rocksdb/db/table_cache.cc

659 lines
24 KiB
C++
Raw Normal View History

// Copyright (c) 2011-present, Facebook, Inc. All rights reserved.
// 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).
//
// 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.
#include "db/table_cache.h"
#include "db/dbformat.h"
Use only "local" range tombstones during Get (#4449) Summary: Previously, range tombstones were accumulated from every level, which was necessary if a range tombstone in a higher level covered a key in a lower level. However, RangeDelAggregator::AddTombstones's complexity is based on the number of tombstones that are currently stored in it, which is wasteful in the Get case, where we only need to know the highest sequence number of range tombstones that cover the key from higher levels, and compute the highest covering sequence number at the current level. This change introduces this optimization, and removes the use of RangeDelAggregator from the Get path. In the benchmark results, the following command was used to initialize the database: ``` ./db_bench -db=/dev/shm/5k-rts -use_existing_db=false -benchmarks=filluniquerandom -write_buffer_size=1048576 -compression_type=lz4 -target_file_size_base=1048576 -max_bytes_for_level_base=4194304 -value_size=112 -key_size=16 -block_size=4096 -level_compaction_dynamic_level_bytes=true -num=5000000 -max_background_jobs=12 -benchmark_write_rate_limit=20971520 -range_tombstone_width=100 -writes_per_range_tombstone=100 -max_num_range_tombstones=50000 -bloom_bits=8 ``` ...and the following command was used to measure read throughput: ``` ./db_bench -db=/dev/shm/5k-rts/ -use_existing_db=true -benchmarks=readrandom -disable_auto_compactions=true -num=5000000 -reads=100000 -threads=32 ``` The filluniquerandom command was only run once, and the resulting database was used to measure read performance before and after the PR. Both binaries were compiled with `DEBUG_LEVEL=0`. Readrandom results before PR: ``` readrandom : 4.544 micros/op 220090 ops/sec; 16.9 MB/s (63103 of 100000 found) ``` Readrandom results after PR: ``` readrandom : 11.147 micros/op 89707 ops/sec; 6.9 MB/s (63103 of 100000 found) ``` So it's actually slower right now, but this PR paves the way for future optimizations (see #4493). ---- Pull Request resolved: https://github.com/facebook/rocksdb/pull/4449 Differential Revision: D10370575 Pulled By: abhimadan fbshipit-source-id: 9a2e152be1ef36969055c0e9eb4beb0d96c11f4d
2018-10-24 21:29:29 +02:00
#include "db/range_tombstone_fragmenter.h"
#include "db/snapshot_impl.h"
#include "db/version_edit.h"
#include "file/filename.h"
#include "monitoring/perf_context_imp.h"
#include "rocksdb/statistics.h"
#include "table/block_based/block_based_table_reader.h"
#include "table/get_context.h"
#include "table/internal_iterator.h"
#include "table/iterator_wrapper.h"
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
#include "table/multiget_context.h"
#include "table/table_builder.h"
#include "table/table_reader.h"
#include "test_util/sync_point.h"
#include "util/cast_util.h"
#include "util/coding.h"
#include "util/file_reader_writer.h"
#include "util/stop_watch.h"
namespace rocksdb {
namespace {
template <class T>
static void DeleteEntry(const Slice& /*key*/, void* value) {
T* typed_value = reinterpret_cast<T*>(value);
delete typed_value;
}
static void UnrefEntry(void* arg1, void* arg2) {
Cache* cache = reinterpret_cast<Cache*>(arg1);
Cache::Handle* h = reinterpret_cast<Cache::Handle*>(arg2);
cache->Release(h);
}
static Slice GetSliceForFileNumber(const uint64_t* file_number) {
return Slice(reinterpret_cast<const char*>(file_number),
sizeof(*file_number));
}
#ifndef ROCKSDB_LITE
void AppendVarint64(IterKey* key, uint64_t v) {
char buf[10];
auto ptr = EncodeVarint64(buf, v);
key->TrimAppend(key->Size(), buf, ptr - buf);
}
#endif // ROCKSDB_LITE
} // namespace
TableCache::TableCache(const ImmutableCFOptions& ioptions,
const EnvOptions& env_options, Cache* const cache,
BlockCacheTracer* const block_cache_tracer)
: ioptions_(ioptions),
env_options_(env_options),
cache_(cache),
immortal_tables_(false),
block_cache_tracer_(block_cache_tracer) {
if (ioptions_.row_cache) {
// If the same cache is shared by multiple instances, we need to
// disambiguate its entries.
