rocksdb/cache/cache_bench.cc

382 lines
11 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).
#ifndef GFLAGS
#include <cstdio>
int main() {
fprintf(stderr, "Please install gflags to run rocksdb tools\n");
return 1;
}
#else
#include <stdio.h>
#include <sys/types.h>
#include <cinttypes>
Revamp cache_bench to resemble a real workload (#6629) Summary: I suspect LRUCache could use some optimization, and to support such an effort, a good benchmarking tool is needed. The existing cache_bench was heavily skewed toward insertion and lookup misses, and did not saturate memory with other work. This change should improve those things to better resemble a real workload. (All below using clang compiler, for some consistency, but not necessarily same version and settings.) The real workload is from production MySQL on RocksDB, filtering stacks containing "LRU", "ShardedCache" or "CacheShard." Lookup inclusive: 66% Insert inclusive: 17% Release inclusive: 15% An alternate simulated workload is MySQL running a LinkBench read test: Lookup inclusive: 54% Insert inclusive: 24% Release inclusive: 21% cache_bench default settings, prior to this change: Lookup inclusive: 35.8% Insert inclusive: 63.6% Release inclusive: 0% cache_bench after this change (intended as somewhat "tighter" workload than average production, more like LinkBench): Lookup inclusive: 52% Insert inclusive: 20% Release inclusive: 26% And top exclusive stacks (portion of stack samples as filtered above): Production MySQL: LRUHandleTable::FindPointer: 25.3% rocksdb::operator==: 15.1% <-- Slice == LRUCacheShard::LRU_Remove: 13.8% ShardedCache::Lookup: 8.9% __pthread_mutex_lock: 7.1% LRUCacheShard::LRU_Insert: 6.3% MurmurHash64A: 4.8% <-- Since upgraded to XXH3p ... Old cache_bench: LRUHandleTable::FindPointer: 23.6% __pthread_mutex_lock: 15.0% __pthread_mutex_unlock_usercnt: 11.7% __lll_lock_wait: 8.6% __lll_unlock_wake: 6.8% LRUCacheShard::LRU_Insert: 6.0% ShardedCache::Lookup: 4.4% LRUCacheShard::LRU_Remove: 2.8% ... rocksdb::operator==: 0.2% <-- Slice == ... New cache_bench: LRUHandleTable::FindPointer: 22.8% __pthread_mutex_unlock_usercnt: 14.3% rocksdb::operator==: 10.5% <-- Slice == LRUCacheShard::LRU_Insert: 9.0% __pthread_mutex_lock: 5.9% LRUCacheShard::LRU_Remove: 5.0% ... ShardedCache::Lookup: 2.9% ... So there's a bit more lock contention in the benchmark than in production, but otherwise looks similar enough to me. At least it's a big improvement over the existing code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6629 Test Plan: No production code changes, ran cache_bench with ASAN Reviewed By: ltamasi Differential Revision: D20824318 Pulled By: pdillinger fbshipit-source-id: 6f8dc5891ead0f87edbed3a615ecd5289d9abe12
2020-04-03 19:24:09 +02:00
#include <limits>
#include "port/port.h"
#include "rocksdb/cache.h"
#include "rocksdb/db.h"
#include "rocksdb/env.h"
Revamp cache_bench to resemble a real workload (#6629) Summary: I suspect LRUCache could use some optimization, and to support such an effort, a good benchmarking tool is needed. The existing cache_bench was heavily skewed toward insertion and lookup misses, and did not saturate memory with other work. This change should improve those things to better resemble a real workload. (All below using clang compiler, for some consistency, but not necessarily same version and settings.) The real workload is from production MySQL on RocksDB, filtering stacks containing "LRU", "ShardedCache" or "CacheShard." Lookup inclusive: 66% Insert inclusive: 17% Release inclusive: 15% An alternate simulated workload is MySQL running a LinkBench read test: Lookup inclusive: 54% Insert inclusive: 24% Release inclusive: 21% cache_bench default settings, prior to this change: Lookup inclusive: 35.8% Insert inclusive: 63.6% Release inclusive: 0% cache_bench after this change (intended as somewhat "tighter" workload than average production, more like LinkBench): Lookup inclusive: 52% Insert inclusive: 20% Release inclusive: 26% And top exclusive stacks (portion of stack samples as filtered above): Production MySQL: LRUHandleTable::FindPointer: 25.3% rocksdb::operator==: 15.1% <-- Slice == LRUCacheShard::LRU_Remove: 13.8% ShardedCache::Lookup: 8.9% __pthread_mutex_lock: 7.1% LRUCacheShard::LRU_Insert: 6.3% MurmurHash64A: 4.8% <-- Since upgraded to XXH3p ... Old cache_bench: LRUHandleTable::FindPointer: 23.6% __pthread_mutex_lock: 15.0% __pthread_mutex_unlock_usercnt: 11.7% __lll_lock_wait: 8.6% __lll_unlock_wake: 6.8% LRUCacheShard::LRU_Insert: 6.0% ShardedCache::Lookup: 4.4% LRUCacheShard::LRU_Remove: 2.8% ... rocksdb::operator==: 0.2% <-- Slice == ... New cache_bench: LRUHandleTable::FindPointer: 22.8% __pthread_mutex_unlock_usercnt: 14.3% rocksdb::operator==: 10.5% <-- Slice == LRUCacheShard::LRU_Insert: 9.0% __pthread_mutex_lock: 5.9% LRUCacheShard::LRU_Remove: 5.0% ... ShardedCache::Lookup: 2.9% ... So there's a bit more lock contention in the benchmark than in production, but otherwise looks similar enough to me. At least it's a big improvement over the existing code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6629 Test Plan: No production code changes, ran cache_bench with ASAN Reviewed By: ltamasi Differential Revision: D20824318 Pulled By: pdillinger fbshipit-source-id: 6f8dc5891ead0f87edbed3a615ecd5289d9abe12
2020-04-03 19:24:09 +02:00
#include "util/coding.h"
#include "util/gflags_compat.h"
