
|
|
db->GetColumnFamilyMetaData(&meta);
|
|
|
|
auto total_size = meta.levels[0].size;
|
|
|
|
assert(total_size <=
|
|
|
|
db->GetOptions().compaction_options_fifo.max_table_files_size);
|
|
|
|
break;
|
2016-10-19 01:30:57 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
// verify smallest/largest seqno and key range of each sorted run
|
|
|
|
auto max_level = num_levels - 1;
|
|
|
|
int level;
|
|
|
|
for (size_t i = 0; i < sorted_runs[k].size(); i++) {
|
|
|
|
level = static_cast<int>(max_level - i);
|
|
|
|
SequenceNumber sorted_run_smallest_seqno = kMaxSequenceNumber;
|
|
|
|
SequenceNumber sorted_run_largest_seqno = 0;
|
|
|
|
std::string sorted_run_smallest_key, sorted_run_largest_key;
|
|
|
|
bool first_key = true;
|
|
|
|
for (auto fileMeta : sorted_runs[k][i]) {
|
|
|
|
sorted_run_smallest_seqno =
|
|
|
|
std::min(sorted_run_smallest_seqno, fileMeta.smallest_seqno);
|
|
|
|
sorted_run_largest_seqno =
|
|
|
|
std::max(sorted_run_largest_seqno, fileMeta.largest_seqno);
|
|
|
|
if (first_key ||
|
|
|
|
db->DefaultColumnFamily()->GetComparator()->Compare(
|
|
|
|
fileMeta.smallestkey, sorted_run_smallest_key) < 0) {
|
|
|
|
sorted_run_smallest_key = fileMeta.smallestkey;
|
|
|
|
}
|
|
|
|
if (first_key ||
|
|
|
|
db->DefaultColumnFamily()->GetComparator()->Compare(
|
|
|
|
fileMeta.largestkey, sorted_run_largest_key) > 0) {
|
|
|
|
sorted_run_largest_key = fileMeta.largestkey;
|
|
|
|
}
|
|
|
|
first_key = false;
|
|
|
|
}
|
|
|
|
if (compaction_style == kCompactionStyleLevel ||
|
|
|
|
(compaction_style == kCompactionStyleUniversal && level > 0)) {
|
|
|
|
SequenceNumber level_smallest_seqno = kMaxSequenceNumber;
|
|
|
|
SequenceNumber level_largest_seqno = 0;
|
|
|
|
for (auto fileMeta : meta.levels[level].files) {
|
|
|
|
level_smallest_seqno =
|
|
|
|
std::min(level_smallest_seqno, fileMeta.smallest_seqno);
|
|
|
|
level_largest_seqno =
|
|
|
|
std::max(level_largest_seqno, fileMeta.largest_seqno);
|
|
|
|
}
|
|
|
|
assert(sorted_run_smallest_key ==
|
|
|
|
meta.levels[level].files.front().smallestkey);
|
|
|
|
assert(sorted_run_largest_key ==
|
|
|
|
meta.levels[level].files.back().largestkey);
|
|
|
|
if (level != static_cast<int>(max_level)) {
|
|
|
|
// compaction at max_level would change sequence number
|
|
|
|
assert(sorted_run_smallest_seqno == level_smallest_seqno);
|
|
|
|
assert(sorted_run_largest_seqno == level_largest_seqno);
|
|
|
|
}
|
|
|
|
} else if (compaction_style == kCompactionStyleUniversal) {
|
|
|
|
// level <= 0 means sorted runs on level 0
|
|
|
|
auto level0_file =
|
|
|
|
meta.levels[0].files[sorted_runs[k].size() - 1 - i];
|
|
|
|
assert(sorted_run_smallest_key == level0_file.smallestkey);
|
|
|
|
assert(sorted_run_largest_key == level0_file.largestkey);
|
|
|
|
if (level != static_cast<int>(max_level)) {
|
|
|
|
assert(sorted_run_smallest_seqno == level0_file.smallest_seqno);
|
|
|
|
assert(sorted_run_largest_seqno == level0_file.largest_seqno);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
// print the size of each sorted_run
|
|
|
|
for (size_t k = 0; k < num_db; k++) {
|
|
|
|
auto db = db_list[k];
|
|
|
|
fprintf(stdout,
|
2017-01-06 23:42:18 +01:00
|
|
|
"---------------------- DB %" ROCKSDB_PRIszt " LSM ---------------------\n", k);
|
2016-10-19 01:30:57 +02:00
|
|
|
db->GetColumnFamilyMetaData(&meta);
|
|
|
|
for (auto& levelMeta : meta.levels) {
|
|
|
|
if (levelMeta.files.empty()) {
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
if (levelMeta.level == 0) {
|
|
|
|
for (auto& fileMeta : levelMeta.files) {
|
2018-10-19 23:43:55 +02:00
|
|
|
fprintf(stdout, "Level[%d]: %s(size: %" ROCKSDB_PRIszt " bytes)\n",
|
2016-11-13 00:43:33 +01:00
|
|
|
levelMeta.level, fileMeta.name.c_str(), fileMeta.size);
|
2016-10-19 01:30:57 +02:00
|
|
|
}
|
|
|
|
} else {
|
2016-11-13 00:43:33 +01:00
|
|
|
fprintf(stdout, "Level[%d]: %s - %s(total size: %" PRIi64 " bytes)\n",
|
2016-10-19 01:30:57 +02:00
|
|
|
levelMeta.level, levelMeta.files.front().name.c_str(),
|
|
|
|
levelMeta.files.back().name.c_str(), levelMeta.size);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
for (size_t i = 0; i < num_db; i++) {
|
|
|
|
db_list[i]->SetOptions(
|
|
|
|
{{"disable_auto_compactions",
|
|
|
|
std::to_string(options_list[i].disable_auto_compactions)},
|
|
|
|
{"level0_slowdown_writes_trigger",
|
|
|
|
std::to_string(options_list[i].level0_slowdown_writes_trigger)},
|
|
|
|
{"level0_stop_writes_trigger",
|
|
|
|
std::to_string(options_list[i].level0_stop_writes_trigger)}});
|
|
|
|
}
|
|
|
|
return Status::OK();
|
|
|
|
#else
|
2018-04-16 02:19:57 +02:00
|
|
|
(void)thread;
|
|
|
|
(void)compaction_style;
|
|
|
|
(void)write_mode;
|
2016-10-19 01:30:57 +02:00
|
|
|
fprintf(stderr, "Rocksdb Lite doesn't support filldeterministic\n");
|
|
|
|
return Status::NotSupported(
|
|
|
|
"Rocksdb Lite doesn't support filldeterministic");
|
|
|
|
#endif // ROCKSDB_LITE
|
|
|
|
}
|
|
|
|
|
2011-08-22 23:08:51 +02:00
|
|
|
void ReadSequential(ThreadState* thread) {
|
2014-08-19 03:15:01 +02:00
|
|
|
if (db_.db != nullptr) {
|
|
|
|
ReadSequential(thread, db_.db);
|
2014-04-11 21:15:09 +02:00
|
|
|
} else {
|
2014-08-19 03:15:01 +02:00
|
|
|
for (const auto& db_with_cfh : multi_dbs_) {
|
|
|
|
ReadSequential(thread, db_with_cfh.db);
|
2014-04-11 21:15:09 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void ReadSequential(ThreadState* thread, DB* db) {
|
2014-09-26 23:20:24 +02:00
|
|
|
ReadOptions options(FLAGS_verify_checksum, true);
|
|
|
|
options.tailing = FLAGS_use_tailing_iterator;
|
|
|
|
|
|
|
|
Iterator* iter = db->NewIterator(options);
|
2014-03-05 02:08:05 +01:00
|
|
|
int64_t i = 0;
|
2011-08-22 23:08:51 +02:00
|
|
|
int64_t bytes = 0;
|
2011-05-21 04:17:43 +02:00
|
|
|
for (iter->SeekToFirst(); i < reads_ && iter->Valid(); iter->Next()) {
|
2011-08-22 23:08:51 +02:00
|
|
|
bytes += iter->key().size() + iter->value().size();
|
2015-12-08 02:33:18 +01:00
|
|
|
thread->stats.FinishedOps(nullptr, db, 1, kRead);
|
2011-03-22 19:32:49 +01:00
|
|
|
++i;
|
2016-12-19 21:25:35 +01:00
|
|
|
|
|
|
|
if (thread->shared->read_rate_limiter.get() != nullptr &&
|
|
|
|
i % 1024 == 1023) {
|
2017-03-17 01:45:28 +01:00
|
|
|
thread->shared->read_rate_limiter->Request(1024, Env::IO_HIGH,
|
2017-06-13 23:51:22 +02:00
|
|
|
nullptr /* stats */,
|
|
|
|
RateLimiter::OpType::kRead);
|
2016-12-19 21:25:35 +01:00
|
|
|
}
|
2011-03-22 19:32:49 +01:00
|
|
|
}
|
2016-12-19 21:25:35 +01:00
|
|
|
|
2011-03-22 19:32:49 +01:00
|
|
|
delete iter;
|
2011-08-22 23:08:51 +02:00
|
|
|
thread->stats.AddBytes(bytes);
|
2016-12-01 22:30:01 +01:00
|
|
|
if (FLAGS_perf_level > rocksdb::PerfLevel::kDisable) {
|
2018-07-20 23:34:07 +02:00
|
|
|
thread->stats.AddMessage(std::string("PERF_CONTEXT:\n") +
|
|
|
|
get_perf_context()->ToString());
|
2016-12-01 22:30:01 +01:00
|
|
|
}
|
2011-03-22 19:32:49 +01:00
|
|
|
}
|
|
|
|
|
2011-08-22 23:08:51 +02:00
|
|
|
void ReadReverse(ThreadState* thread) {
|
2014-08-19 03:15:01 +02:00
|
|
|
if (db_.db != nullptr) {
|
|
|
|
ReadReverse(thread, db_.db);
|
2014-04-11 21:15:09 +02:00
|
|
|
} else {
|
2014-08-19 03:15:01 +02:00
|
|
|
for (const auto& db_with_cfh : multi_dbs_) {
|
|
|
|
ReadReverse(thread, db_with_cfh.db);
|
2014-04-11 21:15:09 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void ReadReverse(ThreadState* thread, DB* db) {
|
|
|
|
Iterator* iter = db->NewIterator(ReadOptions(FLAGS_verify_checksum, true));
|
2014-03-05 02:08:05 +01:00
|
|
|
int64_t i = 0;
|
2011-08-22 23:08:51 +02:00
|
|
|
int64_t bytes = 0;
|
2011-05-21 04:17:43 +02:00
|
|
|
for (iter->SeekToLast(); i < reads_ && iter->Valid(); iter->Prev()) {
|
2011-08-22 23:08:51 +02:00
|
|
|
bytes += iter->key().size() + iter->value().size();
|
2015-12-08 02:33:18 +01:00
|
|
|
thread->stats.FinishedOps(nullptr, db, 1, kRead);
|
2011-03-22 19:32:49 +01:00
|
|
|
++i;
|
2016-12-19 21:25:35 +01:00
|
|
|
if (thread->shared->read_rate_limiter.get() != nullptr &&
|
|
|
|
i % 1024 == 1023) {
|
2017-03-17 01:45:28 +01:00
|
|
|
thread->shared->read_rate_limiter->Request(1024, Env::IO_HIGH,
|
2017-06-13 23:51:22 +02:00
|
|
|
nullptr /* stats */,
|
|
|
|
RateLimiter::OpType::kRead);
|
2016-12-19 21:25:35 +01:00
|
|
|
}
|
2011-03-22 19:32:49 +01:00
|
|
|
}
|
|
|
|
delete iter;
|
2011-08-22 23:08:51 +02:00
|
|
|
thread->stats.AddBytes(bytes);
|
2011-03-22 19:32:49 +01:00
|
|
|
}
|
|
|
|
|
2014-09-25 20:14:01 +02:00
|
|
|
void ReadRandomFast(ThreadState* thread) {
|
|
|
|
int64_t read = 0;
|
|
|
|
int64_t found = 0;
|
2014-09-25 22:34:51 +02:00
|
|
|
int64_t nonexist = 0;
|
2014-09-25 20:14:01 +02:00
|
|
|
ReadOptions options(FLAGS_verify_checksum, true);
|
2015-02-19 23:27:48 +01:00
|
|
|
std::unique_ptr<const char[]> key_guard;
|
|
|
|
Slice key = AllocateKey(&key_guard);
|
2014-09-25 20:14:01 +02:00
|
|
|
std::string value;
|
|
|
|
DB* db = SelectDBWithCfh(thread)->db;
|
|
|
|
|
|
|
|
int64_t pot = 1;
|
|
|
|
while (pot < FLAGS_num) {
|
|
|
|
pot <<= 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
Duration duration(FLAGS_duration, reads_);
|
|
|
|
do {
|
|
|
|
for (int i = 0; i < 100; ++i) {
|
|
|
|
int64_t key_rand = thread->rand.Next() & (pot - 1);
|
|
|
|
GenerateKeyFromInt(key_rand, FLAGS_num, &key);
|
|
|
|
++read;
|
2015-01-22 03:18:15 +01:00
|
|
|
auto status = db->Get(options, key, &value);
|
|
|
|
if (status.ok()) {
|
2014-09-25 20:14:01 +02:00
|
|
|
++found;
|
2015-01-22 03:18:15 +01:00
|
|
|
} else if (!status.IsNotFound()) {
|
2015-01-22 03:23:12 +01:00
|
|
|
fprintf(stderr, "Get returned an error: %s\n",
|
|
|
|
status.ToString().c_str());
|
2015-01-22 03:18:15 +01:00
|
|
|
abort();
|
2014-09-25 20:14:01 +02:00
|
|
|
}
|
2014-09-25 22:34:51 +02:00
|
|
|
if (key_rand >= FLAGS_num) {
|
|
|
|
++nonexist;
|
|
|
|
}
|
2014-09-25 20:14:01 +02:00
|
|
|
}
|
2016-12-19 21:25:35 +01:00
|
|
|
if (thread->shared->read_rate_limiter.get() != nullptr) {
|
2017-06-13 23:51:22 +02:00
|
|
|
thread->shared->read_rate_limiter->Request(
|
|
|
|
100, Env::IO_HIGH, nullptr /* stats */, RateLimiter::OpType::kRead);
|
2016-12-19 21:25:35 +01:00
|
|
|
}
|
|
|
|
|
2015-12-08 02:33:18 +01:00
|
|
|
thread->stats.FinishedOps(nullptr, db, 100, kRead);
|
2014-09-25 20:14:01 +02:00
|
|
|
} while (!duration.Done(100));
|
|
|
|
|
|
|
|
char msg[100];
|
2014-09-25 22:34:51 +02:00
|
|
|
snprintf(msg, sizeof(msg), "(%" PRIu64 " of %" PRIu64 " found, "
|
|
|
|
"issued %" PRIu64 " non-exist keys)\n",
|
|
|
|
found, read, nonexist);
|
2014-09-25 20:14:01 +02:00
|
|
|
|
|
|
|
thread->stats.AddMessage(msg);
|
|
|
|
|
2016-06-30 03:44:22 +02:00
|
|
|
if (FLAGS_perf_level > rocksdb::PerfLevel::kDisable) {
|
2018-07-20 23:34:07 +02:00
|
|
|
thread->stats.AddMessage(std::string("PERF_CONTEXT:\n") +
|
|
|
|
get_perf_context()->ToString());
|
2014-09-25 20:14:01 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2015-03-07 01:38:50 +01:00
|
|
|
int64_t GetRandomKey(Random64* rand) {
|
|
|
|
uint64_t rand_int = rand->Next();
|
|
|
|
int64_t key_rand;
|
|
|
|
if (read_random_exp_range_ == 0) {
|
|
|
|
key_rand = rand_int % FLAGS_num;
|
|
|
|
} else {
|
|
|
|
const uint64_t kBigInt = static_cast<uint64_t>(1U) << 62;
|
|
|
|
long double order = -static_cast<long double>(rand_int % kBigInt) /
|
|
|
|
static_cast<long double>(kBigInt) *
|
|
|
|
read_random_exp_range_;
|
|
|
|
long double exp_ran = std::exp(order);
|
2015-03-11 00:32:12 +01:00
|
|
|
uint64_t rand_num =
|
2015-03-07 01:38:50 +01:00
|
|
|
static_cast<int64_t>(exp_ran * static_cast<long double>(FLAGS_num));
|
2015-03-11 00:32:12 +01:00
|
|
|
// Map to a different number to avoid locality.
|
|
|
|
const uint64_t kBigPrime = 0x5bd1e995;
|
|
|
|
// Overflow is like %(2^64). Will have little impact of results.
|
|
|
|
key_rand = static_cast<int64_t>((rand_num * kBigPrime) % FLAGS_num);
|
2015-03-07 01:38:50 +01:00
|
|
|
}
|
|
|
|
return key_rand;
|
|
|
|
}
|
|
|
|
|
2011-08-22 23:08:51 +02:00
|
|
|
void ReadRandom(ThreadState* thread) {
|
2014-03-28 17:21:20 +01:00
|
|
|
int64_t read = 0;
|
2014-04-08 20:21:09 +02:00
|
|
|
int64_t found = 0;
|
2015-03-30 20:28:25 +02:00
|
|
|
int64_t bytes = 0;
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
2019-04-11 23:24:09 +02:00
|
|
|
int num_keys = 0;
|
|
|
|
int64_t key_rand = GetRandomKey(&thread->rand);
|
2014-04-08 20:21:09 +02:00
|
|
|
ReadOptions options(FLAGS_verify_checksum, true);
|
2015-02-19 23:27:48 +01:00
|
|
|
std::unique_ptr<const char[]> key_guard;
|
|
|
|
Slice key = AllocateKey(&key_guard);
|
2017-03-13 19:44:50 +01:00
|
|
|
PinnableSlice pinnable_val;
|
2013-01-03 21:11:50 +01:00
|
|
|
|
2014-04-08 20:21:09 +02:00
|
|
|
Duration duration(FLAGS_duration, reads_);
|
|
|
|
while (!duration.Done(1)) {
|
2014-08-19 03:15:01 +02:00
|
|
|
DBWithColumnFamilies* db_with_cfh = SelectDBWithCfh(thread);
|
|
|
|
// We use same key_rand as seed for key and column family so that we can
|
|
|
|
// deterministically find the cfh corresponding to a particular key, as it
|
|
|
|
// is done in DoWrite method.
|
|
|
|
GenerateKeyFromInt(key_rand, FLAGS_num, &key);
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
2019-04-11 23:24:09 +02:00
|
|
|
if (entries_per_batch_ > 1 && FLAGS_multiread_stride) {
|
|
|
|
if (++num_keys == entries_per_batch_) {
|
|
|
|
num_keys = 0;
|
|
|
|
key_rand = GetRandomKey(&thread->rand);
|
|
|
|
if ((key_rand + (entries_per_batch_ - 1) * FLAGS_multiread_stride) >=
|
|
|
|
FLAGS_num) {
|
|
|
|
key_rand = FLAGS_num - entries_per_batch_ * FLAGS_multiread_stride;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
key_rand += FLAGS_multiread_stride;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
key_rand = GetRandomKey(&thread->rand);
|
|
|
|
}
|
2014-04-08 20:21:09 +02:00
|
|
|
read++;
|
2014-08-19 03:15:01 +02:00
|
|
|
Status s;
|
|
|
|
if (FLAGS_num_column_families > 1) {
|
2014-12-23 03:39:28 +01:00
|
|
|
s = db_with_cfh->db->Get(options, db_with_cfh->GetCfh(key_rand), key,
|
2017-06-16 01:08:32 +02:00
|
|
|
&pinnable_val);
|
2014-08-19 03:15:01 +02:00
|
|
|
} else {
|
2017-06-16 01:08:32 +02:00
|
|
|
pinnable_val.Reset();
|
|
|
|
s = db_with_cfh->db->Get(options,
|
|
|
|
db_with_cfh->db->DefaultColumnFamily(), key,
|
|
|
|
&pinnable_val);
|
2014-08-19 03:15:01 +02:00
|
|
|
}
|
|
|
|
if (s.ok()) {
|
2014-04-08 20:21:09 +02:00
|
|
|
found++;
|
2017-06-16 01:08:32 +02:00
|
|
|
bytes += key.size() + pinnable_val.size();
|
2015-01-22 03:18:15 +01:00
|
|
|
} else if (!s.IsNotFound()) {
|
2015-01-22 03:23:12 +01:00
|
|
|
fprintf(stderr, "Get returned an error: %s\n", s.ToString().c_str());
|
2015-01-22 03:18:15 +01:00
|
|
|
abort();
|
2013-03-01 20:25:13 +01:00
|
|
|
}
|
2016-12-19 21:25:35 +01:00
|
|
|
|
|
|
|
if (thread->shared->read_rate_limiter.get() != nullptr &&
|
|
|
|
read % 256 == 255) {
|
2017-06-13 23:51:22 +02:00
|
|
|
thread->shared->read_rate_limiter->Request(
|
|
|
|
256, Env::IO_HIGH, nullptr /* stats */, RateLimiter::OpType::kRead);
|
2016-12-19 21:25:35 +01:00
|
|
|
}
|
|
|
|
|
2015-12-08 02:33:18 +01:00
|
|
|
thread->stats.FinishedOps(db_with_cfh, db_with_cfh->db, 1, kRead);
|
2012-04-17 17:36:46 +02:00
|
|
|
}
|
2013-06-12 21:42:21 +02:00
|
|
|
|
2012-04-17 17:36:46 +02:00
|
|
|
char msg[100];
|
2014-05-30 23:31:55 +02:00
|
|
|
snprintf(msg, sizeof(msg), "(%" PRIu64 " of %" PRIu64 " found)\n",
|
2014-03-28 17:21:20 +01:00
|
|
|
found, read);
|
2014-03-05 19:32:54 +01:00
|
|
|
|
2015-03-30 20:28:25 +02:00
|
|
|
thread->stats.AddBytes(bytes);
|
2012-04-17 17:36:46 +02:00
|
|
|
thread->stats.AddMessage(msg);
|
2014-03-05 19:32:54 +01:00
|
|
|
|
2016-06-30 03:44:22 +02:00
|
|
|
if (FLAGS_perf_level > rocksdb::PerfLevel::kDisable) {
|
2018-07-20 23:34:07 +02:00
|
|
|
thread->stats.AddMessage(std::string("PERF_CONTEXT:\n") +
|
|
|
|
get_perf_context()->ToString());
|
2014-03-05 19:32:54 +01:00
|
|
|
}
|
2012-04-17 17:36:46 +02:00
|
|
|
}
|
|
|
|
|
2014-04-08 20:21:09 +02:00
|
|
|
// Calls MultiGet over a list of keys from a random distribution.
|
|
|
|
// Returns the total number of keys found.