PutVarint64(&row_cache_id_, ioptions_.row_cache->NewId());
}
}
TableCache::~TableCache() {
}
TableReader* TableCache::GetTableReaderFromHandle(Cache::Handle* handle) {
return reinterpret_cast<TableReader*>(cache_->Value(handle));
}
void TableCache::ReleaseHandle(Cache::Handle* handle) {
cache_->Release(handle);
}
Status TableCache::GetTableReader(
const EnvOptions& env_options,
const InternalKeyComparator& internal_comparator, const FileDescriptor& fd,
bool sequential_mode, bool record_read_stats, HistogramImpl* file_read_hist,
std::unique_ptr<TableReader>* table_reader,
const SliceTransform* prefix_extractor, bool skip_filters, int level,
bool prefetch_index_and_filter_in_cache) {
std::string fname =
TableFileName(ioptions_.cf_paths, fd.GetNumber(), fd.GetPathId());
std::unique_ptr<RandomAccessFile> file;
Status s = ioptions_.env->NewRandomAccessFile(fname, &file, env_options);
RecordTick(ioptions_.statistics, NO_FILE_OPENS);
if (s.ok()) {
if (!sequential_mode && ioptions_.advise_random_on_open) {
file->Hint(RandomAccessFile::RANDOM);
}
StopWatch sw(ioptions_.env, ioptions_.statistics, TABLE_OPEN_IO_MICROS);
std::unique_ptr<RandomAccessFileReader> file_reader(
new RandomAccessFileReader(
std::move(file), fname, ioptions_.env,
record_read_stats ? ioptions_.statistics : nullptr, SST_READ_MICROS,
file_read_hist, ioptions_.rate_limiter, ioptions_.listeners));
s = ioptions_.table_factory->NewTableReader(
TableReaderOptions(ioptions_, prefix_extractor, env_options,
internal_comparator, skip_filters, immortal_tables_,
level, fd.largest_seqno, block_cache_tracer_),
std::move(file_reader), fd.GetFileSize(), table_reader,
prefetch_index_and_filter_in_cache);
TEST_SYNC_POINT("TableCache::GetTableReader:0");
}
return s;
}
void TableCache::EraseHandle(const FileDescriptor& fd, Cache::Handle* handle) {
ReleaseHandle(handle);
uint64_t number = fd.GetNumber();
Slice key = GetSliceForFileNumber(&number);
cache_->Erase(key);
}
Status TableCache::FindTable(const EnvOptions& env_options,
const InternalKeyComparator& internal_comparator,
const FileDescriptor& fd, Cache::Handle** handle,
const SliceTransform* prefix_extractor,
const bool no_io, bool record_read_stats,
HistogramImpl* file_read_hist, bool skip_filters,
int level,
bool prefetch_index_and_filter_in_cache) {
PERF_TIMER_GUARD_WITH_ENV(find_table_nanos, ioptions_.env);
Status s;
uint64_t number = fd.GetNumber();
Slice key = GetSliceForFileNumber(&number);
*handle = cache_->Lookup(key);
TEST_SYNC_POINT_CALLBACK("TableCache::FindTable:0",
const_cast<bool*>(&no_io));
if (*handle == nullptr) {
if (no_io) { // Don't do IO and return a not-found status
return Status::Incomplete("Table not found in table_cache, no_io is set");
}
std::unique_ptr<TableReader> table_reader;
s = GetTableReader(env_options, internal_comparator, fd,
false /* sequential mode */, record_read_stats,
file_read_hist, &table_reader, prefix_extractor,
skip_filters, level, prefetch_index_and_filter_in_cache);
if (!s.ok()) {
assert(table_reader == nullptr);
RecordTick(ioptions_.statistics, NO_FILE_ERRORS);
// We do not cache error results so that if the error is transient,
// or somebody repairs the file, we recover automatically.
} else {
s = cache_->Insert(key, table_reader.get(), 1, &DeleteEntry<TableReader>,
handle);
if (s.ok()) {
// Release ownership of table reader.
table_reader.release();
}
}
}
return s;
}
InternalIterator* TableCache::NewIterator(
const ReadOptions& options, const EnvOptions& env_options,
const InternalKeyComparator& icomparator, const FileMetaData& file_meta,
RangeDelAggregator* range_del_agg, const SliceTransform* prefix_extractor,
TableReader** table_reader_ptr, HistogramImpl* file_read_hist,
TableReaderCaller caller, Arena* arena, bool skip_filters, int level,
const InternalKey* smallest_compaction_key,
const InternalKey* largest_compaction_key) {
PERF_TIMER_GUARD(new_table_iterator_nanos);
Status s;
TableReader* table_reader = nullptr;
Cache::Handle* handle = nullptr;
if (table_reader_ptr != nullptr) {
*table_reader_ptr = nullptr;
}
bool for_compaction = caller == TableReaderCaller::kCompaction;
auto& fd = file_meta.fd;
table_reader = fd.table_reader;
if (table_reader == nullptr) {
s = FindTable(env_options, icomparator, fd, &handle, prefix_extractor,
options.read_tier == kBlockCacheTier /* no_io */,
!for_compaction /* record_read_stats */, file_read_hist,
skip_filters, level);
if (s.ok()) {
table_reader = GetTableReaderFromHandle(handle);
}
}
InternalIterator* result = nullptr;
if (s.ok()) {
expose a hook to skip tables during iteration Summary: As discussed on the mailing list (["Skipping entire SSTs while iterating"](https://groups.google.com/forum/#!topic/rocksdb/ujHCJVLrHlU)), this patch adds a `table_filter` to `ReadOptions` that allows specifying a callback to be executed during iteration before each table in the database is scanned. The callback is passed the table's properties; the table is scanned iff the callback returns true. This can be used in conjunction with a `TablePropertiesCollector` to dramatically speed up scans by skipping tables that are known to contain irrelevant data for the scan at hand. We're using this [downstream in CockroachDB](https://github.com/cockroachdb/cockroach/blob/master/pkg/storage/engine/db.cc#L2009-L2022) already. With this feature, under ideal conditions, we can reduce the time of an incremental backup in from hours to seconds. FYI, the first commit in this PR fixes a segfault that I unfortunately have not figured out how to reproduce outside of CockroachDB. I'm hoping you accept it on the grounds that it is not correct to return 8-byte aligned memory from a call to `malloc` on some 64-bit platforms; one correct approach is to infer the necessary alignment from `std::max_align_t`, as done here. As noted in the first commit message, the bug is tickled by having a`std::function` in `struct