Revamp cache_bench to resemble a real workload (#6629) Summary: I suspect LRUCache could use some optimization, and to support such an effort, a good benchmarking tool is needed. The existing cache_bench was heavily skewed toward insertion and lookup misses, and did not saturate memory with other work. This change should improve those things to better resemble a real workload. (All below using clang compiler, for some consistency, but not necessarily same version and settings.) The real workload is from production MySQL on RocksDB, filtering stacks containing "LRU", "ShardedCache" or "CacheShard." Lookup inclusive: 66% Insert inclusive: 17% Release inclusive: 15% An alternate simulated workload is MySQL running a LinkBench read test: Lookup inclusive: 54% Insert inclusive: 24% Release inclusive: 21% cache_bench default settings, prior to this change: Lookup inclusive: 35.8% Insert inclusive: 63.6% Release inclusive: 0% cache_bench after this change (intended as somewhat "tighter" workload than average production, more like LinkBench): Lookup inclusive: 52% Insert inclusive: 20% Release inclusive: 26% And top exclusive stacks (portion of stack samples as filtered above): Production MySQL: LRUHandleTable::FindPointer: 25.3% rocksdb::operator==: 15.1% <-- Slice == LRUCacheShard::LRU_Remove: 13.8% ShardedCache::Lookup: 8.9% __pthread_mutex_lock: 7.1% LRUCacheShard::LRU_Insert: 6.3% MurmurHash64A: 4.8% <-- Since upgraded to XXH3p ... Old cache_bench: LRUHandleTable::FindPointer: 23.6% __pthread_mutex_lock: 15.0% __pthread_mutex_unlock_usercnt: 11.7% __lll_lock_wait: 8.6% __lll_unlock_wake: 6.8% LRUCacheShard::LRU_Insert: 6.0% ShardedCache::Lookup: 4.4% LRUCacheShard::LRU_Remove: 2.8% ... rocksdb::operator==: 0.2% <-- Slice == ... New cache_bench: LRUHandleTable::FindPointer: 22.8% __pthread_mutex_unlock_usercnt: 14.3% rocksdb::operator==: 10.5% <-- Slice == LRUCacheShard::LRU_Insert: 9.0% __pthread_mutex_lock: 5.9% LRUCacheShard::LRU_Remove: 5.0% ... ShardedCache::Lookup: 2.9% ... So there's a bit more lock contention in the benchmark than in production, but otherwise looks similar enough to me. At least it's a big improvement over the existing code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6629 Test Plan: No production code changes, ran cache_bench with ASAN Reviewed By: ltamasi Differential Revision: D20824318 Pulled By: pdillinger fbshipit-source-id: 6f8dc5891ead0f87edbed3a615ecd5289d9abe12
2020-04-03 19:24:09 +02:00
#include "util/hash.h"
#include "util/mutexlock.h"
#include "util/random.h"
using GFLAGS_NAMESPACE::ParseCommandLineFlags;
Revamp cache_bench to resemble a real workload (#6629) Summary: I suspect LRUCache could use some optimization, and to support such an effort, a good benchmarking tool is needed. The existing cache_bench was heavily skewed toward insertion and lookup misses, and did not saturate memory with other work. This change should improve those things to better resemble a real workload. (All below using clang compiler, for some consistency, but not necessarily same version and settings.) The real workload is from production MySQL on RocksDB, filtering stacks containing "LRU", "ShardedCache" or "CacheShard." Lookup inclusive: 66% Insert inclusive: 17% Release inclusive: 15% An alternate simulated workload is MySQL running a LinkBench read test: Lookup inclusive: 54% Insert inclusive: 24% Release inclusive: 21% cache_bench default settings, prior to this change: Lookup inclusive: 35.8% Insert inclusive: 63.6% Release inclusive: 0% cache_bench after this change (intended as somewhat "tighter" workload than average production, more like LinkBench): Lookup inclusive: 52% Insert inclusive: 20% Release inclusive: 26% And top exclusive stacks (portion of stack samples as filtered above): Production MySQL: LRUHandleTable::FindPointer: 25.3% rocksdb::operator==: 15.1% <-- Slice == LRUCacheShard::LRU_Remove: 13.8% ShardedCache::Lookup: 8.9% __pthread_mutex_lock: 7.1% LRUCacheShard::LRU_Insert: 6.3% MurmurHash64A: 4.8% <-- Since upgraded to XXH3p ... Old cache_bench: LRUHandleTable::FindPointer: 23.6% __pthread_mutex_lock: 15.0% __pthread_mutex_unlock_usercnt: 11.7% __lll_lock_wait: 8.6% __lll_unlock_wake: 6.8% LRUCacheShard::LRU_Insert: 6.0% ShardedCache::Lookup: 4.4% LRUCacheShard::LRU_Remove: 2.8% ... rocksdb::operator==: 0.2% <-- Slice == ... New cache_bench: LRUHandleTable::FindPointer: 22.8% __pthread_mutex_unlock_usercnt: 14.3% rocksdb::operator==: 10.5% <-- Slice == LRUCacheShard::LRU_Insert: 9.0% __pthread_mutex_lock: 5.9% LRUCacheShard::LRU_Remove: 5.0% ... ShardedCache::Lookup: 2.9% ... So there's a bit more lock contention in the benchmark than in production, but otherwise looks similar enough to me. At least it's a big improvement over the existing code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6629 Test Plan: No production code changes, ran cache_bench with ASAN Reviewed By: ltamasi Differential Revision: D20824318 Pulled By: pdillinger fbshipit-source-id: 6f8dc5891ead0f87edbed3a615ecd5289d9abe12
2020-04-03 19:24:09 +02:00
static constexpr uint32_t KiB = uint32_t{1} << 10;
static constexpr uint32_t MiB = KiB << 10;
static constexpr uint64_t GiB = MiB << 10;
DEFINE_uint32(threads, 16, "Number of concurrent threads to run.");
DEFINE_uint64(cache_size, 1 * GiB,
"Number of bytes to use as a cache of uncompressed data.");
DEFINE_uint32(num_shard_bits, 6, "shard_bits.");
DEFINE_double(resident_ratio, 0.25,
"Ratio of keys fitting in cache to keyspace.");
DEFINE_uint64(ops_per_thread, 0,
"Number of operations per thread. (Default: 5 * keyspace size)");
DEFINE_uint32(value_bytes, 8 * KiB, "Size of each value added.");
DEFINE_uint32(skew, 5, "Degree of skew in key selection");
DEFINE_bool(populate_cache, true, "Populate cache before operations");
DEFINE_uint32(lookup_insert_percent, 87,
"Ratio of lookup (+ insert on not found) to total workload "
"(expressed as a percentage)");
DEFINE_uint32(insert_percent, 2,
"Ratio of insert to total workload (expressed as a percentage)");
DEFINE_uint32(lookup_percent, 10,
"Ratio of lookup to total workload (expressed as a percentage)");
DEFINE_uint32(erase_percent, 1,
"Ratio of erase to total workload (expressed as a percentage)");
DEFINE_bool(use_clock_cache, false, "");
namespace ROCKSDB_NAMESPACE {
class CacheBench;
namespace {
// State shared by all concurrent executions of the same benchmark.
