rocksdb/util/dynamic_bloom.h

215 lines
7.5 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).
#pragma once
MultiGet batching in memtable (#5818) Summary: RocksDB has a MultiGet() API that implements batched key lookup for higher performance (https://github.com/facebook/rocksdb/blob/master/include/rocksdb/db.h#L468). Currently, batching is implemented in BlockBasedTableReader::MultiGet() for SST file lookups. One of the ways it improves performance is by pipelining bloom filter lookups (by prefetching required cachelines for all the keys in the batch, and then doing the probe) and thus hiding the cache miss latency. The same concept can be extended to the memtable as well. This PR involves implementing a pipelined bloom filter lookup in DynamicBloom, and implementing MemTable::MultiGet() that can leverage it. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5818 Test Plan: Existing tests Performance Test: Ran the below command which fills up the memtable and makes sure there are no flushes and then call multiget. Ran it on master and on the new change and see atleast 1% performance improvement across all the test runs I did. Sometimes the improvement was upto 5%. TEST_TMPDIR=/data/users/$USER/benchmarks/feature/ numactl -C 10 ./db_bench -benchmarks="fillseq,multireadrandom" -num=600000 -compression_type="none" -level_compaction_dynamic_level_bytes -write_buffer_size=200000000 -target_file_size_base=200000000 -max_bytes_for_level_base=16777216 -reads=90000 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4 -statistics -memtable_whole_key_filtering=true -memtable_bloom_size_ratio=10 Differential Revision: D17578869 Pulled By: vjnadimpalli fbshipit-source-id: 23dc651d9bf49db11d22375bf435708875a1f192
2019-10-10 18:37:38 +02:00
#include <array>
#include <string>
#include "port/port.h"
MultiGet batching in memtable (#5818) Summary: RocksDB has a MultiGet() API that implements batched key lookup for higher performance (https://github.com/facebook/rocksdb/blob/master/include/rocksdb/db.h#L468). Currently, batching is implemented in BlockBasedTableReader::MultiGet() for SST file lookups. One of the ways it improves performance is by pipelining bloom filter lookups (by prefetching required cachelines for all the keys in the batch, and then doing the probe) and thus hiding the cache miss latency. The same concept can be extended to the memtable as well. This PR involves implementing a pipelined bloom filter lookup in DynamicBloom, and implementing MemTable::MultiGet() that can leverage it. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5818 Test Plan: Existing tests Performance Test: Ran the below command which fills up the memtable and makes sure there are no flushes and then call multiget. Ran it on master and on the new change and see atleast 1% performance improvement across all the test runs I did. Sometimes the improvement was upto 5%. TEST_TMPDIR=/data/users/$USER/benchmarks/feature/ numactl -C 10 ./db_bench -benchmarks="fillseq,multireadrandom" -num=600000 -compression_type="none" -level_compaction_dynamic_level_bytes -write_buffer_size=200000000 -target_file_size_base=200000000 -max_bytes_for_level_base=16777216 -reads=90000 -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 -statistics -memtable_whole_key_filtering=true -memtable_bloom_size_ratio=10 Differential Revision: D17578869 Pulled By: vjnadimpalli fbshipit-source-id: 23dc651d9bf49db11d22375bf435708875a1f192
2019-10-10 18:37:38 +02:00
#include "rocksdb/slice.h"
#include "table/multiget_context.h"
#include "util/hash.h"
#include <atomic>
#include <memory>
namespace ROCKSDB_NAMESPACE {
class Slice;
class Allocator;
class Logger;
// A Bloom filter intended only to be used in memory, never serialized in a way
// that could lead to schema incompatibility. Supports opt-in lock-free
// concurrent access.
Faster new DynamicBloom implementation (for memtable) (#5762) Summary: Since DynamicBloom is now only used in-memory, we're free to change it without schema compatibility issues. The new implementation is drawn from (with manifest permission) https://github.com/pdillinger/wormhashing/blob/303542a767437f56d8b66cea6ebecaac0e6a61e9/bloom_simulation_tests/foo.cc#L613 This has several speed advantages over the prior implementation: * Uses fastrange instead of % * Minimum logic to determine first (and all) probed memory addresses * (Major) Two probes per 64-bit memory fetch/write. * Very fast and effective (murmur-like) hash expansion/re-mixing. (At least on recent CPUs, integer multiplication is very cheap.) While a Bloom filter with 512-bit cache locality has about a 1.15x FP rate penalty (e.g. 0.84% to 0.97%), further restricting to two probes per 64 bits incurs an additional 1.12x FP rate penalty (e.g. 0.97% to 1.09%). Nevertheless, the unit tests show no "mediocre" FP rate samples, unlike the old implementation with more erratic FP rates. Especially for the memtable, we expect speed to outweigh somewhat higher FP rates. For example, a negative table query would have to be 1000x slower than a BF query to justify doubling BF query time to shave 10% off FP rate (working assumption around 1% FP rate). While that seems likely for SSTs, my data suggests a speed factor of roughly 50x for the memtable (vs. BF; ~1.5% lower write throughput when enabling memtable Bloom filter, after this change). Thus, it's probably not worth even 5% more time in the Bloom filter to shave off 1/10th of the Bloom FP rate, or 0.1% in absolute terms, and it's probably at least 20% slower to recoup that much FP rate from this new implementation. Because of this, we do not see a need for a 'locality' option that affects the MemTable Bloom filter and have decoupled the MemTable Bloom filter from Options::bloom_locality. Note that just 3% more memory to the Bloom filter (10.3 bits per key vs. just 10) is able to make up for the ~12% FP rate drop in the new implementation: [] # Nearly "ideal" FP-wise but reasonably fast cache-local implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out 10000000 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out time: 3.29372 sampled_fp_rate: 0.00985956 ... [] # Close match to this new implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out 10000000 6 10.3 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.10072 sampled_fp_rate: 0.00985655 ... [] # Old locality=1 implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out 10000000 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out time: 3.95472 sampled_fp_rate: 0.00988943 ... Also note the dramatic speed improvement vs. alternatives. -- Performance unit test: DynamicBloomTest.concurrent_with_perf is updated to report more precise timing data. (Measure running time of each thread, not just longest running thread, etc.) Results averaged over various sizes enabled with --enable_perf and 20 runs each; old dynamic bloom refers to locality=1, the faster of the old: old dynamic bloom, avg add latency = 65.6468 new dynamic bloom, avg add latency = 44.3809 old dynamic bloom, avg query latency = 50.6485 new dynamic bloom, avg query latency = 43.2186 old avg parallel add latency = 41.678 new avg parallel add latency = 24.5238 old avg parallel hit latency = 14.6322 new avg parallel hit latency = 12.3939 old avg parallel miss latency = 16.7289 new avg parallel miss latency = 12.2134 Tested on a dedicated 64-bit production machine at Facebook. Significant improvement all around. Despite now using std::atomic<uint64_t>, quick before-and-after test on a 32-bit machine (Intel Atom N270, released 2008) shows no regression in performance, in some cases modest improvement. -- Performance integration test (synthetic): with DEBUG_LEVEL=0, used TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=fillrandom,readmissing,readrandom,stats --num=2000000 and optionally with -memtable_whole_key_filtering -memtable_bloom_size_ratio=0.01 300 runs each configuration. Write throughput change by enabling memtable bloom: Old locality=0: -3.06% Old locality=1: -2.37% New: -1.50% conclusion -> seems to substantially close the gap Readmissing throughput change by enabling memtable bloom: Old locality=0: +34.47% Old locality=1: +34.80% New: +33.25% conclusion -> maybe a small new penalty from FP rate Readrandom throughput change by enabling memtable bloom: Old locality=0: +31.54% Old locality=1: +31.13% New: +30.60% conclusion -> maybe also from FP rate (after memtable flush) -- Another conclusion we can draw from this new implementation is that the existing 32-bit hash function is not inherently crippling the Bloom filter speed or accuracy, below about 5 million keys. For speed, the implementation is essentially the same whether starting with 32-bits or 64-bits of hash; it just determines whether the first multiplication after fastrange is a pseudorandom expansion or needed re-mix. Note that this multiplication can occur while memory is fetching. For accuracy, in a standard configuration, you need about 5 million keys before you have about a 1.1x FP penalty due to using a 32-bit hash vs. 64-bit: [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.52069 sampled_fp_rate: 0.0118267 ... [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out time: 2.43871 sampled_fp_rate: 0.0109059 Pull Request resolved: https://github.com/facebook/rocksdb/pull/5762 Differential Revision: D17214194 Pulled By: pdillinger fbshipit-source-id: ad9da031772e985fd6b62a0e1db8e81892520595
2019-09-05 23:57:39 +02:00
//
// This implementation is also intended for applications generally preferring
// speed vs. maximum accuracy: roughly 0.9x BF op latency for 1.1x FP rate.
// For 1% FP rate, that means that the latency of a look-up triggered by an FP
// should be less than roughly 100x the cost of a Bloom filter op.
//
// For simplicity and performance, the current implementation requires
// num_probes to be a multiple of two and <= 10.
//
class DynamicBloom {
public:
// allocator: pass allocator to bloom filter, hence trace the usage of memory
// total_bits: fixed total bits for the bloom
// num_probes: number of hash probes for a single key
// hash_func: customized hash function
// huge_page_tlb_size: if >0, try to allocate bloom bytes from huge page TLB
// within this page size. Need to reserve huge pages for
// it to be allocated, like:
// sysctl -w vm.nr_hugepages=20
// See linux doc Documentation/vm/hugetlbpage.txt
explicit DynamicBloom(Allocator* allocator, uint32_t total_bits,
uint32_t num_probes = 6, size_t huge_page_tlb_size = 0,
Logger* logger = nullptr);
~DynamicBloom() {}
// Assuming single threaded access to this function.
void Add(const Slice& key);
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
// Like Add, but may be called concurrent with other functions.
void AddConcurrently(const Slice& key);
// Assuming single threaded access to this function.
void AddHash(uint32_t hash);
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
// Like AddHash, but may be called concurrent with other functions.