ReadOptions`. That is, the following patch alone is enough to cause RocksDB to segfault when run from CockroachDB on Darwin. ```diff --- a/include/rocksdb/options.h +++ b/include/rocksdb/options.h @@ -1546,6 +1546,13 @@ struct ReadOptions { // Default: false bool ignore_range_deletions; + // A callback to determine whether relevant keys for this scan exist in a + // given table based on the table's properties. The callback is passed the + // properties of each table during iteration. If the callback returns false, + // the table will not be scanned. + // Default: empty (every table will be scanned) + std::function<bool(const TableProperties&)> table_filter; + ReadOptions(); ReadOptions(bool cksum, bool cache); }; ``` /cc danhhz Closes https://github.com/facebook/rocksdb/pull/2265 Differential Revision: D5054262 Pulled By: yiwu-arbug fbshipit-source-id: dd6b28f2bba6cb8466250d8c5c542d3c92785476
2017-10-18 07:09:01 +02:00
if (options.table_filter &&
!options.table_filter(*table_reader->GetTableProperties())) {
result = NewEmptyInternalIterator<Slice>(arena);
expose a hook to skip tables during iteration Summary: As discussed on the mailing list (["Skipping entire SSTs while iterating"](https://groups.google.com/forum/#!topic/rocksdb/ujHCJVLrHlU)), this patch adds a `table_filter` to `ReadOptions` that allows specifying a callback to be executed during iteration before each table in the database is scanned. The callback is passed the table's properties; the table is scanned iff the callback returns true. This can be used in conjunction with a `TablePropertiesCollector` to dramatically speed up scans by skipping tables that are known to contain irrelevant data for the scan at hand. We're using this [downstream in CockroachDB](https://github.com/cockroachdb/cockroach/blob/master/pkg/storage/engine/db.cc#L2009-L2022) already. With this feature, under ideal conditions, we can reduce the time of an incremental backup in from hours to seconds. FYI, the first commit in this PR fixes a segfault that I unfortunately have not figured out how to reproduce outside of CockroachDB. I'm hoping you accept it on the grounds that it is not correct to return 8-byte aligned memory from a call to `malloc` on some 64-bit platforms; one correct approach is to infer the necessary alignment from `std::max_align_t`, as done here. As noted in the first commit message, the bug is tickled by having a`std::function` in `struct ReadOptions`. That is, the following patch alone is enough to cause RocksDB to segfault when run from CockroachDB on Darwin. ```diff --- a/include/rocksdb/options.h +++ b/include/rocksdb/options.h @@ -1546,6 +1546,13 @@ struct ReadOptions { // Default: false bool ignore_range_deletions; + // A callback to determine whether relevant keys for this scan exist in a + // given table based on the table's properties. The callback is passed the + // properties of each table during iteration. If the callback returns false, + // the table will not be scanned. + // Default: empty (every table will be scanned) + std::function<bool(const TableProperties&)> table_filter; + ReadOptions(); ReadOptions(bool cksum, bool cache); }; ``` /cc danhhz Closes https://github.com/facebook/rocksdb/pull/2265 Differential Revision: D5054262 Pulled By: yiwu-arbug fbshipit-source-id: dd6b28f2bba6cb8466250d8c5c542d3c92785476
2017-10-18 07:09:01 +02:00
} else {
result = table_reader->NewIterator(options, prefix_extractor, arena,
skip_filters, caller,
env_options.compaction_readahead_size);
expose a hook to skip tables during iteration Summary: As discussed on the mailing list (["Skipping entire SSTs while iterating"](https://groups.google.com/forum/#!topic/rocksdb/ujHCJVLrHlU)), this patch adds a `table_filter` to `ReadOptions` that allows specifying a callback to be executed during iteration before each table in the database is scanned. The callback is passed the table's properties; the table is scanned iff the callback returns true. This can be used in conjunction with a `TablePropertiesCollector` to dramatically speed up scans by skipping tables that are known to contain irrelevant data for the scan at hand. We're using this [downstream in CockroachDB](https://github.com/cockroachdb/cockroach/blob/master/pkg/storage/engine/db.cc#L2009-L2022) already. With this feature, under ideal conditions, we can reduce the time of an incremental backup in from hours to seconds. FYI, the first commit in this PR fixes a segfault that I unfortunately have not figured out how to reproduce outside of CockroachDB. I'm hoping you accept it on the grounds that it is not correct to return 8-byte aligned memory from a call to `malloc` on some 64-bit platforms; one correct approach is to infer the necessary alignment from `std::max_align_t`, as done here. As noted in the first commit message, the bug is tickled by having a`std::function` in `struct ReadOptions`. That is, the following patch alone is enough to cause RocksDB to segfault when run from CockroachDB on Darwin. ```diff --- a/include/rocksdb/options.h +++ b/include/rocksdb/options.h @@ -1546,6 +1546,13 @@ struct ReadOptions { // Default: false bool ignore_range_deletions; + // A callback to determine whether relevant keys for this scan exist in a + // given table based on the table's properties. The callback is passed the + // properties of each table during iteration. If the callback returns false, + // the table will not be scanned. + // Default: empty (every table will be scanned) + std::function<bool(const TableProperties&)> table_filter; + ReadOptions(); ReadOptions(bool cksum, bool cache); }; ``` /cc danhhz Closes https://github.com/facebook/rocksdb/pull/2265 Differential Revision: D5054262 Pulled By: yiwu-arbug fbshipit-source-id: dd6b28f2bba6cb8466250d8c5c542d3c92785476
2017-10-18 07:09:01 +02:00
}
if (handle != nullptr) {
result->RegisterCleanup(&UnrefEntry, cache_, handle);