class SharedState {
public:
explicit SharedState(CacheBench* cache_bench)
: cv_(&mu_),
num_initialized_(0),
start_(false),
num_done_(0),
Revamp cache_bench to resemble a real workload (#6629) Summary: I suspect LRUCache could use some optimization, and to support such an effort, a good benchmarking tool is needed. The existing cache_bench was heavily skewed toward insertion and lookup misses, and did not saturate memory with other work. This change should improve those things to better resemble a real workload. (All below using clang compiler, for some consistency, but not necessarily same version and settings.) The real workload is from production MySQL on RocksDB, filtering stacks containing "LRU", "ShardedCache" or "CacheShard." Lookup inclusive: 66% Insert inclusive: 17% Release inclusive: 15% An alternate simulated workload is MySQL running a LinkBench read test: Lookup inclusive: 54% Insert inclusive: 24% Release inclusive: 21% cache_bench default settings, prior to this change: Lookup inclusive: 35.8% Insert inclusive: 63.6% Release inclusive: 0% cache_bench after this change (intended as somewhat "tighter" workload than average production, more like LinkBench): Lookup inclusive: 52% Insert inclusive: 20% Release inclusive: 26% And top exclusive stacks (portion of stack samples as filtered above): Production MySQL: LRUHandleTable::FindPointer: 25.3% rocksdb::operator==: 15.1% <-- Slice == LRUCacheShard::LRU_Remove: 13.8% ShardedCache::Lookup: 8.9% __pthread_mutex_lock: 7.1% LRUCacheShard::LRU_Insert: 6.3% MurmurHash64A: 4.8% <-- Since upgraded to XXH3p ... Old cache_bench: LRUHandleTable::FindPointer: 23.6% __pthread_mutex_lock: 15.0% __pthread_mutex_unlock_usercnt: 11.7% __lll_lock_wait: 8.6% __lll_unlock_wake: 6.8% LRUCacheShard::LRU_Insert: 6.0% ShardedCache::Lookup: 4.4% LRUCacheShard::LRU_Remove: 2.8% ... rocksdb::operator==: 0.2% <-- Slice == ... New cache_bench: LRUHandleTable::FindPointer: 22.8% __pthread_mutex_unlock_usercnt: 14.3% rocksdb::operator==: 10.5% <-- Slice == LRUCacheShard::LRU_Insert: 9.0% __pthread_mutex_lock: 5.9% LRUCacheShard::LRU_Remove: 5.0% ... ShardedCache::Lookup: 2.9% ... So there's a bit more lock contention in the benchmark than in production, but otherwise looks similar enough to me. At least it's a big improvement over the existing code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6629 Test Plan: No production code changes, ran cache_bench with ASAN Reviewed By: ltamasi Differential Revision: D20824318 Pulled By: pdillinger fbshipit-source-id: 6f8dc5891ead0f87edbed3a615ecd5289d9abe12
2020-04-03 19:24:09 +02:00
cache_bench_(cache_bench) {}
~SharedState() {}
port::Mutex* GetMutex() {
return &mu_;
}
port::CondVar* GetCondVar() {
return &cv_;
}
CacheBench* GetCacheBench() const {
return cache_bench_;
}
void IncInitialized() {
num_initialized_++;
}
void IncDone() {
num_done_++;
}
Revamp cache_bench to resemble a real workload (#6629) Summary: I suspect LRUCache could use some optimization, and to support such an effort, a good benchmarking tool is needed. The existing cache_bench was heavily skewed toward insertion and lookup misses, and did not saturate memory with other work. This change should improve those things to better resemble a real workload. (All below using clang compiler, for some consistency, but not necessarily same version and settings.) The real workload is from production MySQL on RocksDB, filtering stacks containing "LRU", "ShardedCache" or "CacheShard." Lookup inclusive: 66% Insert inclusive: 17% Release inclusive: 15% An alternate simulated workload is MySQL running a LinkBench read test: Lookup inclusive: 54% Insert inclusive: 24% Release inclusive: 21% cache_bench default settings, prior to this change: Lookup inclusive: 35.8% Insert inclusive: 63.6% Release inclusive: 0% cache_bench after this change (intended as somewhat "tighter" workload than average production, more like LinkBench): Lookup inclusive: 52% Insert inclusive: 20% Release inclusive: 26% And top exclusive stacks (portion of stack samples as filtered above): Production MySQL: LRUHandleTable::FindPointer: 