|
|
|
|
void MultiReadRandom(ThreadState* thread) {
|
|
|
|
int64_t read = 0;
|
2016-12-19 21:25:35 +01:00
|
|
|
int64_t num_multireads = 0;
|
2014-03-05 02:08:05 +01:00
|
|
|
int64_t found = 0;
|
2012-05-19 09:05:48 +02:00
|
|
|
ReadOptions options(FLAGS_verify_checksum, true);
|
2014-04-15 00:42:20 +02:00
|
|
|
std::vector<Slice> keys;
|
2015-02-19 23:27:48 +01:00
|
|
|
std::vector<std::unique_ptr<const char[]> > key_guards;
|
2014-04-08 20:21:09 +02:00
|
|
|
std::vector<std::string> values(entries_per_batch_);
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
2019-04-11 23:24:09 +02:00
|
|
|
PinnableSlice* pin_values = new PinnableSlice[entries_per_batch_];
|
2019-04-12 19:42:06 +02:00
|
|
|
std::unique_ptr<PinnableSlice[]> pin_values_guard(pin_values);
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
2019-04-11 23:24:09 +02:00
|
|
|
std::vector<Status> stat_list(entries_per_batch_);
|
2014-04-29 21:33:57 +02:00
|
|
|
while (static_cast<int64_t>(keys.size()) < entries_per_batch_) {
|
2015-12-29 02:42:32 +01:00
|
|
|
key_guards.push_back(std::unique_ptr<const char[]>());
|
2015-02-19 23:27:48 +01:00
|
|
|
keys.push_back(AllocateKey(&key_guards.back()));
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
2013-03-01 20:25:13 +01:00
|
|
|
Duration duration(FLAGS_duration, reads_);
|
2014-04-08 20:21:09 +02:00
|
|
|
while (!duration.Done(1)) {
|
2014-04-11 21:15:09 +02:00
|
|
|
DB* db = SelectDB(thread);
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
2019-04-11 23:24:09 +02:00
|
|
|
if (FLAGS_multiread_stride) {
|
|
|
|
int64_t key = GetRandomKey(&thread->rand);
|
|
|
|
if ((key + (entries_per_batch_ - 1) * FLAGS_multiread_stride) >=
|
|
|
|
(int64_t)FLAGS_num) {
|
|
|
|
key = FLAGS_num - entries_per_batch_ * FLAGS_multiread_stride;
|
|
|
|
}
|
|
|
|
for (int64_t i = 0; i < entries_per_batch_; ++i) {
|
|
|
|
GenerateKeyFromInt(key, FLAGS_num, &keys[i]);
|
|
|
|
key += FLAGS_multiread_stride;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
for (int64_t i = 0; i < entries_per_batch_; ++i) {
|
|
|
|
GenerateKeyFromInt(GetRandomKey(&thread->rand), FLAGS_num, &keys[i]);
|
|
|
|
}
|
2013-06-12 21:42:21 +02:00
|
|
|
}
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
2019-04-11 23:24:09 +02:00
|
|
|
if (!FLAGS_multiread_batched) {
|
|
|
|
std::vector<Status> statuses = db->MultiGet(options, keys, &values);
|
|
|
|
assert(static_cast<int64_t>(statuses.size()) == entries_per_batch_);
|
|
|
|
|
|
|
|
read += entries_per_batch_;
|
|
|
|
num_multireads++;
|
|
|
|
for (int64_t i = 0; i < entries_per_batch_; ++i) {
|
|
|
|
if (statuses[i].ok()) {
|
|
|
|
++found;
|
|
|
|
} else if (!statuses[i].IsNotFound()) {
|
|
|
|
fprintf(stderr, "MultiGet returned an error: %s\n",
|
|
|
|
statuses[i].ToString().c_str());
|
|
|
|
abort();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
db->MultiGet(options, db->DefaultColumnFamily(), keys.size(),
|
|
|
|
keys.data(), pin_values, stat_list.data());
|
|
|
|
|
|
|
|
read += entries_per_batch_;
|
|
|
|
num_multireads++;
|
|
|
|
for (int64_t i = 0; i < entries_per_batch_; ++i) {
|
|
|
|
if (stat_list[i].ok()) {
|
|
|
|
++found;
|
|
|
|
} else if (!stat_list[i].IsNotFound()) {
|
|
|
|
fprintf(stderr, "MultiGet returned an error: %s\n",
|
|
|
|
stat_list[i].ToString().c_str());
|
|
|
|
abort();
|
|
|
|
}
|
|
|
|
stat_list[i] = Status::OK();
|
|
|
|
pin_values[i].Reset();
|
2013-06-12 21:42:21 +02:00
|
|
|
}
|
|
|
|
}
|
2016-12-19 21:25:35 +01:00
|
|
|
if (thread->shared->read_rate_limiter.get() != nullptr &&
|
|
|
|
num_multireads % 256 == 255) {
|
2017-03-17 01:45:28 +01:00
|
|
|
thread->shared->read_rate_limiter->Request(
|
2017-06-13 23:51:22 +02:00
|
|
|
256 * entries_per_batch_, Env::IO_HIGH, nullptr /* stats */,
|
|
|
|
RateLimiter::OpType::kRead);
|
2016-12-19 21:25:35 +01:00
|
|
|
}
|
2015-12-08 02:33:18 +01:00
|
|
|
thread->stats.FinishedOps(nullptr, db, entries_per_batch_, kRead);
|
2011-05-21 04:17:43 +02:00
|
|
|
}
|
2013-06-12 21:42:21 +02:00
|
|
|
|
|
|
|
char msg[100];
|
2014-03-05 02:08:05 +01:00
|
|
|
snprintf(msg, sizeof(msg), "(%" PRIu64 " of %" PRIu64 " found)",
|
2014-04-08 20:21:09 +02:00
|
|
|
found, read);
|
2013-06-12 21:42:21 +02:00
|
|
|
thread->stats.AddMessage(msg);
|
2011-05-21 04:17:43 +02:00
|
|
|
}
|
|
|
|
|
Generate mixed workload with Get, Put, Seek in db_bench (#4788)
Summary:
Based on the specific workload models (key access distribution, value size distribution, and iterator scan length distribution, the QPS variation), the MixGraph benchmark generate the synthetic workload according to these distributions which can reflect the real-world workload characteristics.
After user enable the tracing function, they will get the trace file. By analyzing the trace file with the trace_analyzer tool, user can generate a set of statistic data files including. The *_accessed_key_stats.txt, *-accessed_value_size_distribution.txt, *-iterator_length_distribution.txt, and *-qps_stats.txt are mainly used to fit the Matlab model fitting. After that, user can get the parameters of the workload distributions (the modeling details are described: [here](https://github.com/facebook/rocksdb/wiki/RocksDB-Trace%2C-Replay%2C-and-Analyzer))
The key access distribution follows the The two-term power model. The probability density function is: `f(x) = ax^{b}+c`. The corresponding parameters are key_dist_a, key_dist_b, and key_dist_c in db_bench
For the value size distribution and iterator scan length distribution, they both follow the Generalized Pareto Distribution. The probability density function is `f(x) = (1/sigma)(1+k*(x-theta)/sigma))^{-1-1/k)`. The parameters are: value_k, value_theta, value_sigma and iter_k, iter_theta, iter_sigma. For more information about the Generalized Pareto Distribution, users can find the [wiki](https://en.wikipedia.org/wiki/Generalized_Pareto_distribution) and [Matalb page](https://www.mathworks.com/help/stats/generalized-pareto-distribution.html)
As for the QPS, it follows the diurnal pattern. So Sine is a good model to fit it. `F(x) = sine_a*sin(sine_b*x + sine_c) + sine_d`. The trace_will tell you the average QPS in the print out resutls, which is sine_d. After user fit the "*-qps_stats.txt" to the Matlab model, user can get the sine_a, sine_b, and sine_c. By using the 4 parameters, user can control the QPS variation including the period, average, changes.
To use the bench mark, user can indicate the following parameters as examples:
```
-benchmarks="mixgraph" -key_dist_a=0.002312 -key_dist_b=0.3467 -value_k=0.9233 -value_sigma=226.4092 -iter_k=2.517 -iter_sigma=14.236 -mix_get_ratio=0.7 -mix_put_ratio=0.25 -mix_seek_ratio=0.05 -sine_mix_rate_interval_milliseconds=500 -sine_a=15000 -sine_b=1 -sine_d=20000
```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4788
Differential Revision: D13573940
Pulled By: sagar0
fbshipit-source-id: e184c27e07b4f1bc0b436c2be36c5090c1fb0222
2019-01-22 19:40:44 +01:00
|
|
|
// THe reverse function of Pareto function
|
|
|
|
int64_t ParetoCdfInversion(double u, double theta, double k, double sigma) {
|
|
|
|
double ret;
|
|
|
|
if (k == 0.0) {
|
|
|
|
ret = theta - sigma * std::log(u);
|
|
|
|
} else {
|
|
|
|
ret = theta + sigma * (std::pow(u, -1 * k) - 1) / k;
|
|
|
|
}
|
|
|
|
return static_cast<int64_t>(ceil(ret));
|
|
|
|
}
|
|
|
|
// inversion of y=ax^b
|
|
|
|
int64_t PowerCdfInversion(double u, double a, double b) {
|
|
|
|
double ret;
|
|
|
|
ret = std::pow((u / a), (1 / b));
|
|
|
|
return static_cast<int64_t>(ceil(ret));
|
|
|
|
}
|
|
|
|
|
|
|
|
// Add the noice to the QPS
|
|
|
|
double AddNoise(double origin, double noise_ratio) {
|
|
|
|
if (noise_ratio < 0.0 || noise_ratio > 1.0) {
|
|
|
|
return origin;
|
|
|
|
}
|
|
|
|
int band_int = static_cast<int>(FLAGS_sine_a);
|
|
|
|
double delta = (rand() % band_int - band_int / 2) * noise_ratio;
|
|
|
|
if (origin + delta < 0) {
|
|
|
|
return origin;
|
|
|
|
} else {
|
|
|
|
return (origin + delta);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// decide the query type
|
|
|
|
// 0 Get, 1 Put, 2 Seek, 3 SeekForPrev, 4 Delete, 5 SingleDelete, 6 merge
|
|
|
|
class QueryDecider {
|
|
|
|
public:
|
|
|
|
std::vector<int> type_;
|
|
|
|
std::vector<double> ratio_;
|
|
|
|
int range_;
|
|
|
|
|
|
|
|
QueryDecider() {}
|
|
|
|
~QueryDecider() {}
|
|
|
|
|
|
|
|
Status Initiate(std::vector<double> ratio_input) {
|
|
|
|
int range_max = 1000;
|
|
|
|
double sum = 0.0;
|
|
|
|
for (auto& ratio : ratio_input) {
|
|
|
|
sum += ratio;
|
|
|
|
}
|
|
|
|
range_ = 0;
|
|
|
|
for (auto& ratio : ratio_input) {
|
|
|
|
range_ += static_cast<int>(ceil(range_max * (ratio / sum)));
|
|
|
|
type_.push_back(range_);
|
|
|
|
ratio_.push_back(ratio / sum);
|
|
|
|
}
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
int GetType(int64_t rand_num) {
|
|
|
|
if (rand_num < 0) {
|
|
|
|
rand_num = rand_num * (-1);
|
|
|
|
}
|
2019-01-23 22:26:03 +01:00
|
|
|
assert(range_ != 0);
|
Generate mixed workload with Get, Put, Seek in db_bench (#4788)
Summary:
Based on the specific workload models (key access distribution, value size distribution, and iterator scan length distribution, the QPS variation), the MixGraph benchmark generate the synthetic workload according to these distributions which can reflect the real-world workload characteristics.
After user enable the tracing function, they will get the trace file. By analyzing the trace file with the trace_analyzer tool, user can generate a set of statistic data files including. The *_accessed_key_stats.txt, *-accessed_value_size_distribution.txt, *-iterator_length_distribution.txt, and *-qps_stats.txt are mainly used to fit the Matlab model fitting. After that, user can get the parameters of the workload distributions (the modeling details are described: [here](https://github.com/facebook/rocksdb/wiki/RocksDB-Trace%2C-Replay%2C-and-Analyzer))
The key access distribution follows the The two-term power model. The probability density function is: `f(x) = ax^{b}+c`. The corresponding parameters are key_dist_a, key_dist_b, and key_dist_c in db_bench
For the value size distribution and iterator scan length distribution, they both follow the Generalized Pareto Distribution. The probability density function is `f(x) = (1/sigma)(1+k*(x-theta)/sigma))^{-1-1/k)`. The parameters are: value_k, value_theta, value_sigma and iter_k, iter_theta, iter_sigma. For more information about the Generalized Pareto Distribution, users can find the [wiki](https://en.wikipedia.org/wiki/Generalized_Pareto_distribution) and [Matalb page](https://www.mathworks.com/help/stats/generalized-pareto-distribution.html)
As for the QPS, it follows the diurnal pattern. So Sine is a good model to fit it. `F(x) = sine_a*sin(sine_b*x + sine_c) + sine_d`. The trace_will tell you the average QPS in the print out resutls, which is sine_d. After user fit the "*-qps_stats.txt" to the Matlab model, user can get the sine_a, sine_b, and sine_c. By using the 4 parameters, user can control the QPS variation including the period, average, changes.
To use the bench mark, user can indicate the following parameters as examples:
```
-benchmarks="mixgraph" -key_dist_a=0.002312 -key_dist_b=0.3467 -value_k=0.9233 -value_sigma=226.4092 -iter_k=2.517 -iter_sigma=14.236 -mix_get_ratio=0.7 -mix_put_ratio=0.25 -mix_seek_ratio=0.05 -sine_mix_rate_interval_milliseconds=500 -sine_a=15000 -sine_b=1 -sine_d=20000
```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4788
Differential Revision: D13573940
Pulled By: sagar0
fbshipit-source-id: e184c27e07b4f1bc0b436c2be36c5090c1fb0222
2019-01-22 19:40:44 +01:00
|
|
|
int pos = static_cast<int>(rand_num % range_);
|
|
|
|
for (int i = 0; i < static_cast<int>(type_.size()); i++) {
|
|
|
|
if (pos < type_[i]) {
|
|
|
|
return i;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
};
|
|
|
|
|
|
|
|
// The graph wokrload mixed with Get, Put, Iterator
|
|
|
|
void MixGraph(ThreadState* thread) {
|
|
|
|
int64_t read = 0; // including single gets and Next of iterators
|
|
|
|
int64_t gets = 0;
|
|
|
|
int64_t puts = 0;
|
|
|
|
int64_t found = 0;
|
|
|
|
int64_t seek = 0;
|
|
|
|
int64_t seek_found = 0;
|
|
|
|
int64_t bytes = 0;
|
2019-03-08 23:07:19 +01:00
|
|
|
const int64_t default_value_max = 1 * 1024 * 1024;
|
2019-01-28 21:23:03 +01:00
|
|
|
int64_t value_max = default_value_max;
|
Generate mixed workload with Get, Put, Seek in db_bench (#4788)
Summary:
Based on the specific workload models (key access distribution, value size distribution, and iterator scan length distribution, the QPS variation), the MixGraph benchmark generate the synthetic workload according to these distributions which can reflect the real-world workload characteristics.
After user enable the tracing function, they will get the trace file. By analyzing the trace file with the trace_analyzer tool, user can generate a set of statistic data files including. The *_accessed_key_stats.txt, *-accessed_value_size_distribution.txt, *-iterator_length_distribution.txt, and *-qps_stats.txt are mainly used to fit the Matlab model fitting. After that, user can get the parameters of the workload distributions (the modeling details are described: [here](https://github.com/facebook/rocksdb/wiki/RocksDB-Trace%2C-Replay%2C-and-Analyzer))
The key access distribution follows the The two-term power model. The probability density function is: `f(x) = ax^{b}+c`. The corresponding parameters are key_dist_a, key_dist_b, and key_dist_c in db_bench
For the value size distribution and iterator scan length distribution, they both follow the Generalized Pareto Distribution. The probability density function is `f(x) = (1/sigma)(1+k*(x-theta)/sigma))^{-1-1/k)`. The parameters are: value_k, value_theta, value_sigma and iter_k, iter_theta, iter_sigma. For more information about the Generalized Pareto Distribution, users can find the [wiki](https://en.wikipedia.org/wiki/Generalized_Pareto_distribution) and [Matalb page](https://www.mathworks.com/help/stats/generalized-pareto-distribution.html)
As for the QPS, it follows the diurnal pattern. So Sine is a good model to fit it. `F(x) = sine_a*sin(sine_b*x + sine_c) + sine_d`. The trace_will tell you the average QPS in the print out resutls, which is sine_d. After user fit the "*-qps_stats.txt" to the Matlab model, user can get the sine_a, sine_b, and sine_c. By using the 4 parameters, user can control the QPS variation including the period, average, changes.
To use the bench mark, user can indicate the following parameters as examples:
```
-benchmarks="mixgraph" -key_dist_a=0.002312 -key_dist_b=0.3467 -value_k=0.9233 -value_sigma=226.4092 -iter_k=2.517 -iter_sigma=14.236 -mix_get_ratio=0.7 -mix_put_ratio=0.25 -mix_seek_ratio=0.05 -sine_mix_rate_interval_milliseconds=500 -sine_a=15000 -sine_b=1 -sine_d=20000
```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4788
Differential Revision: D13573940
Pulled By: sagar0
fbshipit-source-id: e184c27e07b4f1bc0b436c2be36c5090c1fb0222
2019-01-22 19:40:44 +01:00
|
|
|
int64_t scan_len_max = FLAGS_mix_max_scan_len;
|
|
|
|
double write_rate = 1000000.0;
|
|
|
|
double read_rate = 1000000.0;
|
2019-01-30 01:23:21 +01:00
|
|
|
std::vector<double> ratio{FLAGS_mix_get_ratio, FLAGS_mix_put_ratio,
|
|
|
|
FLAGS_mix_seek_ratio};
|
2019-01-28 21:23:03 +01:00
|
|
|
char value_buffer[default_value_max];
|
Generate mixed workload with Get, Put, Seek in db_bench (#4788)
Summary:
Based on the specific workload models (key access distribution, value size distribution, and iterator scan length distribution, the QPS variation), the MixGraph benchmark generate the synthetic workload according to these distributions which can reflect the real-world workload characteristics.
After user enable the tracing function, they will get the trace file. By analyzing the trace file with the trace_analyzer tool, user can generate a set of statistic data files including. The *_accessed_key_stats.txt, *-accessed_value_size_distribution.txt, *-iterator_length_distribution.txt, and *-qps_stats.txt are mainly used to fit the Matlab model fitting. After that, user can get the parameters of the workload distributions (the modeling details are described: [here](https://github.com/facebook/rocksdb/wiki/RocksDB-Trace%2C-Replay%2C-and-Analyzer))
The key access distribution follows the The two-term power model. The probability density function is: `f(x) = ax^{b}+c`. The corresponding parameters are key_dist_a, key_dist_b, and key_dist_c in db_bench
For the value size distribution and iterator scan length distribution, they both follow the Generalized Pareto Distribution. The probability density function is `f(x) = (1/sigma)(1+k*(x-theta)/sigma))^{-1-1/k)`. The parameters are: value_k, value_theta, value_sigma and iter_k, iter_theta, iter_sigma. For more information about the Generalized Pareto Distribution, users can find the [wiki](https://en.wikipedia.org/wiki/Generalized_Pareto_distribution) and [Matalb page](https://www.mathworks.com/help/stats/generalized-pareto-distribution.html)
As for the QPS, it follows the diurnal pattern. So Sine is a good model to fit it. `F(x) = sine_a*sin(sine_b*x + sine_c) + sine_d`. The trace_will tell you the average QPS in the print out resutls, which is sine_d. After user fit the "*-qps_stats.txt" to the Matlab model, user can get the sine_a, sine_b, and sine_c. By using the 4 parameters, user can control the QPS variation including the period, average, changes.
To use the bench mark, user can indicate the following parameters as examples:
```
-benchmarks="mixgraph" -key_dist_a=0.002312 -key_dist_b=0.3467 -value_k=0.9233 -value_sigma=226.4092 -iter_k=2.517 -iter_sigma=14.236 -mix_get_ratio=0.7 -mix_put_ratio=0.25 -mix_seek_ratio=0.05 -sine_mix_rate_interval_milliseconds=500 -sine_a=15000 -sine_b=1 -sine_d=20000
```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4788
Differential Revision: D13573940
Pulled By: sagar0
fbshipit-source-id: e184c27e07b4f1bc0b436c2be36c5090c1fb0222
2019-01-22 19:40:44 +01:00
|
|
|
QueryDecider query;
|
|
|
|
RandomGenerator gen;
|
|
|
|
Status s;
|
2019-01-30 01:23:21 +01:00
|
|
|
if (value_max > FLAGS_mix_max_value_size) {
|
2019-01-28 21:23:03 +01:00
|
|
|
value_max = FLAGS_mix_max_value_size;
|
|
|
|
}
|
Generate mixed workload with Get, Put, Seek in db_bench (#4788)
Summary:
Based on the specific workload models (key access distribution, value size distribution, and iterator scan length distribution, the QPS variation), the MixGraph benchmark generate the synthetic workload according to these distributions which can reflect the real-world workload characteristics.
After user enable the tracing function, they will get the trace file. By analyzing the trace file with the trace_analyzer tool, user can generate a set of statistic data files including. The *_accessed_key_stats.txt, *-accessed_value_size_distribution.txt, *-iterator_length_distribution.txt, and *-qps_stats.txt are mainly used to fit the Matlab model fitting. After that, user can get the parameters of the workload distributions (the modeling details are described: [here](https://github.com/facebook/rocksdb/wiki/RocksDB-Trace%2C-Replay%2C-and-Analyzer))
The key access distribution follows the The two-term power model. The probability density function is: `f(x) = ax^{b}+c`. The corresponding parameters are key_dist_a, key_dist_b, and key_dist_c in db_bench
For the value size distribution and iterator scan length distribution, they both follow the Generalized Pareto Distribution. The probability density function is `f(x) = (1/sigma)(1+k*(x-theta)/sigma))^{-1-1/k)`. The parameters are: value_k, value_theta, value_sigma and iter_k, iter_theta, iter_sigma. For more information about the Generalized Pareto Distribution, users can find the [wiki](https://en.wikipedia.org/wiki/Generalized_Pareto_distribution) and [Matalb page](https://www.mathworks.com/help/stats/generalized-pareto-distribution.html)
As for the QPS, it follows the diurnal pattern. So Sine is a good model to fit it. `F(x) = sine_a*sin(sine_b*x + sine_c) + sine_d`. The trace_will tell you the average QPS in the print out resutls, which is sine_d. After user fit the "*-qps_stats.txt" to the Matlab model, user can get the sine_a, sine_b, and sine_c. By using the 4 parameters, user can control the QPS variation including the period, average, changes.