void AddHashConcurrently(uint32_t hash);
// Multithreaded access to this function is OK
bool MayContain(const Slice& key) const;
Fix major bug with MultiGet, DeleteRange, and memtable Bloom (#9453) Summary: MemTable::MultiGet was not considering range tombstones before querying Bloom filter. This means range tombstones would be skipped for keys (or prefixes) with no other entries in the memtable. This could cause old values for a key (in SST files) to still show up until the range tombstone covering it has been flushed. This is fixed by essentially disabling the memtable Bloom filter when there are any range tombstones. (This could be better optimized in the future, but good enough for now.) Did some other cleanup/optimization in the same code to (more than) offset the cost of checking on range tombstones in more cases. There is now notable improvement when memtable_whole_key_filtering and prefix_extractor are used together (unusual), and this makes MultiGet closer to the Get implementation. Pull Request resolved: https://github.com/facebook/rocksdb/pull/9453 Test Plan: new unit test added. Added memtable Bloom to crash test. Performance testing -------------------- Build WAL-only DB (recovers to memtable): ``` TEST_TMPDIR=/dev/shm/rocksdb ./db_bench -benchmarks=fillrandom -num=1000000 -write_buffer_size=250000000 ``` Query test command, to maximize sensitivity to the changed code: ``` TEST_TMPDIR=/dev/shm/rocksdb ./db_bench -use_existing_db -readonly -benchmarks=multireadrandom -num=10000000 -write_buffer_size=250000000 -memtable_bloom_size_ratio=0.015 -multiread_batched -batch_size=24 -threads=8 -memtable_whole_key_filtering=$MWKF -prefix_size=$PXS ``` (Note -num here is 10x larger for mostly memtable misses) Before & after run simultaneously, average over 10 iterations per data point, ops/sec. MWKF=0 PXS=0 (Bloom disabled) Before: 5724844 After: 6722066 MWKF=0 PXS=7 (prefixes hardly unique; Bloom not useful) Before: 9981319 After: 10237990 MWKF=0 PXS=8 (prefixes unique; Bloom useful) Before: 12081715 After: 12117603 MWKF=1 PXS=0 (whole key Bloom useful) Before: 11944354 After: 12096085 MWKF=1 PXS=7 (whole key Bloom useful in new version; prefixes not useful in old version) Before: 9444299 After: 11826029 MWKF=1 PXS=7 (whole key Bloom useful in new version; prefixes useful in old version) Before: 11784465 After: 11778591 Only in this last case is the 'before' *slightly* faster, perhaps because hashing prefixes is slightly faster than hashing whole keys. Otherwise, 'after' is faster. Reviewed By: ajkr Differential Revision: D33805025 Pulled By: pdillinger fbshipit-source-id: 597523cae4f4eafdf6ae6bb2bc6cb46f83b017bf
2022-01-27 23:53:39 +01:00
void MayContain(int num_keys, Slice* keys, bool* may_match) const;
MultiGet batching in memtable (#5818) Summary: RocksDB has a MultiGet() API that implements batched key lookup for higher performance (https://github.com/facebook/rocksdb/blob/master/include/rocksdb/db.h#L468). Currently, batching is implemented in BlockBasedTableReader::MultiGet() for SST file lookups. One of the ways it improves performance is by pipelining bloom filter lookups (by prefetching required cachelines for all the keys in the batch, and then doing the probe) and thus hiding the cache miss latency. The same concept can be extended to the memtable as well. This PR involves implementing a pipelined bloom filter lookup in DynamicBloom, and implementing MemTable::MultiGet() that can leverage it. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5818 Test Plan: Existing tests Performance Test: Ran the below command which fills up the memtable and makes sure there are no flushes and then call multiget. Ran it on master and on the new change and see atleast 1% performance improvement across all the test runs I did. Sometimes the improvement was upto 5%. TEST_TMPDIR=/data/users/$USER/benchmarks/feature/ numactl -C 10 ./db_bench -benchmarks="fillseq,multireadrandom" -num=600000 -compression_type="none" -level_compaction_dynamic_level_bytes -write_buffer_size=200000000 -target_file_size_base=200000000 -max_bytes_for_level_base=16777216 -reads=90000 -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 -statistics -memtable_whole_key_filtering=true -memtable_bloom_size_ratio=10 Differential Revision: D17578869 Pulled By: vjnadimpalli fbshipit-source-id: 23dc651d9bf49db11d22375bf435708875a1f192
2019-10-10 18:37:38 +02:00
// Multithreaded access to this function is OK
bool MayContainHash(uint32_t hash) const;
void Prefetch(uint32_t h);
private:
Faster new DynamicBloom implementation (for memtable) (#5762) Summary: Since DynamicBloom is now only used in-memory, we're free to change it without schema compatibility issues. The new implementation is drawn from (with manifest permission) https://github.com/pdillinger/wormhashing/blob/303542a767437f56d8b66cea6ebecaac0e6a61e9/bloom_simulation_tests/foo.cc#L613 This has several speed advantages over the prior implementation: * Uses fastrange instead of % * Minimum logic to determine first (and all) probed memory addresses * (Major) Two probes per 64-bit memory fetch/write. * Very fast and effective (murmur-like) hash expansion/re-mixing. (At least on recent CPUs, integer multiplication is very cheap.) While a Bloom filter with 512-bit cache locality has about a 1.15x FP rate penalty (e.g. 0.84% to 0.97%), further restricting to two probes per 64 bits incurs an additional 1.12x FP rate penalty (e.g. 0.97% to 1.09%). Nevertheless, the unit tests show no "mediocre" FP rate samples, unlike the old implementation with more erratic FP rates. Especially for the memtable, we expect speed to outweigh somewhat higher FP rates. For example, a negative table query would have to be 1000x slower than a BF query to justify doubling BF query time to shave 10% off FP rate (working assumption around 1% FP rate). While that seems likely for SSTs, my data suggests a speed factor of roughly 50x for the memtable (vs. BF; ~1.5% lower write throughput when enabling memtable Bloom filter, after this change). Thus, it's probably not worth even 5% more time in the Bloom filter to shave off 1/10th of the Bloom FP rate, or 0.1% in absolute terms, and it's probably at least 20% slower to recoup that much FP rate from this new implementation. Because of this, we do not see a need for a 'locality' option that affects the MemTable Bloom filter and have decoupled the MemTable Bloom filter from Options::bloom_locality. Note that just 3% more memory to the Bloom filter (10.3 bits per key vs. just 10) is able to make up for the ~12% FP rate drop in the new implementation: [] # Nearly "ideal" FP-wise but reasonably fast cache-local implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out 10000000 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out time: 3.29372 sampled_fp_rate: 0.00985956 ... [] # Close match to this new implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out 10000000 6 10.3 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.10072 sampled_fp_rate: 0.00985655 ... [] # Old locality=1 implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out 10000000 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out time: 3.95472 sampled_fp_rate: 0.00988943 ... Also note the dramatic speed improvement vs. alternatives. -- Performance unit test: DynamicBloomTest.concurrent_with_perf is updated to report more precise timing data. (Measure running time of each thread, not just longest running thread, etc.) Results averaged over various sizes enabled with --enable_perf and 20 runs each; old dynamic bloom refers to locality=1, the faster of the old: old dynamic bloom, avg add latency = 65.6468 new dynamic bloom, avg add latency = 44.3809 old dynamic bloom, avg query latency = 50.6485 new dynamic bloom, avg query latency = 43.2186 old avg parallel add latency = 41.678 new avg parallel add latency = 24.5238 old avg parallel hit latency = 14.6322 new avg parallel hit latency = 12.3939 old avg parallel miss latency = 16.7289 new avg parallel miss latency = 12.2134 Tested on a dedicated 64-bit production machine at Facebook. Significant improvement all around. Despite now using std::atomic<uint64_t>, quick before-and-after test on a 32-bit machine (Intel Atom N270, released 2008) shows no regression in performance, in some cases modest improvement. -- Performance integration test (synthetic): with DEBUG_LEVEL=0, used TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=fillrandom,readmissing,readrandom,stats --num=2000000 and optionally with -memtable_whole_key_filtering -memtable_bloom_size_ratio=0.01 300 runs each configuration. Write throughput change by enabling memtable bloom: Old locality=0: -3.06% Old locality=1: -2.37% New: -1.50% conclusion -> seems to substantially close the gap Readmissing throughput change by enabling memtable bloom: Old locality=0: +34.47% Old locality=1: +34.80% New: +33.25% conclusion -> maybe a small new penalty from FP rate Readrandom throughput change by enabling memtable bloom: Old locality=0: +31.54% Old locality=1: +31.13% New: +30.60% conclusion -> maybe also from FP rate (after memtable flush) -- Another conclusion we can draw from this new implementation is that the existing 32-bit hash function is not inherently crippling the Bloom filter speed or accuracy, below about 5 million keys. For speed, the implementation is essentially the same whether starting with 32-bits or 64-bits of hash; it just determines whether the first multiplication after fastrange is a pseudorandom expansion or needed re-mix. Note that this multiplication can occur while memory is fetching. For accuracy, in a standard configuration, you need about 5 million keys before you have about a 1.1x FP penalty due to using a 32-bit hash vs. 64-bit: [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.52069 sampled_fp_rate: 0.0118267 ... [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out time: 2.43871 sampled_fp_rate: 0.0109059 Pull Request resolved: https://github.com/facebook/rocksdb/pull/5762 Differential Revision: D17214194 Pulled By: pdillinger fbshipit-source-id: ad9da031772e985fd6b62a0e1db8e81892520595
2019-09-05 23:57:39 +02:00
// Length of the structure, in 64-bit words. For this structure, "word"
// will always refer to 64-bit words.
uint32_t kLen;
// We make the k probes in pairs, two for each 64-bit read/write. Thus,
// this stores k/2, the number of words to double-probe.
const uint32_t kNumDoubleProbes;
Faster new DynamicBloom implementation (for memtable) (#5762) Summary: Since DynamicBloom is now only used in-memory, we're free to change it without schema compatibility issues. The new implementation is drawn from (with manifest permission) https://github.com/pdillinger/wormhashing/blob/303542a767437f56d8b66cea6ebecaac0e6a61e9/bloom_simulation_tests/foo.cc#L613 This has several speed advantages over the prior implementation: * Uses fastrange instead of % * Minimum logic to determine first (and all) probed memory addresses * (Major) Two probes per 64-bit memory fetch/write. * Very fast and effective (murmur-like) hash expansion/re-mixing. (At least on recent CPUs, integer multiplication is very cheap.) While a Bloom filter with 512-bit cache locality has about a 1.15x FP rate penalty (e.g. 0.84% to 0.97%), further restricting to two probes per 64 bits incurs an additional 1.12x FP rate penalty (e.g. 0.97% to 1.09%). Nevertheless, the unit tests show no "mediocre" FP rate samples, unlike the old implementation with more erratic FP rates. Especially for the memtable, we expect speed to outweigh somewhat higher FP rates. For example, a negative table query would have to be 1000x slower than a BF query to justify doubling BF query time to shave 10% off FP rate (working assumption around 1% FP rate). While that seems likely for SSTs, my data suggests a speed factor of roughly 50x for the memtable (vs. BF; ~1.5% lower write throughput when enabling memtable Bloom filter, after this change). Thus, it's probably not worth even 5% more time in the Bloom filter to shave off 1/10th of the Bloom FP rate, or 0.1% in absolute terms, and it's probably at least 20% slower to recoup that much FP rate from this new implementation. Because of this, we do not see a need for a 'locality' option that affects the MemTable Bloom filter and have decoupled the MemTable Bloom filter from Options::bloom_locality. Note that just 3% more memory to the Bloom filter (10.3 bits per key vs. just 10) is able to make up for the ~12% FP rate drop in the new implementation: [] # Nearly "ideal" FP-wise but reasonably fast cache-local implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out 10000000 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out time: 3.29372 sampled_fp_rate: 0.00985956 ... [] # Close match to this new implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out 10000000 6 10.3 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.10072 sampled_fp_rate: 0.00985655 ... [] # Old locality=1 implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out 10000000 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out time: 3.95472 sampled_fp_rate: 0.00988943 ... Also note the dramatic speed improvement vs. alternatives. -- Performance unit test: DynamicBloomTest.concurrent_with_perf is updated to report more precise timing data. (Measure running time of each thread, not just longest running thread, etc.) Results averaged over various sizes enabled with --enable_perf and 20 runs each; old dynamic bloom refers to locality=1, the faster of the old: old dynamic bloom, avg add latency = 65.6468 new dynamic bloom, avg add latency = 44.3809 old dynamic bloom, avg query latency = 50.6485 new dynamic bloom, avg query latency = 43.2186 old avg parallel add latency = 41.678 new avg parallel add latency = 24.5238 old avg parallel hit latency = 14.6322 new avg parallel hit latency = 12.3939 old avg parallel miss latency = 16.7289 new avg parallel miss latency = 12.2134 Tested on a dedicated 64-bit production machine at Facebook. Significant improvement all around. Despite now using std::atomic<uint64_t>, quick before-and-after test on a 32-bit machine (Intel Atom N270, released 2008) shows no regression in performance, in some cases modest improvement. -- Performance integration test (synthetic): with DEBUG_LEVEL=0, used TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=fillrandom,readmissing,readrandom,stats --num=2000000 and optionally with -memtable_whole_key_filtering -memtable_bloom_size_ratio=0.01 300 runs each configuration. Write throughput change by enabling memtable bloom: Old locality=0: -3.06% Old locality=1: -2.37% New: -1.50% conclusion -> seems to substantially close the gap Readmissing throughput change by enabling memtable bloom: Old locality=0: +34.47% Old locality=1: +34.80% New: +33.25% conclusion -> maybe a small new penalty from FP rate Readrandom throughput change by enabling memtable bloom: Old locality=0: +31.54% Old locality=1: +31.13% New: +30.60% conclusion -> maybe also from FP rate (after memtable flush) -- Another conclusion we can draw from this new implementation is that the existing 32-bit hash function is not inherently crippling the Bloom filter speed or accuracy, below about 5 million keys. For speed, the implementation is essentially the same whether starting with 32-bits or 64-bits of hash; it just determines whether the first multiplication after fastrange is a pseudorandom expansion or needed re-mix. Note that this multiplication can occur while memory is fetching. For accuracy, in a standard configuration, you need about 5 million keys before you have about a 1.1x FP penalty due to using a 32-bit hash vs. 64-bit: [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.52069 sampled_fp_rate: 0.0118267 ... [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out time: 2.43871 sampled_fp_rate: 0.0109059 Pull Request resolved: https://github.com/facebook/rocksdb/pull/5762 Differential Revision: D17214194 Pulled By: pdillinger fbshipit-source-id: ad9da031772e985fd6b62a0e1db8e81892520595
2019-09-05 23:57:39 +02:00
std::atomic<uint64_t>* data_;
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
// or_func(ptr, mask) should effect *ptr |= mask with the appropriate
// concurrency safety, working with bytes.