handle = nullptr; // prevent from releasing below
}
if (for_compaction) {
table_reader->SetupForCompaction();
}
if (table_reader_ptr != nullptr) {
*table_reader_ptr = table_reader;
}
}
if (s.ok() && range_del_agg != nullptr && !options.ignore_range_deletions) {
Recommit "Avoid adding tombstones of the same file to RangeDelAggregator multiple times" Summary: The origin commit #3635 will hurt performance for users who aren't using range deletions, because unneeded std::set operations, so it was reverted by commit 44653c7b7aabe821e671946e732dda7ae6b43d1b. (see #3672) To fix this, move the set to and add a check in , i.e., file will be added only if is non-nullptr. The db_bench command which find the performance regression: > ./db_bench --benchmarks=fillrandom,seekrandomwhilewriting --threads=1 --num=1000000 --reads=150000 --key_size=66 > --value_size=1262 --statistics=0 --compression_ratio=0.5 --histogram=1 --seek_nexts=1 --stats_per_interval=1 > --stats_interval_seconds=600 --max_background_flushes=4 --num_multi_db=1 --max_background_compactions=16 --seed=1522388277 > -write_buffer_size=1048576 --level0_file_num_compaction_trigger=10000 --compression_type=none Before and after the modification, I re-run this command on the machine, the results of are as follows: **fillrandom** Table | P50 | P75 | P99 | P99.9 | P99.99 | ---- | --- | --- | --- | ----- | ------ | before commit | 5.92 | 8.57 | 19.63 | 980.97 | 12196.00 | after commit | 5.91 | 8.55 | 19.34 | 965.56 | 13513.56 | **seekrandomwhilewriting** Table | P50 | P75 | P99 | P99.9 | P99.99 | ---- | --- | --- | --- | ----- | ------ | before commit | 1418.62 | 1867.01 | 3823.28 | 4980.99 | 9240.00 | after commit | 1450.54 | 1880.61 | 3962.87 | 5429.60 | 7542.86 | Closes https://github.com/facebook/rocksdb/pull/3800 Differential Revision: D7874245 Pulled By: ajkr fbshipit-source-id: 2e8bec781b3f7399246babd66395c88619534a17
2018-05-05 01:37:39 +02:00
if (range_del_agg->AddFile(fd.GetNumber())) {
std::unique_ptr<FragmentedRangeTombstoneIterator> range_del_iter(
static_cast<FragmentedRangeTombstoneIterator*>(
table_reader->NewRangeTombstoneIterator(options)));
Recommit "Avoid adding tombstones of the same file to RangeDelAggregator multiple times" Summary: The origin commit #3635 will hurt performance for users who aren't using range deletions, because unneeded std::set operations, so it was reverted by commit 44653c7b7aabe821e671946e732dda7ae6b43d1b. (see #3672) To fix this, move the set to and add a check in , i.e., file will be added only if is non-nullptr. The db_bench command which find the performance regression: > ./db_bench --benchmarks=fillrandom,seekrandomwhilewriting --threads=1 --num=1000000 --reads=150000 --key_size=66 > --value_size=1262 --statistics=0 --compression_ratio=0.5 --histogram=1 --seek_nexts=1 --stats_per_interval=1 > --stats_interval_seconds=600 --max_background_flushes=4 --num_multi_db=1 --max_background_compactions=16 --seed=1522388277 > -write_buffer_size=1048576 --level0_file_num_compaction_trigger=10000 --compression_type=none Before and after the modification, I re-run this command on the machine, the results of are as follows: **fillrandom** Table | P50 | P75 | P99 | P99.9 | P99.99 | ---- | --- | --- | --- | ----- | ------ | before commit | 5.92 | 8.57 | 19.63 | 980.97 | 12196.00 | after commit | 5.91 | 8.55 | 19.34 | 965.56 | 13513.56 | **seekrandomwhilewriting** Table | P50 | P75 | P99 | P99.9 | P99.99 | ---- | --- | --- | --- | ----- | ------ | before commit | 1418.62 | 1867.01 | 3823.28 | 4980.99 | 9240.00 | after commit | 1450.54 | 1880.61 | 3962.87 | 5429.60 | 7542.86 | Closes https://github.com/facebook/rocksdb/pull/3800 Differential Revision: D7874245 Pulled By: ajkr fbshipit-source-id: 2e8bec781b3f7399246babd66395c88619534a17
2018-05-05 01:37:39 +02:00
if (range_del_iter != nullptr) {
s = range_del_iter->status();
}
if (s.ok()) {
const InternalKey* smallest = &file_meta.smallest;
const InternalKey* largest = &file_meta.largest;
if (smallest_compaction_key != nullptr) {
smallest = smallest_compaction_key;
}
if (largest_compaction_key != nullptr) {
largest = largest_compaction_key;
}
range_del_agg->AddTombstones(std::move(range_del_iter), smallest,
largest);
Recommit "Avoid adding tombstones of the same file to RangeDelAggregator multiple times" Summary: The origin commit #3635 will hurt performance for users who aren't using range deletions, because unneeded std::set operations, so it was reverted by commit 44653c7b7aabe821e671946e732dda7ae6b43d1b. (see #3672) To fix this, move the set to and add a check in , i.e., file will be added only if is non-nullptr. The db_bench command which find the performance regression: > ./db_bench --benchmarks=fillrandom,seekrandomwhilewriting --threads=1 --num=1000000 --reads=150000 --key_size=66 > --value_size=1262 --statistics=0 --compression_ratio=0.5 --histogram=1 --seek_nexts=1 --stats_per_interval=1 > --stats_interval_seconds=600 --max_background_flushes=4 --num_multi_db=1 --max_background_compactions=16 --seed=1522388277 > -write_buffer_size=1048576 --level0_file_num_compaction_trigger=10000 --compression_type=none Before and after the modification, I re-run this command on the machine, the results of are as follows: **fillrandom** Table | P50 | P75 | P99 | P99.9 | P99.99 | ---- | --- | --- | --- | ----- | ------ | before commit | 5.92 | 8.57 | 19.63 | 980.97 | 12196.00 | after commit | 5.91 | 8.55 | 19.34 | 965.56 | 13513.56 | **seekrandomwhilewriting** Table | P50 | P75 | P99 | P99.9 | P99.99 | ---- | --- | --- | --- | ----- | ------ | before commit | 1418.62 | 1867.01 | 3823.28 | 4980.99 | 9240.00 | after commit | 1450.54 | 1880.61 | 3962.87 | 5429.60 | 7542.86 | Closes https://github.com/facebook/rocksdb/pull/3800 Differential Revision: D7874245 Pulled By: ajkr fbshipit-source-id: 2e8bec781b3f7399246babd66395c88619534a17