25.3% rocksdb::operator==: 15.1% <-- Slice == LRUCacheShard::LRU_Remove: 13.8% ShardedCache::Lookup: 8.9% __pthread_mutex_lock: 7.1% LRUCacheShard::LRU_Insert: 6.3% MurmurHash64A: 4.8% <-- Since upgraded to XXH3p ... Old cache_bench: LRUHandleTable::FindPointer: 23.6% __pthread_mutex_lock: 15.0% __pthread_mutex_unlock_usercnt: 11.7% __lll_lock_wait: 8.6% __lll_unlock_wake: 6.8% LRUCacheShard::LRU_Insert: 6.0% ShardedCache::Lookup: 4.4% LRUCacheShard::LRU_Remove: 2.8% ... rocksdb::operator==: 0.2% <-- Slice == ... New cache_bench: LRUHandleTable::FindPointer: 22.8% __pthread_mutex_unlock_usercnt: 14.3% rocksdb::operator==: 10.5% <-- Slice == LRUCacheShard::LRU_Insert: 9.0% __pthread_mutex_lock: 5.9% LRUCacheShard::LRU_Remove: 5.0% ... ShardedCache::Lookup: 2.9% ... So there's a bit more lock contention in the benchmark than in production, but otherwise looks similar enough to me. At least it's a big improvement over the existing code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6629 Test Plan: No production code changes, ran cache_bench with ASAN Reviewed By: ltamasi Differential Revision: D20824318 Pulled By: pdillinger fbshipit-source-id: 6f8dc5891ead0f87edbed3a615ecd5289d9abe12
2020-04-03 19:24:09 +02:00
bool AllInitialized() const { return num_initialized_ >= FLAGS_threads; }
Revamp cache_bench to resemble a real workload (#6629) Summary: I suspect LRUCache could use some optimization, and to support such an effort, a good benchmarking tool is needed. The existing cache_bench was heavily skewed toward insertion and lookup misses, and did not saturate memory with other work. This change should improve those things to better resemble a real workload. (All below using clang compiler, for some consistency, but not necessarily same version and settings.) The real workload is from production MySQL on RocksDB, filtering stacks containing "LRU", "ShardedCache" or "CacheShard." Lookup inclusive: 66% Insert inclusive: 17% Release inclusive: 15% An alternate simulated workload is MySQL running a LinkBench read test: Lookup inclusive: 54% Insert inclusive: 24% Release inclusive: 21% cache_bench default settings, prior to this change: Lookup inclusive: 35.8% Insert inclusive: 63.6% Release inclusive: 0% cache_bench after this change (intended as somewhat "tighter" workload than average production, more like LinkBench): Lookup inclusive: 52% Insert inclusive: 20% Release inclusive: 26% And top exclusive stacks (portion of stack samples as filtered above): Production MySQL: LRUHandleTable::FindPointer: 25.3% rocksdb::operator==: 15.1% <-- Slice == LRUCacheShard::LRU_Remove: 13.8% ShardedCache::Lookup: 8.9% __pthread_mutex_lock: 7.1% LRUCacheShard::LRU_Insert: 6.3% MurmurHash64A: 4.8% <-- Since upgraded to XXH3p ... Old cache_bench: LRUHandleTable::FindPointer: 23.6% __pthread_mutex_lock: 15.0% __pthread_mutex_unlock_usercnt: 11.7% __lll_lock_wait: 8.6% __lll_unlock_wake: 6.8% LRUCacheShard::LRU_Insert: 6.0% ShardedCache::Lookup: 4.4% LRUCacheShard::LRU_Remove: 2.8% ... rocksdb::operator==: 0.2% <-- Slice == ... New cache_bench: LRUHandleTable::FindPointer: 22.8% __pthread_mutex_unlock_usercnt: 14.3% rocksdb::operator==: 10.5% <-- Slice == LRUCacheShard::LRU_Insert: 9.0% __pthread_mutex_lock: 5.9% LRUCacheShard::LRU_Remove: 5.0% ... ShardedCache::Lookup: 2.9% ... So there's a bit more lock contention in the benchmark than in production, but otherwise looks similar enough to me. At least it's a big improvement over the existing code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6629 Test Plan: No production code changes, ran cache_bench with ASAN Reviewed By: ltamasi Differential Revision: D20824318 Pulled By: pdillinger fbshipit-source-id: 6f8dc5891ead0f87edbed3a615ecd5289d9abe12
2020-04-03 19:24:09 +02:00
bool AllDone() const { return num_done_ >= FLAGS_threads; }
void SetStart() {
start_ = true;
}
bool Started() const {
return start_;
}
private:
port::Mutex mu_;
port::CondVar cv_;
uint64_t num_initialized_;
bool start_;
uint64_t num_done_;
CacheBench* cache_bench_;
};
// Per-thread state for concurrent executions of the same benchmark.