template <typename OrFunc>
void AddHash(uint32_t hash, const OrFunc& or_func);
MultiGet batching in memtable (#5818) Summary: RocksDB has a MultiGet() API that implements batched key lookup for higher performance (https://github.com/facebook/rocksdb/blob/master/include/rocksdb/db.h#L468). Currently, batching is implemented in BlockBasedTableReader::MultiGet() for SST file lookups. One of the ways it improves performance is by pipelining bloom filter lookups (by prefetching required cachelines for all the keys in the batch, and then doing the probe) and thus hiding the cache miss latency. The same concept can be extended to the memtable as well. This PR involves implementing a pipelined bloom filter lookup in DynamicBloom, and implementing MemTable::MultiGet() that can leverage it. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5818 Test Plan: Existing tests Performance Test: Ran the below command which fills up the memtable and makes sure there are no flushes and then call multiget. Ran it on master and on the new change and see atleast 1% performance improvement across all the test runs I did. Sometimes the improvement was upto 5%. TEST_TMPDIR=/data/users/$USER/benchmarks/feature/ numactl -C 10 ./db_bench -benchmarks="fillseq,multireadrandom" -num=600000 -compression_type="none" -level_compaction_dynamic_level_bytes -write_buffer_size=200000000 -target_file_size_base=200000000 -max_bytes_for_level_base=16777216 -reads=90000 -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 -statistics -memtable_whole_key_filtering=true -memtable_bloom_size_ratio=10 Differential Revision: D17578869 Pulled By: vjnadimpalli fbshipit-source-id: 23dc651d9bf49db11d22375bf435708875a1f192
2019-10-10 18:37:38 +02:00
bool DoubleProbe(uint32_t h32, size_t a) const;
};
inline void DynamicBloom::Add(const Slice& key) { AddHash(BloomHash(key)); }
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
inline void DynamicBloom::AddConcurrently(const Slice& key) {
AddHashConcurrently(BloomHash(key));
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
}
inline void DynamicBloom::AddHash(uint32_t hash) {
Faster new DynamicBloom implementation (for memtable) (#5762) Summary: Since DynamicBloom is now only used in-memory, we're free to change it without schema compatibility issues. The new implementation is drawn from (with manifest permission) https://github.com/pdillinger/wormhashing/blob/303542a767437f56d8b66cea6ebecaac0e6a61e9/bloom_simulation_tests/foo.cc#L613 This has several speed advantages over the prior implementation: * Uses fastrange instead of % * Minimum logic to determine first (and all) probed memory addresses * (Major) Two probes per 64-bit memory fetch/write. * Very fast and effective (murmur-like) hash expansion/re-mixing. (At least on recent CPUs, integer multiplication is very cheap.) While a Bloom filter with 512-bit cache locality has about a 1.15x FP rate penalty (e.g. 0.84% to 0.97%), further restricting to two probes per 64 bits incurs an additional 1.12x FP rate penalty (e.g. 0.97% to 1.09%). Nevertheless, the unit tests show no "mediocre" FP rate samples, unlike the old implementation with more erratic FP rates. Especially for the memtable, we expect speed to outweigh somewhat higher FP rates. For example, a negative table query would have to be 1000x slower than a BF query to justify doubling BF query time to shave 10% off FP rate (working assumption around 1% FP rate). While that seems likely for SSTs, my data suggests a speed factor of roughly 50x for the memtable (vs. BF; ~1.5% lower write throughput when enabling memtable Bloom filter, after this change). Thus, it's probably not worth even 5% more time in the Bloom filter to shave off 1/10th of the Bloom FP rate, or 0.1% in absolute terms, and it's probably at least 20% slower to recoup that much FP rate from this new implementation. Because of this, we do not see a need for a 'locality' option that affects the MemTable Bloom filter and have decoupled the MemTable Bloom filter from Options::bloom_locality. Note that just 3% more memory to the Bloom filter (10.3 bits per key vs. just 10) is able to make up for the ~12% FP rate drop in the new implementation: [] # Nearly "ideal" FP-wise but reasonably fast cache-local implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out 10000000 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out time: 3.29372 sampled_fp_rate: 0.00985956 ... [] # Close match to this new implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out 10000000 6 10.3 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.10072 sampled_fp_rate: 0.00985655 ... [] # Old locality=1 implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out 10000000 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out time: 3.95472 sampled_fp_rate: 0.00988943 ... Also note the dramatic speed improvement vs. alternatives. -- Performance unit test: DynamicBloomTest.concurrent_with_perf is updated to report more precise timing data. (Measure running time of each thread, not just longest running thread, etc.) Results averaged over various sizes enabled with --enable_perf and 20 runs each; old dynamic bloom refers to locality=1, the faster of the old: old dynamic bloom, avg add latency = 65.6468 new dynamic bloom, avg add latency = 44.3809 old dynamic bloom, avg query latency = 50.6485 new dynamic bloom, avg query latency = 43.2186 old avg parallel add latency = 41.678 new avg parallel add latency = 24.5238 old avg parallel hit latency = 14.6322 new avg parallel hit latency = 12.3939 old avg parallel miss latency = 16.7289 new avg parallel miss latency = 12.2134 Tested on a dedicated 64-bit production machine at Facebook. Significant improvement all around. Despite now using std::atomic<uint64_t>, quick before-and-after test on a 32-bit machine (Intel Atom N270, released 2008) shows no regression in performance, in some cases modest improvement. -- Performance integration test (synthetic): with DEBUG_LEVEL=0, used TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=fillrandom,readmissing,readrandom,stats --num=2000000 and optionally with -memtable_whole_key_filtering -memtable_bloom_size_ratio=0.01 300 runs each configuration. Write throughput change by enabling memtable bloom: Old locality=0: -3.06% Old locality=1: -2.37% New: -1.50% conclusion -> seems to substantially close the gap Readmissing throughput change by enabling memtable bloom: Old locality=0: +34.47% Old locality=1: +34.80% New: +33.25% conclusion -> maybe a small new penalty from FP rate Readrandom throughput change by enabling memtable bloom: Old locality=0: +31.54% Old locality=1: +31.13% New: +30.60% conclusion -> maybe also from FP rate (after memtable flush) -- Another conclusion we can draw from this new implementation is that the existing 32-bit hash function is not inherently crippling the Bloom filter speed or accuracy, below about 5 million keys. For speed, the implementation is essentially the same whether starting with 32-bits or 64-bits of hash; it just determines whether the first multiplication after fastrange is a pseudorandom expansion or needed re-mix. Note that this multiplication can occur while memory is fetching. For accuracy, in a standard configuration, you need about 5 million keys before you have about a 1.1x FP penalty due to using a 32-bit hash vs. 64-bit: [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.52069 sampled_fp_rate: 0.0118267 ... [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out time: 2.43871 sampled_fp_rate: 0.0109059 Pull Request resolved: https://github.com/facebook/rocksdb/pull/5762 Differential Revision: D17214194 Pulled By: pdillinger fbshipit-source-id: ad9da031772e985fd6b62a0e1db8e81892520595
2019-09-05 23:57:39 +02:00
AddHash(hash, [](std::atomic<uint64_t>* ptr, uint64_t mask) {
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
ptr->store(ptr->load(std::memory_order_relaxed) | mask,
std::memory_order_relaxed);
});
}
inline void DynamicBloom::AddHashConcurrently(uint32_t hash) {
Faster new DynamicBloom implementation (for memtable) (#5762) Summary: Since DynamicBloom is now only used in-memory, we're free to change it without schema compatibility issues. The new implementation is drawn from (with manifest permission) https://github.com/pdillinger/wormhashing/blob/303542a767437f56d8b66cea6ebecaac0e6a61e9/bloom_simulation_tests/foo.cc#L613 This has several speed advantages over the prior implementation: * Uses fastrange instead of % * Minimum logic to determine first (and all) probed memory addresses * (Major) Two probes per 64-bit memory fetch/write. * Very fast and effective (murmur-like) hash expansion/re-mixing. (At least on recent CPUs, integer multiplication is very cheap.) While a Bloom filter with 512-bit cache locality has about a 1.15x FP rate penalty (e.g. 0.84% to 0.97%), further restricting to two probes per 64 bits incurs an additional 1.12x FP rate penalty (e.g. 0.97% to 1.09%). Nevertheless, the unit tests show no "mediocre" FP rate samples, unlike the old implementation with more erratic FP rates. Especially for the memtable, we expect speed to outweigh somewhat higher FP rates. For example, a negative table query would have to be 1000x slower than a BF query to justify doubling BF query time to shave 10% off FP rate (working assumption around 1% FP rate). While that seems likely for SSTs, my data suggests a speed factor of roughly 50x for the memtable (vs. BF; ~1.5% lower write throughput when enabling memtable Bloom filter, after this change). Thus, it's probably not worth even 5% more time in the Bloom filter to shave off 1/10th of the Bloom FP rate, or 0.1% in absolute terms, and it's probably at least 20% slower to recoup that much FP rate from this new implementation. Because of this, we do not see a need for a 'locality' option that affects the MemTable Bloom filter and have decoupled the MemTable Bloom filter from Options::bloom_locality. Note that just 3% more memory to the Bloom filter (10.3 bits per key vs. just 10) is able to make up for the ~12% FP rate drop in the new implementation: [] # Nearly "ideal" FP-wise but reasonably fast cache-local implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out 10000000 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out time: 3.29372 sampled_fp_rate: 0.00985956 ... [] # Close match to this new implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out 10000000 6 10.3 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.10072 sampled_fp_rate: 0.00985655 ... [] # Old locality=1 implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out 10000000 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out time: 3.95472 sampled_fp_rate: 0.00988943 ... Also note the dramatic speed improvement vs. alternatives. -- Performance unit test: DynamicBloomTest.concurrent_with_perf is updated to report more precise timing data. (Measure running time of each thread, not just longest running thread, etc.) Results averaged over various sizes enabled with --enable_perf and 20 runs each; old dynamic bloom refers to locality=1, the faster of the old: old dynamic bloom, avg add latency = 65.6468 new dynamic bloom, avg add latency = 44.3809 old dynamic bloom, avg query latency = 50.6485 new dynamic bloom, avg query latency = 43.2186 old avg parallel add latency = 41.678 new avg parallel add latency = 24.5238 old avg parallel hit latency = 14.6322 new avg parallel hit latency = 12.3939 old avg parallel miss latency = 16.7289 new avg parallel miss latency = 12.2134 Tested on a dedicated 64-bit production machine at Facebook. Significant improvement all around. Despite now using std::atomic<uint64_t>, quick before-and-after test on a 32-bit machine (Intel Atom N270, released 2008) shows no regression in performance, in some cases modest improvement. -- Performance integration test (synthetic): with DEBUG_LEVEL=0, used TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=fillrandom,readmissing,readrandom,stats --num=2000000 and optionally with -memtable_whole_key_filtering -memtable_bloom_size_ratio=0.01 300 runs each configuration. Write throughput change by enabling memtable bloom: Old locality=0: -3.06% Old locality=1: -2.37% New: -1.50% conclusion -> seems to substantially close the gap Readmissing throughput change by enabling memtable bloom: Old locality=0: +34.47% Old locality=1: +34.80% New: +33.25% conclusion -> maybe a small new penalty from FP rate Readrandom throughput change by enabling memtable bloom: Old locality=0: +31.54% Old locality=1: +31.13% New: +30.60% conclusion -> maybe also from FP rate (after memtable flush) -- Another conclusion we can draw from this new implementation is that the existing 32-bit hash function is not inherently crippling the Bloom filter speed or accuracy, below about 5 million keys. For speed, the implementation is essentially the same whether starting with 32-bits or 64-bits of hash; it just determines whether the first multiplication after fastrange is a pseudorandom expansion or needed re-mix. Note that this multiplication can occur while memory is fetching. For accuracy, in a standard configuration, you need about 5 million keys before you have about a 1.1x FP penalty due to using a 32-bit hash vs. 64-bit: [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.52069 sampled_fp_rate: 0.0118267 ... [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out time: 2.43871 sampled_fp_rate: 0.0109059 Pull Request resolved: https://github.com/facebook/rocksdb/pull/5762 Differential Revision: D17214194 Pulled By: pdillinger fbshipit-source-id: ad9da031772e985fd6b62a0e1db8e81892520595
2019-09-05 23:57:39 +02:00
AddHash(hash, [](std::atomic<uint64_t>* ptr, uint64_t mask) {
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
// Happens-before between AddHash and MaybeContains is handled by
// access to versions_->LastSequence(), so all we have to do here is
// avoid races (so we don't give the compiler a license to mess up
// our code) and not lose bits. std::memory_order_relaxed is enough
// for that.