2018-05-05 01:37:39 +02:00
}
Compaction Support for Range Deletion Summary: This diff introduces RangeDelAggregator, which takes ownership of iterators provided to it via AddTombstones(). The tombstones are organized in a two-level map (snapshot stripe -> begin key -> tombstone). Tombstone creation avoids data copy by holding Slices returned by the iterator, which remain valid thanks to pinning. For compaction, we create a hierarchical range tombstone iterator with structure matching the iterator over compaction input data. An aggregator based on that iterator is used by CompactionIterator to determine which keys are covered by range tombstones. In case of merge operand, the same aggregator is used by MergeHelper. Upon finishing each file in the compaction, relevant range tombstones are added to the output file's range tombstone metablock and file boundaries are updated accordingly. To check whether a key is covered by range tombstone, RangeDelAggregator::ShouldDelete() considers tombstones in the key's snapshot stripe. When this function is used outside of compaction, it also checks newer stripes, which can contain covering tombstones. Currently the intra-stripe check involves a linear scan; however, in the future we plan to collapse ranges within a stripe such that binary search can be used. RangeDelAggregator::AddToBuilder() adds all range tombstones in the table's key-range to a new table's range tombstone meta-block. Since range tombstones may fall in the gap between files, we may need to extend some files' key-ranges. The strategy is (1) first file extends as far left as possible and other files do not extend left, (2) all files extend right until either the start of the next file or the end of the last range tombstone in the gap, whichever comes first. One other notable change is adding release/move semantics to ScopedArenaIterator such that it can be used to transfer ownership of an arena-allocated iterator, similar to how unique_ptr is used for malloc'd data. Depends on D61473 Test Plan: compaction_iterator_test, mock_table, end-to-end tests in D63927 Reviewers: sdong, IslamAbdelRahman, wanning, yhchiang, lightmark Reviewed By: lightmark Subscribers: andrewkr, dhruba, leveldb Differential Revision: https://reviews.facebook.net/D62205
2016-10-18 21:04:56 +02:00
}
}
if (handle != nullptr) {
ReleaseHandle(handle);
}
if (!s.ok()) {
assert(result == nullptr);
result = NewErrorInternalIterator<Slice>(s, arena);
Compaction Support for Range Deletion Summary: This diff introduces RangeDelAggregator, which takes ownership of iterators provided to it via AddTombstones(). The tombstones are organized in a two-level map (snapshot stripe -> begin key -> tombstone). Tombstone creation avoids data copy by holding Slices returned by the iterator, which remain valid thanks to pinning. For compaction, we create a hierarchical range tombstone iterator with structure matching the iterator over compaction input data. An aggregator based on that iterator is used by CompactionIterator to determine which keys are covered by range tombstones. In case of merge operand, the same aggregator is used by MergeHelper. Upon finishing each file in the compaction, relevant range tombstones are added to the output file's range tombstone metablock and file boundaries are updated accordingly. To check whether a key is covered by range tombstone, RangeDelAggregator::ShouldDelete() considers tombstones in the key's snapshot stripe. When this function is used outside of compaction, it also checks newer stripes, which can contain covering tombstones. Currently the intra-stripe check involves a linear scan; however, in the future we plan to collapse ranges within a stripe such that binary search can be used. RangeDelAggregator::AddToBuilder() adds all range tombstones in the table's key-range to a new table's range tombstone meta-block. Since range tombstones may fall in the gap between files, we may need to extend some files' key-ranges. The strategy is (1) first file extends as far left as possible and other files do not extend left, (2) all files extend right until either the start of the next file or the end of the last range tombstone in the gap, whichever comes first. One other notable change is adding release/move semantics to ScopedArenaIterator such that it can be used to transfer ownership of an arena-allocated iterator, similar to how unique_ptr is used for malloc'd data. Depends on D61473 Test Plan: compaction_iterator_test, mock_table, end-to-end tests in D63927 Reviewers: sdong, IslamAbdelRahman, wanning, yhchiang, lightmark Reviewed By: lightmark Subscribers: andrewkr, dhruba, leveldb Differential Revision: https://reviews.facebook.net/D62205
2016-10-18 21:04:56 +02:00
}
return result;
}
Status TableCache::GetRangeTombstoneIterator(
const ReadOptions& options,
const InternalKeyComparator& internal_comparator,
const FileMetaData& file_meta,
std::unique_ptr<FragmentedRangeTombstoneIterator>* out_iter) {
const FileDescriptor& fd = file_meta.fd;
Status s;
TableReader* t = fd.table_reader;
Cache::Handle* handle = nullptr;
if (t == nullptr) {
s = FindTable(env_options_, internal_comparator, fd, &handle);
if (s.ok()) {
t = GetTableReaderFromHandle(handle);
}
}
if (s.ok()) {
out_iter->reset(t->NewRangeTombstoneIterator(options));
assert(out_iter);
}
return s;
}
#ifndef ROCKSDB_LITE
void TableCache::CreateRowCacheKeyPrefix(
const ReadOptions& options,
const FileDescriptor& fd, const Slice& internal_key,
GetContext* get_context, IterKey& row_cache_key) {
uint64_t fd_number = fd.GetNumber();
// We use the user key as cache key instead of the internal key,
// otherwise the whole cache would be invalidated every time the
// sequence key increases. However, to support caching snapshot
// reads, we append the sequence number (incremented by 1 to
// distinguish from 0) only in this case.