struct ThreadState {
uint32_t tid;
Revamp cache_bench to resemble a real workload (#6629) Summary: I suspect LRUCache could use some optimization, and to support such an effort, a good benchmarking tool is needed. The existing cache_bench was heavily skewed toward insertion and lookup misses, and did not saturate memory with other work. This change should improve those things to better resemble a real workload. (All below using clang compiler, for some consistency, but not necessarily same version and settings.) The real workload is from production MySQL on RocksDB, filtering stacks containing "LRU", "ShardedCache" or "CacheShard." Lookup inclusive: 66% Insert inclusive: 17% Release inclusive: 15% An alternate simulated workload is MySQL running a LinkBench read test: Lookup inclusive: 54% Insert inclusive: 24% Release inclusive: 21% cache_bench default settings, prior to this change: Lookup inclusive: 35.8% Insert inclusive: 63.6% Release inclusive: 0% cache_bench after this change (intended as somewhat "tighter" workload than average production, more like LinkBench): Lookup inclusive: 52% Insert inclusive: 20% Release inclusive: 26% And top exclusive stacks (portion of stack samples as filtered above): Production MySQL: LRUHandleTable::FindPointer: 25.3% rocksdb::operator==: 15.1% <-- Slice == LRUCacheShard::LRU_Remove: 13.8% ShardedCache::Lookup: 8.9% __pthread_mutex_lock: 7.1% LRUCacheShard::LRU_Insert: 6.3% MurmurHash64A: 4.8% <-- Since upgraded to XXH3p ... Old cache_bench: LRUHandleTable::FindPointer: 23.6% __pthread_mutex_lock: 15.0% __pthread_mutex_unlock_usercnt: 11.7% __lll_lock_wait: 8.6% __lll_unlock_wake: 6.8% LRUCacheShard::LRU_Insert: 6.0% ShardedCache::Lookup: 4.4% LRUCacheShard::LRU_Remove: 2.8% ... rocksdb::operator==: 0.2% <-- Slice == ... New cache_bench: LRUHandleTable::FindPointer: 22.8% __pthread_mutex_unlock_usercnt: 14.3% rocksdb::operator==: 10.5% <-- Slice == LRUCacheShard::LRU_Insert: 9.0% __pthread_mutex_lock: 5.9% LRUCacheShard::LRU_Remove: 5.0% ... ShardedCache::Lookup: 2.9% ... So there's a bit more lock contention in the benchmark than in production, but otherwise looks similar enough to me. At least it's a big improvement over the existing code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6629 Test Plan: No production code changes, ran cache_bench with ASAN Reviewed By: ltamasi Differential Revision: D20824318 Pulled By: pdillinger fbshipit-source-id: 6f8dc5891ead0f87edbed3a615ecd5289d9abe12
2020-04-03 19:24:09 +02:00
Random64 rnd;
SharedState* shared;
ThreadState(uint32_t index, SharedState* _shared)
: tid(index), rnd(1000 + index), shared(_shared) {}
};
Revamp cache_bench to resemble a real workload (#6629) Summary: I suspect LRUCache could use some optimization, and to support such an effort, a good benchmarking tool is needed. The existing cache_bench was heavily skewed toward insertion and lookup misses, and did not saturate memory with other work. This change should improve those things to better resemble a real workload. (All below using clang compiler, for some consistency, but not necessarily same version and settings.) The real workload is from production MySQL on RocksDB, filtering stacks containing "LRU", "ShardedCache" or "CacheShard." Lookup inclusive: 66% Insert inclusive: 17% Release inclusive: 15% An alternate simulated workload is MySQL running a LinkBench read test: Lookup inclusive: 54% Insert inclusive: 24% Release inclusive: 21% cache_bench default settings, prior to this change: Lookup inclusive: 35.8% Insert inclusive: 63.6% Release inclusive: 0% cache_bench after this change (intended as somewhat "tighter" workload than average production, more like LinkBench): Lookup inclusive: 52% Insert inclusive: 20% Release inclusive: 26% And top exclusive stacks (portion of stack samples as filtered above): Production MySQL: LRUHandleTable::FindPointer: 25.3% rocksdb::operator==: 15.1% <-- Slice == LRUCacheShard::LRU_Remove: 13.8% ShardedCache::Lookup: 8.9% __pthread_mutex_lock: 7.1% LRUCacheShard::LRU_Insert: 6.3% MurmurHash64A: 4.8% <-- Since upgraded to XXH3p ... Old cache_bench: LRUHandleTable::FindPointer: 23.6% __pthread_mutex_lock: 15.0% __pthread_mutex_unlock_usercnt: 11.7% __lll_lock_wait: 8.6% __lll_unlock_wake: 6.8% LRUCacheShard::LRU_Insert: 6.0% ShardedCache::Lookup: 4.4% LRUCacheShard::LRU_Remove: 2.8% ... rocksdb::operator==: 0.2% <-- Slice == ... New cache_bench: LRUHandleTable::FindPointer: 22.8% __pthread_mutex_unlock_usercnt: 14.3% rocksdb::operator==: 10.5% <-- Slice == LRUCacheShard::LRU_Insert: 9.0% __pthread_mutex_lock: 5.9% LRUCacheShard::LRU_Remove: 5.0% ... ShardedCache::Lookup: 2.9% ... So there's a bit more lock contention in the benchmark than in production, but otherwise looks similar enough to me. At least it's a big improvement over the existing code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6629 Test Plan: No production code changes, ran cache_bench with ASAN Reviewed By: ltamasi Differential Revision: D20824318 Pulled By: pdillinger fbshipit-source-id: 6f8dc5891ead0f87edbed3a615ecd5289d9abe12
2020-04-03 19:24:09 +02:00
struct KeyGen {
char key_data[27];
Slice GetRand(Random64& rnd, uint64_t max_key) {
uint64_t raw = rnd.Next();
// Skew according to setting