To use the bench mark, user can indicate the following parameters as examples:
```
-benchmarks="mixgraph" -key_dist_a=0.002312 -key_dist_b=0.3467 -value_k=0.9233 -value_sigma=226.4092 -iter_k=2.517 -iter_sigma=14.236 -mix_get_ratio=0.7 -mix_put_ratio=0.25 -mix_seek_ratio=0.05 -sine_mix_rate_interval_milliseconds=500 -sine_a=15000 -sine_b=1 -sine_d=20000
```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4788
Differential Revision: D13573940
Pulled By: sagar0
fbshipit-source-id: e184c27e07b4f1bc0b436c2be36c5090c1fb0222
2019-01-22 19:40:44 +01:00
|
|
|
|
|
|
|
ReadOptions options(FLAGS_verify_checksum, true);
|
|
|
|
std::unique_ptr<const char[]> key_guard;
|
|
|
|
Slice key = AllocateKey(&key_guard);
|
|
|
|
PinnableSlice pinnable_val;
|
|
|
|
query.Initiate(ratio);
|
|
|
|
|
|
|
|
// the limit of qps initiation
|
|
|
|
if (FLAGS_sine_a != 0 || FLAGS_sine_d != 0) {
|
|
|
|
thread->shared->read_rate_limiter.reset(NewGenericRateLimiter(
|
|
|
|
read_rate, 100000 /* refill_period_us */, 10 /* fairness */,
|
|
|
|
RateLimiter::Mode::kReadsOnly));
|
|
|
|
thread->shared->write_rate_limiter.reset(
|
|
|
|
NewGenericRateLimiter(write_rate));
|
|
|
|
}
|
|
|
|
|
|
|
|
Duration duration(FLAGS_duration, reads_);
|
|
|
|
while (!duration.Done(1)) {
|
|
|
|
DBWithColumnFamilies* db_with_cfh = SelectDBWithCfh(thread);
|
|
|
|
int64_t rand_v, key_rand, key_seed;
|
|
|
|
rand_v = GetRandomKey(&thread->rand) % FLAGS_num;
|
|
|
|
double u = static_cast<double>(rand_v) / FLAGS_num;
|
|
|
|
key_seed = PowerCdfInversion(u, FLAGS_key_dist_a, FLAGS_key_dist_b);
|
|
|
|
Random64 rand(key_seed);
|
|
|
|
key_rand = static_cast<int64_t>(rand.Next()) % FLAGS_num;
|
|
|
|
GenerateKeyFromInt(key_rand, FLAGS_num, &key);
|
|
|
|
int query_type = query.GetType(rand_v);
|
|
|
|
|
|
|
|
// change the qps
|
|
|
|
uint64_t now = FLAGS_env->NowMicros();
|
|
|
|
uint64_t usecs_since_last;
|
|
|
|
if (now > thread->stats.GetSineInterval()) {
|
|
|
|
usecs_since_last = now - thread->stats.GetSineInterval();
|
|
|
|
} else {
|
|
|
|
usecs_since_last = 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (usecs_since_last >
|
|
|
|
(FLAGS_sine_mix_rate_interval_milliseconds * uint64_t{1000})) {
|
|
|
|
double usecs_since_start =
|
|
|
|
static_cast<double>(now - thread->stats.GetStart());
|
|
|
|
thread->stats.ResetSineInterval();
|
|
|
|
double mix_rate_with_noise = AddNoise(
|
|
|
|
SineRate(usecs_since_start / 1000000.0), FLAGS_sine_mix_rate_noise);
|
|
|
|
read_rate = mix_rate_with_noise * (query.ratio_[0] + query.ratio_[2]);
|
|
|
|
write_rate =
|
|
|
|
mix_rate_with_noise * query.ratio_[1] * FLAGS_mix_ave_kv_size;
|
|
|
|
|
|
|
|
thread->shared->write_rate_limiter.reset(
|
|
|
|
NewGenericRateLimiter(write_rate));
|
|
|
|
thread->shared->read_rate_limiter.reset(NewGenericRateLimiter(
|
|
|
|
read_rate,
|
|
|
|
FLAGS_sine_mix_rate_interval_milliseconds * uint64_t{1000}, 10,
|
|
|
|
RateLimiter::Mode::kReadsOnly));
|
|
|
|
}
|
|
|
|
// Start the query
|
|
|
|
if (query_type == 0) {
|
|
|
|
// the Get query
|
|
|
|
gets++;
|
|
|
|
read++;
|
|
|
|
if (FLAGS_num_column_families > 1) {
|
|
|
|
s = db_with_cfh->db->Get(options, db_with_cfh->GetCfh(key_rand), key,
|
|
|
|
&pinnable_val);
|
|
|
|
} else {
|
|
|
|
pinnable_val.Reset();
|
|
|
|
s = db_with_cfh->db->Get(options,
|
|
|
|
db_with_cfh->db->DefaultColumnFamily(), key,
|
|
|
|
&pinnable_val);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (s.ok()) {
|
|
|
|
found++;
|
|
|
|
bytes += key.size() + pinnable_val.size();
|
|
|
|
} else if (!s.IsNotFound()) {
|
|
|
|
fprintf(stderr, "Get returned an error: %s\n", s.ToString().c_str());
|
|
|
|
abort();
|
|
|
|
}
|
|
|
|
|
|
|
|
if (thread->shared->read_rate_limiter.get() != nullptr &&
|
|
|
|
read % 256 == 255) {
|
|
|
|
thread->shared->read_rate_limiter->Request(
|
|
|
|
256, Env::IO_HIGH, nullptr /* stats */,
|
|
|
|
RateLimiter::OpType::kRead);
|
|
|
|
}
|
2019-03-08 23:07:19 +01:00
|
|
|
thread->stats.FinishedOps(db_with_cfh, db_with_cfh->db, 1, kRead);
|
Generate mixed workload with Get, Put, Seek in db_bench (#4788)
Summary:
Based on the specific workload models (key access distribution, value size distribution, and iterator scan length distribution, the QPS variation), the MixGraph benchmark generate the synthetic workload according to these distributions which can reflect the real-world workload characteristics.
After user enable the tracing function, they will get the trace file. By analyzing the trace file with the trace_analyzer tool, user can generate a set of statistic data files including. The *_accessed_key_stats.txt, *-accessed_value_size_distribution.txt, *-iterator_length_distribution.txt, and *-qps_stats.txt are mainly used to fit the Matlab model fitting. After that, user can get the parameters of the workload distributions (the modeling details are described: [here](https://github.com/facebook/rocksdb/wiki/RocksDB-Trace%2C-Replay%2C-and-Analyzer))
The key access distribution follows the The two-term power model. The probability density function is: `f(x) = ax^{b}+c`. The corresponding parameters are key_dist_a, key_dist_b, and key_dist_c in db_bench
For the value size distribution and iterator scan length distribution, they both follow the Generalized Pareto Distribution. The probability density function is `f(x) = (1/sigma)(1+k*(x-theta)/sigma))^{-1-1/k)`. The parameters are: value_k, value_theta, value_sigma and iter_k, iter_theta, iter_sigma. For more information about the Generalized Pareto Distribution, users can find the [wiki](https://en.wikipedia.org/wiki/Generalized_Pareto_distribution) and [Matalb page](https://www.mathworks.com/help/stats/generalized-pareto-distribution.html)
As for the QPS, it follows the diurnal pattern. So Sine is a good model to fit it. `F(x) = sine_a*sin(sine_b*x + sine_c) + sine_d`. The trace_will tell you the average QPS in the print out resutls, which is sine_d. After user fit the "*-qps_stats.txt" to the Matlab model, user can get the sine_a, sine_b, and sine_c. By using the 4 parameters, user can control the QPS variation including the period, average, changes.
To use the bench mark, user can indicate the following parameters as examples:
```
-benchmarks="mixgraph" -key_dist_a=0.002312 -key_dist_b=0.3467 -value_k=0.9233 -value_sigma=226.4092 -iter_k=2.517 -iter_sigma=14.236 -mix_get_ratio=0.7 -mix_put_ratio=0.25 -mix_seek_ratio=0.05 -sine_mix_rate_interval_milliseconds=500 -sine_a=15000 -sine_b=1 -sine_d=20000
```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4788
Differential Revision: D13573940
Pulled By: sagar0
fbshipit-source-id: e184c27e07b4f1bc0b436c2be36c5090c1fb0222
2019-01-22 19:40:44 +01:00
|
|
|
} else if (query_type == 1) {
|
|
|
|
// the Put query
|
|
|
|
puts++;
|
|
|
|
int64_t value_size = ParetoCdfInversion(
|
|
|
|
u, FLAGS_value_theta, FLAGS_value_k, FLAGS_value_sigma);
|
|
|
|
if (value_size < 0) {
|
|
|
|
value_size = 10;
|
|
|
|
} else if (value_size > value_max) {
|
|
|
|
value_size = value_size % value_max;
|
|
|
|
}
|
2019-01-23 07:41:49 +01:00
|
|
|
s = db_with_cfh->db->Put(
|
|
|
|
write_options_, key,
|
|
|
|
gen.Generate(static_cast<unsigned int>(value_size)));
|
Generate mixed workload with Get, Put, Seek in db_bench (#4788)
Summary:
Based on the specific workload models (key access distribution, value size distribution, and iterator scan length distribution, the QPS variation), the MixGraph benchmark generate the synthetic workload according to these distributions which can reflect the real-world workload characteristics.
After user enable the tracing function, they will get the trace file. By analyzing the trace file with the trace_analyzer tool, user can generate a set of statistic data files including. The *_accessed_key_stats.txt, *-accessed_value_size_distribution.txt, *-iterator_length_distribution.txt, and *-qps_stats.txt are mainly used to fit the Matlab model fitting. After that, user can get the parameters of the workload distributions (the modeling details are described: [here](https://github.com/facebook/rocksdb/wiki/RocksDB-Trace%2C-Replay%2C-and-Analyzer))
The key access distribution follows the The two-term power model. The probability density function is: `f(x) = ax^{b}+c`. The corresponding parameters are key_dist_a, key_dist_b, and key_dist_c in db_bench
For the value size distribution and iterator scan length distribution, they both follow the Generalized Pareto Distribution. The probability density function is `f(x) = (1/sigma)(1+k*(x-theta)/sigma))^{-1-1/k)`. The parameters are: value_k, value_theta, value_sigma and iter_k, iter_theta, iter_sigma. For more information about the Generalized Pareto Distribution, users can find the [wiki](https://en.wikipedia.org/wiki/Generalized_Pareto_distribution) and [Matalb page](https://www.mathworks.com/help/stats/generalized-pareto-distribution.html)
As for the QPS, it follows the diurnal pattern. So Sine is a good model to fit it. `F(x) = sine_a*sin(sine_b*x + sine_c) + sine_d`. The trace_will tell you the average QPS in the print out resutls, which is sine_d. After user fit the "*-qps_stats.txt" to the Matlab model, user can get the sine_a, sine_b, and sine_c. By using the 4 parameters, user can control the QPS variation including the period, average, changes.
To use the bench mark, user can indicate the following parameters as examples:
```
-benchmarks="mixgraph" -key_dist_a=0.002312 -key_dist_b=0.3467 -value_k=0.9233 -value_sigma=226.4092 -iter_k=2.517 -iter_sigma=14.236 -mix_get_ratio=0.7 -mix_put_ratio=0.25 -mix_seek_ratio=0.05 -sine_mix_rate_interval_milliseconds=500 -sine_a=15000 -sine_b=1 -sine_d=20000
```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4788
Differential Revision: D13573940
Pulled By: sagar0
fbshipit-source-id: e184c27e07b4f1bc0b436c2be36c5090c1fb0222
2019-01-22 19:40:44 +01:00
|
|
|
if (!s.ok()) {
|
|
|
|
fprintf(stderr, "put error: %s\n", s.ToString().c_str());
|
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (thread->shared->write_rate_limiter) {
|
|
|
|
thread->shared->write_rate_limiter->Request(
|
|
|
|
key.size() + value_size, Env::IO_HIGH, nullptr /*stats*/,
|
|
|
|
RateLimiter::OpType::kWrite);
|
|
|
|
}
|
2019-03-08 23:07:19 +01:00
|
|
|
thread->stats.FinishedOps(db_with_cfh, db_with_cfh->db, 1, kWrite);
|
Generate mixed workload with Get, Put, Seek in db_bench (#4788)
Summary:
Based on the specific workload models (key access distribution, value size distribution, and iterator scan length distribution, the QPS variation), the MixGraph benchmark generate the synthetic workload according to these distributions which can reflect the real-world workload characteristics.
After user enable the tracing function, they will get the trace file. By analyzing the trace file with the trace_analyzer tool, user can generate a set of statistic data files including. The *_accessed_key_stats.txt, *-accessed_value_size_distribution.txt, *-iterator_length_distribution.txt, and *-qps_stats.txt are mainly used to fit the Matlab model fitting. After that, user can get the parameters of the workload distributions (the modeling details are described: [here](https://github.com/facebook/rocksdb/wiki/RocksDB-Trace%2C-Replay%2C-and-Analyzer))
The key access distribution follows the The two-term power model. The probability density function is: `f(x) = ax^{b}+c`. The corresponding parameters are key_dist_a, key_dist_b, and key_dist_c in db_bench
For the value size distribution and iterator scan length distribution, they both follow the Generalized Pareto Distribution. The probability density function is `f(x) = (1/sigma)(1+k*(x-theta)/sigma))^{-1-1/k)`. The parameters are: value_k, value_theta, value_sigma and iter_k, iter_theta, iter_sigma. For more information about the Generalized Pareto Distribution, users can find the [wiki](https://en.wikipedia.org/wiki/Generalized_Pareto_distribution) and [Matalb page](https://www.mathworks.com/help/stats/generalized-pareto-distribution.html)
As for the QPS, it follows the diurnal pattern. So Sine is a good model to fit it. `F(x) = sine_a*sin(sine_b*x + sine_c) + sine_d`. The trace_will tell you the average QPS in the print out resutls, which is sine_d. After user fit the "*-qps_stats.txt" to the Matlab model, user can get the sine_a, sine_b, and sine_c. By using the 4 parameters, user can control the QPS variation including the period, average, changes.
To use the bench mark, user can indicate the following parameters as examples:
```
-benchmarks="mixgraph" -key_dist_a=0.002312 -key_dist_b=0.3467 -value_k=0.9233 -value_sigma=226.4092 -iter_k=2.517 -iter_sigma=14.236 -mix_get_ratio=0.7 -mix_put_ratio=0.25 -mix_seek_ratio=0.05 -sine_mix_rate_interval_milliseconds=500 -sine_a=15000 -sine_b=1 -sine_d=20000
```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4788
Differential Revision: D13573940
Pulled By: sagar0
fbshipit-source-id: e184c27e07b4f1bc0b436c2be36c5090c1fb0222
2019-01-22 19:40:44 +01:00
|
|
|
} else if (query_type == 2) {
|
|
|
|
// Seek query
|
|
|
|
if (db_with_cfh->db != nullptr) {
|
|
|
|
Iterator* single_iter = nullptr;
|
|
|
|
single_iter = db_with_cfh->db->NewIterator(options);
|
|
|
|
if (single_iter != nullptr) {
|
|
|
|
single_iter->Seek(key);
|
|
|
|
seek++;
|
|
|
|
read++;
|
|
|
|
if (single_iter->Valid() && single_iter->key().compare(key) == 0) {
|
|
|
|
seek_found++;
|
|
|
|
}
|
|
|
|
int64_t scan_length =
|
|
|
|
ParetoCdfInversion(u, FLAGS_iter_theta, FLAGS_iter_k,
|
|
|
|
FLAGS_iter_sigma) %
|
|
|
|
scan_len_max;
|
|
|
|
for (int64_t j = 0; j < scan_length && single_iter->Valid(); j++) {
|
|
|
|
Slice value = single_iter->value();
|
|
|
|
memcpy(value_buffer, value.data(),
|
|
|
|
std::min(value.size(), sizeof(value_buffer)));
|
|
|
|
bytes += single_iter->key().size() + single_iter->value().size();
|
|
|
|
single_iter->Next();
|
|
|
|
assert(single_iter->status().ok());
|
|
|
|
}
|
|
|
|
}
|
|
|
|
delete single_iter;
|
|
|
|
}
|
2019-03-08 23:07:19 +01:00
|
|
|
thread->stats.FinishedOps(db_with_cfh, db_with_cfh->db, 1, kSeek);
|
Generate mixed workload with Get, Put, Seek in db_bench (#4788)
Summary:
Based on the specific workload models (key access distribution, value size distribution, and iterator scan length distribution, the QPS variation), the MixGraph benchmark generate the synthetic workload according to these distributions which can reflect the real-world workload characteristics.
After user enable the tracing function, they will get the trace file. By analyzing the trace file with the trace_analyzer tool, user can generate a set of statistic data files including. The *_accessed_key_stats.txt, *-accessed_value_size_distribution.txt, *-iterator_length_distribution.txt, and *-qps_stats.txt are mainly used to fit the Matlab model fitting. After that, user can get the parameters of the workload distributions (the modeling details are described: [here](https://github.com/facebook/rocksdb/wiki/RocksDB-Trace%2C-Replay%2C-and-Analyzer))
The key access distribution follows the The two-term power model. The probability density function is: `f(x) = ax^{b}+c`. The corresponding parameters are key_dist_a, key_dist_b, and key_dist_c in db_bench
For the value size distribution and iterator scan length distribution, they both follow the Generalized Pareto Distribution. The probability density function is `f(x) = (1/sigma)(1+k*(x-theta)/sigma))^{-1-1/k)`. The parameters are: value_k, value_theta, value_sigma and iter_k, iter_theta, iter_sigma. For more information about the Generalized Pareto Distribution, users can find the [wiki](https://en.wikipedia.org/wiki/Generalized_Pareto_distribution) and [Matalb page](https://www.mathworks.com/help/stats/generalized-pareto-distribution.html)
As for the QPS, it follows the diurnal pattern. So Sine is a good model to fit it. `F(x) = sine_a*sin(sine_b*x + sine_c) + sine_d`. The trace_will tell you the average QPS in the print out resutls, which is sine_d. After user fit the "*-qps_stats.txt" to the Matlab model, user can get the sine_a, sine_b, and sine_c. By using the 4 parameters, user can control the QPS variation including the period, average, changes.
To use the bench mark, user can indicate the following parameters as examples:
```
-benchmarks="mixgraph" -key_dist_a=0.002312 -key_dist_b=0.3467 -value_k=0.9233 -value_sigma=226.4092 -iter_k=2.517 -iter_sigma=14.236 -mix_get_ratio=0.7 -mix_put_ratio=0.25 -mix_seek_ratio=0.05 -sine_mix_rate_interval_milliseconds=500 -sine_a=15000 -sine_b=1 -sine_d=20000
```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4788
Differential Revision: D13573940
Pulled By: sagar0
fbshipit-source-id: e184c27e07b4f1bc0b436c2be36c5090c1fb0222
2019-01-22 19:40:44 +01:00
|
|
|
}
|
|
|
|
}
|
2019-01-28 19:57:57 +01:00
|
|
|
char msg[256];
|
Generate mixed workload with Get, Put, Seek in db_bench (#4788)
Summary:
Based on the specific workload models (key access distribution, value size distribution, and iterator scan length distribution, the QPS variation), the MixGraph benchmark generate the synthetic workload according to these distributions which can reflect the real-world workload characteristics.
After user enable the tracing function, they will get the trace file. By analyzing the trace file with the trace_analyzer tool, user can generate a set of statistic data files including. The *_accessed_key_stats.txt, *-accessed_value_size_distribution.txt, *-iterator_length_distribution.txt, and *-qps_stats.txt are mainly used to fit the Matlab model fitting. After that, user can get the parameters of the workload distributions (the modeling details are described: [here](https://github.com/facebook/rocksdb/wiki/RocksDB-Trace%2C-Replay%2C-and-Analyzer))
The key access distribution follows the The two-term power model. The probability density function is: `f(x) = ax^{b}+c`. The corresponding parameters are key_dist_a, key_dist_b, and key_dist_c in db_bench
For the value size distribution and iterator scan length distribution, they both follow the Generalized Pareto Distribution. The probability density function is `f(x) = (1/sigma)(1+k*(x-theta)/sigma))^{-1-1/k)`. The parameters are: value_k, value_theta, value_sigma and iter_k, iter_theta, iter_sigma. For more information about the Generalized Pareto Distribution, users can find the [wiki](https://en.wikipedia.org/wiki/Generalized_Pareto_distribution) and [Matalb page](https://www.mathworks.com/help/stats/generalized-pareto-distribution.html)
As for the QPS, it follows the diurnal pattern. So Sine is a good model to fit it. `F(x) = sine_a*sin(sine_b*x + sine_c) + sine_d`. The trace_will tell you the average QPS in the print out resutls, which is sine_d. After user fit the "*-qps_stats.txt" to the Matlab model, user can get the sine_a, sine_b, and sine_c. By using the 4 parameters, user can control the QPS variation including the period, average, changes.