// If the snapshot is larger than the largest seqno in the file,
// all data should be exposed to the snapshot, so we treat it
// the same as there is no snapshot. The exception is that if
// a seq-checking callback is registered, some internal keys
// may still be filtered out.
uint64_t seq_no = 0;
// Maybe we can include the whole file ifsnapshot == fd.largest_seqno.
if (options.snapshot != nullptr &&
(get_context->has_callback() ||
static_cast_with_check<const SnapshotImpl, const Snapshot>(
options.snapshot)
->GetSequenceNumber() <= fd.largest_seqno)) {
// We should consider to use options.snapshot->GetSequenceNumber()
// instead of GetInternalKeySeqno(k), which will make the code
// easier to understand.
seq_no = 1 + GetInternalKeySeqno(internal_key);
}
// Compute row cache key.
row_cache_key.TrimAppend(row_cache_key.Size(), row_cache_id_.data(),
row_cache_id_.size());
AppendVarint64(&row_cache_key, fd_number);
AppendVarint64(&row_cache_key, seq_no);
}
bool TableCache::GetFromRowCache(
const Slice& user_key, IterKey& row_cache_key,
size_t prefix_size, GetContext* get_context) {
bool found = false;
row_cache_key.TrimAppend(prefix_size, user_key.data(),
user_key.size());
if (auto row_handle =
ioptions_.row_cache->Lookup(row_cache_key.GetUserKey())) {
// Cleanable routine to release the cache entry
Cleanable value_pinner;
auto release_cache_entry_func = [](void* cache_to_clean,
void* cache_handle) {
((Cache*)cache_to_clean)->Release((Cache::Handle*)cache_handle);
};
auto found_row_cache_entry = static_cast<const std::string*>(
ioptions_.row_cache->Value(row_handle));
// If it comes here value is located on the cache.
// found_row_cache_entry points to the value on cache,
// and value_pinner has cleanup procedure for the cached entry.
// After replayGetContextLog() returns, get_context.pinnable_slice_
// will point to cache entry buffer (or a copy based on that) and
// cleanup routine under value_pinner will be delegated to
// get_context.pinnable_slice_. Cache entry is released when
// get_context.pinnable_slice_ is reset.
value_pinner.RegisterCleanup(release_cache_entry_func,
ioptions_.row_cache.get(), row_handle);
replayGetContextLog(*found_row_cache_entry, user_key, get_context,
&value_pinner);
RecordTick(ioptions_.statistics, ROW_CACHE_HIT);
found = true;
} else {
RecordTick(ioptions_.statistics, ROW_CACHE_MISS);
}
return found;
}
#endif // ROCKSDB_LITE
Status TableCache::Get(const ReadOptions& options,
const InternalKeyComparator& internal_comparator,
const FileMetaData& file_meta, const Slice& k,
GetContext* get_context,
const SliceTransform* prefix_extractor,
HistogramImpl* file_read_hist, bool skip_filters,
int level) {
auto& fd = file_meta.fd;
std::string* row_cache_entry = nullptr;
bool done = false;
#ifndef ROCKSDB_LITE
IterKey row_cache_key;
std::string row_cache_entry_buffer;
// Check row cache if enabled. Since row cache does not currently store
// sequence numbers, we cannot use it if we need to fetch the sequence.