for (uint32_t i = 0; i < FLAGS_skew; ++i) {
raw = std::min(raw, rnd.Next());
}
uint64_t key = fastrange64(raw, max_key);
// Variable size and alignment
size_t off = key % 8;
key_data[0] = char{42};
EncodeFixed64(key_data + 1, key);
key_data[9] = char{11};
EncodeFixed64(key_data + 10, key);
key_data[18] = char{4};
EncodeFixed64(key_data + 19, key);
return Slice(&key_data[off], sizeof(key_data) - off);
}
};
char* createValue(Random64& rnd) {
char* rv = new char[FLAGS_value_bytes];
// Fill with some filler data, and take some CPU time
for (uint32_t i = 0; i < FLAGS_value_bytes; i += 8) {
EncodeFixed64(rv + i, rnd.Next());
}
return rv;
}
void deleter(const Slice& /*key*/, void* value) {
delete[] static_cast<char*>(value);
}
} // namespace
class CacheBench {
Revamp cache_bench to resemble a real workload (#6629) Summary: I suspect LRUCache could use some optimization, and to support such an effort, a good benchmarking tool is needed. The existing cache_bench was heavily skewed toward insertion and lookup misses, and did not saturate memory with other work. This change should improve those things to better resemble a real workload. (All below using clang compiler, for some consistency, but not necessarily same version and settings.) The real workload is from production MySQL on RocksDB, filtering stacks containing "LRU", "ShardedCache" or "CacheShard." Lookup inclusive: 66% Insert inclusive: 17% Release inclusive: 15% An alternate simulated workload is MySQL running a LinkBench read test: Lookup inclusive: 54% Insert inclusive: 24% Release inclusive: 21% cache_bench default settings, prior to this change: Lookup inclusive: 35.8% Insert inclusive: 63.6% Release inclusive: 0% cache_bench after this change (intended as somewhat "tighter" workload than average production, more like LinkBench): Lookup inclusive: 52% Insert inclusive: 20% Release inclusive: 26% And top exclusive stacks (portion of stack samples as filtered above): Production MySQL: LRUHandleTable::FindPointer: 25.3% rocksdb::operator==: 15.1% <-- Slice == LRUCacheShard::LRU_Remove: 13.8% ShardedCache::Lookup: 8.9% __pthread_mutex_lock: 7.1% LRUCacheShard::LRU_Insert: 6.3% MurmurHash64A: 4.8% <-- Since upgraded to XXH3p ... Old cache_bench: LRUHandleTable::FindPointer: 23.6% __pthread_mutex_lock: 15.0% __pthread_mutex_unlock_usercnt: 11.7% __lll_lock_wait: 8.6% __lll_unlock_wake: 6.8% LRUCacheShard::LRU_Insert: 6.0% ShardedCache::Lookup: 4.4% LRUCacheShard::LRU_Remove: 2.8% ... rocksdb::operator==: 0.2% <-- Slice == ... New cache_bench: LRUHandleTable::FindPointer: 22.8% __pthread_mutex_unlock_usercnt: 14.3% rocksdb::operator==: 10.5% <-- Slice == LRUCacheShard::LRU_Insert: 9.0% __pthread_mutex_lock: 5.9% LRUCacheShard::LRU_Remove: 5.0% ... ShardedCache::Lookup: 2.9% ... So there's a bit more lock contention in the benchmark than in production, but otherwise looks similar enough to me. At least it's a big improvement over the existing code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6629 Test Plan: No production code changes, ran cache_bench with ASAN Reviewed By: ltamasi Differential Revision: D20824318 Pulled By: pdillinger fbshipit-source-id: 6f8dc5891ead0f87edbed3a615ecd5289d9abe12
2020-04-03 19:24:09 +02:00
static constexpr uint64_t kHundredthUint64 =
std::numeric_limits<uint64_t>::max() / 100U;
public:
Revamp cache_bench to resemble a real workload (#6629) Summary: I suspect LRUCache could use some optimization, and to support such an effort, a good benchmarking tool is needed. The existing cache_bench was heavily skewed toward insertion and lookup misses, and did not saturate memory with other work. This change should improve those things to better resemble a real workload. (All below using clang compiler, for some consistency, but not necessarily same version and settings.) The real workload is from production MySQL on RocksDB, filtering stacks containing "LRU", "ShardedCache" or "CacheShard." Lookup inclusive: 66% Insert inclusive: 17% Release inclusive: 15% An alternate simulated workload is MySQL running a LinkBench read test: Lookup inclusive: 54% Insert inclusive: 24% Release inclusive: 21% cache_bench default settings, prior to this change: Lookup inclusive: 35.8% Insert inclusive: 63.6% Release inclusive: 0% cache_bench after this change (intended as somewhat "tighter" workload than average production, more like LinkBench): Lookup inclusive: 52% Insert inclusive: 20% Release inclusive: 26% And top exclusive stacks (portion of stack samples as filtered above): Production MySQL: LRUHandleTable::FindPointer: 25.3% rocksdb::operator==: 15.1% <-- Slice == LRUCacheShard::LRU_Remove: 13.8% ShardedCache::Lookup: 8.9% __pthread_mutex_lock: 7.1% LRUCacheShard::LRU_Insert: 6.3% MurmurHash64A: 4.8% <-- Since upgraded to XXH3p ... Old cache_bench: LRUHandleTable::FindPointer: 23.6% __pthread_mutex_lock: 15.0% __pthread_mutex_unlock_usercnt: 11.7% __lll_lock_wait: 8.6% __lll_unlock_wake: 6.8% LRUCacheShard::LRU_Insert: 6.0% ShardedCache::Lookup: 4.4% LRUCacheShard::LRU_Remove: 2.8% ... rocksdb::operator==: 0.2% <-- Slice == ... New cache_bench: LRUHandleTable::FindPointer: 22.8% __pthread_mutex_unlock_usercnt: 14.3% rocksdb::operator==: 10.5% <-- Slice == LRUCacheShard::LRU_Insert: 9.0% __pthread_mutex_lock: 5.9% LRUCacheShard::LRU_Remove: 5.0% ... ShardedCache::Lookup: 2.9% ... So there's a bit more lock contention in the benchmark than in production, but otherwise looks similar enough to me. At least it's a big improvement over the existing code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6629 Test Plan: No production code changes, ran cache_bench with ASAN Reviewed By: ltamasi Differential Revision: D20824318 Pulled By: pdillinger fbshipit-source-id: 6f8dc5891ead0f87edbed3a615ecd5289d9abe12