To use the bench mark, user can indicate the following parameters as examples:
```
-benchmarks="mixgraph" -key_dist_a=0.002312 -key_dist_b=0.3467 -value_k=0.9233 -value_sigma=226.4092 -iter_k=2.517 -iter_sigma=14.236 -mix_get_ratio=0.7 -mix_put_ratio=0.25 -mix_seek_ratio=0.05 -sine_mix_rate_interval_milliseconds=500 -sine_a=15000 -sine_b=1 -sine_d=20000
```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4788
Differential Revision: D13573940
Pulled By: sagar0
fbshipit-source-id: e184c27e07b4f1bc0b436c2be36c5090c1fb0222
2019-01-22 19:40:44 +01:00
|
|
|
snprintf(msg, sizeof(msg),
|
|
|
|
"( Gets:%" PRIu64 " Puts:%" PRIu64 " Seek:%" PRIu64 " of %" PRIu64
|
|
|
|
" in %" PRIu64 " found)\n",
|
|
|
|
gets, puts, seek, found, read);
|
|
|
|
|
|
|
|
thread->stats.AddBytes(bytes);
|
|
|
|
thread->stats.AddMessage(msg);
|
|
|
|
|
|
|
|
if (FLAGS_perf_level > rocksdb::PerfLevel::kDisable) {
|
|
|
|
thread->stats.AddMessage(std::string("PERF_CONTEXT:\n") +
|
|
|
|
get_perf_context()->ToString());
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2014-02-03 22:13:36 +01:00
|
|
|
void IteratorCreation(ThreadState* thread) {
|
|
|
|
Duration duration(FLAGS_duration, reads_);
|
|
|
|
ReadOptions options(FLAGS_verify_checksum, true);
|
|
|
|
while (!duration.Done(1)) {
|
2014-04-11 21:15:09 +02:00
|
|
|
DB* db = SelectDB(thread);
|
|
|
|
Iterator* iter = db->NewIterator(options);
|
2014-02-03 22:13:36 +01:00
|
|
|
delete iter;
|
2015-12-08 02:33:18 +01:00
|
|
|
thread->stats.FinishedOps(nullptr, db, 1, kOthers);
|
2014-02-03 22:13:36 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2014-04-10 19:15:59 +02:00
|
|
|
void IteratorCreationWhileWriting(ThreadState* thread) {
|
|
|
|
if (thread->tid > 0) {
|
|
|
|
IteratorCreation(thread);
|
|
|
|
} else {
|
2015-12-08 02:33:18 +01:00
|
|
|
BGWriter(thread, kWrite);
|
2014-04-10 19:15:59 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2012-04-17 17:36:46 +02:00
|
|
|
void SeekRandom(ThreadState* thread) {
|
2014-04-08 20:21:09 +02:00
|
|
|
int64_t read = 0;
|
2014-03-05 02:08:05 +01:00
|
|
|
int64_t found = 0;
|
2015-03-30 20:28:25 +02:00
|
|
|
int64_t bytes = 0;
|
2014-04-08 20:21:09 +02:00
|
|
|
ReadOptions options(FLAGS_verify_checksum, true);
|
|
|
|
options.tailing = FLAGS_use_tailing_iterator;
|
2014-04-11 21:15:09 +02:00
|
|
|
|
|
|
|
Iterator* single_iter = nullptr;
|
|
|
|
std::vector<Iterator*> multi_iters;
|
2014-08-19 03:15:01 +02:00
|
|
|
if (db_.db != nullptr) {
|
|
|
|
single_iter = db_.db->NewIterator(options);
|
2014-04-11 21:15:09 +02:00
|
|
|
} else {
|
2014-08-19 03:15:01 +02:00
|
|
|
for (const auto& db_with_cfh : multi_dbs_) {
|
|
|
|
multi_iters.push_back(db_with_cfh.db->NewIterator(options));
|
2014-04-11 21:15:09 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2015-02-19 23:27:48 +01:00
|
|
|
std::unique_ptr<const char[]> key_guard;
|
|
|
|
Slice key = AllocateKey(&key_guard);
|
2014-04-08 20:21:09 +02:00
|
|
|
|
2018-11-14 19:39:54 +01:00
|
|
|
std::unique_ptr<const char[]> upper_bound_key_guard;
|
|
|
|
Slice upper_bound = AllocateKey(&upper_bound_key_guard);
|
|
|
|
std::unique_ptr<const char[]> lower_bound_key_guard;
|
|
|
|
Slice lower_bound = AllocateKey(&lower_bound_key_guard);
|
|
|
|
|
2014-04-08 20:21:09 +02:00
|
|
|
Duration duration(FLAGS_duration, reads_);
|
2014-10-17 19:20:25 +02:00
|
|
|
char value_buffer[256];
|
2013-03-01 20:25:13 +01:00
|
|
|
while (!duration.Done(1)) {
|
2018-11-14 19:39:54 +01:00
|
|
|
int64_t seek_pos = thread->rand.Next() % FLAGS_num;
|
|
|
|
GenerateKeyFromInt((uint64_t)seek_pos, FLAGS_num, &key);
|
|
|
|
if (FLAGS_max_scan_distance != 0) {
|
|
|
|
if (FLAGS_reverse_iterator) {
|
|
|
|
GenerateKeyFromInt(
|
2018-12-13 23:12:02 +01:00
|
|
|
static_cast<uint64_t>(std::max(
|
|
|
|
static_cast<int64_t>(0), seek_pos - FLAGS_max_scan_distance)),
|
2018-11-14 19:39:54 +01:00
|
|
|
FLAGS_num, &lower_bound);
|
|
|
|
options.iterate_lower_bound = &lower_bound;
|
|
|
|
} else {
|
|
|
|
GenerateKeyFromInt(
|
|
|
|
(uint64_t)std::min(FLAGS_num, seek_pos + FLAGS_max_scan_distance),
|
|
|
|
FLAGS_num, &upper_bound);
|
|
|
|
options.iterate_upper_bound = &upper_bound;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2015-05-01 23:17:45 +02:00
|
|
|
if (!FLAGS_use_tailing_iterator) {
|
|
|
|
if (db_.db != nullptr) {
|
|
|
|
delete single_iter;
|
|
|
|
single_iter = db_.db->NewIterator(options);
|
|
|
|
} else {
|
|
|
|
for (auto iter : multi_iters) {
|
|
|
|
delete iter;
|
|
|
|
}
|
|
|
|
multi_iters.clear();
|
|
|
|
for (const auto& db_with_cfh : multi_dbs_) {
|
|
|
|
multi_iters.push_back(db_with_cfh.db->NewIterator(options));
|
2014-05-30 23:09:22 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2014-04-11 21:15:09 +02:00
|
|
|
// Pick a Iterator to use
|
|
|
|
Iterator* iter_to_use = single_iter;
|
|
|
|
if (single_iter == nullptr) {
|
|
|
|
iter_to_use = multi_iters[thread->rand.Next() % multi_iters.size()];
|
|
|
|
}
|
|
|
|
|
|
|
|
iter_to_use->Seek(key);
|
2014-04-08 20:21:09 +02:00
|
|
|
read++;
|
2014-04-11 21:15:09 +02:00
|
|
|
if (iter_to_use->Valid() && iter_to_use->key().compare(key) == 0) {
|
2014-04-08 20:21:09 +02:00
|
|
|
found++;
|
|
|
|
}
|
2014-10-17 19:20:25 +02:00
|
|
|
|
|
|
|
for (int j = 0; j < FLAGS_seek_nexts && iter_to_use->Valid(); ++j) {
|
|
|
|
// Copy out iterator's value to make sure we read them.
|
|
|
|
Slice value = iter_to_use->value();
|
|
|
|
memcpy(value_buffer, value.data(),
|
|
|
|
std::min(value.size(), sizeof(value_buffer)));
|
2015-03-30 20:28:25 +02:00
|
|
|
bytes += iter_to_use->key().size() + iter_to_use->value().size();
|
2015-03-18 21:50:52 +01:00
|
|
|
|
|
|
|
if (!FLAGS_reverse_iterator) {
|
|
|
|
iter_to_use->Next();
|
|
|
|
} else {
|
|
|
|
iter_to_use->Prev();
|
|
|
|
}
|
2014-10-17 19:20:25 +02:00
|
|
|
assert(iter_to_use->status().ok());
|
|
|
|
}
|
|
|
|
|
2016-12-19 21:25:35 +01:00
|
|
|
if (thread->shared->read_rate_limiter.get() != nullptr &&
|
|
|
|
read % 256 == 255) {
|
2017-06-13 23:51:22 +02:00
|
|
|
thread->shared->read_rate_limiter->Request(
|
|
|
|
256, Env::IO_HIGH, nullptr /* stats */, RateLimiter::OpType::kRead);
|
2016-12-19 21:25:35 +01:00
|
|
|
}
|
|
|
|
|
2015-12-08 02:33:18 +01:00
|
|
|
thread->stats.FinishedOps(&db_, db_.db, 1, kSeek);
|
2012-04-17 17:36:46 +02:00
|
|
|
}
|
2014-04-11 21:15:09 +02:00
|
|
|
delete single_iter;
|
|
|
|
for (auto iter : multi_iters) {
|
|
|
|
delete iter;
|
|
|
|
}
|
2014-04-08 20:21:09 +02:00
|
|
|
|
2012-04-17 17:36:46 +02:00
|
|
|
char msg[100];
|
2014-05-30 23:31:55 +02:00
|
|
|
snprintf(msg, sizeof(msg), "(%" PRIu64 " of %" PRIu64 " found)\n",
|
2014-04-08 20:21:09 +02:00
|
|
|
found, read);
|
2015-03-30 20:28:25 +02:00
|
|
|
thread->stats.AddBytes(bytes);
|
2012-04-17 17:36:46 +02:00
|
|
|
thread->stats.AddMessage(msg);
|
2016-06-30 03:44:22 +02:00
|
|
|
if (FLAGS_perf_level > rocksdb::PerfLevel::kDisable) {
|
2018-07-20 23:34:07 +02:00
|
|
|
thread->stats.AddMessage(std::string("PERF_CONTEXT:\n") +
|
|
|
|
get_perf_context()->ToString());
|
2014-05-30 23:31:55 +02:00
|
|
|
}
|
2012-04-17 17:36:46 +02:00
|
|
|
}
|
2014-04-11 18:47:20 +02:00
|
|
|
|
|
|
|
void SeekRandomWhileWriting(ThreadState* thread) {
|
|
|
|
if (thread->tid > 0) {
|
|
|
|
SeekRandom(thread);
|
|
|
|
} else {
|
2015-12-08 02:33:18 +01:00
|
|
|
BGWriter(thread, kWrite);
|
2014-04-11 18:47:20 +02:00
|
|
|
}
|
|
|
|
}
|
2012-04-17 17:36:46 +02:00
|
|
|
|
2015-03-30 20:28:25 +02:00
|
|
|
void SeekRandomWhileMerging(ThreadState* thread) {
|
|
|
|
if (thread->tid > 0) {
|
|
|
|
SeekRandom(thread);
|
|
|
|
} else {
|
|
|
|
BGWriter(thread, kMerge);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2012-04-17 17:36:46 +02:00
|
|
|
void DoDelete(ThreadState* thread, bool seq) {
|
|
|
|
WriteBatch batch;
|
2016-06-28 23:10:24 +02:00
|
|
|
Duration duration(seq ? 0 : FLAGS_duration, deletes_);
|
2014-04-08 20:21:09 +02:00
|
|
|
int64_t i = 0;
|
2015-02-19 23:27:48 +01:00
|
|
|
std::unique_ptr<const char[]> key_guard;
|
|
|
|
Slice key = AllocateKey(&key_guard);
|
2014-04-08 20:21:09 +02:00
|
|
|
|
2013-03-01 20:25:13 +01:00
|
|
|
while (!duration.Done(entries_per_batch_)) {
|
2014-04-11 21:15:09 +02:00
|
|
|
DB* db = SelectDB(thread);
|
2012-04-17 17:36:46 +02:00
|
|
|
batch.Clear();
|
2014-04-08 20:21:09 +02:00
|
|
|
for (int64_t j = 0; j < entries_per_batch_; ++j) {
|
|
|
|
const int64_t k = seq ? i + j : (thread->rand.Next() % FLAGS_num);
|
|
|
|
GenerateKeyFromInt(k, FLAGS_num, &key);
|
2014-03-05 02:08:05 +01:00
|
|
|
batch.Delete(key);
|
2012-04-17 17:36:46 +02:00
|
|
|
}
|
2014-04-11 21:15:09 +02:00
|
|
|
auto s = db->Write(write_options_, &batch);
|
2015-12-08 02:33:18 +01:00
|
|
|
thread->stats.FinishedOps(nullptr, db, entries_per_batch_, kDelete);
|
2012-04-17 17:36:46 +02:00
|
|
|
if (!s.ok()) {
|
|
|
|
fprintf(stderr, "del error: %s\n", s.ToString().c_str());
|
|
|
|
exit(1);
|
|
|
|
}
|
2014-04-08 20:21:09 +02:00
|
|
|
i += entries_per_batch_;
|
2012-04-17 17:36:46 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void DeleteSeq(ThreadState* thread) {
|
|
|
|
DoDelete(thread, true);
|
|
|
|
}
|
|
|
|
|
|
|
|
void DeleteRandom(ThreadState* thread) {
|
|
|
|
DoDelete(thread, false);
|
|
|
|
}
|
|
|
|
|
2011-09-01 21:08:02 +02:00
|
|
|
void ReadWhileWriting(ThreadState* thread) {
|
|
|
|
if (thread->tid > 0) {
|
|
|
|
ReadRandom(thread);
|
|
|
|
} else {
|
2015-12-08 02:33:18 +01:00
|
|
|
BGWriter(thread, kWrite);
|
2014-04-10 19:15:59 +02:00
|
|
|
}
|
|
|
|
}
|
2011-09-01 21:08:02 +02:00
|
|
|
|
2015-03-18 21:50:52 +01:00
|
|
|
void ReadWhileMerging(ThreadState* thread) {
|
|
|
|
if (thread->tid > 0) {
|
|
|
|
ReadRandom(thread);
|
|
|
|
} else {
|
|
|
|
BGWriter(thread, kMerge);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2015-12-08 02:33:18 +01:00
|
|
|
void BGWriter(ThreadState* thread, enum OperationType write_merge) {
|
2014-04-10 19:15:59 +02:00
|
|
|
// Special thread that keeps writing until other threads are done.
|
|
|
|
RandomGenerator gen;
|
2015-03-30 20:28:25 +02:00
|
|
|
int64_t bytes = 0;
|
2014-04-10 19:15:59 +02:00
|
|
|
|
2016-01-04 21:01:27 +01:00
|
|
|
std::unique_ptr<RateLimiter> write_rate_limiter;
|
|
|
|
if (FLAGS_benchmark_write_rate_limit > 0) {
|
|
|
|
write_rate_limiter.reset(
|
|
|
|
NewGenericRateLimiter(FLAGS_benchmark_write_rate_limit));
|
|
|
|
}
|
2014-04-10 19:15:59 +02:00
|
|
|
|
|
|
|
// Don't merge stats from this thread with the readers.
|
|
|
|
thread->stats.SetExcludeFromMerge();
|
|
|
|
|
2015-02-19 23:27:48 +01:00
|
|
|
std::unique_ptr<const char[]> key_guard;
|
|
|
|
Slice key = AllocateKey(&key_guard);
|
2017-04-07 20:25:42 +02:00
|
|
|
uint32_t written = 0;
|
|
|
|
bool hint_printed = false;
|
2014-04-10 19:15:59 +02:00
|
|
|
|
|
|
|
while (true) {
|
2014-04-11 21:15:09 +02:00
|
|
|
DB* db = SelectDB(thread);
|
2014-04-10 19:15:59 +02:00
|
|
|
{
|
|
|
|
MutexLock l(&thread->shared->mu);
|
2017-04-07 20:25:42 +02:00
|
|
|
if (FLAGS_finish_after_writes && written == writes_) {
|
|
|
|
fprintf(stderr, "Exiting the writer after %u writes...\n", written);
|
|
|
|
break;
|
|
|
|
}
|
2014-04-10 19:15:59 +02:00
|
|
|
if (thread->shared->num_done + 1 >= thread->shared->num_initialized) {
|
|
|
|
// Other threads have finished
|
2017-04-07 20:25:42 +02:00
|
|
|
if (FLAGS_finish_after_writes) {
|
|
|
|
// Wait for the writes to be finished
|
|
|
|
if (!hint_printed) {
|
|
|
|
fprintf(stderr, "Reads are finished. Have %d more writes to do\n",
|
|
|
|
(int)writes_ - written);
|
|
|
|
hint_printed = true;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
// Finish the write immediately
|
|
|
|
break;
|
|
|
|
}
|
2011-09-01 21:08:02 +02:00
|
|
|
}
|
2014-04-10 19:15:59 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
GenerateKeyFromInt(thread->rand.Next() % FLAGS_num, FLAGS_num, &key);
|
2015-03-18 21:50:52 +01:00
|
|
|
Status s;
|
|
|
|
|
2015-12-08 02:33:18 +01:00
|
|
|
if (write_merge == kWrite) {
|
2016-08-01 21:28:51 +02:00
|
|
|
s = db->Put(write_options_, key, gen.Generate(value_size_));
|
2015-03-18 21:50:52 +01:00
|
|
|
} else {
|
2016-08-01 21:28:51 +02:00
|
|
|
s = db->Merge(write_options_, key, gen.Generate(value_size_));
|
2015-03-18 21:50:52 +01:00
|
|
|
}
|
2017-04-07 20:25:42 +02:00
|
|
|
written++;
|
2015-03-18 21:50:52 +01:00
|
|
|
|
2014-04-10 19:15:59 +02:00
|
|
|
if (!s.ok()) {
|
2015-03-18 21:50:52 +01:00
|
|
|
fprintf(stderr, "put or merge error: %s\n", s.ToString().c_str());
|
2014-04-10 19:15:59 +02:00
|
|
|
exit(1);
|
|
|
|
}
|
2015-03-30 20:28:25 +02:00
|
|
|
bytes += key.size() + value_size_;
|
2015-12-08 02:33:18 +01:00
|
|
|
thread->stats.FinishedOps(&db_, db_.db, 1, kWrite);
|
2013-04-17 19:56:39 +02:00
|
|
|
|
2016-01-04 21:01:27 +01:00
|
|
|
if (FLAGS_benchmark_write_rate_limit > 0) {
|
|
|
|
write_rate_limiter->Request(
|
2017-03-17 01:45:28 +01:00
|
|
|
entries_per_batch_ * (value_size_ + key_size_), Env::IO_HIGH,
|
2017-06-13 23:51:22 +02:00
|
|
|
nullptr /* stats */, RateLimiter::OpType::kWrite);
|
2011-09-01 21:08:02 +02:00
|
|
|
}
|
|
|
|
}
|
2015-03-30 20:28:25 +02:00
|
|
|
thread->stats.AddBytes(bytes);
|
2011-09-01 21:08:02 +02:00
|
|
|
}
|
|
|
|
|
2018-05-25 00:45:49 +02:00
|
|
|
void ReadWhileScanning(ThreadState* thread) {
|
|
|
|
if (thread->tid > 0) {
|
|
|
|
ReadRandom(thread);
|
|
|
|
} else {
|
|
|
|
BGScan(thread);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void BGScan(ThreadState* thread) {
|
|
|
|
if (FLAGS_num_multi_db > 0) {
|
|
|
|
fprintf(stderr, "Not supporting multiple DBs.\n");
|
|
|
|
abort();
|
|
|
|
}
|
|
|
|
assert(db_.db != nullptr);
|
|
|
|
ReadOptions read_options;
|
|
|
|
Iterator* iter = db_.db->NewIterator(read_options);
|
|
|
|
|
2018-05-31 02:45:02 +02:00
|
|
|
fprintf(stderr, "num reads to do %" PRIu64 "\n", reads_);
|
2018-05-25 00:45:49 +02:00
|
|
|
Duration duration(FLAGS_duration, reads_);
|
|
|
|
uint64_t num_seek_to_first = 0;
|
|
|
|
uint64_t num_next = 0;
|
|
|
|
while (!duration.Done(1)) {
|
|
|
|
if (!iter->Valid()) {
|
|
|
|
iter->SeekToFirst();
|
|
|
|
num_seek_to_first++;
|
|
|
|
} else if (!iter->status().ok()) {
|
|
|
|
fprintf(stderr, "Iterator error: %s\n",
|
|
|
|
iter->status().ToString().c_str());
|
|
|
|
abort();
|
|
|
|
} else {
|
|
|
|
iter->Next();
|
|
|
|
num_next++;
|
|
|
|
}
|
|
|
|
|
|
|
|
thread->stats.FinishedOps(&db_, db_.db, 1, kSeek);
|
|
|
|
}
|
|
|
|
delete iter;
|
|
|
|
}
|
|
|
|
|
2013-02-20 02:51:06 +01:00
|
|
|
// Given a key K and value V, this puts (K+"0", V), (K+"1", V), (K+"2", V)
|
2013-06-12 21:42:21 +02:00
|
|
|
// in DB atomically i.e in a single batch. Also refer GetMany.
|
2014-04-11 21:15:09 +02:00
|
|
|
Status PutMany(DB* db, const WriteOptions& writeoptions, const Slice& key,
|
|
|
|
const Slice& value) {
|
2013-02-20 02:51:06 +01:00
|
|
|
std::string suffixes[3] = {"2", "1", "0"};
|
|
|
|
std::string keys[3];
|
|
|
|
|
|
|
|
WriteBatch batch;
|
|
|
|
Status s;
|
|
|
|
for (int i = 0; i < 3; i++) {
|
|
|
|
keys[i] = key.ToString() + suffixes[i];
|
|
|
|
batch.Put(keys[i], value);
|
|
|
|
}
|
|
|
|
|
2014-04-11 21:15:09 +02:00
|
|
|
s = db->Write(writeoptions, &batch);
|
2013-02-20 02:51:06 +01:00
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
// Given a key K, this deletes (K+"0", V), (K+"1", V), (K+"2", V)
|
2013-06-12 21:42:21 +02:00
|
|
|
// in DB atomically i.e in a single batch. Also refer GetMany.
|
2014-04-11 21:15:09 +02:00
|
|
|
Status DeleteMany(DB* db, const WriteOptions& writeoptions,
|
|
|
|
const Slice& key) {
|
2013-02-20 02:51:06 +01:00
|
|
|
std::string suffixes[3] = {"1", "2", "0"};
|
|
|
|
std::string keys[3];
|
|
|
|
|
|
|
|
WriteBatch batch;
|
|
|
|
Status s;
|
|
|
|
for (int i = 0; i < 3; i++) {
|
|
|
|
keys[i] = key.ToString() + suffixes[i];
|
|
|
|
batch.Delete(keys[i]);
|
|
|
|
}
|
|
|
|
|
2014-04-11 21:15:09 +02:00
|
|
|
s = db->Write(writeoptions, &batch);
|
2013-02-20 02:51:06 +01:00
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Given a key K and value V, this gets values for K+"0", K+"1" and K+"2"
|
|
|
|
// in the same snapshot, and verifies that all the values are identical.
|
2013-06-12 21:42:21 +02:00
|
|
|
// ASSUMES that PutMany was used to put (K, V) into the DB.
|
2014-04-11 21:15:09 +02:00
|
|
|
Status GetMany(DB* db, const ReadOptions& readoptions, const Slice& key,
|
|
|
|
std::string* value) {
|
2013-02-20 02:51:06 +01:00
|
|
|
std::string suffixes[3] = {"0", "1", "2"};
|
|
|
|
std::string keys[3];
|
|
|
|
Slice key_slices[3];
|
|
|
|
std::string values[3];
|
|
|
|
ReadOptions readoptionscopy = readoptions;
|
2014-04-11 21:15:09 +02:00
|
|
|
readoptionscopy.snapshot = db->GetSnapshot();
|
2013-02-20 02:51:06 +01:00
|
|
|
Status s;
|
|
|
|
for (int i = 0; i < 3; i++) {
|
|
|
|
keys[i] = key.ToString() + suffixes[i];
|
|
|
|
key_slices[i] = keys[i];
|
2014-04-11 21:15:09 +02:00
|
|
|
s = db->Get(readoptionscopy, key_slices[i], value);
|
2013-02-20 02:51:06 +01:00
|
|
|
if (!s.ok() && !s.IsNotFound()) {
|
|
|
|
fprintf(stderr, "get error: %s\n", s.ToString().c_str());
|
|
|
|
values[i] = "";
|
|
|
|
// we continue after error rather than exiting so that we can
|
|
|
|
// find more errors if any
|
|
|
|
} else if (s.IsNotFound()) {
|
|
|
|
values[i] = "";
|
|
|
|
} else {
|
|
|
|
values[i] = *value;
|
|
|
|
}
|
|
|
|
}
|
2014-04-11 21:15:09 +02:00
|
|
|
db->ReleaseSnapshot(readoptionscopy.snapshot);
|
2013-02-20 02:51:06 +01:00
|
|
|
|
|
|
|
if ((values[0] != values[1]) || (values[1] != values[2])) {
|
|
|
|
fprintf(stderr, "inconsistent values for key %s: %s, %s, %s\n",
|
|
|
|
key.ToString().c_str(), values[0].c_str(), values[1].c_str(),
|
|
|
|
values[2].c_str());
|
|
|
|
// we continue after error rather than exiting so that we can
|
|
|
|
// find more errors if any
|
|
|
|
}
|
|
|
|
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Differs from readrandomwriterandom in the following ways:
|
2013-06-12 21:42:21 +02:00
|
|
|
// (a) Uses GetMany/PutMany to read/write key values. Refer to those funcs.