if (ioptions_.row_cache && !get_context->NeedToReadSequence()) {
auto user_key = ExtractUserKey(k);
CreateRowCacheKeyPrefix(options, fd, k, get_context, row_cache_key);
done = GetFromRowCache(user_key, row_cache_key, row_cache_key.Size(),
get_context);
if (!done) {
row_cache_entry = &row_cache_entry_buffer;
}
}
#endif // ROCKSDB_LITE
Status s;
TableReader* t = fd.table_reader;
Cache::Handle* handle = nullptr;
if (!done && s.ok()) {
if (t == nullptr) {
s = FindTable(
env_options_, internal_comparator, fd, &handle, prefix_extractor,
options.read_tier == kBlockCacheTier /* no_io */,
true /* record_read_stats */, file_read_hist, skip_filters, level);
if (s.ok()) {
t = GetTableReaderFromHandle(handle);
}
}
Use only "local" range tombstones during Get (#4449) Summary: Previously, range tombstones were accumulated from every level, which was necessary if a range tombstone in a higher level covered a key in a lower level. However, RangeDelAggregator::AddTombstones's complexity is based on the number of tombstones that are currently stored in it, which is wasteful in the Get case, where we only need to know the highest sequence number of range tombstones that cover the key from higher levels, and compute the highest covering sequence number at the current level. This change introduces this optimization, and removes the use of RangeDelAggregator from the Get path. In the benchmark results, the following command was used to initialize the database: ``` ./db_bench -db=/dev/shm/5k-rts -use_existing_db=false -benchmarks=filluniquerandom -write_buffer_size=1048576 -compression_type=lz4 -target_file_size_base=1048576 -max_bytes_for_level_base=4194304 -value_size=112 -key_size=16 -block_size=4096 -level_compaction_dynamic_level_bytes=true -num=5000000 -max_background_jobs=12 -benchmark_write_rate_limit=20971520 -range_tombstone_width=100 -writes_per_range_tombstone=100 -max_num_range_tombstones=50000 -bloom_bits=8 ``` ...and the following command was used to measure read throughput: ``` ./db_bench -db=/dev/shm/5k-rts/ -use_existing_db=true -benchmarks=readrandom -disable_auto_compactions=true -num=5000000 -reads=100000 -threads=32 ``` The filluniquerandom command was only run once, and the resulting database was used to measure read performance before and after the PR. Both binaries were compiled with `DEBUG_LEVEL=0`. Readrandom results before PR: ``` readrandom : 4.544 micros/op 220090 ops/sec; 16.9 MB/s (63103 of 100000 found) ``` Readrandom results after PR: ``` readrandom : 11.147 micros/op 89707 ops/sec; 6.9 MB/s (63103 of 100000 found) ``` So it's actually slower right now, but this PR paves the way for future optimizations (see #4493). ---- Pull Request resolved: https://github.com/facebook/rocksdb/pull/4449 Differential Revision: D10370575 Pulled By: abhimadan fbshipit-source-id: 9a2e152be1ef36969055c0e9eb4beb0d96c11f4d
2018-10-24 21:29:29 +02:00
SequenceNumber* max_covering_tombstone_seq =
get_context->max_covering_tombstone_seq();
if (s.ok() && max_covering_tombstone_seq != nullptr &&
!options.ignore_range_deletions) {
std::unique_ptr<FragmentedRangeTombstoneIterator> range_del_iter(
t->NewRangeTombstoneIterator(options));
if (range_del_iter != nullptr) {
*max_covering_tombstone_seq = std::max(
*max_covering_tombstone_seq,
range_del_iter->MaxCoveringTombstoneSeqnum(ExtractUserKey(k)));
}
}
if (s.ok()) {
get_context->SetReplayLog(row_cache_entry); // nullptr if no cache.
s = t->Get(options, k, get_context, prefix_extractor, skip_filters);
get_context->SetReplayLog(nullptr);
} else if (options.read_tier == kBlockCacheTier && s.IsIncomplete()) {
// Couldn't find Table in cache but treat as kFound if no_io set
get_context->MarkKeyMayExist();
s = Status::OK();
done = true;
}
}
#ifndef ROCKSDB_LITE
// Put the replay log in row cache only if something was found.
if (!done && s.ok() && row_cache_entry && !row_cache_entry->empty()) {
size_t charge =
row_cache_key.Size() + row_cache_entry->size() + sizeof(std::string);
void* row_ptr = new std::string(std::move(*row_cache_entry));
ioptions_.row_cache->Insert(row_cache_key.GetUserKey(), row_ptr, charge,
&DeleteEntry<std::string>);
}
#endif // ROCKSDB_LITE
if (handle != nullptr) {
ReleaseHandle(handle);
}
return s;
}
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
// Batched version of TableCache::MultiGet.
Status TableCache::MultiGet(const ReadOptions& options,
const InternalKeyComparator& internal_comparator,
const FileMetaData& file_meta,
const MultiGetContext::Range* mget_range,
const SliceTransform* prefix_extractor,
HistogramImpl* file_read_hist, bool skip_filters,
int level) {
auto& fd = file_meta.fd;
Status s;
TableReader* t = fd.table_reader;
Cache::Handle* handle = nullptr;
MultiGetRange table_range(*mget_range, mget_range->begin(), mget_range->end());
#ifndef ROCKSDB_LITE
autovector<std::string, MultiGetContext::MAX_BATCH_SIZE> row_cache_entries;
IterKey row_cache_key;
size_t row_cache_key_prefix_size = 0;
KeyContext& first_key = *table_range.begin();
bool lookup_row_cache = ioptions_.row_cache &&
!first_key.get_context->NeedToReadSequence();
// Check row cache if enabled. Since row cache does not currently store
// sequence numbers, we cannot use it if we need to fetch the sequence.