if ((mask & ptr->load(std::memory_order_relaxed)) != mask) {
ptr->fetch_or(mask, std::memory_order_relaxed);
}
});
}
inline bool DynamicBloom::MayContain(const Slice& key) const {
return (MayContainHash(BloomHash(key)));
}
Fix major bug with MultiGet, DeleteRange, and memtable Bloom (#9453) Summary: MemTable::MultiGet was not considering range tombstones before querying Bloom filter. This means range tombstones would be skipped for keys (or prefixes) with no other entries in the memtable. This could cause old values for a key (in SST files) to still show up until the range tombstone covering it has been flushed. This is fixed by essentially disabling the memtable Bloom filter when there are any range tombstones. (This could be better optimized in the future, but good enough for now.) Did some other cleanup/optimization in the same code to (more than) offset the cost of checking on range tombstones in more cases. There is now notable improvement when memtable_whole_key_filtering and prefix_extractor are used together (unusual), and this makes MultiGet closer to the Get implementation. Pull Request resolved: https://github.com/facebook/rocksdb/pull/9453 Test Plan: new unit test added. Added memtable Bloom to crash test. Performance testing -------------------- Build WAL-only DB (recovers to memtable): ``` TEST_TMPDIR=/dev/shm/rocksdb ./db_bench -benchmarks=fillrandom -num=1000000 -write_buffer_size=250000000 ``` Query test command, to maximize sensitivity to the changed code: ``` TEST_TMPDIR=/dev/shm/rocksdb ./db_bench -use_existing_db -readonly -benchmarks=multireadrandom -num=10000000 -write_buffer_size=250000000 -memtable_bloom_size_ratio=0.015 -multiread_batched -batch_size=24 -threads=8 -memtable_whole_key_filtering=$MWKF -prefix_size=$PXS ``` (Note -num here is 10x larger for mostly memtable misses) Before & after run simultaneously, average over 10 iterations per data point, ops/sec. MWKF=0 PXS=0 (Bloom disabled) Before: 5724844 After: 6722066 MWKF=0 PXS=7 (prefixes hardly unique; Bloom not useful) Before: 9981319 After: 10237990 MWKF=0 PXS=8 (prefixes unique; Bloom useful) Before: 12081715 After: 12117603 MWKF=1 PXS=0 (whole key Bloom useful) Before: 11944354 After: 12096085 MWKF=1 PXS=7 (whole key Bloom useful in new version; prefixes not useful in old version) Before: 9444299 After: 11826029 MWKF=1 PXS=7 (whole key Bloom useful in new version; prefixes useful in old version) Before: 11784465 After: 11778591 Only in this last case is the 'before' *slightly* faster, perhaps because hashing prefixes is slightly faster than hashing whole keys. Otherwise, 'after' is faster. Reviewed By: ajkr Differential Revision: D33805025 Pulled By: pdillinger fbshipit-source-id: 597523cae4f4eafdf6ae6bb2bc6cb46f83b017bf
2022-01-27 23:53:39 +01:00
inline void DynamicBloom::MayContain(int num_keys, Slice* keys,
MultiGet batching in memtable (#5818) Summary: RocksDB has a MultiGet() API that implements batched key lookup for higher performance (https://github.com/facebook/rocksdb/blob/master/include/rocksdb/db.h#L468). Currently, batching is implemented in BlockBasedTableReader::MultiGet() for SST file lookups. One of the ways it improves performance is by pipelining bloom filter lookups (by prefetching required cachelines for all the keys in the batch, and then doing the probe) and thus hiding the cache miss latency. The same concept can be extended to the memtable as well. This PR involves implementing a pipelined bloom filter lookup in DynamicBloom, and implementing MemTable::MultiGet() that can leverage it. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5818 Test Plan: Existing tests Performance Test: Ran the below command which fills up the memtable and makes sure there are no flushes and then call multiget. Ran it on master and on the new change and see atleast 1% performance improvement across all the test runs I did. Sometimes the improvement was upto 5%. TEST_TMPDIR=/data/users/$USER/benchmarks/feature/ numactl -C 10 ./db_bench -benchmarks="fillseq,multireadrandom" -num=600000 -compression_type="none" -level_compaction_dynamic_level_bytes -write_buffer_size=200000000 -target_file_size_base=200000000 -max_bytes_for_level_base=16777216 -reads=90000 -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 -statistics -memtable_whole_key_filtering=true -memtable_bloom_size_ratio=10 Differential Revision: D17578869 Pulled By: vjnadimpalli fbshipit-source-id: 23dc651d9bf49db11d22375bf435708875a1f192
2019-10-10 18:37:38 +02:00
bool* may_match) const {
std::array<uint32_t, MultiGetContext::MAX_BATCH_SIZE> hashes;
std::array<size_t, MultiGetContext::MAX_BATCH_SIZE> byte_offsets;
for (int i = 0; i < num_keys; ++i) {
Fix major bug with MultiGet, DeleteRange, and memtable Bloom (#9453) Summary: MemTable::MultiGet was not considering range tombstones before querying Bloom filter. This means range tombstones would be skipped for keys (or prefixes) with no other entries in the memtable. This could cause old values for a key (in SST files) to still show up until the range tombstone covering it has been flushed. This is fixed by essentially disabling the memtable Bloom filter when there are any range tombstones. (This could be better optimized in the future, but good enough for now.) Did some other cleanup/optimization in the same code to (more than) offset the cost of checking on range tombstones in more cases. There is now notable improvement when memtable_whole_key_filtering and prefix_extractor are used together (unusual), and this makes MultiGet closer to the Get implementation. Pull Request resolved: https://github.com/facebook/rocksdb/pull/9453 Test Plan: new unit test added. Added memtable Bloom to crash test. Performance testing -------------------- Build WAL-only DB (recovers to memtable): ``` TEST_TMPDIR=/dev/shm/rocksdb ./db_bench -benchmarks=fillrandom -num=1000000 -write_buffer_size=250000000 ``` Query test command, to maximize sensitivity to the changed code: ``` TEST_TMPDIR=/dev/shm/rocksdb ./db_bench -use_existing_db -readonly -benchmarks=multireadrandom -num=10000000 -write_buffer_size=250000000 -memtable_bloom_size_ratio=0.015 -multiread_batched -batch_size=24 -threads=8 -memtable_whole_key_filtering=$MWKF -prefix_size=$PXS ``` (Note -num here is 10x larger for mostly memtable misses) Before & after run simultaneously, average over 10 iterations per data point, ops/sec. MWKF=0 PXS=0 (Bloom disabled) Before: 5724844 After: 6722066 MWKF=0 PXS=7 (prefixes hardly unique; Bloom not useful) Before: 9981319 After: 10237990 MWKF=0 PXS=8 (prefixes unique; Bloom useful) Before: 12081715 After: 12117603 MWKF=1 PXS=0 (whole key Bloom useful) Before: 11944354 After: 12096085 MWKF=1 PXS=7 (whole key Bloom useful in new version; prefixes not useful in old version) Before: 9444299 After: 11826029 MWKF=1 PXS=7 (whole key Bloom useful in new version; prefixes useful in old version) Before: 11784465 After: 11778591 Only in this last case is the 'before' *slightly* faster, perhaps because hashing prefixes is slightly faster than hashing whole keys. Otherwise, 'after' is faster. Reviewed By: ajkr Differential Revision: D33805025 Pulled By: pdillinger fbshipit-source-id: 597523cae4f4eafdf6ae6bb2bc6cb46f83b017bf
2022-01-27 23:53:39 +01:00
hashes[i] = BloomHash(keys[i]);
size_t a = FastRange32(kLen, hashes[i]);
MultiGet batching in memtable (#5818) Summary: RocksDB has a MultiGet() API that implements batched key lookup for higher performance (https://github.com/facebook/rocksdb/blob/master/include/rocksdb/db.h#L468). Currently, batching is implemented in BlockBasedTableReader::MultiGet() for SST file lookups. One of the ways it improves performance is by pipelining bloom filter lookups (by prefetching required cachelines for all the keys in the batch, and then doing the probe) and thus hiding the cache miss latency. The same concept can be extended to the memtable as well. This PR involves implementing a pipelined bloom filter lookup in DynamicBloom, and implementing MemTable::MultiGet() that can leverage it. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5818 Test Plan: Existing tests Performance Test: Ran the below command which fills up the memtable and makes sure there are no flushes and then call multiget. Ran it on master and on the new change and see atleast 1% performance improvement across all the test runs I did. Sometimes the improvement was upto 5%. TEST_TMPDIR=/data/users/$USER/benchmarks/feature/ numactl -C 10 ./db_bench -benchmarks="fillseq,multireadrandom" -num=600000 -compression_type="none" -level_compaction_dynamic_level_bytes -write_buffer_size=200000000 -target_file_size_base=200000000 -max_bytes_for_level_base=16777216 -reads=90000 -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 -statistics -memtable_whole_key_filtering=true -memtable_bloom_size_ratio=10 Differential Revision: D17578869 Pulled By: vjnadimpalli fbshipit-source-id: 23dc651d9bf49db11d22375bf435708875a1f192
2019-10-10 18:37:38 +02:00
PREFETCH(data_ + a, 0, 3);
byte_offsets[i] = a;
}
for (int i = 0; i < num_keys; i++) {
may_match[i] = DoubleProbe(hashes[i], byte_offsets[i]);
}
}
#if defined(_MSC_VER)
#pragma warning(push)
// local variable is initialized but not referenced
#pragma warning(disable : 4189)
#endif
Faster new DynamicBloom implementation (for memtable) (#5762) Summary: Since DynamicBloom is now only used in-memory, we're free to change it without schema compatibility issues. The new implementation is drawn from (with manifest permission) https://github.com/pdillinger/wormhashing/blob/303542a767437f56d8b66cea6ebecaac0e6a61e9/bloom_simulation_tests/foo.cc#L613 This has several speed advantages over the prior implementation: * Uses fastrange instead of % * Minimum logic to determine first (and all) probed memory addresses * (Major) Two probes per 64-bit memory fetch/write. * Very fast and effective (murmur-like) hash expansion/re-mixing. (At least on recent CPUs, integer multiplication is very cheap.) While a Bloom filter with 512-bit cache locality has about a 1.15x FP rate penalty (e.g. 0.84% to 0.97%), further restricting to two probes per 64 bits incurs an additional 1.12x FP rate penalty (e.g. 0.97% to 1.09%). Nevertheless, the unit tests show no "mediocre" FP rate samples, unlike the old implementation with more erratic FP rates. Especially for the memtable, we expect speed to outweigh somewhat higher FP rates. For example, a negative table query would have to be 1000x slower than a BF query to justify doubling BF query time to shave 10% off FP rate (working assumption around 1% FP rate). While that seems likely for SSTs, my data suggests a speed factor of roughly 50x for the memtable (vs. BF; ~1.5% lower write throughput when enabling memtable Bloom filter, after this change). Thus, it's probably not worth even 5% more time in the Bloom filter to shave off 1/10th of the Bloom FP rate, or 0.1% in absolute terms, and it's probably at least 20% slower to recoup that much FP rate from this new implementation. Because of this, we do not see a need for a 'locality' option that affects the MemTable Bloom filter and have decoupled the MemTable Bloom filter from Options::bloom_locality. Note that just 3% more memory to the Bloom filter (10.3 bits per key vs. just 10) is able to make up for the ~12% FP rate drop in the new implementation: [] # Nearly "ideal" FP-wise but reasonably fast cache-local implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out 10000000 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out time: 3.29372 sampled_fp_rate: 0.00985956 ... [] # Close match to this new implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out 10000000 6 10.3 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.10072 sampled_fp_rate: 0.00985655 ... [] # Old locality=1 implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out 10000000 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out time: 3.95472 sampled_fp_rate: 0.00988943 ... Also note the dramatic speed improvement vs. alternatives. -- Performance unit test: DynamicBloomTest.concurrent_with_perf is updated to report more precise timing data. (Measure running time of each thread, not just longest running thread, etc.) Results averaged over various sizes enabled with --enable_perf and 20 runs each; old dynamic bloom refers to locality=1, the faster of the old: old dynamic bloom, avg add latency = 65.6468 new dynamic bloom, avg add latency = 44.3809 old dynamic bloom, avg query latency = 50.6485 new dynamic bloom, avg query latency = 43.2186 old avg parallel add latency = 41.678 new avg parallel add latency = 24.5238 old avg parallel hit latency = 14.6322 new avg parallel hit latency = 12.3939 old avg parallel miss latency = 16.7289 new avg parallel miss latency = 12.2134 Tested on a dedicated 64-bit production machine at Facebook. Significant improvement all around. Despite now using std::atomic<uint64_t>, quick before-and-after test on a 32-bit machine (Intel Atom N270, released 2008) shows no regression in performance, in some cases modest improvement. -- Performance integration test (synthetic): with DEBUG_LEVEL=0, used TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=fillrandom,readmissing,readrandom,stats --num=2000000 and optionally with -memtable_whole_key_filtering -memtable_bloom_size_ratio=0.01 300 runs each configuration. Write throughput change by enabling memtable bloom: Old locality=0: -3.06% Old locality=1: -2.37% New: -1.50% conclusion -> seems to substantially close the gap Readmissing throughput change by enabling memtable bloom: Old locality=0: +34.47% Old locality=1: +34.80% New: +33.25% conclusion -> maybe a small new penalty from FP rate Readrandom throughput change by enabling memtable bloom: Old locality=0: +31.54% Old locality=1: +31.13% New: +30.60% conclusion -> maybe also from FP rate (after memtable flush) -- Another conclusion we can draw from this new implementation is that the existing 32-bit hash function is not inherently crippling the Bloom filter speed or accuracy, below about 5 million keys. For speed, the implementation is essentially the same whether starting with 32-bits or 64-bits of hash; it just determines whether the first multiplication after fastrange is a pseudorandom expansion or needed re-mix. Note that this multiplication can occur while memory is fetching. For accuracy, in a standard configuration, you need about 5 million keys before you have about a 1.1x FP penalty due to using a 32-bit hash vs. 64-bit: [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.52069 sampled_fp_rate: 0.0118267 ... [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out time: 2.43871 sampled_fp_rate: 0.0109059 Pull Request resolved: https://github.com/facebook/rocksdb/pull/5762 Differential Revision: D17214194 Pulled By: pdillinger fbshipit-source-id: ad9da031772e985fd6b62a0e1db8e81892520595
2019-09-05 23:57:39 +02:00
inline void DynamicBloom::Prefetch(uint32_t h32) {
size_t a = FastRange32(kLen, h32);