|
2013-02-20 02:51:06 +01:00
|
|
|
// (b) Does deletes as well (per FLAGS_deletepercent)
|
|
|
|
// (c) In order to achieve high % of 'found' during lookups, and to do
|
|
|
|
// multiple writes (including puts and deletes) it uses upto
|
|
|
|
// FLAGS_numdistinct distinct keys instead of FLAGS_num distinct keys.
|
2013-06-12 21:42:21 +02:00
|
|
|
// (d) Does not have a MultiGet option.
|
2013-02-20 02:51:06 +01:00
|
|
|
void RandomWithVerify(ThreadState* thread) {
|
|
|
|
ReadOptions options(FLAGS_verify_checksum, true);
|
|
|
|
RandomGenerator gen;
|
|
|
|
std::string value;
|
2014-03-05 02:08:05 +01:00
|
|
|
int64_t found = 0;
|
2013-02-20 02:51:06 +01:00
|
|
|
int get_weight = 0;
|
|
|
|
int put_weight = 0;
|
|
|
|
int delete_weight = 0;
|
2014-03-05 02:08:05 +01:00
|
|
|
int64_t gets_done = 0;
|
|
|
|
int64_t puts_done = 0;
|
|
|
|
int64_t deletes_done = 0;
|
2013-06-12 21:42:21 +02:00
|
|
|
|
2015-02-19 23:27:48 +01:00
|
|
|
std::unique_ptr<const char[]> key_guard;
|
|
|
|
Slice key = AllocateKey(&key_guard);
|
2014-04-08 20:21:09 +02:00
|
|
|
|
2013-02-20 02:51:06 +01:00
|
|
|
// the number of iterations is the larger of read_ or write_
|
2014-03-05 02:08:05 +01:00
|
|
|
for (int64_t i = 0; i < readwrites_; i++) {
|
2014-04-11 21:15:09 +02:00
|
|
|
DB* db = SelectDB(thread);
|
2013-02-20 02:51:06 +01:00
|
|
|
if (get_weight == 0 && put_weight == 0 && delete_weight == 0) {
|
2013-06-12 21:42:21 +02:00
|
|
|
// one batch completed, reinitialize for next batch
|
2013-02-20 02:51:06 +01:00
|
|
|
get_weight = FLAGS_readwritepercent;
|
|
|
|
delete_weight = FLAGS_deletepercent;
|
|
|
|
put_weight = 100 - get_weight - delete_weight;
|
|
|
|
}
|
2014-04-08 20:21:09 +02:00
|
|
|
GenerateKeyFromInt(thread->rand.Next() % FLAGS_numdistinct,
|
|
|
|
FLAGS_numdistinct, &key);
|
2013-02-20 02:51:06 +01:00
|
|
|
if (get_weight > 0) {
|
|
|
|
// do all the gets first
|
2014-04-11 21:15:09 +02:00
|
|
|
Status s = GetMany(db, options, key, &value);
|
2013-02-20 02:51:06 +01:00
|
|
|
if (!s.ok() && !s.IsNotFound()) {
|
2013-06-12 21:42:21 +02:00
|
|
|
fprintf(stderr, "getmany error: %s\n", s.ToString().c_str());
|
2013-02-20 02:51:06 +01:00
|
|
|
// we continue after error rather than exiting so that we can
|
|
|
|
// find more errors if any
|
|
|
|
} else if (!s.IsNotFound()) {
|
|
|
|
found++;
|
|
|
|
}
|
|
|
|
get_weight--;
|
|
|
|
gets_done++;
|
2015-12-08 02:33:18 +01:00
|
|
|
thread->stats.FinishedOps(&db_, db_.db, 1, kRead);
|
2013-02-20 02:51:06 +01:00
|
|
|
} else if (put_weight > 0) {
|
|
|
|
// then do all the corresponding number of puts
|
|
|
|
// for all the gets we have done earlier
|
2014-04-11 21:15:09 +02:00
|
|
|
Status s = PutMany(db, write_options_, key, gen.Generate(value_size_));
|
2013-02-20 02:51:06 +01:00
|
|
|
if (!s.ok()) {
|
2013-06-12 21:42:21 +02:00
|
|
|
fprintf(stderr, "putmany error: %s\n", s.ToString().c_str());
|
2013-02-20 02:51:06 +01:00
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
put_weight--;
|
|
|
|
puts_done++;
|
2015-12-08 02:33:18 +01:00
|
|
|
thread->stats.FinishedOps(&db_, db_.db, 1, kWrite);
|
2013-02-20 02:51:06 +01:00
|
|
|
} else if (delete_weight > 0) {
|
2014-04-11 21:15:09 +02:00
|
|
|
Status s = DeleteMany(db, write_options_, key);
|
2013-02-20 02:51:06 +01:00
|
|
|
if (!s.ok()) {
|
2013-06-12 21:42:21 +02:00
|
|
|
fprintf(stderr, "deletemany error: %s\n", s.ToString().c_str());
|
2013-02-20 02:51:06 +01:00
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
delete_weight--;
|
|
|
|
deletes_done++;
|
2015-12-08 02:33:18 +01:00
|
|
|
thread->stats.FinishedOps(&db_, db_.db, 1, kDelete);
|
2013-02-20 02:51:06 +01:00
|
|
|
}
|
|
|
|
}
|
2017-07-11 19:28:18 +02:00
|
|
|
char msg[128];
|
2013-08-23 07:37:13 +02:00
|
|
|
snprintf(msg, sizeof(msg),
|
2014-03-05 02:08:05 +01:00
|
|
|
"( get:%" PRIu64 " put:%" PRIu64 " del:%" PRIu64 " total:%" \
|
|
|
|
PRIu64 " found:%" PRIu64 ")",
|
2013-02-20 02:51:06 +01:00
|
|
|
gets_done, puts_done, deletes_done, readwrites_, found);
|
|
|
|
thread->stats.AddMessage(msg);
|
|
|
|
}
|
|
|
|
|
2013-08-13 22:58:02 +02:00
|
|
|
// This is different from ReadWhileWriting because it does not use
|
2012-09-26 17:06:19 +02:00
|
|
|
// an extra thread.
|
2012-08-31 08:06:17 +02:00
|
|
|
void ReadRandomWriteRandom(ThreadState* thread) {
|
|
|
|
ReadOptions options(FLAGS_verify_checksum, true);
|
|
|
|
RandomGenerator gen;
|
|
|
|
std::string value;
|
2014-03-05 02:08:05 +01:00
|
|
|
int64_t found = 0;
|
2012-08-31 08:06:17 +02:00
|
|
|
int get_weight = 0;
|
|
|
|
int put_weight = 0;
|
2014-03-05 02:08:05 +01:00
|
|
|
int64_t reads_done = 0;
|
|
|
|
int64_t writes_done = 0;
|
2013-03-01 20:25:13 +01:00
|
|
|
Duration duration(FLAGS_duration, readwrites_);
|
|
|
|
|
2015-02-19 23:27:48 +01:00
|
|
|
std::unique_ptr<const char[]> key_guard;
|
|
|
|
Slice key = AllocateKey(&key_guard);
|
2014-04-08 20:21:09 +02:00
|
|
|
|
2012-08-31 08:06:17 +02:00
|
|
|
// the number of iterations is the larger of read_ or write_
|
2013-03-01 20:25:13 +01:00
|
|
|
while (!duration.Done(1)) {
|
2014-04-11 21:15:09 +02:00
|
|
|
DB* db = SelectDB(thread);
|
2014-04-08 20:21:09 +02:00
|
|
|
GenerateKeyFromInt(thread->rand.Next() % FLAGS_num, FLAGS_num, &key);
|
2012-08-31 08:06:17 +02:00
|
|
|
if (get_weight == 0 && put_weight == 0) {
|
2013-08-13 22:58:02 +02:00
|
|
|
// one batch completed, reinitialize for next batch
|
2012-08-31 08:06:17 +02:00
|
|
|
get_weight = FLAGS_readwritepercent;
|
|
|
|
put_weight = 100 - get_weight;
|
|
|
|
}
|
|
|
|
if (get_weight > 0) {
|
|
|
|
// do all the gets first
|
2014-04-11 21:15:09 +02:00
|
|
|
Status s = db->Get(options, key, &value);
|
2013-02-20 02:51:06 +01:00
|
|
|
if (!s.ok() && !s.IsNotFound()) {
|
|
|
|
fprintf(stderr, "get error: %s\n", s.ToString().c_str());
|
|
|
|
// we continue after error rather than exiting so that we can
|
|
|
|
// find more errors if any
|
|
|
|
} else if (!s.IsNotFound()) {
|
|
|
|
found++;
|
|
|
|
}
|
2012-08-31 08:06:17 +02:00
|
|
|
get_weight--;
|
|
|
|
reads_done++;
|
2015-12-08 02:33:18 +01:00
|
|
|
thread->stats.FinishedOps(nullptr, db, 1, kRead);
|
2012-08-31 08:06:17 +02:00
|
|
|
} else if (put_weight > 0) {
|
|
|
|
// then do all the corresponding number of puts
|
|
|
|
// for all the gets we have done earlier
|
2014-04-11 21:15:09 +02:00
|
|
|
Status s = db->Put(write_options_, key, gen.Generate(value_size_));
|
2012-08-31 08:06:17 +02:00
|
|
|
if (!s.ok()) {
|
|
|
|
fprintf(stderr, "put error: %s\n", s.ToString().c_str());
|
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
put_weight--;
|
|
|
|
writes_done++;
|
2015-12-08 02:33:18 +01:00
|
|
|
thread->stats.FinishedOps(nullptr, db, 1, kWrite);
|
2012-08-31 08:06:17 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
char msg[100];
|
2014-03-05 02:08:05 +01:00
|
|
|
snprintf(msg, sizeof(msg), "( reads:%" PRIu64 " writes:%" PRIu64 \
|
|
|
|
" total:%" PRIu64 " found:%" PRIu64 ")",
|
2013-02-20 02:51:06 +01:00
|
|
|
reads_done, writes_done, readwrites_, found);
|
2012-08-31 08:06:17 +02:00
|
|
|
thread->stats.AddMessage(msg);
|
|
|
|
}
|
|
|
|
|
2013-03-01 20:25:13 +01:00
|
|
|
//
|
|
|
|
// Read-modify-write for random keys
|
|
|
|
void UpdateRandom(ThreadState* thread) {
|
|
|
|
ReadOptions options(FLAGS_verify_checksum, true);
|
|
|
|
RandomGenerator gen;
|
|
|
|
std::string value;
|
2014-03-05 02:08:05 +01:00
|
|
|
int64_t found = 0;
|
2015-03-30 20:28:25 +02:00
|
|
|
int64_t bytes = 0;
|
2013-03-01 20:25:13 +01:00
|
|
|
Duration duration(FLAGS_duration, readwrites_);
|
|
|
|
|
2015-02-19 23:27:48 +01:00
|
|
|
std::unique_ptr<const char[]> key_guard;
|
|
|
|
Slice key = AllocateKey(&key_guard);
|
2013-03-01 20:25:13 +01:00
|
|
|
// the number of iterations is the larger of read_ or write_
|
|
|
|
while (!duration.Done(1)) {
|
2014-04-11 21:15:09 +02:00
|
|
|
DB* db = SelectDB(thread);
|
2014-04-08 20:21:09 +02:00
|
|
|
GenerateKeyFromInt(thread->rand.Next() % FLAGS_num, FLAGS_num, &key);
|
2013-03-01 20:25:13 +01:00
|
|
|
|
2015-01-22 03:18:15 +01:00
|
|
|
auto status = db->Get(options, key, &value);
|
|
|
|
if (status.ok()) {
|
|
|
|
++found;
|
2015-03-30 20:28:25 +02:00
|
|
|
bytes += key.size() + value.size();
|
2015-01-22 03:18:15 +01:00
|
|
|
} else if (!status.IsNotFound()) {
|
2015-01-22 03:23:12 +01:00
|
|
|
fprintf(stderr, "Get returned an error: %s\n",
|
|
|
|
status.ToString().c_str());
|
2015-01-22 03:18:15 +01:00
|
|
|
abort();
|
2013-03-01 20:25:13 +01:00
|
|
|
}
|
|
|
|
|
2017-07-14 01:25:53 +02:00
|
|
|
if (thread->shared->write_rate_limiter) {
|
|
|
|
thread->shared->write_rate_limiter->Request(
|
|
|
|
key.size() + value_size_, Env::IO_HIGH, nullptr /*stats*/,
|
|
|
|
RateLimiter::OpType::kWrite);
|
|
|
|
}
|
|
|
|
|
2014-04-11 21:15:09 +02:00
|
|
|
Status s = db->Put(write_options_, key, gen.Generate(value_size_));
|
2013-03-01 20:25:13 +01:00
|
|
|
if (!s.ok()) {
|
|
|
|
fprintf(stderr, "put error: %s\n", s.ToString().c_str());
|
|
|
|
exit(1);
|
|
|
|
}
|
2015-03-30 20:28:25 +02:00
|
|
|
bytes += key.size() + value_size_;
|
2015-12-08 02:33:18 +01:00
|
|
|
thread->stats.FinishedOps(nullptr, db, 1, kUpdate);
|
2013-03-01 20:25:13 +01:00
|
|
|
}
|
|
|
|
char msg[100];
|
2013-08-23 07:37:13 +02:00
|
|
|
snprintf(msg, sizeof(msg),
|
2014-03-05 02:08:05 +01:00
|
|
|
"( updates:%" PRIu64 " found:%" PRIu64 ")", readwrites_, found);
|
2015-03-30 20:28:25 +02:00
|
|
|
thread->stats.AddBytes(bytes);
|
2013-03-01 20:25:13 +01:00
|
|
|
thread->stats.AddMessage(msg);
|
|
|
|
}
|
|
|
|
|
2018-03-06 19:20:52 +01:00
|
|
|
// Read-XOR-write for random keys. Xors the existing value with a randomly
|
|
|
|
// generated value, and stores the result. Assuming A in the array of bytes
|
|
|
|
// representing the existing value, we generate an array B of the same size,
|
|
|
|
// then compute C = A^B as C[i]=A[i]^B[i], and store C
|
|
|
|
void XORUpdateRandom(ThreadState* thread) {
|
|
|
|
ReadOptions options(FLAGS_verify_checksum, true);
|
|
|
|
RandomGenerator gen;
|
|
|
|
std::string existing_value;
|
|
|
|
int64_t found = 0;
|
|
|
|
Duration duration(FLAGS_duration, readwrites_);
|
|
|
|
|
|
|
|
BytesXOROperator xor_operator;
|
|
|
|
|
|
|
|
std::unique_ptr<const char[]> key_guard;
|
|
|
|
Slice key = AllocateKey(&key_guard);
|
|
|
|
// the number of iterations is the larger of read_ or write_
|
|
|
|
while (!duration.Done(1)) {
|
|
|
|
DB* db = SelectDB(thread);
|
|
|
|
GenerateKeyFromInt(thread->rand.Next() % FLAGS_num, FLAGS_num, &key);
|
|
|
|
|
|
|
|
auto status = db->Get(options, key, &existing_value);
|
|
|
|
if (status.ok()) {
|
|
|
|
++found;
|
|
|
|
} else if (!status.IsNotFound()) {
|
|
|
|
fprintf(stderr, "Get returned an error: %s\n",
|
|
|
|
status.ToString().c_str());
|
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
|
|
|
|
Slice value = gen.Generate(value_size_);
|
|
|
|
std::string new_value;
|
|
|
|
|
|
|
|
if (status.ok()) {
|
|
|
|
Slice existing_value_slice = Slice(existing_value);
|
|
|
|
xor_operator.XOR(&existing_value_slice, value, &new_value);
|
|
|
|
} else {
|
|
|
|
xor_operator.XOR(nullptr, value, &new_value);
|
|
|
|
}
|
|
|
|
|
|
|
|
Status s = db->Put(write_options_, key, Slice(new_value));
|
|
|
|
if (!s.ok()) {
|
|
|
|
fprintf(stderr, "put error: %s\n", s.ToString().c_str());
|
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
thread->stats.FinishedOps(nullptr, db, 1);
|
|
|
|
}
|
|
|
|
char msg[100];
|
|
|
|
snprintf(msg, sizeof(msg),
|
|
|
|
"( updates:%" PRIu64 " found:%" PRIu64 ")", readwrites_, found);
|
|
|
|
thread->stats.AddMessage(msg);
|
|
|
|
}
|
|
|
|
|
Benchmarking for Merge Operator
Summary:
Updated db_bench and utilities/merge_operators.h to allow for dynamic benchmarking
of merge operators in db_bench. Added a new test (--benchmarks=mergerandom), which performs
a bunch of random Merge() operations over random keys. Also added a "--merge_operator=" flag
so that the tester can easily benchmark different merge operators. Currently supports
the PutOperator and UInt64Add operator. Support for stringappend or list append may come later.
Test Plan:
1. make db_bench
2. Test the PutOperator (simulating Put) as follows:
./db_bench --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom --merge_operator=put
--threads=2
3. Test the UInt64AddOperator (simulating numeric addition) similarly:
./db_bench --value_size=8 --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom
--merge_operator=uint64add --threads=2
Reviewers: haobo, dhruba, zshao, MarkCallaghan
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11535
2013-08-16 02:13:07 +02:00
|
|
|
// Read-modify-write for random keys.
|
|
|
|
// Each operation causes the key grow by value_size (simulating an append).
|
|
|
|
// Generally used for benchmarking against merges of similar type
|
|
|
|
void AppendRandom(ThreadState* thread) {
|
|
|
|
ReadOptions options(FLAGS_verify_checksum, true);
|
|
|
|
RandomGenerator gen;
|
|
|
|
std::string value;
|
2014-03-05 02:08:05 +01:00
|
|
|
int64_t found = 0;
|
2015-03-30 20:28:25 +02:00
|
|
|
int64_t bytes = 0;
|
Benchmarking for Merge Operator
Summary:
Updated db_bench and utilities/merge_operators.h to allow for dynamic benchmarking
of merge operators in db_bench. Added a new test (--benchmarks=mergerandom), which performs
a bunch of random Merge() operations over random keys. Also added a "--merge_operator=" flag
so that the tester can easily benchmark different merge operators. Currently supports
the PutOperator and UInt64Add operator. Support for stringappend or list append may come later.
Test Plan:
1. make db_bench
2. Test the PutOperator (simulating Put) as follows:
./db_bench --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom --merge_operator=put
--threads=2
3. Test the UInt64AddOperator (simulating numeric addition) similarly:
./db_bench --value_size=8 --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom
--merge_operator=uint64add --threads=2
Reviewers: haobo, dhruba, zshao, MarkCallaghan
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11535
2013-08-16 02:13:07 +02:00
|
|
|
|
2015-02-19 23:27:48 +01:00
|
|
|
std::unique_ptr<const char[]> key_guard;
|
|
|
|
Slice key = AllocateKey(&key_guard);
|
Benchmarking for Merge Operator
Summary:
Updated db_bench and utilities/merge_operators.h to allow for dynamic benchmarking
of merge operators in db_bench. Added a new test (--benchmarks=mergerandom), which performs
a bunch of random Merge() operations over random keys. Also added a "--merge_operator=" flag
so that the tester can easily benchmark different merge operators. Currently supports
the PutOperator and UInt64Add operator. Support for stringappend or list append may come later.
Test Plan:
1. make db_bench
2. Test the PutOperator (simulating Put) as follows:
./db_bench --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom --merge_operator=put
--threads=2
3. Test the UInt64AddOperator (simulating numeric addition) similarly:
./db_bench --value_size=8 --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom
--merge_operator=uint64add --threads=2
Reviewers: haobo, dhruba, zshao, MarkCallaghan
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11535
2013-08-16 02:13:07 +02:00
|
|
|
// The number of iterations is the larger of read_ or write_
|
|
|
|
Duration duration(FLAGS_duration, readwrites_);
|
|
|
|
while (!duration.Done(1)) {
|
2014-04-11 21:15:09 +02:00
|
|
|
DB* db = SelectDB(thread);
|
2014-04-08 20:21:09 +02:00
|
|
|
GenerateKeyFromInt(thread->rand.Next() % FLAGS_num, FLAGS_num, &key);
|
Benchmarking for Merge Operator
Summary:
Updated db_bench and utilities/merge_operators.h to allow for dynamic benchmarking
of merge operators in db_bench. Added a new test (--benchmarks=mergerandom), which performs
a bunch of random Merge() operations over random keys. Also added a "--merge_operator=" flag
so that the tester can easily benchmark different merge operators. Currently supports
the PutOperator and UInt64Add operator. Support for stringappend or list append may come later.
Test Plan:
1. make db_bench
2. Test the PutOperator (simulating Put) as follows:
./db_bench --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom --merge_operator=put
--threads=2
3. Test the UInt64AddOperator (simulating numeric addition) similarly:
./db_bench --value_size=8 --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom
--merge_operator=uint64add --threads=2
Reviewers: haobo, dhruba, zshao, MarkCallaghan
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11535
2013-08-16 02:13:07 +02:00
|
|
|
|
2015-01-22 03:18:15 +01:00
|
|
|
auto status = db->Get(options, key, &value);
|
|
|
|
if (status.ok()) {
|
|
|
|
++found;
|
2015-03-30 20:28:25 +02:00
|
|
|
bytes += key.size() + value.size();
|
2015-01-22 03:18:15 +01:00
|
|
|
} else if (!status.IsNotFound()) {
|
2015-01-22 03:23:12 +01:00
|
|
|
fprintf(stderr, "Get returned an error: %s\n",
|
|
|
|
status.ToString().c_str());
|
2015-01-22 03:18:15 +01:00
|
|
|
abort();
|
Benchmarking for Merge Operator
Summary:
Updated db_bench and utilities/merge_operators.h to allow for dynamic benchmarking
of merge operators in db_bench. Added a new test (--benchmarks=mergerandom), which performs
a bunch of random Merge() operations over random keys. Also added a "--merge_operator=" flag
so that the tester can easily benchmark different merge operators. Currently supports
the PutOperator and UInt64Add operator. Support for stringappend or list append may come later.