2020-04-03 19:24:09 +02:00
CacheBench()
: max_key_(static_cast<uint64_t>(FLAGS_cache_size / FLAGS_resident_ratio /
FLAGS_value_bytes)),
lookup_insert_threshold_(kHundredthUint64 *
FLAGS_lookup_insert_percent),
insert_threshold_(lookup_insert_threshold_ +
kHundredthUint64 * FLAGS_insert_percent),
lookup_threshold_(insert_threshold_ +
kHundredthUint64 * FLAGS_lookup_percent),
erase_threshold_(lookup_threshold_ +
kHundredthUint64 * FLAGS_erase_percent) {
if (erase_threshold_ != 100U * kHundredthUint64) {
fprintf(stderr, "Percentages must add to 100.\n");
exit(1);
}
if (FLAGS_use_clock_cache) {
cache_ = NewClockCache(FLAGS_cache_size, FLAGS_num_shard_bits);
if (!cache_) {
fprintf(stderr, "Clock cache not supported.\n");
exit(1);
}
} else {
cache_ = NewLRUCache(FLAGS_cache_size, FLAGS_num_shard_bits);
}
Revamp cache_bench to resemble a real workload (#6629) Summary: I suspect LRUCache could use some optimization, and to support such an effort, a good benchmarking tool is needed. The existing cache_bench was heavily skewed toward insertion and lookup misses, and did not saturate memory with other work. This change should improve those things to better resemble a real workload. (All below using clang compiler, for some consistency, but not necessarily same version and settings.) The real workload is from production MySQL on RocksDB, filtering stacks containing "LRU", "ShardedCache" or "CacheShard." Lookup inclusive: 66% Insert inclusive: 17% Release inclusive: 15% An alternate simulated workload is MySQL running a LinkBench read test: Lookup inclusive: 54% Insert inclusive: 24% Release inclusive: 21% cache_bench default settings, prior to this change: Lookup inclusive: 35.8% Insert inclusive: 63.6% Release inclusive: 0% cache_bench after this change (intended as somewhat "tighter" workload than average production, more like LinkBench): Lookup inclusive: 52% Insert inclusive: 20% Release inclusive: 26% And top exclusive stacks (portion of stack samples as filtered above): Production MySQL: LRUHandleTable::FindPointer: 25.3% rocksdb::operator==: 15.1% <-- Slice == LRUCacheShard::LRU_Remove: 13.8% ShardedCache::Lookup: 8.9% __pthread_mutex_lock: 7.1% LRUCacheShard::LRU_Insert: 6.3% MurmurHash64A: 4.8% <-- Since upgraded to XXH3p ... Old cache_bench: LRUHandleTable::FindPointer: 23.6% __pthread_mutex_lock: 15.0% __pthread_mutex_unlock_usercnt: 11.7% __lll_lock_wait: 8.6% __lll_unlock_wake: 6.8% LRUCacheShard::LRU_Insert: 6.0% ShardedCache::Lookup: 4.4% LRUCacheShard::LRU_Remove: 2.8% ... rocksdb::operator==: 0.2% <-- Slice == ... New cache_bench: LRUHandleTable::FindPointer: 22.8% __pthread_mutex_unlock_usercnt: 14.3% rocksdb::operator==: 10.5% <-- Slice == LRUCacheShard::LRU_Insert: 9.0% __pthread_mutex_lock: 5.9% LRUCacheShard::LRU_Remove: 5.0% ... ShardedCache::Lookup: 2.9% ... So there's a bit more lock contention in the benchmark than in production, but otherwise looks similar enough to me. At least it's a big improvement over the existing code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6629 Test Plan: No production code changes, ran cache_bench with ASAN Reviewed By: ltamasi Differential Revision: D20824318 Pulled By: pdillinger fbshipit-source-id: 6f8dc5891ead0f87edbed3a615ecd5289d9abe12
2020-04-03 19:24:09 +02:00
if (FLAGS_ops_per_thread == 0) {
FLAGS_ops_per_thread = 5 * max_key_;
}
}
~CacheBench() {}
void PopulateCache() {
Revamp cache_bench to resemble a real workload (#6629) Summary: I suspect LRUCache could use some optimization, and to support such an effort, a good benchmarking tool is needed. The existing cache_bench was heavily skewed toward insertion and lookup misses, and did not saturate memory with other work. This change should improve those things to better resemble a real workload. (All below using clang compiler, for some consistency, but not necessarily same version and settings.) The real workload is from production MySQL on RocksDB, filtering stacks containing "LRU", "ShardedCache" or "CacheShard." Lookup inclusive: 66% Insert inclusive: 17% Release inclusive: 15% An alternate simulated workload is MySQL running a LinkBench read test: Lookup inclusive: 54% Insert inclusive: 24% Release inclusive: 21% cache_bench default settings, prior to this change: Lookup inclusive: 35.8% Insert inclusive: 63.6% Release inclusive: 0% cache_bench after this change (intended as somewhat "tighter" workload than average production, more like LinkBench): Lookup inclusive: 52% Insert inclusive: 20% Release inclusive: 26% And top exclusive stacks (portion of stack samples as filtered above): Production MySQL: LRUHandleTable::FindPointer: 25.3% rocksdb::operator==: 15.1% <-- Slice == LRUCacheShard::LRU_Remove: 13.8% ShardedCache::Lookup: 