Faster new DynamicBloom implementation (for memtable) (#5762) Summary: Since DynamicBloom is now only used in-memory, we're free to change it without schema compatibility issues. The new implementation is drawn from (with manifest permission) https://github.com/pdillinger/wormhashing/blob/303542a767437f56d8b66cea6ebecaac0e6a61e9/bloom_simulation_tests/foo.cc#L613 This has several speed advantages over the prior implementation: * Uses fastrange instead of % * Minimum logic to determine first (and all) probed memory addresses * (Major) Two probes per 64-bit memory fetch/write. * Very fast and effective (murmur-like) hash expansion/re-mixing. (At least on recent CPUs, integer multiplication is very cheap.) While a Bloom filter with 512-bit cache locality has about a 1.15x FP rate penalty (e.g. 0.84% to 0.97%), further restricting to two probes per 64 bits incurs an additional 1.12x FP rate penalty (e.g. 0.97% to 1.09%). Nevertheless, the unit tests show no "mediocre" FP rate samples, unlike the old implementation with more erratic FP rates. Especially for the memtable, we expect speed to outweigh somewhat higher FP rates. For example, a negative table query would have to be 1000x slower than a BF query to justify doubling BF query time to shave 10% off FP rate (working assumption around 1% FP rate). While that seems likely for SSTs, my data suggests a speed factor of roughly 50x for the memtable (vs. BF; ~1.5% lower write throughput when enabling memtable Bloom filter, after this change). Thus, it's probably not worth even 5% more time in the Bloom filter to shave off 1/10th of the Bloom FP rate, or 0.1% in absolute terms, and it's probably at least 20% slower to recoup that much FP rate from this new implementation. Because of this, we do not see a need for a 'locality' option that affects the MemTable Bloom filter and have decoupled the MemTable Bloom filter from Options::bloom_locality. Note that just 3% more memory to the Bloom filter (10.3 bits per key vs. just 10) is able to make up for the ~12% FP rate drop in the new implementation: [] # Nearly "ideal" FP-wise but reasonably fast cache-local implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out 10000000 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out time: 3.29372 sampled_fp_rate: 0.00985956 ... [] # Close match to this new implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out 10000000 6 10.3 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.10072 sampled_fp_rate: 0.00985655 ... [] # Old locality=1 implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out 10000000 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out time: 3.95472 sampled_fp_rate: 0.00988943 ... Also note the dramatic speed improvement vs. alternatives. -- Performance unit test: DynamicBloomTest.concurrent_with_perf is updated to report more precise timing data. (Measure running time of each thread, not just longest running thread, etc.) Results averaged over various sizes enabled with --enable_perf and 20 runs each; old dynamic bloom refers to locality=1, the faster of the old: old dynamic bloom, avg add latency = 65.6468 new dynamic bloom, avg add latency = 44.3809 old dynamic bloom, avg query latency = 50.6485 new dynamic bloom, avg query latency = 43.2186 old avg parallel add latency = 41.678 new avg parallel add latency = 24.5238 old avg parallel hit latency = 14.6322 new avg parallel hit latency = 12.3939 old avg parallel miss latency = 16.7289 new avg parallel miss latency = 12.2134 Tested on a dedicated 64-bit production machine at Facebook. Significant improvement all around. Despite now using std::atomic<uint64_t>, quick before-and-after test on a 32-bit machine (Intel Atom N270, released 2008) shows no regression in performance, in some cases modest improvement. -- Performance integration test (synthetic): with DEBUG_LEVEL=0, used TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=fillrandom,readmissing,readrandom,stats --num=2000000 and optionally with -memtable_whole_key_filtering -memtable_bloom_size_ratio=0.01 300 runs each configuration. Write throughput change by enabling memtable bloom: Old locality=0: -3.06% Old locality=1: -2.37% New: -1.50% conclusion -> seems to substantially close the gap Readmissing throughput change by enabling memtable bloom: Old locality=0: +34.47% Old locality=1: +34.80% New: +33.25% conclusion -> maybe a small new penalty from FP rate Readrandom throughput change by enabling memtable bloom: Old locality=0: +31.54% Old locality=1: +31.13% New: +30.60% conclusion -> maybe also from FP rate (after memtable flush) -- Another conclusion we can draw from this new implementation is that the existing 32-bit hash function is not inherently crippling the Bloom filter speed or accuracy, below about 5 million keys. For speed, the implementation is essentially the same whether starting with 32-bits or 64-bits of hash; it just determines whether the first multiplication after fastrange is a pseudorandom expansion or needed re-mix. Note that this multiplication can occur while memory is fetching. For accuracy, in a standard configuration, you need about 5 million keys before you have about a 1.1x FP penalty due to using a 32-bit hash vs. 64-bit: [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.52069 sampled_fp_rate: 0.0118267 ... [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out time: 2.43871 sampled_fp_rate: 0.0109059 Pull Request resolved: https://github.com/facebook/rocksdb/pull/5762 Differential Revision: D17214194 Pulled By: pdillinger fbshipit-source-id: ad9da031772e985fd6b62a0e1db8e81892520595
2019-09-05 23:57:39 +02:00
PREFETCH(data_ + a, 0, 3);
}
#if defined(_MSC_VER)
#pragma warning(pop)
#endif
Faster new DynamicBloom implementation (for memtable) (#5762) Summary: Since DynamicBloom is now only used in-memory, we're free to change it without schema compatibility issues. The new implementation is drawn from (with manifest permission) https://github.com/pdillinger/wormhashing/blob/303542a767437f56d8b66cea6ebecaac0e6a61e9/bloom_simulation_tests/foo.cc#L613 This has several speed advantages over the prior implementation: * Uses fastrange instead of % * Minimum logic to determine first (and all) probed memory addresses * (Major) Two probes per 64-bit memory fetch/write. * Very fast and effective (murmur-like) hash expansion/re-mixing. (At least on recent CPUs, integer multiplication is very cheap.) While a Bloom filter with 512-bit cache locality has about a 1.15x FP rate penalty (e.g. 0.84% to 0.97%), further restricting to two probes per 64 bits incurs an additional 1.12x FP rate penalty (e.g. 0.97% to 1.09%). Nevertheless, the unit tests show no "mediocre" FP rate samples, unlike the old implementation with more erratic FP rates. Especially for the memtable, we expect speed to outweigh somewhat higher FP rates. For example, a negative table query would have to be 1000x slower than a BF query to justify doubling BF query time to shave 10% off FP rate (working assumption around 1% FP rate). While that seems likely for SSTs, my data suggests a speed factor of roughly 50x for the memtable (vs. BF; ~1.5% lower write throughput when enabling memtable Bloom filter, after this change). Thus, it's probably not worth even 5% more time in the Bloom filter to shave off 1/10th of the Bloom FP rate, or 0.1% in absolute terms, and it's probably at least 20% slower to recoup that much FP rate from this new implementation. Because of this, we do not see a need for a 'locality' option that affects the MemTable Bloom filter and have decoupled the MemTable Bloom filter from Options::bloom_locality. Note that just 3% more memory to the Bloom filter (10.3 bits per key vs. just 10) is able to make up for the ~12% FP rate drop in the new implementation: [] # Nearly "ideal" FP-wise but reasonably fast cache-local implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out 10000000 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out time: 3.29372 sampled_fp_rate: 0.00985956 ... [] # Close match to this new implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out 10000000 6 10.3 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.10072 sampled_fp_rate: 0.00985655 ... [] # Old locality=1 implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out 10000000 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out time: 3.95472 sampled_fp_rate: 0.00988943 ... Also note the dramatic speed improvement vs. alternatives. -- Performance unit test: DynamicBloomTest.concurrent_with_perf is updated to report more precise timing data. (Measure running time of each thread, not just longest running thread, etc.) Results averaged over various sizes enabled with --enable_perf and 20 runs each; old dynamic bloom refers to locality=1, the faster of the old: old dynamic bloom, avg add latency = 65.6468 new dynamic bloom, avg add latency = 44.3809 old dynamic bloom, avg query latency = 50.6485 new dynamic bloom, avg query latency = 43.2186 old avg parallel add latency = 41.678 new avg parallel add latency = 24.5238 old avg parallel hit latency = 14.6322 new avg parallel hit latency = 12.3939 old avg parallel miss latency = 16.7289 new avg parallel miss latency = 12.2134 Tested on a dedicated 64-bit production machine at Facebook. Significant improvement all around. Despite now using std::atomic<uint64_t>, quick before-and-after test on a 32-bit machine (Intel Atom N270, released 2008) shows no regression in performance, in some cases modest improvement. -- Performance integration test (synthetic): with DEBUG_LEVEL=0, used TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=fillrandom,readmissing,readrandom,stats --num=2000000 and optionally with -memtable_whole_key_filtering -memtable_bloom_size_ratio=0.01 300 runs each configuration. Write throughput change by enabling memtable bloom: Old locality=0: -3.06% Old locality=1: -2.37% New: -1.50% conclusion -> seems to substantially close the gap Readmissing throughput change by enabling memtable bloom: Old locality=0: +34.47% Old locality=1: +34.80% New: +33.25% conclusion -> maybe a small new penalty from FP rate Readrandom throughput change by enabling memtable bloom: Old locality=0: +31.54% Old locality=1: +31.13% New: +30.60% conclusion -> maybe also from FP rate (after memtable flush) -- Another conclusion we can draw from this new implementation is that the existing 32-bit hash function is not inherently crippling the Bloom filter speed or accuracy, below about 5 million keys. For speed, the implementation is essentially the same whether starting with 32-bits or 64-bits of hash; it just determines whether the first multiplication after fastrange is a pseudorandom expansion or needed re-mix. Note that this multiplication can occur while memory is fetching. For accuracy, in a standard configuration, you need about 5 million keys before you have about a 1.1x FP penalty due to using a 32-bit hash vs. 64-bit: [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.52069 sampled_fp_rate: 0.0118267 ... [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out time: 2.43871 sampled_fp_rate: 0.0109059 Pull Request resolved: https://github.com/facebook/rocksdb/pull/5762 Differential Revision: D17214194 Pulled By: pdillinger fbshipit-source-id: ad9da031772e985fd6b62a0e1db8e81892520595
2019-09-05 23:57:39 +02:00
// Speed hacks in this implementation:
// * Uses fastrange instead of %
// * Minimum logic to determine first (and all) probed memory addresses.
// (Uses constant bit-xor offsets from the starting probe address.)
// * (Major) Two probes per 64-bit memory fetch/write.
// Code simplification / optimization: only allow even number of probes.
// * Very fast and effective (murmur-like) hash expansion/re-mixing. (At
// least on recent CPUs, integer multiplication is very cheap. Each 64-bit
// remix provides five pairs of bit addresses within a uint64_t.)
// Code simplification / optimization: only allow up to 10 probes, from a
// single 64-bit remix.
//
// The FP rate penalty for this implementation, vs. standard Bloom filter, is
// roughly 1.12x on top of the 1.15x penalty for a 512-bit cache-local Bloom.
// This implementation does not explicitly use the cache line size, but is
// effectively cache-local (up to 16 probes) because of the bit-xor offsetting.
//
// NB: could easily be upgraded to support a 64-bit hash and
// total_bits > 2^32 (512MB). (The latter is a bad idea without the former,
// because of false positives.)
inline bool DynamicBloom::MayContainHash(uint32_t h32) const {
size_t a = FastRange32(kLen, h32);
Faster new DynamicBloom implementation (for memtable) (#5762) Summary: Since DynamicBloom is now only used in-memory, we're free to change it without schema compatibility issues. The new implementation is drawn from (with manifest permission) https://github.com/pdillinger/wormhashing/blob/303542a767437f56d8b66cea6ebecaac0e6a61e9/bloom_simulation_tests/foo.cc#L613 This has several speed advantages over the prior implementation: * Uses fastrange instead of % * Minimum logic to determine first (and all) probed memory addresses * (Major) Two probes per 64-bit memory fetch/write. * Very fast and effective (murmur-like) hash expansion/re-mixing. (At least on recent CPUs, integer multiplication is very cheap.) While a Bloom filter with 512-bit cache locality has about a 1.15x FP rate penalty (e.g. 0.84% to 0.97%), further restricting to two probes per 64 bits incurs an additional 1.12x FP rate penalty (e.g. 0.97% to 1.09%). Nevertheless, the unit tests show no "mediocre" FP rate samples, unlike the old implementation with more erratic FP rates. Especially for the memtable, we expect speed to outweigh somewhat higher FP rates. For example, a negative table query would have to be 1000x slower than a BF query to justify doubling BF query time to shave 10% off FP rate (working assumption around 1% FP rate). While that seems likely for SSTs, my data suggests a speed factor of roughly 50x for the memtable (vs. BF; ~1.5% lower write throughput when enabling memtable Bloom filter, after this change). Thus, it's probably not worth even 5% more time in the Bloom filter to shave off 1/10th of the Bloom FP rate, or 0.1% in absolute terms, and it's probably at least 20% slower to recoup that much FP rate from this new implementation. Because of this, we do not see a need for a 'locality' option that affects the MemTable Bloom filter and have decoupled the MemTable Bloom filter from Options::bloom_locality. Note that just 3% more memory to the Bloom filter (10.3 bits per key vs. just 10) is able to make up for the ~12% FP rate drop in the new implementation: [] # Nearly "ideal" FP-wise but reasonably fast cache-local implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out 10000000 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out time: 3.29372 sampled_fp_rate: 0.00985956 ... [] # Close match to this new implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out 10000000 6 10.3 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.10072 sampled_fp_rate: 0.00985655 ... [] # Old locality=1 implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out 10000000 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out time: 3.95472 sampled_fp_rate: 0.00988943 ... Also note the dramatic speed improvement vs. alternatives. -- Performance unit test: DynamicBloomTest.concurrent_with_perf is updated to report more precise timing data. (Measure running time of each thread, not just longest running thread, etc.) Results averaged over various sizes enabled with --enable_perf and 20 runs each; old dynamic bloom refers to locality=1, the faster of the old: old dynamic bloom, avg add latency = 65.6468 new dynamic bloom, avg add latency = 44.3809 old dynamic bloom, avg query latency = 50.6485 new dynamic bloom, avg query latency = 43.2186 old avg parallel add latency = 41.678 new avg parallel add latency = 24.5238 old avg parallel hit latency = 14.6322 new avg parallel hit latency = 12.3939 old avg parallel miss latency = 16.7289 new avg parallel miss latency = 12.2134 Tested on a dedicated 64-bit production machine at Facebook. Significant improvement all around. Despite now using std::atomic<uint64_t>, quick before-and-after test on a 32-bit machine (Intel Atom N270, released 2008) shows no regression in performance, in some cases modest improvement. -- Performance integration test (synthetic): with DEBUG_LEVEL=0, used TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=fillrandom,readmissing,readrandom,stats --num=2000000 and optionally with -memtable_whole_key_filtering -memtable_bloom_size_ratio=0.01 300 runs each configuration. Write throughput change by enabling memtable bloom: Old locality=0: -3.06% Old locality=1: -2.37% New: -1.50% conclusion -> seems to substantially close the gap Readmissing throughput change by enabling memtable bloom: Old locality=0: +34.47% Old locality=1: +34.80% New: +33.25% conclusion -> maybe a small new penalty from FP rate Readrandom throughput change by enabling memtable bloom: Old locality=0: +31.54% Old locality=1: +31.13% New: +30.60% conclusion -> maybe also from FP rate (after memtable flush) -- Another conclusion we can draw from this new implementation is that the existing 32-bit hash function is not inherently crippling the Bloom filter speed or accuracy, below about 5 million keys. For speed, the implementation is essentially the same whether starting with 32-bits or 64-bits of hash; it just determines whether the first multiplication after fastrange is a pseudorandom expansion or needed re-mix. Note that this multiplication can occur while memory is fetching. For accuracy, in a standard configuration, you need about 5 million keys before you have about a 1.1x FP penalty due to using a 32-bit hash vs. 64-bit: [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.52069 sampled_fp_rate: 0.0118267 ... [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out time: 2.43871 sampled_fp_rate: 0.0109059 Pull Request resolved: https://github.com/facebook/rocksdb/pull/5762 Differential Revision: D17214194 Pulled By: pdillinger fbshipit-source-id: ad9da031772e985fd6b62a0e1db8e81892520595