Test Plan:
1. make db_bench
2. Test the PutOperator (simulating Put) as follows:
./db_bench --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom --merge_operator=put
--threads=2
3. Test the UInt64AddOperator (simulating numeric addition) similarly:
./db_bench --value_size=8 --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom
--merge_operator=uint64add --threads=2
Reviewers: haobo, dhruba, zshao, MarkCallaghan
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11535
2013-08-16 02:13:07 +02:00
|
|
|
} else {
|
|
|
|
// If not existing, then just assume an empty string of data
|
|
|
|
value.clear();
|
|
|
|
}
|
|
|
|
|
|
|
|
// Update the value (by appending data)
|
|
|
|
Slice operand = gen.Generate(value_size_);
|
|
|
|
if (value.size() > 0) {
|
2015-08-11 20:46:15 +02:00
|
|
|
// Use a delimiter to match the semantics for StringAppendOperator
|
Benchmarking for Merge Operator
Summary:
Updated db_bench and utilities/merge_operators.h to allow for dynamic benchmarking
of merge operators in db_bench. Added a new test (--benchmarks=mergerandom), which performs
a bunch of random Merge() operations over random keys. Also added a "--merge_operator=" flag
so that the tester can easily benchmark different merge operators. Currently supports
the PutOperator and UInt64Add operator. Support for stringappend or list append may come later.
Test Plan:
1. make db_bench
2. Test the PutOperator (simulating Put) as follows:
./db_bench --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom --merge_operator=put
--threads=2
3. Test the UInt64AddOperator (simulating numeric addition) similarly:
./db_bench --value_size=8 --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom
--merge_operator=uint64add --threads=2
Reviewers: haobo, dhruba, zshao, MarkCallaghan
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11535
2013-08-16 02:13:07 +02:00
|
|
|
value.append(1,',');
|
|
|
|
}
|
|
|
|
value.append(operand.data(), operand.size());
|
|
|
|
|
|
|
|
// Write back to the database
|
2014-04-11 21:15:09 +02:00
|
|
|
Status s = db->Put(write_options_, key, value);
|
Benchmarking for Merge Operator
Summary:
Updated db_bench and utilities/merge_operators.h to allow for dynamic benchmarking
of merge operators in db_bench. Added a new test (--benchmarks=mergerandom), which performs
a bunch of random Merge() operations over random keys. Also added a "--merge_operator=" flag
so that the tester can easily benchmark different merge operators. Currently supports
the PutOperator and UInt64Add operator. Support for stringappend or list append may come later.
Test Plan:
1. make db_bench
2. Test the PutOperator (simulating Put) as follows:
./db_bench --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom --merge_operator=put
--threads=2
3. Test the UInt64AddOperator (simulating numeric addition) similarly:
./db_bench --value_size=8 --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom
--merge_operator=uint64add --threads=2
Reviewers: haobo, dhruba, zshao, MarkCallaghan
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11535
2013-08-16 02:13:07 +02:00
|
|
|
if (!s.ok()) {
|
|
|
|
fprintf(stderr, "put error: %s\n", s.ToString().c_str());
|
|
|
|
exit(1);
|
|
|
|
}
|
2015-03-30 20:28:25 +02:00
|
|
|
bytes += key.size() + value.size();
|
2015-12-08 02:33:18 +01:00
|
|
|
thread->stats.FinishedOps(nullptr, db, 1, kUpdate);
|
Benchmarking for Merge Operator
Summary:
Updated db_bench and utilities/merge_operators.h to allow for dynamic benchmarking
of merge operators in db_bench. Added a new test (--benchmarks=mergerandom), which performs
a bunch of random Merge() operations over random keys. Also added a "--merge_operator=" flag
so that the tester can easily benchmark different merge operators. Currently supports
the PutOperator and UInt64Add operator. Support for stringappend or list append may come later.
Test Plan:
1. make db_bench
2. Test the PutOperator (simulating Put) as follows:
./db_bench --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom --merge_operator=put
--threads=2
3. Test the UInt64AddOperator (simulating numeric addition) similarly:
./db_bench --value_size=8 --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom
--merge_operator=uint64add --threads=2
Reviewers: haobo, dhruba, zshao, MarkCallaghan
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11535
2013-08-16 02:13:07 +02:00
|
|
|
}
|
2014-04-08 20:21:09 +02:00
|
|
|
|
Benchmarking for Merge Operator
Summary:
Updated db_bench and utilities/merge_operators.h to allow for dynamic benchmarking
of merge operators in db_bench. Added a new test (--benchmarks=mergerandom), which performs
a bunch of random Merge() operations over random keys. Also added a "--merge_operator=" flag
so that the tester can easily benchmark different merge operators. Currently supports
the PutOperator and UInt64Add operator. Support for stringappend or list append may come later.
Test Plan:
1. make db_bench
2. Test the PutOperator (simulating Put) as follows:
./db_bench --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom --merge_operator=put
--threads=2
3. Test the UInt64AddOperator (simulating numeric addition) similarly:
./db_bench --value_size=8 --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom
--merge_operator=uint64add --threads=2
Reviewers: haobo, dhruba, zshao, MarkCallaghan
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11535
2013-08-16 02:13:07 +02:00
|
|
|
char msg[100];
|
2014-03-05 02:08:05 +01:00
|
|
|
snprintf(msg, sizeof(msg), "( updates:%" PRIu64 " found:%" PRIu64 ")",
|
|
|
|
readwrites_, found);
|
2015-03-30 20:28:25 +02:00
|
|
|
thread->stats.AddBytes(bytes);
|
Benchmarking for Merge Operator
Summary:
Updated db_bench and utilities/merge_operators.h to allow for dynamic benchmarking
of merge operators in db_bench. Added a new test (--benchmarks=mergerandom), which performs
a bunch of random Merge() operations over random keys. Also added a "--merge_operator=" flag
so that the tester can easily benchmark different merge operators. Currently supports
the PutOperator and UInt64Add operator. Support for stringappend or list append may come later.
Test Plan:
1. make db_bench
2. Test the PutOperator (simulating Put) as follows:
./db_bench --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom --merge_operator=put
--threads=2
3. Test the UInt64AddOperator (simulating numeric addition) similarly:
./db_bench --value_size=8 --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom
--merge_operator=uint64add --threads=2
Reviewers: haobo, dhruba, zshao, MarkCallaghan
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11535
2013-08-16 02:13:07 +02:00
|
|
|
thread->stats.AddMessage(msg);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Read-modify-write for random keys (using MergeOperator)
|
|
|
|
// The merge operator to use should be defined by FLAGS_merge_operator
|
|
|
|
// Adjust FLAGS_value_size so that the keys are reasonable for this operator
|
|
|
|
// Assumes that the merge operator is non-null (i.e.: is well-defined)
|
|
|
|
//
|
|
|
|
// For example, use FLAGS_merge_operator="uint64add" and FLAGS_value_size=8
|
|
|
|
// to simulate random additions over 64-bit integers using merge.
|
2014-01-11 02:33:56 +01:00
|
|
|
//
|
|
|
|
// The number of merges on the same key can be controlled by adjusting
|
|
|
|
// FLAGS_merge_keys.
|
Benchmarking for Merge Operator
Summary:
Updated db_bench and utilities/merge_operators.h to allow for dynamic benchmarking
of merge operators in db_bench. Added a new test (--benchmarks=mergerandom), which performs
a bunch of random Merge() operations over random keys. Also added a "--merge_operator=" flag
so that the tester can easily benchmark different merge operators. Currently supports
the PutOperator and UInt64Add operator. Support for stringappend or list append may come later.
Test Plan:
1. make db_bench
2. Test the PutOperator (simulating Put) as follows:
./db_bench --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom --merge_operator=put
--threads=2
3. Test the UInt64AddOperator (simulating numeric addition) similarly:
./db_bench --value_size=8 --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom
--merge_operator=uint64add --threads=2
Reviewers: haobo, dhruba, zshao, MarkCallaghan
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11535
2013-08-16 02:13:07 +02:00
|
|
|
void MergeRandom(ThreadState* thread) {
|
|
|
|
RandomGenerator gen;
|
2015-03-30 20:28:25 +02:00
|
|
|
int64_t bytes = 0;
|
2015-02-19 23:27:48 +01:00
|
|
|
std::unique_ptr<const char[]> key_guard;
|
|
|
|
Slice key = AllocateKey(&key_guard);
|
Benchmarking for Merge Operator
Summary:
Updated db_bench and utilities/merge_operators.h to allow for dynamic benchmarking
of merge operators in db_bench. Added a new test (--benchmarks=mergerandom), which performs
a bunch of random Merge() operations over random keys. Also added a "--merge_operator=" flag
so that the tester can easily benchmark different merge operators. Currently supports
the PutOperator and UInt64Add operator. Support for stringappend or list append may come later.
Test Plan:
1. make db_bench
2. Test the PutOperator (simulating Put) as follows:
./db_bench --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom --merge_operator=put
--threads=2
3. Test the UInt64AddOperator (simulating numeric addition) similarly:
./db_bench --value_size=8 --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom
--merge_operator=uint64add --threads=2
Reviewers: haobo, dhruba, zshao, MarkCallaghan
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11535
2013-08-16 02:13:07 +02:00
|
|
|
// The number of iterations is the larger of read_ or write_
|
|
|
|
Duration duration(FLAGS_duration, readwrites_);
|
|
|
|
while (!duration.Done(1)) {
|
2018-07-21 00:44:00 +02:00
|
|
|
DBWithColumnFamilies* db_with_cfh = SelectDBWithCfh(thread);
|
|
|
|
int64_t key_rand = thread->rand.Next() % merge_keys_;
|
|
|
|
GenerateKeyFromInt(key_rand, merge_keys_, &key);
|
Benchmarking for Merge Operator
Summary:
Updated db_bench and utilities/merge_operators.h to allow for dynamic benchmarking
of merge operators in db_bench. Added a new test (--benchmarks=mergerandom), which performs
a bunch of random Merge() operations over random keys. Also added a "--merge_operator=" flag
so that the tester can easily benchmark different merge operators. Currently supports
the PutOperator and UInt64Add operator. Support for stringappend or list append may come later.
Test Plan:
1. make db_bench
2. Test the PutOperator (simulating Put) as follows:
./db_bench --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom --merge_operator=put
--threads=2
3. Test the UInt64AddOperator (simulating numeric addition) similarly:
./db_bench --value_size=8 --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom
--merge_operator=uint64add --threads=2
Reviewers: haobo, dhruba, zshao, MarkCallaghan
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11535
2013-08-16 02:13:07 +02:00
|
|
|
|
2018-07-21 00:44:00 +02:00
|
|
|
Status s;
|
|
|
|
if (FLAGS_num_column_families > 1) {
|
|
|
|
s = db_with_cfh->db->Merge(write_options_,
|
|
|
|
db_with_cfh->GetCfh(key_rand), key,
|
|
|
|
gen.Generate(value_size_));
|
|
|
|
} else {
|
|
|
|
s = db_with_cfh->db->Merge(write_options_,
|
|
|
|
db_with_cfh->db->DefaultColumnFamily(), key,
|
|
|
|
gen.Generate(value_size_));
|
|
|
|
}
|
Benchmarking for Merge Operator
Summary:
Updated db_bench and utilities/merge_operators.h to allow for dynamic benchmarking
of merge operators in db_bench. Added a new test (--benchmarks=mergerandom), which performs
a bunch of random Merge() operations over random keys. Also added a "--merge_operator=" flag
so that the tester can easily benchmark different merge operators. Currently supports
the PutOperator and UInt64Add operator. Support for stringappend or list append may come later.
Test Plan:
1. make db_bench
2. Test the PutOperator (simulating Put) as follows:
./db_bench --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom --merge_operator=put
--threads=2
3. Test the UInt64AddOperator (simulating numeric addition) similarly:
./db_bench --value_size=8 --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom
--merge_operator=uint64add --threads=2
Reviewers: haobo, dhruba, zshao, MarkCallaghan
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11535
2013-08-16 02:13:07 +02:00
|
|
|
|
|
|
|
if (!s.ok()) {
|
|
|
|
fprintf(stderr, "merge error: %s\n", s.ToString().c_str());
|
|
|
|
exit(1);
|
|
|
|
}
|
2015-03-30 20:28:25 +02:00
|
|
|
bytes += key.size() + value_size_;
|
2018-07-21 00:44:00 +02:00
|
|
|
thread->stats.FinishedOps(nullptr, db_with_cfh->db, 1, kMerge);
|
Benchmarking for Merge Operator
Summary:
Updated db_bench and utilities/merge_operators.h to allow for dynamic benchmarking
of merge operators in db_bench. Added a new test (--benchmarks=mergerandom), which performs
a bunch of random Merge() operations over random keys. Also added a "--merge_operator=" flag
so that the tester can easily benchmark different merge operators. Currently supports
the PutOperator and UInt64Add operator. Support for stringappend or list append may come later.
Test Plan:
1. make db_bench
2. Test the PutOperator (simulating Put) as follows:
./db_bench --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom --merge_operator=put
--threads=2
3. Test the UInt64AddOperator (simulating numeric addition) similarly:
./db_bench --value_size=8 --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom
--merge_operator=uint64add --threads=2
Reviewers: haobo, dhruba, zshao, MarkCallaghan
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11535
2013-08-16 02:13:07 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
// Print some statistics
|
|
|
|
char msg[100];
|
2014-03-05 02:08:05 +01:00
|
|
|
snprintf(msg, sizeof(msg), "( updates:%" PRIu64 ")", readwrites_);
|
2015-03-30 20:28:25 +02:00
|
|
|
thread->stats.AddBytes(bytes);
|
Benchmarking for Merge Operator
Summary:
Updated db_bench and utilities/merge_operators.h to allow for dynamic benchmarking
of merge operators in db_bench. Added a new test (--benchmarks=mergerandom), which performs
a bunch of random Merge() operations over random keys. Also added a "--merge_operator=" flag
so that the tester can easily benchmark different merge operators. Currently supports
the PutOperator and UInt64Add operator. Support for stringappend or list append may come later.
Test Plan:
1. make db_bench
2. Test the PutOperator (simulating Put) as follows:
./db_bench --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom --merge_operator=put
--threads=2
3. Test the UInt64AddOperator (simulating numeric addition) similarly:
./db_bench --value_size=8 --benchmarks=fillrandom,readrandom,updaterandom,readrandom,mergerandom,readrandom
--merge_operator=uint64add --threads=2
Reviewers: haobo, dhruba, zshao, MarkCallaghan
Reviewed By: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11535
2013-08-16 02:13:07 +02:00
|
|
|
thread->stats.AddMessage(msg);
|
|
|
|
}
|
|
|
|
|
2014-01-11 02:33:56 +01:00
|
|
|
// Read and merge random keys. The amount of reads and merges are controlled
|
|
|
|
// by adjusting FLAGS_num and FLAGS_mergereadpercent. The number of distinct
|
|
|
|
// keys (and thus also the number of reads and merges on the same key) can be
|
|
|
|
// adjusted with FLAGS_merge_keys.
|
|
|
|
//
|
|
|
|
// As with MergeRandom, the merge operator to use should be defined by
|
|
|
|
// FLAGS_merge_operator.
|
|
|
|
void ReadRandomMergeRandom(ThreadState* thread) {
|
|
|
|
ReadOptions options(FLAGS_verify_checksum, true);
|
|
|
|
RandomGenerator gen;
|
|
|
|
std::string value;
|
2014-03-05 02:08:05 +01:00
|
|
|
int64_t num_hits = 0;
|
|
|
|
int64_t num_gets = 0;
|
|
|
|
int64_t num_merges = 0;
|
2014-01-11 02:33:56 +01:00
|
|
|
size_t max_length = 0;
|
|
|
|
|
2015-02-19 23:27:48 +01:00
|
|
|
std::unique_ptr<const char[]> key_guard;
|
|
|
|
Slice key = AllocateKey(&key_guard);
|
2014-01-11 02:33:56 +01:00
|
|
|
// the number of iterations is the larger of read_ or write_
|
|
|
|
Duration duration(FLAGS_duration, readwrites_);
|
|
|
|
while (!duration.Done(1)) {
|
2014-04-11 21:15:09 +02:00
|
|
|
DB* db = SelectDB(thread);
|
2014-04-08 20:21:09 +02:00
|
|
|
GenerateKeyFromInt(thread->rand.Next() % merge_keys_, merge_keys_, &key);
|
2014-01-11 02:33:56 +01:00
|
|
|
|
|
|
|
bool do_merge = int(thread->rand.Next() % 100) < FLAGS_mergereadpercent;
|
|
|
|
|
|
|
|
if (do_merge) {
|
2014-04-11 21:15:09 +02:00
|
|
|
Status s = db->Merge(write_options_, key, gen.Generate(value_size_));
|
2014-01-11 02:33:56 +01:00
|
|
|
if (!s.ok()) {
|
|
|
|
fprintf(stderr, "merge error: %s\n", s.ToString().c_str());
|
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
num_merges++;
|
2015-12-08 02:33:18 +01:00
|
|
|
thread->stats.FinishedOps(nullptr, db, 1, kMerge);
|
2014-01-11 02:33:56 +01:00
|
|
|
} else {
|
2014-04-11 21:15:09 +02:00
|
|
|
Status s = db->Get(options, key, &value);
|
2014-01-11 02:33:56 +01:00
|
|
|
if (value.length() > max_length)
|
|
|
|
max_length = value.length();
|
|
|
|
|
|
|
|
if (!s.ok() && !s.IsNotFound()) {
|
|
|
|
fprintf(stderr, "get error: %s\n", s.ToString().c_str());
|
|
|
|
// we continue after error rather than exiting so that we can
|
|
|
|
// find more errors if any
|
|
|
|
} else if (!s.IsNotFound()) {
|
|
|
|
num_hits++;
|
|
|
|
}
|
|
|
|
num_gets++;
|
2015-12-08 02:33:18 +01:00
|
|
|
thread->stats.FinishedOps(nullptr, db, 1, kRead);
|
2014-01-11 02:33:56 +01:00
|
|
|
}
|
|
|
|
}
|
2014-04-08 20:21:09 +02:00
|
|
|
|
2014-01-11 02:33:56 +01:00
|
|
|
char msg[100];
|
|
|
|
snprintf(msg, sizeof(msg),
|
2015-07-13 21:11:05 +02:00
|
|
|
"(reads:%" PRIu64 " merges:%" PRIu64 " total:%" PRIu64
|
|
|
|
" hits:%" PRIu64 " maxlength:%" ROCKSDB_PRIszt ")",
|
2014-01-11 02:33:56 +01:00
|
|
|
num_gets, num_merges, readwrites_, num_hits, max_length);
|
|
|
|
thread->stats.AddMessage(msg);
|
|
|
|
}
|
|
|
|
|
SkipListRep::LookaheadIterator
Summary:
This diff introduces the `lookahead` argument to `SkipListFactory()`. This is an
optimization for the tailing use case which includes many seeks. E.g. consider
the following operations on a skip list iterator:
Seek(x), Next(), Next(), Seek(x+2), Next(), Seek(x+3), Next(), Next(), ...
If `lookahead` is positive, `SkipListRep` will return an iterator which also
keeps track of the previously visited node. Seek() then first does a linear
search starting from that node (up to `lookahead` steps). As in the tailing
example above, this may require fewer than ~log(n) comparisons as with regular
skip list search.
Test Plan:
Added a new benchmark (`fillseekseq`) which simulates the usage pattern. It
first writes N records (with consecutive keys), then measures how much time it
takes to read them by calling `Seek()` and `Next()`.
$ time ./db_bench -num 10000000 -benchmarks fillseekseq -prefix_size 1 \
-key_size 8 -write_buffer_size $[1024*1024*1024] -value_size 50 \
-seekseq_next 2 -skip_list_lookahead=0
[...]
DB path: [/dev/shm/rocksdbtest/dbbench]
fillseekseq : 0.389 micros/op 2569047 ops/sec;
real 0m21.806s
user 0m12.106s
sys 0m9.672s
$ time ./db_bench [...] -skip_list_lookahead=2
[...]
DB path: [/dev/shm/rocksdbtest/dbbench]
fillseekseq : 0.153 micros/op 6540684 ops/sec;
real 0m19.469s
user 0m10.192s
sys 0m9.252s
Reviewers: ljin, sdong, igor
Reviewed By: igor
Subscribers: dhruba, leveldb, march, lovro
Differential Revision: https://reviews.facebook.net/D23997
2014-09-24 00:52:28 +02:00
|
|
|
void WriteSeqSeekSeq(ThreadState* thread) {
|
|
|
|
writes_ = FLAGS_num;
|
|
|
|
DoWrite(thread, SEQUENTIAL);
|
|
|
|
// exclude writes from the ops/sec calculation
|
|
|
|
thread->stats.Start(thread->tid);
|
|
|
|
|
|
|
|
DB* db = SelectDB(thread);
|
|
|
|
std::unique_ptr<Iterator> iter(
|
|
|
|
db->NewIterator(ReadOptions(FLAGS_verify_checksum, true)));
|
|
|
|
|
2015-02-19 23:27:48 +01:00
|
|
|
std::unique_ptr<const char[]> key_guard;
|
|
|
|
Slice key = AllocateKey(&key_guard);
|
SkipListRep::LookaheadIterator
Summary:
This diff introduces the `lookahead` argument to `SkipListFactory()`. This is an
optimization for the tailing use case which includes many seeks. E.g. consider
the following operations on a skip list iterator:
Seek(x), Next(), Next(), Seek(x+2), Next(), Seek(x+3), Next(), Next(), ...
If `lookahead` is positive, `SkipListRep` will return an iterator which also
keeps track of the previously visited node. Seek() then first does a linear
search starting from that node (up to `lookahead` steps). As in the tailing
example above, this may require fewer than ~log(n) comparisons as with regular
skip list search.
Test Plan:
Added a new benchmark (`fillseekseq`) which simulates the usage pattern. It
first writes N records (with consecutive keys), then measures how much time it
takes to read them by calling `Seek()` and `Next()`.
$ time ./db_bench -num 10000000 -benchmarks fillseekseq -prefix_size 1 \
-key_size 8 -write_buffer_size $[1024*1024*1024] -value_size 50 \
-seekseq_next 2 -skip_list_lookahead=0
[...]
DB path: [/dev/shm/rocksdbtest/dbbench]
fillseekseq : 0.389 micros/op 2569047 ops/sec;
real 0m21.806s
user 0m12.106s
sys 0m9.672s
$ time ./db_bench [...] -skip_list_lookahead=2
[...]