if (lookup_row_cache) {
GetContext* first_context = first_key.get_context;
CreateRowCacheKeyPrefix(options, fd, first_key.ikey, first_context,
row_cache_key);
row_cache_key_prefix_size = row_cache_key.Size();
for (auto miter = table_range.begin(); miter != table_range.end(); ++miter) {
const Slice& user_key = miter->ukey;;
GetContext* get_context = miter->get_context;
if (GetFromRowCache(user_key, row_cache_key, row_cache_key_prefix_size,
get_context)) {
table_range.SkipKey(miter);
} else {
row_cache_entries.emplace_back();
get_context->SetReplayLog(&(row_cache_entries.back()));
}
}
}
#endif // ROCKSDB_LITE
// Check that table_range is not empty. Its possible all keys may have been
// found in the row cache and thus the range may now be empty
if (s.ok() && !table_range.empty()) {
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 (t == nullptr) {
s = FindTable(
env_options_, internal_comparator, fd, &handle, prefix_extractor,
options.read_tier == kBlockCacheTier /* no_io */,
true /* record_read_stats */, file_read_hist, skip_filters, level);
if (s.ok()) {
t = GetTableReaderFromHandle(handle);
assert(t);
}
}
if (s.ok() && !options.ignore_range_deletions) {
std::unique_ptr<FragmentedRangeTombstoneIterator> range_del_iter(
t->NewRangeTombstoneIterator(options));
if (range_del_iter != nullptr) {
for (auto iter = table_range.begin(); iter != table_range.end();
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
++iter) {
SequenceNumber* max_covering_tombstone_seq =
iter->get_context->max_covering_tombstone_seq();
*max_covering_tombstone_seq =
std::max(*max_covering_tombstone_seq,
range_del_iter->MaxCoveringTombstoneSeqnum(iter->ukey));
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 (s.ok()) {
t->MultiGet(options, &table_range, prefix_extractor, skip_filters);
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
} else if (options.read_tier == kBlockCacheTier && s.IsIncomplete()) {
for (auto iter = table_range.begin(); iter != table_range.end(); ++iter) {
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
Status* status = iter->s;
if (status->IsIncomplete()) {
// Couldn't find Table in cache but treat as kFound if no_io set
iter->get_context->MarkKeyMayExist();
s = Status::OK();
}
}
}
}
#ifndef ROCKSDB_LITE
if (lookup_row_cache) {
size_t row_idx = 0;
for (auto miter = table_range.begin(); miter != table_range.end(); ++miter) {
std::string& row_cache_entry = row_cache_entries[row_idx++];
const Slice& user_key = miter->ukey;;
GetContext* get_context = miter->get_context;
get_context->SetReplayLog(nullptr);
// Compute row cache key.
row_cache_key.TrimAppend(row_cache_key_prefix_size, user_key.data(),
user_key.size());
// Put the replay log in row cache only if something was found.
if (s.ok() && !row_cache_entry.empty()) {
size_t charge =
row_cache_key.Size() + row_cache_entry.size() + sizeof(std::string);
void* row_ptr = new std::string(std::move(row_cache_entry));
ioptions_.row_cache->Insert(row_cache_key.GetUserKey(), row_ptr, charge,
&DeleteEntry<std::string>);
}
}
}
#endif // ROCKSDB_LITE
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 (handle != nullptr) {
ReleaseHandle(handle);
}
return s;
}
Status TableCache::GetTableProperties(
const EnvOptions& env_options,
const InternalKeyComparator& internal_comparator, const FileDescriptor& fd,
std::shared_ptr<const TableProperties>* properties,
const SliceTransform* prefix_extractor, bool no_io) {
Status s;
auto table_reader = fd.table_reader;
// table already been pre-loaded?
if (table_reader) {
*properties = table_reader->GetTableProperties();
return s;
}
Cache::Handle* table_handle = nullptr;
s = FindTable(env_options, internal_comparator, fd, &table_handle,
prefix_extractor, no_io);
if (!s.ok()) {
return s;
}
assert(table_handle);
auto table = GetTableReaderFromHandle(table_handle);
*properties = table->GetTableProperties();
ReleaseHandle(table_handle);
return s;
}
size_t TableCache::GetMemoryUsageByTableReader(
const EnvOptions& env_options,
const InternalKeyComparator& internal_comparator, const FileDescriptor& fd,
const SliceTransform* prefix_extractor) {
Status s;
auto table_reader = fd.table_reader;
// table already been pre-loaded?
if (table_reader) {
return table_reader->ApproximateMemoryUsage();
}
Cache::Handle* table_handle = nullptr;
s = FindTable(env_options, internal_comparator, fd, &table_handle,
prefix_extractor, true);
if (!s.ok()) {
return 0;
}
assert(table_handle);
auto table = GetTableReaderFromHandle(table_handle);
auto ret = table->ApproximateMemoryUsage();
ReleaseHandle(table_handle);
return ret;
}
void TableCache::Evict(Cache* cache, uint64_t file_number) {
cache->Erase(GetSliceForFileNumber(&file_number));
}
uint64_t TableCache::ApproximateOffsetOf(
const Slice& key, const FileDescriptor& fd, TableReaderCaller caller,
const InternalKeyComparator& internal_comparator,
const SliceTransform* prefix_extractor) {
uint64_t result = 0;
TableReader* table_reader = fd.table_reader;
Cache::Handle* table_handle = nullptr;
if (table_reader == nullptr) {
const bool for_compaction = (caller == TableReaderCaller::kCompaction);
Status s = FindTable(env_options_, internal_comparator, fd, &table_handle,
prefix_extractor, false /* no_io */,
!for_compaction /* record_read_stats */);
if (s.ok()) {
table_reader = GetTableReaderFromHandle(table_handle);
}
}
if (table_reader != nullptr) {
result = table_reader->ApproximateOffsetOf(key, caller);
}
if (table_handle != nullptr) {
ReleaseHandle(table_handle);
}
return result;
}
uint64_t TableCache::ApproximateSize(
const Slice& start, const Slice& end, const FileDescriptor& fd,
TableReaderCaller caller, const InternalKeyComparator& internal_comparator,
const SliceTransform* prefix_extractor) {
uint64_t result = 0;
TableReader* table_reader = fd.table_reader;
Cache::Handle* table_handle = nullptr;
if (table_reader == nullptr) {
const bool for_compaction = (caller == TableReaderCaller::kCompaction);
Status s = FindTable(env_options_, internal_comparator, fd, &table_handle,
prefix_extractor, false /* no_io */,
!for_compaction /* record_read_stats */);
if (s.ok()) {
table_reader = GetTableReaderFromHandle(table_handle);
}
}
if (table_reader != nullptr) {
result = table_reader->ApproximateSize(start, end, caller);
}
if (table_handle != nullptr) {
ReleaseHandle(table_handle);
}
return result;
}
} // namespace rocksdb