2019-09-05 23:57:39 +02:00
PREFETCH(data_ + a, 0, 3);
MultiGet batching in memtable (#5818) Summary: RocksDB has a MultiGet() API that implements batched key lookup for higher performance (https://github.com/facebook/rocksdb/blob/master/include/rocksdb/db.h#L468). Currently, batching is implemented in BlockBasedTableReader::MultiGet() for SST file lookups. One of the ways it improves performance is by pipelining bloom filter lookups (by prefetching required cachelines for all the keys in the batch, and then doing the probe) and thus hiding the cache miss latency. The same concept can be extended to the memtable as well. This PR involves implementing a pipelined bloom filter lookup in DynamicBloom, and implementing MemTable::MultiGet() that can leverage it. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5818 Test Plan: Existing tests Performance Test: Ran the below command which fills up the memtable and makes sure there are no flushes and then call multiget. Ran it on master and on the new change and see atleast 1% performance improvement across all the test runs I did. Sometimes the improvement was upto 5%. TEST_TMPDIR=/data/users/$USER/benchmarks/feature/ numactl -C 10 ./db_bench -benchmarks="fillseq,multireadrandom" -num=600000 -compression_type="none" -level_compaction_dynamic_level_bytes -write_buffer_size=200000000 -target_file_size_base=200000000 -max_bytes_for_level_base=16777216 -reads=90000 -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 -statistics -memtable_whole_key_filtering=true -memtable_bloom_size_ratio=10 Differential Revision: D17578869 Pulled By: vjnadimpalli fbshipit-source-id: 23dc651d9bf49db11d22375bf435708875a1f192
2019-10-10 18:37:38 +02:00
return DoubleProbe(h32, a);
}
inline bool DynamicBloom::DoubleProbe(uint32_t h32, size_t byte_offset) const {
Faster new DynamicBloom implementation (for memtable) (#5762) Summary: Since DynamicBloom is now only used in-memory, we're free to change it without schema compatibility issues. The new implementation is drawn from (with manifest permission) https://github.com/pdillinger/wormhashing/blob/303542a767437f56d8b66cea6ebecaac0e6a61e9/bloom_simulation_tests/foo.cc#L613 This has several speed advantages over the prior implementation: * Uses fastrange instead of % * Minimum logic to determine first (and all) probed memory addresses * (Major) Two probes per 64-bit memory fetch/write. * Very fast and effective (murmur-like) hash expansion/re-mixing. (At least on recent CPUs, integer multiplication is very cheap.) While a Bloom filter with 512-bit cache locality has about a 1.15x FP rate penalty (e.g. 0.84% to 0.97%), further restricting to two probes per 64 bits incurs an additional 1.12x FP rate penalty (e.g. 0.97% to 1.09%). Nevertheless, the unit tests show no "mediocre" FP rate samples, unlike the old implementation with more erratic FP rates. Especially for the memtable, we expect speed to outweigh somewhat higher FP rates. For example, a negative table query would have to be 1000x slower than a BF query to justify doubling BF query time to shave 10% off FP rate (working assumption around 1% FP rate). While that seems likely for SSTs, my data suggests a speed factor of roughly 50x for the memtable (vs. BF; ~1.5% lower write throughput when enabling memtable Bloom filter, after this change). Thus, it's probably not worth even 5% more time in the Bloom filter to shave off 1/10th of the Bloom FP rate, or 0.1% in absolute terms, and it's probably at least 20% slower to recoup that much FP rate from this new implementation. Because of this, we do not see a need for a 'locality' option that affects the MemTable Bloom filter and have decoupled the MemTable Bloom filter from Options::bloom_locality. Note that just 3% more memory to the Bloom filter (10.3 bits per key vs. just 10) is able to make up for the ~12% FP rate drop in the new implementation: [] # Nearly "ideal" FP-wise but reasonably fast cache-local implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out 10000000 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out time: 3.29372 sampled_fp_rate: 0.00985956 ... [] # Close match to this new implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out 10000000 6 10.3 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.10072 sampled_fp_rate: 0.00985655 ... [] # Old locality=1 implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out 10000000 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out time: 3.95472 sampled_fp_rate: 0.00988943 ... Also note the dramatic speed improvement vs. alternatives. -- Performance unit test: DynamicBloomTest.concurrent_with_perf is updated to report more precise timing data. (Measure running time of each thread, not just longest running thread, etc.) Results averaged over various sizes enabled with --enable_perf and 20 runs each; old dynamic bloom refers to locality=1, the faster of the old: old dynamic bloom, avg add latency = 65.6468 new dynamic bloom, avg add latency = 44.3809 old dynamic bloom, avg query latency = 50.6485 new dynamic bloom, avg query latency = 43.2186 old avg parallel add latency = 41.678 new avg parallel add latency = 24.5238 old avg parallel hit latency = 14.6322 new avg parallel hit latency = 12.3939 old avg parallel miss latency = 16.7289 new avg parallel miss latency = 12.2134 Tested on a dedicated 64-bit production machine at Facebook. Significant improvement all around. Despite now using std::atomic<uint64_t>, quick before-and-after test on a 32-bit machine (Intel Atom N270, released 2008) shows no regression in performance, in some cases modest improvement. -- Performance integration test (synthetic): with DEBUG_LEVEL=0, used TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=fillrandom,readmissing,readrandom,stats --num=2000000 and optionally with -memtable_whole_key_filtering -memtable_bloom_size_ratio=0.01 300 runs each configuration. Write throughput change by enabling memtable bloom: Old locality=0: -3.06% Old locality=1: -2.37% New: -1.50% conclusion -> seems to substantially close the gap Readmissing throughput change by enabling memtable bloom: Old locality=0: +34.47% Old locality=1: +34.80% New: +33.25% conclusion -> maybe a small new penalty from FP rate Readrandom throughput change by enabling memtable bloom: Old locality=0: +31.54% Old locality=1: +31.13% New: +30.60% conclusion -> maybe also from FP rate (after memtable flush) -- Another conclusion we can draw from this new implementation is that the existing 32-bit hash function is not inherently crippling the Bloom filter speed or accuracy, below about 5 million keys. For speed, the implementation is essentially the same whether starting with 32-bits or 64-bits of hash; it just determines whether the first multiplication after fastrange is a pseudorandom expansion or needed re-mix. Note that this multiplication can occur while memory is fetching. For accuracy, in a standard configuration, you need about 5 million keys before you have about a 1.1x FP penalty due to using a 32-bit hash vs. 64-bit: [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.52069 sampled_fp_rate: 0.0118267 ... [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out time: 2.43871 sampled_fp_rate: 0.0109059 Pull Request resolved: https://github.com/facebook/rocksdb/pull/5762 Differential Revision: D17214194 Pulled By: pdillinger fbshipit-source-id: ad9da031772e985fd6b62a0e1db8e81892520595
2019-09-05 23:57:39 +02:00
// Expand/remix with 64-bit golden ratio
uint64_t h = 0x9e3779b97f4a7c13ULL * h32;
for (unsigned i = 0;; ++i) {
// Two bit probes per uint64_t probe
uint64_t mask =
((uint64_t)1 << (h & 63)) | ((uint64_t)1 << ((h >> 6) & 63));
MultiGet batching in memtable (#5818) Summary: RocksDB has a MultiGet() API that implements batched key lookup for higher performance (https://github.com/facebook/rocksdb/blob/master/include/rocksdb/db.h#L468). Currently, batching is implemented in BlockBasedTableReader::MultiGet() for SST file lookups. One of the ways it improves performance is by pipelining bloom filter lookups (by prefetching required cachelines for all the keys in the batch, and then doing the probe) and thus hiding the cache miss latency. The same concept can be extended to the memtable as well. This PR involves implementing a pipelined bloom filter lookup in DynamicBloom, and implementing MemTable::MultiGet() that can leverage it. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5818 Test Plan: Existing tests Performance Test: Ran the below command which fills up the memtable and makes sure there are no flushes and then call multiget. Ran it on master and on the new change and see atleast 1% performance improvement across all the test runs I did. Sometimes the improvement was upto 5%. TEST_TMPDIR=/data/users/$USER/benchmarks/feature/ numactl -C 10 ./db_bench -benchmarks="fillseq,multireadrandom" -num=600000 -compression_type="none" -level_compaction_dynamic_level_bytes -write_buffer_size=200000000 -target_file_size_base=200000000 -max_bytes_for_level_base=16777216 -reads=90000 -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 -statistics -memtable_whole_key_filtering=true -memtable_bloom_size_ratio=10 Differential Revision: D17578869 Pulled By: vjnadimpalli fbshipit-source-id: 23dc651d9bf49db11d22375bf435708875a1f192
2019-10-10 18:37:38 +02:00
uint64_t val = data_[byte_offset ^ i].load(std::memory_order_relaxed);
Faster new DynamicBloom implementation (for memtable) (#5762) Summary: Since DynamicBloom is now only used in-memory, we're free to change it without schema compatibility issues. The new implementation is drawn from (with manifest permission) https://github.com/pdillinger/wormhashing/blob/303542a767437f56d8b66cea6ebecaac0e6a61e9/bloom_simulation_tests/foo.cc#L613 This has several speed advantages over the prior implementation: * Uses fastrange instead of % * Minimum logic to determine first (and all) probed memory addresses * (Major) Two probes per 64-bit memory fetch/write. * Very fast and effective (murmur-like) hash expansion/re-mixing. (At least on recent CPUs, integer multiplication is very cheap.) While a Bloom filter with 512-bit cache locality has about a 1.15x FP rate penalty (e.g. 0.84% to 0.97%), further restricting to two probes per 64 bits incurs an additional 1.12x FP rate penalty (e.g. 0.97% to 1.09%). Nevertheless, the unit tests show no "mediocre" FP rate samples, unlike the old implementation with more erratic FP rates. Especially for the memtable, we expect speed to outweigh somewhat higher FP rates. For example, a negative table query would have to be 1000x slower than a BF query to justify doubling BF query time to shave 10% off FP rate (working assumption around 1% FP rate). While that seems likely for SSTs, my data suggests a speed factor of roughly 50x for the memtable (vs. BF; ~1.5% lower write throughput when enabling memtable Bloom filter, after this change). Thus, it's probably not worth even 5% more time in the Bloom filter to shave off 1/10th of the Bloom FP rate, or 0.1% in absolute terms, and it's probably at least 20% slower to recoup that much FP rate from this new implementation. Because of this, we do not see a need for a 'locality' option that affects the MemTable Bloom filter and have decoupled the MemTable Bloom filter from Options::bloom_locality. Note that just 3% more memory to the Bloom filter (10.3 bits per key vs. just 10) is able to make up for the ~12% FP rate drop in the new implementation: [] # Nearly "ideal" FP-wise but reasonably fast cache-local implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out 10000000 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out time: 3.29372 sampled_fp_rate: 0.00985956 ... [] # Close match to this new implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out 10000000 6 10.3 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.10072 sampled_fp_rate: 0.00985655 ... [] # Old locality=1 implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out 10000000 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out time: 3.95472 sampled_fp_rate: 0.00988943 ... Also note the dramatic speed improvement vs. alternatives. -- Performance unit test: DynamicBloomTest.concurrent_with_perf is updated to report more precise timing data. (Measure running time of each thread, not just longest running thread, etc.) Results averaged over various sizes enabled with --enable_perf and 20 runs each; old dynamic bloom refers to locality=1, the faster of the old: old dynamic bloom, avg add latency = 65.6468 new dynamic bloom, avg add latency = 44.3809 old dynamic bloom, avg query latency = 50.6485 new dynamic bloom, avg query latency = 43.2186 old avg parallel add latency = 41.678 new avg parallel add latency = 24.5238 old avg parallel hit latency = 14.6322 new avg parallel hit latency = 12.3939 old avg parallel miss latency = 16.7289 new avg parallel miss latency = 12.2134 Tested on a dedicated 64-bit production machine at Facebook. Significant improvement all around. Despite now using std::atomic<uint64_t>, quick before-and-after test on a 32-bit machine (Intel Atom N270, released 2008) shows no regression in performance, in some cases modest improvement. -- Performance integration test (synthetic): with DEBUG_LEVEL=0, used TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=fillrandom,readmissing,readrandom,stats --num=2000000 and optionally with -memtable_whole_key_filtering -memtable_bloom_size_ratio=0.01 300 runs each configuration. Write throughput change by enabling memtable bloom: Old locality=0: -3.06% Old locality=1: -2.37% New: -1.50% conclusion -> seems to substantially close the gap Readmissing throughput change by enabling memtable bloom: Old locality=0: +34.47% Old locality=1: +34.80% New: +33.25% conclusion -> maybe a small new penalty from FP rate Readrandom throughput change by enabling memtable bloom: Old locality=0: +31.54% Old locality=1: +31.13% New: +30.60% conclusion -> maybe also from FP rate (after memtable flush) -- Another conclusion we can draw from this new implementation is that the existing 32-bit hash function is not inherently crippling the Bloom filter speed or accuracy, below about 5 million keys. For speed, the implementation is essentially the same whether starting with 32-bits or 64-bits of hash; it just determines whether the first multiplication after fastrange is a pseudorandom expansion or needed re-mix. Note that this multiplication can occur while memory is fetching. For accuracy, in a standard configuration, you need about 5 million keys before you have about a 1.1x FP penalty due to using a 32-bit hash vs. 64-bit: [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.52069 sampled_fp_rate: 0.0118267 ... [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out time: 2.43871 sampled_fp_rate: 0.0109059 Pull Request resolved: https://github.com/facebook/rocksdb/pull/5762 Differential Revision: D17214194 Pulled By: pdillinger fbshipit-source-id: ad9da031772e985fd6b62a0e1db8e81892520595