DB path: [/dev/shm/rocksdbtest/dbbench]
fillseekseq : 0.153 micros/op 6540684 ops/sec;
real 0m19.469s
user 0m10.192s
sys 0m9.252s
Reviewers: ljin, sdong, igor
Reviewed By: igor
Subscribers: dhruba, leveldb, march, lovro
Differential Revision: https://reviews.facebook.net/D23997
2014-09-24 00:52:28 +02:00
|
|
|
for (int64_t i = 0; i < FLAGS_num; ++i) {
|
|
|
|
GenerateKeyFromInt(i, FLAGS_num, &key);
|
|
|
|
iter->Seek(key);
|
|
|
|
assert(iter->Valid() && iter->key() == key);
|
2015-12-08 02:33:18 +01:00
|
|
|
thread->stats.FinishedOps(nullptr, db, 1, kSeek);
|
SkipListRep::LookaheadIterator
Summary:
This diff introduces the `lookahead` argument to `SkipListFactory()`. This is an
optimization for the tailing use case which includes many seeks. E.g. consider
the following operations on a skip list iterator:
Seek(x), Next(), Next(), Seek(x+2), Next(), Seek(x+3), Next(), Next(), ...
If `lookahead` is positive, `SkipListRep` will return an iterator which also
keeps track of the previously visited node. Seek() then first does a linear
search starting from that node (up to `lookahead` steps). As in the tailing
example above, this may require fewer than ~log(n) comparisons as with regular
skip list search.
Test Plan:
Added a new benchmark (`fillseekseq`) which simulates the usage pattern. It
first writes N records (with consecutive keys), then measures how much time it
takes to read them by calling `Seek()` and `Next()`.
$ time ./db_bench -num 10000000 -benchmarks fillseekseq -prefix_size 1 \
-key_size 8 -write_buffer_size $[1024*1024*1024] -value_size 50 \
-seekseq_next 2 -skip_list_lookahead=0
[...]
DB path: [/dev/shm/rocksdbtest/dbbench]
fillseekseq : 0.389 micros/op 2569047 ops/sec;
real 0m21.806s
user 0m12.106s
sys 0m9.672s
$ time ./db_bench [...] -skip_list_lookahead=2
[...]
DB path: [/dev/shm/rocksdbtest/dbbench]
fillseekseq : 0.153 micros/op 6540684 ops/sec;
real 0m19.469s
user 0m10.192s
sys 0m9.252s
Reviewers: ljin, sdong, igor
Reviewed By: igor
Subscribers: dhruba, leveldb, march, lovro
Differential Revision: https://reviews.facebook.net/D23997
2014-09-24 00:52:28 +02:00
|
|
|
|
2014-10-17 19:20:25 +02:00
|
|
|
for (int j = 0; j < FLAGS_seek_nexts && i + 1 < FLAGS_num; ++j) {
|
2015-03-18 21:50:52 +01:00
|
|
|
if (!FLAGS_reverse_iterator) {
|
|
|
|
iter->Next();
|
|
|
|
} else {
|
|
|
|
iter->Prev();
|
|
|
|
}
|
SkipListRep::LookaheadIterator
Summary:
This diff introduces the `lookahead` argument to `SkipListFactory()`. This is an
optimization for the tailing use case which includes many seeks. E.g. consider
the following operations on a skip list iterator:
Seek(x), Next(), Next(), Seek(x+2), Next(), Seek(x+3), Next(), Next(), ...
If `lookahead` is positive, `SkipListRep` will return an iterator which also
keeps track of the previously visited node. Seek() then first does a linear
search starting from that node (up to `lookahead` steps). As in the tailing
example above, this may require fewer than ~log(n) comparisons as with regular
skip list search.
Test Plan:
Added a new benchmark (`fillseekseq`) which simulates the usage pattern. It
first writes N records (with consecutive keys), then measures how much time it
takes to read them by calling `Seek()` and `Next()`.
$ time ./db_bench -num 10000000 -benchmarks fillseekseq -prefix_size 1 \
-key_size 8 -write_buffer_size $[1024*1024*1024] -value_size 50 \
-seekseq_next 2 -skip_list_lookahead=0
[...]
DB path: [/dev/shm/rocksdbtest/dbbench]
fillseekseq : 0.389 micros/op 2569047 ops/sec;
real 0m21.806s
user 0m12.106s
sys 0m9.672s
$ time ./db_bench [...] -skip_list_lookahead=2
[...]
DB path: [/dev/shm/rocksdbtest/dbbench]
fillseekseq : 0.153 micros/op 6540684 ops/sec;
real 0m19.469s
user 0m10.192s
sys 0m9.252s
Reviewers: ljin, sdong, igor
Reviewed By: igor
Subscribers: dhruba, leveldb, march, lovro
Differential Revision: https://reviews.facebook.net/D23997
2014-09-24 00:52:28 +02:00
|
|
|
GenerateKeyFromInt(++i, FLAGS_num, &key);
|
|
|
|
assert(iter->Valid() && iter->key() == key);
|
2015-12-08 02:33:18 +01:00
|
|
|
thread->stats.FinishedOps(nullptr, db, 1, kSeek);
|
SkipListRep::LookaheadIterator
Summary:
This diff introduces the `lookahead` argument to `SkipListFactory()`. This is an
optimization for the tailing use case which includes many seeks. E.g. consider
the following operations on a skip list iterator:
Seek(x), Next(), Next(), Seek(x+2), Next(), Seek(x+3), Next(), Next(), ...
If `lookahead` is positive, `SkipListRep` will return an iterator which also
keeps track of the previously visited node. Seek() then first does a linear
search starting from that node (up to `lookahead` steps). As in the tailing
example above, this may require fewer than ~log(n) comparisons as with regular
skip list search.
Test Plan:
Added a new benchmark (`fillseekseq`) which simulates the usage pattern. It
first writes N records (with consecutive keys), then measures how much time it
takes to read them by calling `Seek()` and `Next()`.
$ time ./db_bench -num 10000000 -benchmarks fillseekseq -prefix_size 1 \
-key_size 8 -write_buffer_size $[1024*1024*1024] -value_size 50 \
-seekseq_next 2 -skip_list_lookahead=0
[...]
DB path: [/dev/shm/rocksdbtest/dbbench]
fillseekseq : 0.389 micros/op 2569047 ops/sec;
real 0m21.806s
user 0m12.106s
sys 0m9.672s
$ time ./db_bench [...] -skip_list_lookahead=2
[...]
DB path: [/dev/shm/rocksdbtest/dbbench]
fillseekseq : 0.153 micros/op 6540684 ops/sec;
real 0m19.469s
user 0m10.192s
sys 0m9.252s
Reviewers: ljin, sdong, igor
Reviewed By: igor
Subscribers: dhruba, leveldb, march, lovro
Differential Revision: https://reviews.facebook.net/D23997
2014-09-24 00:52:28 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
iter->Seek(key);
|
|
|
|
assert(iter->Valid() && iter->key() == key);
|
2015-12-08 02:33:18 +01:00
|
|
|
thread->stats.FinishedOps(nullptr, db, 1, kSeek);
|
SkipListRep::LookaheadIterator
Summary:
This diff introduces the `lookahead` argument to `SkipListFactory()`. This is an
optimization for the tailing use case which includes many seeks. E.g. consider
the following operations on a skip list iterator:
Seek(x), Next(), Next(), Seek(x+2), Next(), Seek(x+3), Next(), Next(), ...
If `lookahead` is positive, `SkipListRep` will return an iterator which also
keeps track of the previously visited node. Seek() then first does a linear
search starting from that node (up to `lookahead` steps). As in the tailing
example above, this may require fewer than ~log(n) comparisons as with regular
skip list search.
Test Plan:
Added a new benchmark (`fillseekseq`) which simulates the usage pattern. It
first writes N records (with consecutive keys), then measures how much time it
takes to read them by calling `Seek()` and `Next()`.
$ time ./db_bench -num 10000000 -benchmarks fillseekseq -prefix_size 1 \
-key_size 8 -write_buffer_size $[1024*1024*1024] -value_size 50 \
-seekseq_next 2 -skip_list_lookahead=0
[...]
DB path: [/dev/shm/rocksdbtest/dbbench]
fillseekseq : 0.389 micros/op 2569047 ops/sec;
real 0m21.806s
user 0m12.106s
sys 0m9.672s
$ time ./db_bench [...] -skip_list_lookahead=2
[...]
DB path: [/dev/shm/rocksdbtest/dbbench]
fillseekseq : 0.153 micros/op 6540684 ops/sec;
real 0m19.469s
user 0m10.192s
sys 0m9.252s
Reviewers: ljin, sdong, igor
Reviewed By: igor
Subscribers: dhruba, leveldb, march, lovro
Differential Revision: https://reviews.facebook.net/D23997
2014-09-24 00:52:28 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2015-10-14 21:43:00 +02:00
|
|
|
#ifndef ROCKSDB_LITE
|
2015-05-29 23:36:35 +02:00
|
|
|
// This benchmark stress tests Transactions. For a given --duration (or
|
|
|
|
// total number of --writes, a Transaction will perform a read-modify-write
|
|
|
|
// to increment the value of a key in each of N(--transaction-sets) sets of
|
|
|
|
// keys (where each set has --num keys). If --threads is set, this will be
|
|
|
|
// done in parallel.
|
|
|
|
//
|
|
|
|
// To test transactions, use --transaction_db=true. Not setting this
|
|
|
|
// parameter
|
|
|
|
// will run the same benchmark without transactions.
|
|
|
|
//
|
|
|
|
// RandomTransactionVerify() will then validate the correctness of the results
|
|
|
|
// by checking if the sum of all keys in each set is the same.
|
|
|
|
void RandomTransaction(ThreadState* thread) {
|
|
|
|
ReadOptions options(FLAGS_verify_checksum, true);
|
|
|
|
Duration duration(FLAGS_duration, readwrites_);
|
|
|
|
ReadOptions read_options(FLAGS_verify_checksum, true);
|
2016-03-15 18:57:33 +01:00
|
|
|
uint16_t num_prefix_ranges = static_cast<uint16_t>(FLAGS_transaction_sets);
|
2016-03-03 20:20:25 +01:00
|
|
|
uint64_t transactions_done = 0;
|
2015-05-29 23:36:35 +02:00
|
|
|
|
|
|
|
if (num_prefix_ranges == 0 || num_prefix_ranges > 9999) {
|
|
|
|
fprintf(stderr, "invalid value for transaction_sets\n");
|
|
|
|
abort();
|
|
|
|
}
|
|
|
|
|
2016-03-03 20:20:25 +01:00
|
|
|
TransactionOptions txn_options;
|
|
|
|
txn_options.lock_timeout = FLAGS_transaction_lock_timeout;
|
|
|
|
txn_options.set_snapshot = FLAGS_transaction_set_snapshot;
|
|
|
|
|
|
|
|
RandomTransactionInserter inserter(&thread->rand, write_options_,
|
|
|
|
read_options, FLAGS_num,
|
|
|
|
num_prefix_ranges);
|
|
|
|
|
2015-05-29 23:36:35 +02:00
|
|
|
if (FLAGS_num_multi_db > 1) {
|
|
|
|
fprintf(stderr,
|
|
|
|
"Cannot run RandomTransaction benchmark with "
|
|
|
|
"FLAGS_multi_db > 1.");
|
|
|
|
abort();
|
|
|
|
}
|
|
|
|
|
|
|
|
while (!duration.Done(1)) {
|
2016-03-03 20:20:25 +01:00
|
|
|
bool success;
|
2015-05-29 23:36:35 +02:00
|
|
|
|
2016-03-03 20:20:25 +01:00
|
|
|
// RandomTransactionInserter will attempt to insert a key for each
|
|
|
|
// # of FLAGS_transaction_sets
|
Pessimistic Transactions
Summary:
Initial implementation of Pessimistic Transactions. This diff contains the api changes discussed in D38913. This diff is pretty large, so let me know if people would prefer to meet up to discuss it.
MyRocks folks: please take a look at the API in include/rocksdb/utilities/transaction[_db].h and let me know if you have any issues.
Also, you'll notice a couple of TODOs in the implementation of RollbackToSavePoint(). After chatting with Siying, I'm going to send out a separate diff for an alternate implementation of this feature that implements the rollback inside of WriteBatch/WriteBatchWithIndex. We can then decide which route is preferable.
Next, I'm planning on doing some perf testing and then integrating this diff into MongoRocks for further testing.
Test Plan: Unit tests, db_bench parallel testing.
Reviewers: igor, rven, sdong, yhchiang, yoshinorim
Reviewed By: sdong
Subscribers: hermanlee4, maykov, spetrunia, leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D40869
2015-05-26 02:37:33 +02:00
|
|
|
if (FLAGS_optimistic_transaction_db) {
|
2016-03-03 20:20:25 +01:00
|
|
|
success = inserter.OptimisticTransactionDBInsert(db_.opt_txn_db);
|
Pessimistic Transactions
Summary:
Initial implementation of Pessimistic Transactions. This diff contains the api changes discussed in D38913. This diff is pretty large, so let me know if people would prefer to meet up to discuss it.
MyRocks folks: please take a look at the API in include/rocksdb/utilities/transaction[_db].h and let me know if you have any issues.
Also, you'll notice a couple of TODOs in the implementation of RollbackToSavePoint(). After chatting with Siying, I'm going to send out a separate diff for an alternate implementation of this feature that implements the rollback inside of WriteBatch/WriteBatchWithIndex. We can then decide which route is preferable.
Next, I'm planning on doing some perf testing and then integrating this diff into MongoRocks for further testing.
Test Plan: Unit tests, db_bench parallel testing.
Reviewers: igor, rven, sdong, yhchiang, yoshinorim
Reviewed By: sdong
Subscribers: hermanlee4, maykov, spetrunia, leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D40869
2015-05-26 02:37:33 +02:00
|
|
|
} else if (FLAGS_transaction_db) {
|
|
|
|
TransactionDB* txn_db = reinterpret_cast<TransactionDB*>(db_.db);
|
2016-03-03 20:20:25 +01:00
|
|
|
success = inserter.TransactionDBInsert(txn_db, txn_options);
|
2015-05-29 23:36:35 +02:00
|
|
|
} else {
|
2016-03-03 20:20:25 +01:00
|
|
|
success = inserter.DBInsert(db_.db);
|
2015-05-29 23:36:35 +02:00
|
|
|
}
|
|
|
|
|
2016-03-03 20:20:25 +01:00
|
|
|
if (!success) {
|
|
|
|
fprintf(stderr, "Unexpected error: %s\n",
|
|
|
|
inserter.GetLastStatus().ToString().c_str());
|
|
|
|
abort();
|
2015-08-29 01:09:34 +02:00
|
|
|
}
|
|
|
|
|
2016-03-03 20:20:25 +01:00
|
|
|
thread->stats.FinishedOps(nullptr, db_.db, 1, kOthers);
|
2015-05-29 23:36:35 +02:00
|
|
|
transactions_done++;
|
|
|
|
}
|
|
|
|
|
|
|
|
char msg[100];
|
Pessimistic Transactions
Summary:
Initial implementation of Pessimistic Transactions. This diff contains the api changes discussed in D38913. This diff is pretty large, so let me know if people would prefer to meet up to discuss it.
MyRocks folks: please take a look at the API in include/rocksdb/utilities/transaction[_db].h and let me know if you have any issues.
Also, you'll notice a couple of TODOs in the implementation of RollbackToSavePoint(). After chatting with Siying, I'm going to send out a separate diff for an alternate implementation of this feature that implements the rollback inside of WriteBatch/WriteBatchWithIndex. We can then decide which route is preferable.
Next, I'm planning on doing some perf testing and then integrating this diff into MongoRocks for further testing.
Test Plan: Unit tests, db_bench parallel testing.
Reviewers: igor, rven, sdong, yhchiang, yoshinorim
Reviewed By: sdong
Subscribers: hermanlee4, maykov, spetrunia, leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D40869
2015-05-26 02:37:33 +02:00
|
|
|
if (FLAGS_optimistic_transaction_db || FLAGS_transaction_db) {
|
2015-05-29 23:36:35 +02:00
|
|
|
snprintf(msg, sizeof(msg),
|
|
|
|
"( transactions:%" PRIu64 " aborts:%" PRIu64 ")",
|
2016-03-03 20:20:25 +01:00
|
|
|
transactions_done, inserter.GetFailureCount());
|
2015-05-29 23:36:35 +02:00
|
|
|
} else {
|
|
|
|
snprintf(msg, sizeof(msg), "( batches:%" PRIu64 " )", transactions_done);
|
|
|
|
}
|
|
|
|
thread->stats.AddMessage(msg);
|
|
|
|
|
2016-06-30 03:44:22 +02:00
|
|
|
if (FLAGS_perf_level > rocksdb::PerfLevel::kDisable) {
|
2018-07-20 23:34:07 +02:00
|
|
|
thread->stats.AddMessage(std::string("PERF_CONTEXT:\n") +
|
|
|
|
get_perf_context()->ToString());
|
2015-05-29 23:36:35 +02:00
|
|
|
}
|
2017-10-17 03:29:50 +02:00
|
|
|
thread->stats.AddBytes(static_cast<int64_t>(inserter.GetBytesInserted()));
|
2015-05-29 23:36:35 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
// Verifies consistency of data after RandomTransaction() has been run.
|
|
|
|
// Since each iteration of RandomTransaction() incremented a key in each set
|
|
|
|
// by the same value, the sum of the keys in each set should be the same.
|
|
|
|
void RandomTransactionVerify() {
|
Pessimistic Transactions
Summary:
Initial implementation of Pessimistic Transactions. This diff contains the api changes discussed in D38913. This diff is pretty large, so let me know if people would prefer to meet up to discuss it.
MyRocks folks: please take a look at the API in include/rocksdb/utilities/transaction[_db].h and let me know if you have any issues.
Also, you'll notice a couple of TODOs in the implementation of RollbackToSavePoint(). After chatting with Siying, I'm going to send out a separate diff for an alternate implementation of this feature that implements the rollback inside of WriteBatch/WriteBatchWithIndex. We can then decide which route is preferable.
Next, I'm planning on doing some perf testing and then integrating this diff into MongoRocks for further testing.
Test Plan: Unit tests, db_bench parallel testing.
Reviewers: igor, rven, sdong, yhchiang, yoshinorim
Reviewed By: sdong
Subscribers: hermanlee4, maykov, spetrunia, leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D40869
2015-05-26 02:37:33 +02:00
|
|
|
if (!FLAGS_transaction_db && !FLAGS_optimistic_transaction_db) {
|
2015-05-29 23:36:35 +02:00
|
|
|
// transactions not used, nothing to verify.
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2016-03-03 20:20:25 +01:00
|
|
|
Status s =
|
2016-03-15 18:57:33 +01:00
|
|
|
RandomTransactionInserter::Verify(db_.db,
|
|
|
|
static_cast<uint16_t>(FLAGS_transaction_sets));
|
2015-05-29 23:36:35 +02:00
|
|
|
|
2016-03-03 20:20:25 +01:00
|
|
|
if (s.ok()) {
|
|
|
|
fprintf(stdout, "RandomTransactionVerify Success.\n");
|
|
|
|
} else {
|
|
|
|
fprintf(stdout, "RandomTransactionVerify FAILED!!\n");
|
2015-05-29 23:36:35 +02:00
|
|
|
}
|
|
|
|
}
|
2015-10-14 21:43:00 +02:00
|
|
|
#endif // ROCKSDB_LITE
|
2015-05-29 23:36:35 +02:00
|
|
|
|
Support for SingleDelete()
Summary:
This patch fixes #7460559. It introduces SingleDelete as a new database
operation. This operation can be used to delete keys that were never
overwritten (no put following another put of the same key). If an overwritten
key is single deleted the behavior is undefined. Single deletion of a
non-existent key has no effect but multiple consecutive single deletions are
not allowed (see limitations).
In contrast to the conventional Delete() operation, the deletion entry is
removed along with the value when the two are lined up in a compaction. Note:
The semantics are similar to @igor's prototype that allowed to have this
behavior on the granularity of a column family (
https://reviews.facebook.net/D42093 ). This new patch, however, is more
aggressive when it comes to removing tombstones: It removes the SingleDelete
together with the value whenever there is no snapshot between them while the
older patch only did this when the sequence number of the deletion was older
than the earliest snapshot.
Most of the complex additions are in the Compaction Iterator, all other changes
should be relatively straightforward. The patch also includes basic support for
single deletions in db_stress and db_bench.
Limitations:
- Not compatible with cuckoo hash tables
- Single deletions cannot be used in combination with merges and normal
deletions on the same key (other keys are not affected by this)
- Consecutive single deletions are currently not allowed (and older version of
this patch supported this so it could be resurrected if needed)
Test Plan: make all check
Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor
Reviewed By: igor
Subscribers: maykov, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D43179
2015-09-17 20:42:56 +02:00
|
|
|
// Writes and deletes random keys without overwriting keys.
|
|
|
|
//
|
|
|
|
// This benchmark is intended to partially replicate the behavior of MyRocks
|
|
|
|
// secondary indices: All data is stored in keys and updates happen by
|
|
|
|
// deleting the old version of the key and inserting the new version.
|
|
|
|
void RandomReplaceKeys(ThreadState* thread) {
|
|
|
|
std::unique_ptr<const char[]> key_guard;
|
|
|
|
Slice key = AllocateKey(&key_guard);
|
|
|
|
std::vector<uint32_t> counters(FLAGS_numdistinct, 0);
|
|
|
|
size_t max_counter = 50;
|
|
|
|
RandomGenerator gen;
|
|
|
|
|
|
|
|
Status s;
|
|
|
|
DB* db = SelectDB(thread);
|
|
|
|
for (int64_t i = 0; i < FLAGS_numdistinct; i++) {
|
|
|
|
GenerateKeyFromInt(i * max_counter, FLAGS_num, &key);
|
|
|
|
s = db->Put(write_options_, key, gen.Generate(value_size_));
|
|
|
|
if (!s.ok()) {
|
|
|
|
fprintf(stderr, "Operation failed: %s\n", s.ToString().c_str());
|
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
db->GetSnapshot();
|
|
|
|
|
|
|
|
std::default_random_engine generator;
|
|
|
|
std::normal_distribution<double> distribution(FLAGS_numdistinct / 2.0,
|
|
|
|
FLAGS_stddev);
|
|
|
|
Duration duration(FLAGS_duration, FLAGS_num);
|
|
|
|
while (!duration.Done(1)) {
|
|
|
|
int64_t rnd_id = static_cast<int64_t>(distribution(generator));
|
|
|
|
int64_t key_id = std::max(std::min(FLAGS_numdistinct - 1, rnd_id),
|
|
|
|
static_cast<int64_t>(0));
|
|
|
|
GenerateKeyFromInt(key_id * max_counter + counters[key_id], FLAGS_num,
|
|
|
|
&key);
|
|
|
|
s = FLAGS_use_single_deletes ? db->SingleDelete(write_options_, key)
|
|
|
|
: db->Delete(write_options_, key);
|
|
|
|
if (s.ok()) {
|
|
|
|
counters[key_id] = (counters[key_id] + 1) % max_counter;
|
|
|
|
GenerateKeyFromInt(key_id * max_counter + counters[key_id], FLAGS_num,
|
|
|
|
&key);
|
|
|
|
s = db->Put(write_options_, key, Slice());
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!s.ok()) {
|
|
|
|
fprintf(stderr, "Operation failed: %s\n", s.ToString().c_str());
|
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
|
2015-12-08 02:33:18 +01:00
|
|
|
thread->stats.FinishedOps(nullptr, db, 1, kOthers);
|
Support for SingleDelete()
Summary:
This patch fixes #7460559. It introduces SingleDelete as a new database
operation. This operation can be used to delete keys that were never
overwritten (no put following another put of the same key). If an overwritten
key is single deleted the behavior is undefined. Single deletion of a
non-existent key has no effect but multiple consecutive single deletions are
not allowed (see limitations).