2019-09-05 23:57:39 +02:00
if (i + 1 >= kNumDoubleProbes) {
return (val & mask) == mask;
} else if ((val & mask) != mask) {
return false;
}
Faster new DynamicBloom implementation (for memtable) (#5762) Summary: Since DynamicBloom is now only used in-memory, we're free to change it without schema compatibility issues. The new implementation is drawn from (with manifest permission) https://github.com/pdillinger/wormhashing/blob/303542a767437f56d8b66cea6ebecaac0e6a61e9/bloom_simulation_tests/foo.cc#L613 This has several speed advantages over the prior implementation: * Uses fastrange instead of % * Minimum logic to determine first (and all) probed memory addresses * (Major) Two probes per 64-bit memory fetch/write. * Very fast and effective (murmur-like) hash expansion/re-mixing. (At least on recent CPUs, integer multiplication is very cheap.) While a Bloom filter with 512-bit cache locality has about a 1.15x FP rate penalty (e.g. 0.84% to 0.97%), further restricting to two probes per 64 bits incurs an additional 1.12x FP rate penalty (e.g. 0.97% to 1.09%). Nevertheless, the unit tests show no "mediocre" FP rate samples, unlike the old implementation with more erratic FP rates. Especially for the memtable, we expect speed to outweigh somewhat higher FP rates. For example, a negative table query would have to be 1000x slower than a BF query to justify doubling BF query time to shave 10% off FP rate (working assumption around 1% FP rate). While that seems likely for SSTs, my data suggests a speed factor of roughly 50x for the memtable (vs. BF; ~1.5% lower write throughput when enabling memtable Bloom filter, after this change). Thus, it's probably not worth even 5% more time in the Bloom filter to shave off 1/10th of the Bloom FP rate, or 0.1% in absolute terms, and it's probably at least 20% slower to recoup that much FP rate from this new implementation. Because of this, we do not see a need for a 'locality' option that affects the MemTable Bloom filter and have decoupled the MemTable Bloom filter from Options::bloom_locality. Note that just 3% more memory to the Bloom filter (10.3 bits per key vs. just 10) is able to make up for the ~12% FP rate drop in the new implementation: [] # Nearly "ideal" FP-wise but reasonably fast cache-local implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out 10000000 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out time: 3.29372 sampled_fp_rate: 0.00985956 ... [] # Close match to this new implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out 10000000 6 10.3 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.10072 sampled_fp_rate: 0.00985655 ... [] # Old locality=1 implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out 10000000 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out time: 3.95472 sampled_fp_rate: 0.00988943 ... Also note the dramatic speed improvement vs. alternatives. -- Performance unit test: DynamicBloomTest.concurrent_with_perf is updated to report more precise timing data. (Measure running time of each thread, not just longest running thread, etc.) Results averaged over various sizes enabled with --enable_perf and 20 runs each; old dynamic bloom refers to locality=1, the faster of the old: old dynamic bloom, avg add latency = 65.6468 new dynamic bloom, avg add latency = 44.3809 old dynamic bloom, avg query latency = 50.6485 new dynamic bloom, avg query latency = 43.2186 old avg parallel add latency = 41.678 new avg parallel add latency = 24.5238 old avg parallel hit latency = 14.6322 new avg parallel hit latency = 12.3939 old avg parallel miss latency = 16.7289 new avg parallel miss latency = 12.2134 Tested on a dedicated 64-bit production machine at Facebook. Significant improvement all around. Despite now using std::atomic<uint64_t>, quick before-and-after test on a 32-bit machine (Intel Atom N270, released 2008) shows no regression in performance, in some cases modest improvement. -- Performance integration test (synthetic): with DEBUG_LEVEL=0, used TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=fillrandom,readmissing,readrandom,stats --num=2000000 and optionally with -memtable_whole_key_filtering -memtable_bloom_size_ratio=0.01 300 runs each configuration. Write throughput change by enabling memtable bloom: Old locality=0: -3.06% Old locality=1: -2.37% New: -1.50% conclusion -> seems to substantially close the gap Readmissing throughput change by enabling memtable bloom: Old locality=0: +34.47% Old locality=1: +34.80% New: +33.25% conclusion -> maybe a small new penalty from FP rate Readrandom throughput change by enabling memtable bloom: Old locality=0: +31.54% Old locality=1: +31.13% New: +30.60% conclusion -> maybe also from FP rate (after memtable flush) -- Another conclusion we can draw from this new implementation is that the existing 32-bit hash function is not inherently crippling the Bloom filter speed or accuracy, below about 5 million keys. For speed, the implementation is essentially the same whether starting with 32-bits or 64-bits of hash; it just determines whether the first multiplication after fastrange is a pseudorandom expansion or needed re-mix. Note that this multiplication can occur while memory is fetching. For accuracy, in a standard configuration, you need about 5 million keys before you have about a 1.1x FP penalty due to using a 32-bit hash vs. 64-bit: [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.52069 sampled_fp_rate: 0.0118267 ... [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out time: 2.43871 sampled_fp_rate: 0.0109059 Pull Request resolved: https://github.com/facebook/rocksdb/pull/5762 Differential Revision: D17214194 Pulled By: pdillinger fbshipit-source-id: ad9da031772e985fd6b62a0e1db8e81892520595
2019-09-05 23:57:39 +02:00
h = (h >> 12) | (h << 52);
}
}
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
template <typename OrFunc>
Faster new DynamicBloom implementation (for memtable) (#5762) Summary: Since DynamicBloom is now only used in-memory, we're free to change it without schema compatibility issues. The new implementation is drawn from (with manifest permission) https://github.com/pdillinger/wormhashing/blob/303542a767437f56d8b66cea6ebecaac0e6a61e9/bloom_simulation_tests/foo.cc#L613 This has several speed advantages over the prior implementation: * Uses fastrange instead of % * Minimum logic to determine first (and all) probed memory addresses * (Major) Two probes per 64-bit memory fetch/write. * Very fast and effective (murmur-like) hash expansion/re-mixing. (At least on recent CPUs, integer multiplication is very cheap.) While a Bloom filter with 512-bit cache locality has about a 1.15x FP rate penalty (e.g. 0.84% to 0.97%), further restricting to two probes per 64 bits incurs an additional 1.12x FP rate penalty (e.g. 0.97% to 1.09%). Nevertheless, the unit tests show no "mediocre" FP rate samples, unlike the old implementation with more erratic FP rates. Especially for the memtable, we expect speed to outweigh somewhat higher FP rates. For example, a negative table query would have to be 1000x slower than a BF query to justify doubling BF query time to shave 10% off FP rate (working assumption around 1% FP rate). While that seems likely for SSTs, my data suggests a speed factor of roughly 50x for the memtable (vs. BF; ~1.5% lower write throughput when enabling memtable Bloom filter, after this change). Thus, it's probably not worth even 5% more time in the Bloom filter to shave off 1/10th of the Bloom FP rate, or 0.1% in absolute terms, and it's probably at least 20% slower to recoup that much FP rate from this new implementation. Because of this, we do not see a need for a 'locality' option that affects the MemTable Bloom filter and have decoupled the MemTable Bloom filter from Options::bloom_locality. Note that just 3% more memory to the Bloom filter (10.3 bits per key vs. just 10) is able to make up for the ~12% FP rate drop in the new implementation: [] # Nearly "ideal" FP-wise but reasonably fast cache-local implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out 10000000 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out time: 3.29372 sampled_fp_rate: 0.00985956 ... [] # Close match to this new implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out 10000000 6 10.3 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.10072 sampled_fp_rate: 0.00985655 ... [] # Old locality=1 implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out 10000000 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out time: 3.95472 sampled_fp_rate: 0.00988943 ... Also note the dramatic speed improvement vs. alternatives. -- Performance unit test: DynamicBloomTest.concurrent_with_perf is updated to report more precise timing data. (Measure running time of each thread, not just longest running thread, etc.) Results averaged over various sizes enabled with --enable_perf and 20 runs each; old dynamic bloom refers to locality=1, the faster of the old: old dynamic bloom, avg add latency = 65.6468 new dynamic bloom, avg add latency = 44.3809 old dynamic bloom, avg query latency = 50.6485 new dynamic bloom, avg query latency = 43.2186 old avg parallel add latency = 41.678 new avg parallel add latency = 24.5238 old avg parallel hit latency = 14.6322 new avg parallel hit latency = 12.3939 old avg parallel miss latency = 16.7289 new avg parallel miss latency = 12.2134 Tested on a dedicated 64-bit production machine at Facebook. Significant improvement all around. Despite now using std::atomic<uint64_t>, quick before-and-after test on a 32-bit machine (Intel Atom N270, released 2008) shows no regression in performance, in some cases modest improvement. -- Performance integration test (synthetic): with DEBUG_LEVEL=0, used TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=fillrandom,readmissing,readrandom,stats --num=2000000 and optionally with -memtable_whole_key_filtering -memtable_bloom_size_ratio=0.01 300 runs each configuration. Write throughput change by enabling memtable bloom: Old locality=0: -3.06% Old locality=1: -2.37% New: -1.50% conclusion -> seems to substantially close the gap Readmissing throughput change by enabling memtable bloom: Old locality=0: +34.47% Old locality=1: +34.80% New: +33.25% conclusion -> maybe a small new penalty from FP rate Readrandom throughput change by enabling memtable bloom: Old locality=0: +31.54% Old locality=1: +31.13% New: +30.60% conclusion -> maybe also from FP rate (after memtable flush) -- Another conclusion we can draw from this new implementation is that the existing 32-bit hash function is not inherently crippling the Bloom filter speed or accuracy, below about 5 million keys. For speed, the implementation is essentially the same whether starting with 32-bits or 64-bits of hash; it just determines whether the first multiplication after fastrange is a pseudorandom expansion or needed re-mix. Note that this multiplication can occur while memory is fetching. For accuracy, in a standard configuration, you need about 5 million keys before you have about a 1.1x FP penalty due to using a 32-bit hash vs. 64-bit: [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.52069 sampled_fp_rate: 0.0118267 ... [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out time: 2.43871 sampled_fp_rate: 0.0109059 Pull Request resolved: https://github.com/facebook/rocksdb/pull/5762 Differential Revision: D17214194 Pulled By: pdillinger fbshipit-source-id: ad9da031772e985fd6b62a0e1db8e81892520595
2019-09-05 23:57:39 +02:00
inline void DynamicBloom::AddHash(uint32_t h32, const OrFunc& or_func) {
size_t a = FastRange32(kLen, h32);
Faster new DynamicBloom implementation (for memtable) (#5762) Summary: Since DynamicBloom is now only used in-memory, we're free to change it without schema compatibility issues. The new implementation is drawn from (with manifest permission) https://github.com/pdillinger/wormhashing/blob/303542a767437f56d8b66cea6ebecaac0e6a61e9/bloom_simulation_tests/foo.cc#L613 This has several speed advantages over the prior implementation: * Uses fastrange instead of % * Minimum logic to determine first (and all) probed memory addresses * (Major) Two probes per 64-bit memory fetch/write. * Very fast and effective (murmur-like) hash expansion/re-mixing. (At least on recent CPUs, integer multiplication is very cheap.) While a Bloom filter with 512-bit cache locality has about a 1.15x FP rate penalty (e.g. 0.84% to 0.97%), further restricting to two probes per 64 bits incurs an additional 1.12x FP rate penalty (e.g. 0.97% to 1.09%). Nevertheless, the unit tests show no "mediocre" FP rate samples, unlike the old implementation with more erratic FP rates. Especially for the memtable, we expect speed to outweigh somewhat higher FP rates. For example, a negative table query would have to be 1000x slower than a BF query to justify doubling BF query time to shave 10% off FP rate (working assumption around 1% FP rate). While that seems likely for SSTs, my data suggests a speed factor of roughly 50x for the memtable (vs. BF; ~1.5% lower write throughput when enabling memtable Bloom filter, after this change). Thus, it's probably not worth even 5% more time in the Bloom filter to shave off 1/10th of the Bloom FP rate, or 0.1% in absolute terms, and it's probably at least 20% slower to recoup that much FP rate from this new implementation. Because of this, we do not see a need for a 'locality' option that affects the MemTable Bloom filter and have decoupled the MemTable Bloom filter from Options::bloom_locality. Note that just 3% more memory to the Bloom filter (10.3 bits per key vs. just 10) is able to make up for the ~12% FP rate drop in the new implementation: [] # Nearly "ideal" FP-wise but reasonably fast cache-local implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out 10000000 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out time: 3.29372 sampled_fp_rate: 0.00985956 ... [] # Close match to this new implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out 10000000 6 10.3 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.10072 sampled_fp_rate: 0.00985655 ... [] # Old locality=1 implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out 10000000 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out time: 3.95472 sampled_fp_rate: 0.00988943 ... Also note the dramatic speed improvement vs. alternatives. -- Performance unit test: DynamicBloomTest.concurrent_with_perf is updated to report more precise timing data. (Measure running time of each thread, not just longest running thread, etc.) Results averaged over various sizes enabled with --enable_perf and 20 runs each; old dynamic bloom refers to locality=1, the faster of the old: old dynamic bloom, avg add latency = 65.6468 new dynamic bloom, avg add latency = 44.3809 old dynamic bloom, avg query latency = 50.6485 new dynamic bloom, avg query latency = 43.2186 old avg parallel add latency = 41.678 new avg parallel add latency = 24.5238 old avg parallel hit latency = 14.6322 new avg parallel hit latency = 12.3939 old avg parallel miss latency = 16.7289 new avg parallel miss latency = 12.2134 Tested on a dedicated 64-bit production machine at Facebook. Significant improvement all around. Despite now using std::atomic<uint64_t>, quick before-and-after test on a 32-bit machine (Intel Atom N270, released 2008) shows no regression in performance, in some cases modest improvement. -- Performance integration test (synthetic): with DEBUG_LEVEL=0, used TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=fillrandom,readmissing,readrandom,stats --num=2000000 and optionally with -memtable_whole_key_filtering -memtable_bloom_size_ratio=0.01 300 runs each configuration. Write throughput change by enabling memtable bloom: Old locality=0: -3.06% Old locality=1: -2.37% New: -1.50% conclusion -> seems to substantially close the gap Readmissing throughput change by enabling memtable bloom: Old locality=0: +34.47% Old locality=1: +34.80% New: +33.25% conclusion -> maybe a small new penalty from FP rate Readrandom throughput change by enabling memtable bloom: Old locality=0: +31.54% Old locality=1: +31.13% New: +30.60% conclusion -> maybe also from FP rate (after memtable flush) -- Another conclusion we can draw from this new implementation is that the existing 32-bit hash function is not inherently crippling the Bloom filter speed or accuracy, below about 5 million keys. For speed, the implementation is essentially the same whether starting with 32-bits or 64-bits of hash; it just determines whether the first multiplication after fastrange is a pseudorandom expansion or needed re-mix. Note that this multiplication can occur while memory is fetching. For accuracy, in a standard configuration, you need about 5 million keys before you have about a 1.1x FP penalty due to using a 32-bit hash vs. 64-bit: [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.52069 sampled_fp_rate: 0.0118267 ... [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out time: 2.43871 sampled_fp_rate: 0.0109059 Pull Request resolved: https://github.com/facebook/rocksdb/pull/5762 Differential Revision: D17214194 Pulled By: pdillinger fbshipit-source-id: ad9da031772e985fd6b62a0e1db8e81892520595
2019-09-05 23:57:39 +02:00
PREFETCH(data_ + a, 0, 3);
// Expand/remix with 64-bit golden ratio
uint64_t h = 0x9e3779b97f4a7c13ULL * h32;
for (unsigned i = 0;; ++i) {
// Two bit probes per uint64_t probe
uint64_t mask =
((uint64_t)1 << (h & 63)) | ((uint64_t)1 << ((h >> 6) & 63));
Faster new DynamicBloom implementation (for memtable) (#5762) Summary: Since DynamicBloom is now only used in-memory, we're free to change it without schema compatibility issues. The new implementation is drawn from (with manifest permission) https://github.com/pdillinger/wormhashing/blob/303542a767437f56d8b66cea6ebecaac0e6a61e9/bloom_simulation_tests/foo.cc#L613 This has several speed advantages over the prior implementation: * Uses fastrange instead of % * Minimum logic to determine first (and all) probed memory addresses * (Major) Two probes per 64-bit memory fetch/write. * Very fast and effective (murmur-like) hash expansion/re-mixing. (At least on recent CPUs, integer multiplication is very cheap.) While a Bloom filter with 512-bit cache locality has about a 1.15x FP rate penalty (e.g. 0.84% to 0.97%), further restricting to two probes per 64 bits incurs an additional 1.12x FP rate penalty (e.g. 0.97% to 1.09%). Nevertheless, the unit tests show no "mediocre" FP rate samples, unlike the old implementation with more erratic FP rates. Especially for the memtable, we expect speed to outweigh somewhat higher FP rates. For example, a negative table query would have to be 1000x slower than a BF query to justify doubling BF query time to shave 10% off FP rate (working assumption around 1% FP rate). While that seems likely for SSTs, my data suggests a speed factor of roughly 50x for the memtable (vs. BF; ~1.5% lower write throughput when enabling memtable Bloom filter, after this change). Thus, it's probably not worth even 5% more time in the Bloom filter to shave off 1/10th of the Bloom FP rate, or 0.1% in absolute terms, and it's probably at least 20% slower to recoup that much FP rate from this new implementation. Because of this, we do not see a need for a 'locality' option that affects the MemTable Bloom filter and have decoupled the MemTable Bloom filter from Options::bloom_locality. Note that just 3% more memory to the Bloom filter (10.3 bits per key vs. just 10) is able to make up for the ~12% FP rate drop in the new implementation: [] # Nearly "ideal" FP-wise but reasonably fast cache-local implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out 10000000 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out time: 3.29372 sampled_fp_rate: 0.00985956 ... [] # Close match to this new implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out 10000000 6 10.3 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.10072 sampled_fp_rate: 0.00985655 ... [] # Old locality=1 implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out 10000000 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out time: 3.95472 sampled_fp_rate: 0.00988943 ... Also note the dramatic speed improvement vs. alternatives. -- Performance unit test: DynamicBloomTest.concurrent_with_perf is updated to report more precise timing data. (Measure running time of each thread, not just longest running thread, etc.) Results averaged over various sizes enabled with --enable_perf and 20 runs each; old dynamic bloom refers to locality=1, the faster of the old: old dynamic bloom, avg add latency = 65.6468 new dynamic bloom, avg add latency = 44.3809 old dynamic bloom, avg query latency = 50.6485 new dynamic bloom, avg query latency = 43.2186 old avg parallel add latency = 41.678 new avg parallel add latency = 24.5238 old avg parallel hit latency = 14.6322 new avg parallel hit latency = 12.3939 old avg parallel miss latency = 16.7289 new avg parallel miss latency = 12.2134 Tested on a dedicated 64-bit production machine at Facebook. Significant improvement all around. Despite now using std::atomic<uint64_t>, quick before-and-after test on a 32-bit machine (Intel Atom N270, released 2008) shows no regression in performance, in some cases modest improvement. -- Performance integration test (synthetic): with DEBUG_LEVEL=0, used TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=fillrandom,readmissing,readrandom,stats --num=2000000 and optionally with -memtable_whole_key_filtering -memtable_bloom_size_ratio=0.01 300 runs each configuration. Write throughput change by enabling memtable bloom: Old locality=0: -3.06% Old locality=1: -2.37% New: -1.50% conclusion -> seems to substantially close the gap Readmissing throughput change by enabling memtable bloom: Old locality=0: +34.47% Old locality=1: +34.80% New: +33.25% conclusion -> maybe a small new penalty from FP rate Readrandom throughput change by enabling memtable bloom: Old locality=0: +31.54% Old locality=1: +31.13% New: +30.60% conclusion -> maybe also from FP rate (after memtable flush) -- Another conclusion we can draw from this new implementation is that the existing 32-bit hash function is not inherently crippling the Bloom filter speed or accuracy, below about 5 million keys. For speed, the implementation is essentially the same whether starting with 32-bits or 64-bits of hash; it just determines whether the first multiplication after fastrange is a pseudorandom expansion or needed re-mix. Note that this multiplication can occur while memory is fetching. For accuracy, in a standard configuration, you need about 5 million keys before you have about a 1.1x FP penalty due to using a 32-bit hash vs. 64-bit: [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.52069 sampled_fp_rate: 0.0118267 ... [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out time: 2.43871 sampled_fp_rate: 0.0109059 Pull Request resolved: https://github.com/facebook/rocksdb/pull/5762 Differential Revision: D17214194 Pulled By: pdillinger fbshipit-source-id: ad9da031772e985fd6b62a0e1db8e81892520595
2019-09-05 23:57:39 +02:00
or_func(&data_[a ^ i], mask);
if (i + 1 >= kNumDoubleProbes) {
return;
}
Faster new DynamicBloom implementation (for memtable) (#5762) Summary: Since DynamicBloom is now only used in-memory, we're free to change it without schema compatibility issues. The new implementation is drawn from (with manifest permission) https://github.com/pdillinger/wormhashing/blob/303542a767437f56d8b66cea6ebecaac0e6a61e9/bloom_simulation_tests/foo.cc#L613 This has several speed advantages over the prior implementation: * Uses fastrange instead of % * Minimum logic to determine first (and all) probed memory addresses * (Major) Two probes per 64-bit memory fetch/write. * Very fast and effective (murmur-like) hash expansion/re-mixing. (At least on recent CPUs, integer multiplication is very cheap.) While a Bloom filter with 512-bit cache locality has about a 1.15x FP rate penalty (e.g. 0.84% to 0.97%), further restricting to two probes per 64 bits incurs an additional 1.12x FP rate penalty (e.g. 0.97% to 1.09%). Nevertheless, the unit tests show no "mediocre" FP rate samples, unlike the old implementation with more erratic FP rates. Especially for the memtable, we expect speed to outweigh somewhat higher FP rates. For example, a negative table query would have to be 1000x slower than a BF query to justify doubling BF query time to shave 10% off FP rate (working assumption around 1% FP rate). While that seems likely for SSTs, my data suggests a speed factor of roughly 50x for the memtable (vs. BF; ~1.5% lower write throughput when enabling memtable Bloom filter, after this change). Thus, it's probably not worth even 5% more time in the Bloom filter to shave off 1/10th of the Bloom FP rate, or 0.1% in absolute terms, and it's probably at least 20% slower to recoup that much FP rate from this new implementation. Because of this, we do not see a need for a 'locality' option that affects the MemTable Bloom filter and have decoupled the MemTable Bloom filter from Options::bloom_locality. Note that just 3% more memory to the Bloom filter (10.3 bits per key vs. just 10) is able to make up for the ~12% FP rate drop in the new implementation: [] # Nearly "ideal" FP-wise but reasonably fast cache-local implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out 10000000 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_WORM64_FROM32_any.out time: 3.29372 sampled_fp_rate: 0.00985956 ... [] # Close match to this new implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out 10000000 6 10.3 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.10072 sampled_fp_rate: 0.00985655 ... [] # Old locality=1 implementation [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out 10000000 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_ROCKSDB_DYNAMIC_any.out time: 3.95472 sampled_fp_rate: 0.00988943 ... Also note the dramatic speed improvement vs. alternatives. -- Performance unit test: DynamicBloomTest.concurrent_with_perf is updated to report more precise timing data. (Measure running time of each thread, not just longest running thread, etc.) Results averaged over various sizes enabled with --enable_perf and 20 runs each; old dynamic bloom refers to locality=1, the faster of the old: old dynamic bloom, avg add latency = 65.6468 new dynamic bloom, avg add latency = 44.3809 old dynamic bloom, avg query latency = 50.6485 new dynamic bloom, avg query latency = 43.2186 old avg parallel add latency = 41.678 new avg parallel add latency = 24.5238 old avg parallel hit latency = 14.6322 new avg parallel hit latency = 12.3939 old avg parallel miss latency = 16.7289 new avg parallel miss latency = 12.2134 Tested on a dedicated 64-bit production machine at Facebook. Significant improvement all around. Despite now using std::atomic<uint64_t>, quick before-and-after test on a 32-bit machine (Intel Atom N270, released 2008) shows no regression in performance, in some cases modest improvement. -- Performance integration test (synthetic): with DEBUG_LEVEL=0, used TEST_TMPDIR=/dev/shm ./db_bench --benchmarks=fillrandom,readmissing,readrandom,stats --num=2000000 and optionally with -memtable_whole_key_filtering -memtable_bloom_size_ratio=0.01 300 runs each configuration. Write throughput change by enabling memtable bloom: Old locality=0: -3.06% Old locality=1: -2.37% New: -1.50% conclusion -> seems to substantially close the gap Readmissing throughput change by enabling memtable bloom: Old locality=0: +34.47% Old locality=1: +34.80% New: +33.25% conclusion -> maybe a small new penalty from FP rate Readrandom throughput change by enabling memtable bloom: Old locality=0: +31.54% Old locality=1: +31.13% New: +30.60% conclusion -> maybe also from FP rate (after memtable flush) -- Another conclusion we can draw from this new implementation is that the existing 32-bit hash function is not inherently crippling the Bloom filter speed or accuracy, below about 5 million keys. For speed, the implementation is essentially the same whether starting with 32-bits or 64-bits of hash; it just determines whether the first multiplication after fastrange is a pseudorandom expansion or needed re-mix. Note that this multiplication can occur while memory is fetching. For accuracy, in a standard configuration, you need about 5 million keys before you have about a 1.1x FP penalty due to using a 32-bit hash vs. 64-bit: [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_FROM32_any.out time: 2.52069 sampled_fp_rate: 0.0118267 ... [~/wormhashing/bloom_simulation_tests] ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out $((5 * 1000 * 1000 * 10)) 6 10 $RANDOM 100000000 ./foo_gcc_IMPL_CACHE_MUL64_BLOCK_any.out time: 2.43871 sampled_fp_rate: 0.0109059 Pull Request resolved: https://github.com/facebook/rocksdb/pull/5762 Differential Revision: D17214194 Pulled By: pdillinger fbshipit-source-id: ad9da031772e985fd6b62a0e1db8e81892520595
2019-09-05 23:57:39 +02:00
h = (h >> 12) | (h << 52);
}
}
} // namespace ROCKSDB_NAMESPACE