In contrast to the conventional Delete() operation, the deletion entry is
removed along with the value when the two are lined up in a compaction. Note:
The semantics are similar to @igor's prototype that allowed to have this
behavior on the granularity of a column family (
https://reviews.facebook.net/D42093 ). This new patch, however, is more
aggressive when it comes to removing tombstones: It removes the SingleDelete
together with the value whenever there is no snapshot between them while the
older patch only did this when the sequence number of the deletion was older
than the earliest snapshot.
Most of the complex additions are in the Compaction Iterator, all other changes
should be relatively straightforward. The patch also includes basic support for
single deletions in db_stress and db_bench.
Limitations:
- Not compatible with cuckoo hash tables
- Single deletions cannot be used in combination with merges and normal
deletions on the same key (other keys are not affected by this)
- Consecutive single deletions are currently not allowed (and older version of
this patch supported this so it could be resurrected if needed)
Test Plan: make all check
Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor
Reviewed By: igor
Subscribers: maykov, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D43179
2015-09-17 20:42:56 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
char msg[200];
|
|
|
|
snprintf(msg, sizeof(msg),
|
|
|
|
"use single deletes: %d, "
|
|
|
|
"standard deviation: %lf\n",
|
|
|
|
FLAGS_use_single_deletes, FLAGS_stddev);
|
|
|
|
thread->stats.AddMessage(msg);
|
|
|
|
}
|
|
|
|
|
2016-08-01 21:28:51 +02:00
|
|
|
void TimeSeriesReadOrDelete(ThreadState* thread, bool do_deletion) {
|
|
|
|
ReadOptions options(FLAGS_verify_checksum, true);
|
|
|
|
int64_t read = 0;
|
|
|
|
int64_t found = 0;
|
|
|
|
int64_t bytes = 0;
|
|
|
|
|
|
|
|
Iterator* iter = nullptr;
|
|
|
|
// Only work on single database
|
|
|
|
assert(db_.db != nullptr);
|
|
|
|
iter = db_.db->NewIterator(options);
|
|
|
|
|
|
|
|
std::unique_ptr<const char[]> key_guard;
|
|
|
|
Slice key = AllocateKey(&key_guard);
|
|
|
|
|
|
|
|
char value_buffer[256];
|
|
|
|
while (true) {
|
|
|
|
{
|
|
|
|
MutexLock l(&thread->shared->mu);
|
|
|
|
if (thread->shared->num_done >= 1) {
|
|
|
|
// Write thread have finished
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (!FLAGS_use_tailing_iterator) {
|
|
|
|
delete iter;
|
|
|
|
iter = db_.db->NewIterator(options);
|
|
|
|
}
|
|
|
|
// Pick a Iterator to use
|
|
|
|
|
|
|
|
int64_t key_id = thread->rand.Next() % FLAGS_key_id_range;
|
|
|
|
GenerateKeyFromInt(key_id, FLAGS_num, &key);
|
|
|
|
// Reset last 8 bytes to 0
|
|
|
|
char* start = const_cast<char*>(key.data());
|
|
|
|
start += key.size() - 8;
|
|
|
|
memset(start, 0, 8);
|
|
|
|
++read;
|
|
|
|
|
|
|
|
bool key_found = false;
|
|
|
|
// Seek the prefix
|
|
|
|
for (iter->Seek(key); iter->Valid() && iter->key().starts_with(key);
|
|
|
|
iter->Next()) {
|
|
|
|
key_found = true;
|
|
|
|
// Copy out iterator's value to make sure we read them.
|
|
|
|
if (do_deletion) {
|
|
|
|
bytes += iter->key().size();
|
|
|
|
if (KeyExpired(timestamp_emulator_.get(), iter->key())) {
|
|
|
|
thread->stats.FinishedOps(&db_, db_.db, 1, kDelete);
|
|
|
|
db_.db->Delete(write_options_, iter->key());
|
|
|
|
} else {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
bytes += iter->key().size() + iter->value().size();
|
|
|
|
thread->stats.FinishedOps(&db_, db_.db, 1, kRead);
|
|
|
|
Slice value = iter->value();
|
|
|
|
memcpy(value_buffer, value.data(),
|
|
|
|
std::min(value.size(), sizeof(value_buffer)));
|
|
|
|
|
|
|
|
assert(iter->status().ok());
|
|
|
|
}
|
|
|
|
}
|
|
|
|
found += key_found;
|
2016-12-19 21:25:35 +01:00
|
|
|
|
|
|
|
if (thread->shared->read_rate_limiter.get() != nullptr) {
|
2017-06-13 23:51:22 +02:00
|
|
|
thread->shared->read_rate_limiter->Request(
|
|
|
|
1, Env::IO_HIGH, nullptr /* stats */, RateLimiter::OpType::kRead);
|
2016-12-19 21:25:35 +01:00
|
|
|
}
|
2016-08-01 21:28:51 +02:00
|
|
|
}
|
|
|
|
delete iter;
|
|
|
|
|
|
|
|
char msg[100];
|
|
|
|
snprintf(msg, sizeof(msg), "(%" PRIu64 " of %" PRIu64 " found)", found,
|
|
|
|
read);
|
|
|
|
thread->stats.AddBytes(bytes);
|
|
|
|
thread->stats.AddMessage(msg);
|
|
|
|
if (FLAGS_perf_level > rocksdb::PerfLevel::kDisable) {
|
2018-07-20 23:34:07 +02:00
|
|
|
thread->stats.AddMessage(std::string("PERF_CONTEXT:\n") +
|
|
|
|
get_perf_context()->ToString());
|
2016-08-01 21:28:51 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void TimeSeriesWrite(ThreadState* thread) {
|
|
|
|
// Special thread that keeps writing until other threads are done.
|
|
|
|
RandomGenerator gen;
|
|
|
|
int64_t bytes = 0;
|
|
|
|
|
|
|
|
// Don't merge stats from this thread with the readers.
|
|
|
|
thread->stats.SetExcludeFromMerge();
|
|
|
|
|
|
|
|
std::unique_ptr<RateLimiter> write_rate_limiter;
|
|
|
|
if (FLAGS_benchmark_write_rate_limit > 0) {
|
|
|
|
write_rate_limiter.reset(
|
|
|
|
NewGenericRateLimiter(FLAGS_benchmark_write_rate_limit));
|
|
|
|
}
|
|
|
|
|
|
|
|
std::unique_ptr<const char[]> key_guard;
|
|
|
|
Slice key = AllocateKey(&key_guard);
|
|
|
|
|
|
|
|
Duration duration(FLAGS_duration, writes_);
|
|
|
|
while (!duration.Done(1)) {
|
|
|
|
DB* db = SelectDB(thread);
|
|
|
|
|
|
|
|
uint64_t key_id = thread->rand.Next() % FLAGS_key_id_range;
|
|
|
|
// Write key id
|
|
|
|
GenerateKeyFromInt(key_id, FLAGS_num, &key);
|
|
|
|
// Write timestamp
|
|
|
|
|
|
|
|
char* start = const_cast<char*>(key.data());
|
|
|
|
char* pos = start + 8;
|
|
|
|
int bytes_to_fill =
|
|
|
|
std::min(key_size_ - static_cast<int>(pos - start), 8);
|
|
|
|
uint64_t timestamp_value = timestamp_emulator_->Get();
|
|
|
|
if (port::kLittleEndian) {
|
|
|
|
for (int i = 0; i < bytes_to_fill; ++i) {
|
|
|
|
pos[i] = (timestamp_value >> ((bytes_to_fill - i - 1) << 3)) & 0xFF;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
memcpy(pos, static_cast<void*>(×tamp_value), bytes_to_fill);
|
|
|
|
}
|
|
|
|
|
|
|
|
timestamp_emulator_->Inc();
|
|
|
|
|
|
|
|
Status s;
|
|
|
|
|
|
|
|
s = db->Put(write_options_, key, gen.Generate(value_size_));
|
|
|
|
|
|
|
|
if (!s.ok()) {
|
|
|
|
fprintf(stderr, "put error: %s\n", s.ToString().c_str());
|
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
bytes = key.size() + value_size_;
|
|
|
|
thread->stats.FinishedOps(&db_, db_.db, 1, kWrite);
|
|
|
|
thread->stats.AddBytes(bytes);
|
|
|
|
|
|
|
|
if (FLAGS_benchmark_write_rate_limit > 0) {
|
|
|
|
write_rate_limiter->Request(
|
2017-03-17 01:45:28 +01:00
|
|
|
entries_per_batch_ * (value_size_ + key_size_), Env::IO_HIGH,
|
2017-06-13 23:51:22 +02:00
|
|
|
nullptr /* stats */, RateLimiter::OpType::kWrite);
|
2016-08-01 21:28:51 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void TimeSeries(ThreadState* thread) {
|
|
|
|
if (thread->tid > 0) {
|
|
|
|
bool do_deletion = FLAGS_expire_style == "delete" &&
|
|
|
|
thread->tid <= FLAGS_num_deletion_threads;
|
|
|
|
TimeSeriesReadOrDelete(thread, do_deletion);
|
|
|
|
} else {
|
|
|
|
TimeSeriesWrite(thread);
|
|
|
|
thread->stats.Stop();
|
|
|
|
thread->stats.Report("timeseries write");
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2011-08-22 23:08:51 +02:00
|
|
|
void Compact(ThreadState* thread) {
|
2014-04-11 21:15:09 +02:00
|
|
|
DB* db = SelectDB(thread);
|
2017-11-07 19:42:47 +01:00
|
|
|
CompactRangeOptions cro;
|
|
|
|
cro.bottommost_level_compaction = BottommostLevelCompaction::kForce;
|
|
|
|
db->CompactRange(cro, nullptr, nullptr);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
2017-05-31 01:33:26 +02:00
|
|
|
void CompactAll() {
|
|
|
|
if (db_.db != nullptr) {
|
|
|
|
db_.db->CompactRange(CompactRangeOptions(), nullptr, nullptr);
|
|
|
|
}
|
|
|
|
for (const auto& db_with_cfh : multi_dbs_) {
|
|
|
|
db_with_cfh.db->CompactRange(CompactRangeOptions(), nullptr, nullptr);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2017-04-19 01:30:51 +02:00
|
|
|
void ResetStats() {
|
|
|
|
if (db_.db != nullptr) {
|
|
|
|
db_.db->ResetStats();
|
|
|
|
}
|
|
|
|
for (const auto& db_with_cfh : multi_dbs_) {
|
|
|
|
db_with_cfh.db->ResetStats();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2012-04-17 17:36:46 +02:00
|
|
|
void PrintStats(const char* key) {
|
2014-08-19 03:15:01 +02:00
|
|
|
if (db_.db != nullptr) {
|
|
|
|
PrintStats(db_.db, key, false);
|
2014-04-11 21:15:09 +02:00
|
|
|
}
|
2014-08-19 03:15:01 +02:00
|
|
|
for (const auto& db_with_cfh : multi_dbs_) {
|
|
|
|
PrintStats(db_with_cfh.db, key, true);
|
2014-04-11 21:15:09 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void PrintStats(DB* db, const char* key, bool print_header = false) {
|
|
|
|
if (print_header) {
|
|
|
|
fprintf(stdout, "\n==== DB: %s ===\n", db->GetName().c_str());
|
|
|
|
}
|
2011-04-12 21:38:58 +02:00
|
|
|
std::string stats;
|
2014-04-11 21:15:09 +02:00
|
|
|
if (!db->GetProperty(key, &stats)) {
|
2011-08-22 23:08:51 +02:00
|
|
|
stats = "(failed)";
|
2011-04-12 21:38:58 +02:00
|
|
|
}
|
2011-08-22 23:08:51 +02:00
|
|
|
fprintf(stdout, "\n%s\n", stats.c_str());
|
2011-04-12 21:38:58 +02:00
|
|
|
}
|
2018-08-01 09:14:43 +02:00
|
|
|
|
|
|
|
void Replay(ThreadState* thread) {
|
|
|
|
if (db_.db != nullptr) {
|
|
|
|
Replay(thread, &db_);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void Replay(ThreadState* /*thread*/, DBWithColumnFamilies* db_with_cfh) {
|
|
|
|
Status s;
|
2018-11-09 20:17:34 +01:00
|
|
|
std::unique_ptr<TraceReader> trace_reader;
|
2018-08-01 09:14:43 +02:00
|
|
|
s = NewFileTraceReader(FLAGS_env, EnvOptions(), FLAGS_trace_file,
|
|
|
|
&trace_reader);
|
|
|
|
if (!s.ok()) {
|
|
|
|
fprintf(
|
|
|
|
stderr,
|
|
|
|
"Encountered an error creating a TraceReader from the trace file. "
|
|
|
|
"Error: %s\n",
|
|
|
|
s.ToString().c_str());
|
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
Replayer replayer(db_with_cfh->db, db_with_cfh->cfh,
|
|
|
|
std::move(trace_reader));
|
|
|
|
s = replayer.Replay();
|
|
|
|
if (s.ok()) {
|
|
|
|
fprintf(stdout, "Replay started from trace_file: %s\n",
|
|
|
|
FLAGS_trace_file.c_str());
|
|
|
|
} else {
|
|
|
|
fprintf(stderr, "Starting replay failed. Error: %s\n",
|
|
|
|
s.ToString().c_str());
|
|
|
|
}
|
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
};
|
|
|
|
|
2016-02-16 15:17:31 +01:00
|
|
|
int db_bench_tool(int argc, char** argv) {
|
2014-04-23 15:11:35 +02:00
|
|
|
rocksdb::port::InstallStackTraceHandler();
|
2016-06-03 01:24:14 +02:00
|
|
|
static bool initialized = false;
|
|
|
|
if (!initialized) {
|
|
|
|
SetUsageMessage(std::string("\nUSAGE:\n") + std::string(argv[0]) +
|
|
|
|
" [OPTIONS]...");
|
|
|
|
initialized = true;
|
|
|
|
}
|
2014-05-09 02:25:13 +02:00
|
|
|
ParseCommandLineFlags(&argc, &argv, true);
|
2013-10-24 16:43:14 +02:00
|
|
|
FLAGS_compaction_style_e = (rocksdb::CompactionStyle) FLAGS_compaction_style;
|
2017-01-27 00:18:07 +01:00
|
|
|
#ifndef ROCKSDB_LITE
|
|
|
|
if (FLAGS_statistics && !FLAGS_statistics_string.empty()) {
|
|
|
|
fprintf(stderr,
|
|
|
|
"Cannot provide both --statistics and --statistics_string.\n");
|
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
if (!FLAGS_statistics_string.empty()) {
|
|
|
|
std::unique_ptr<Statistics> custom_stats_guard;
|
|
|
|
dbstats.reset(NewCustomObject<Statistics>(FLAGS_statistics_string,
|
|
|
|
&custom_stats_guard));
|
|
|
|
custom_stats_guard.release();
|
|
|
|
if (dbstats == nullptr) {
|
|
|
|
fprintf(stderr, "No Statistics registered matching string: %s\n",
|
|
|
|
FLAGS_statistics_string.c_str());
|
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
#endif // ROCKSDB_LITE
|
2013-10-24 16:43:14 +02:00
|
|
|
if (FLAGS_statistics) {
|
|
|
|
dbstats = rocksdb::CreateDBStatistics();
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
2019-02-28 19:14:19 +01:00
|
|
|
if (dbstats) {
|
2019-03-01 19:39:00 +01:00
|
|
|
dbstats->set_stats_level(static_cast<StatsLevel>(FLAGS_stats_level));
|
2019-02-28 19:14:19 +01:00
|
|
|
}
|
2015-09-22 02:16:31 +02:00
|
|
|
FLAGS_compaction_pri_e = (rocksdb::CompactionPri)FLAGS_compaction_pri;
|
2011-03-18 23:37:00 +01:00
|
|
|
|
2014-11-21 17:05:28 +01:00
|
|
|
std::vector<std::string> fanout = rocksdb::StringSplit(
|
|
|
|
FLAGS_max_bytes_for_level_multiplier_additional, ',');
|
2015-12-16 00:26:20 +01:00
|
|
|
for (size_t j = 0; j < fanout.size(); j++) {
|
2013-10-24 16:43:14 +02:00
|
|
|
FLAGS_max_bytes_for_level_multiplier_additional_v.push_back(
|
2015-06-08 20:43:55 +02:00
|
|
|
#ifndef CYGWIN
|
|
|
|
std::stoi(fanout[j]));
|
|
|
|
#else
|
|
|
|
stoi(fanout[j]));
|
|
|
|
#endif
|
2013-10-24 16:43:14 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
FLAGS_compression_type_e =
|
|
|
|
StringToCompressionType(FLAGS_compression_type.c_str());
|
|
|
|
|
2016-06-10 02:53:03 +02:00
|
|
|
#ifndef ROCKSDB_LITE
|
|
|
|
std::unique_ptr<Env> custom_env_guard;
|
|
|
|
if (!FLAGS_hdfs.empty() && !FLAGS_env_uri.empty()) {
|
|
|
|
fprintf(stderr, "Cannot provide both --hdfs and --env_uri.\n");
|
|
|
|
exit(1);
|
|
|
|
} else if (!FLAGS_env_uri.empty()) {
|
2017-01-26 00:54:09 +01:00
|
|
|
FLAGS_env = NewCustomObject<Env>(FLAGS_env_uri, &custom_env_guard);
|
2016-06-10 02:53:03 +02:00
|
|
|
if (FLAGS_env == nullptr) {
|
|
|
|
fprintf(stderr, "No Env registered for URI: %s\n", FLAGS_env_uri.c_str());
|
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
#endif // ROCKSDB_LITE
|
2019-03-01 20:14:28 +01:00
|
|
|
if (FLAGS_use_existing_keys && !FLAGS_use_existing_db) {
|
|
|
|
fprintf(stderr,
|
|
|
|
"`-use_existing_db` must be true for `-use_existing_keys` to be "
|
|
|
|
"settable\n");
|
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
|
2013-10-24 16:43:14 +02:00
|
|
|
if (!FLAGS_hdfs.empty()) {
|
|
|
|
FLAGS_env = new rocksdb::HdfsEnv(FLAGS_hdfs);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!strcasecmp(FLAGS_compaction_fadvice.c_str(), "NONE"))
|
|
|
|
FLAGS_compaction_fadvice_e = rocksdb::Options::NONE;
|
|
|
|
else if (!strcasecmp(FLAGS_compaction_fadvice.c_str(), "NORMAL"))
|
|
|
|
FLAGS_compaction_fadvice_e = rocksdb::Options::NORMAL;
|
|
|
|
else if (!strcasecmp(FLAGS_compaction_fadvice.c_str(), "SEQUENTIAL"))
|
|
|
|
FLAGS_compaction_fadvice_e = rocksdb::Options::SEQUENTIAL;
|
|
|
|
else if (!strcasecmp(FLAGS_compaction_fadvice.c_str(), "WILLNEED"))
|
|
|
|
FLAGS_compaction_fadvice_e = rocksdb::Options::WILLNEED;
|
|
|
|
else {
|
|
|
|
fprintf(stdout, "Unknown compaction fadvice:%s\n",
|
|
|
|
FLAGS_compaction_fadvice.c_str());
|
|
|
|
}
|
|
|
|
|
|
|
|
FLAGS_rep_factory = StringToRepFactory(FLAGS_memtablerep.c_str());
|
|
|
|
|
2017-08-04 06:30:37 +02:00
|
|
|
// Note options sanitization may increase thread pool sizes according to
|
|
|
|
// max_background_flushes/max_background_compactions/max_background_jobs
|
|
|
|
FLAGS_env->SetBackgroundThreads(FLAGS_num_high_pri_threads,
|
2014-10-10 05:07:12 +02:00
|
|
|
rocksdb::Env::Priority::HIGH);
|
2017-08-04 00:36:28 +02:00
|
|
|
FLAGS_env->SetBackgroundThreads(FLAGS_num_bottom_pri_threads,
|
|
|
|
rocksdb::Env::Priority::BOTTOM);
|
2017-08-04 06:30:37 +02:00
|
|
|
FLAGS_env->SetBackgroundThreads(FLAGS_num_low_pri_threads,
|
|
|
|
rocksdb::Env::Priority::LOW);
|
2014-10-10 05:07:12 +02:00
|
|
|
|
2012-08-27 08:45:35 +02:00
|
|
|
// Choose a location for the test database if none given with --db=<path>
|
2013-10-24 16:43:14 +02:00
|
|
|
if (FLAGS_db.empty()) {
|
|
|
|
std::string default_db_path;
|
|
|
|
rocksdb::Env::Default()->GetTestDirectory(&default_db_path);
|
|
|
|
default_db_path += "/dbbench";
|
|
|
|
FLAGS_db = default_db_path;
|
2012-08-27 08:45:35 +02:00
|
|
|
}
|
|
|
|
|
2015-03-30 21:58:32 +02:00
|
|
|
if (FLAGS_stats_interval_seconds > 0) {
|
|
|
|
// When both are set then FLAGS_stats_interval determines the frequency
|
|
|
|
// at which the timer is checked for FLAGS_stats_interval_seconds
|
|
|
|
FLAGS_stats_interval = 1000;
|
|
|
|
}
|
|
|
|
|
2013-10-04 06:49:15 +02:00
|
|
|
rocksdb::Benchmark benchmark;
|
2011-03-18 23:37:00 +01:00
|
|
|
benchmark.Run();
|
2018-10-24 20:36:19 +02:00
|
|
|
|
2018-10-25 05:13:06 +02:00
|
|
|
#ifndef ROCKSDB_LITE
|
2018-10-24 20:36:19 +02:00
|
|
|
if (FLAGS_print_malloc_stats) {
|
|
|
|
std::string stats_string;
|
|
|
|
rocksdb::DumpMallocStats(&stats_string);
|
|
|
|
fprintf(stdout, "Malloc stats:\n%s\n", stats_string.c_str());
|
|
|
|
}
|
2018-10-25 05:13:06 +02:00
|
|
|
#endif // ROCKSDB_LITE
|
2018-10-24 20:36:19 +02:00
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
return 0;
|
|
|
|
}
|
2016-02-16 15:17:31 +01:00
|
|
|
} // namespace rocksdb
|
2016-02-16 21:16:47 +01:00
|
|
|
#endif
|