8.9% __pthread_mutex_lock: 7.1% LRUCacheShard::LRU_Insert: 6.3% MurmurHash64A: 4.8% <-- Since upgraded to XXH3p ... Old cache_bench: LRUHandleTable::FindPointer: 23.6% __pthread_mutex_lock: 15.0% __pthread_mutex_unlock_usercnt: 11.7% __lll_lock_wait: 8.6% __lll_unlock_wake: 6.8% LRUCacheShard::LRU_Insert: 6.0% ShardedCache::Lookup: 4.4% LRUCacheShard::LRU_Remove: 2.8% ... rocksdb::operator==: 0.2% <-- Slice == ... New cache_bench: LRUHandleTable::FindPointer: 22.8% __pthread_mutex_unlock_usercnt: 14.3% rocksdb::operator==: 10.5% <-- Slice == LRUCacheShard::LRU_Insert: 9.0% __pthread_mutex_lock: 5.9% LRUCacheShard::LRU_Remove: 5.0% ... ShardedCache::Lookup: 2.9% ... So there's a bit more lock contention in the benchmark than in production, but otherwise looks similar enough to me. At least it's a big improvement over the existing code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6629 Test Plan: No production code changes, ran cache_bench with ASAN Reviewed By: ltamasi Differential Revision: D20824318 Pulled By: pdillinger fbshipit-source-id: 6f8dc5891ead0f87edbed3a615ecd5289d9abe12
2020-04-03 19:24:09 +02:00
Random64 rnd(1);
KeyGen keygen;
for (uint64_t i = 0; i < 2 * FLAGS_cache_size; i += FLAGS_value_bytes) {
cache_->Insert(keygen.GetRand(rnd, max_key_), createValue(rnd),
FLAGS_value_bytes, &deleter);
}
}
bool Run() {
ROCKSDB_NAMESPACE::Env* env = ROCKSDB_NAMESPACE::Env::Default();
PrintEnv();
SharedState shared(this);
Revamp cache_bench to resemble a real workload (#6629) Summary: I suspect LRUCache could use some optimization, and to support such an effort, a good benchmarking tool is needed. The existing cache_bench was heavily skewed toward insertion and lookup misses, and did not saturate memory with other work. This change should improve those things to better resemble a real workload. (All below using clang compiler, for some consistency, but not necessarily same version and settings.) The real workload is from production MySQL on RocksDB, filtering stacks containing "LRU", "ShardedCache" or "CacheShard." Lookup inclusive: 66% Insert inclusive: 17% Release inclusive: 15% An alternate simulated workload is MySQL running a LinkBench read test: Lookup inclusive: 54% Insert inclusive: 24% Release inclusive: 21% cache_bench default settings, prior to this change: Lookup inclusive: 35.8% Insert inclusive: 63.6% Release inclusive: 0% cache_bench after this change (intended as somewhat "tighter" workload than average production, more like LinkBench): Lookup inclusive: 52% Insert inclusive: 20% Release inclusive: 26% And top exclusive stacks (portion of stack samples as filtered above): Production MySQL: LRUHandleTable::FindPointer: 25.3% rocksdb::operator==: 15.1% <-- Slice == LRUCacheShard::LRU_Remove: 13.8% ShardedCache::Lookup: 8.9% __pthread_mutex_lock: 7.1% LRUCacheShard::LRU_Insert: 6.3% MurmurHash64A: 4.8% <-- Since upgraded to XXH3p ... Old cache_bench: LRUHandleTable::FindPointer: 23.6% __pthread_mutex_lock: 15.0% __pthread_mutex_unlock_usercnt: 11.7% __lll_lock_wait: 8.6% __lll_unlock_wake: 6.8% LRUCacheShard::LRU_Insert: 6.0% ShardedCache::Lookup: 4.4% LRUCacheShard::LRU_Remove: 2.8% ... rocksdb::operator==: 0.2% <-- Slice == ... New cache_bench: LRUHandleTable::FindPointer: 22.8% __pthread_mutex_unlock_usercnt: 14.3% rocksdb::operator==: 10.5% <-- Slice == LRUCacheShard::LRU_Insert: 9.0% __pthread_mutex_lock: 5.9% LRUCacheShard::LRU_Remove: 5.0% ... ShardedCache::Lookup: 2.9% ... So there's a bit more lock contention in the benchmark than in production, but otherwise looks similar enough to me. At least it's a big improvement over the existing code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6629 Test Plan: No production code changes, ran cache_bench with ASAN Reviewed By: ltamasi Differential Revision: D20824318 Pulled By: pdillinger fbshipit-source-id: 6f8dc5891ead0f87edbed3a615ecd5289d9abe12
2020-04-03 19:24:09 +02:00
std::vector<std::unique_ptr<ThreadState> > threads(FLAGS_threads);
for (uint32_t i = 0; i < FLAGS_threads; i++) {
threads[i].reset(new ThreadState(i, &shared));
env->StartThread(ThreadBody, threads[i].get());
}
{
MutexLock l(shared.GetMutex());
while (!shared.AllInitialized()) {
shared.GetCondVar()->Wait();
}
// Record start time
uint64_t start_time = env->NowMicros();
// Start all threads
shared.SetStart();
shared.GetCondVar()->SignalAll();
// Wait threads to complete
while (!shared.AllDone()) {
shared.GetCondVar()->Wait();
}
// Record end time
uint64_t end_time = env->NowMicros();
double elapsed = static_cast<double>(end_time - start_time) * 1e-6;
uint32_t qps = static_cast<uint32_t>(
static_cast<double>(FLAGS_threads * FLAGS_ops_per_thread) / elapsed);
fprintf(stdout, "Complete in %.3f s; QPS = %u\n", elapsed, qps);
}
return true;
}
private:
std::shared_ptr<Cache> cache_;
Revamp cache_bench to resemble a real workload (#6629) Summary: I suspect LRUCache could use some optimization, and to support such an effort, a good benchmarking tool is needed. The existing cache_bench was heavily skewed toward insertion and lookup misses, and did not saturate memory with other work. This change should improve those things to better resemble a real workload. (All below using clang compiler, for some consistency, but not necessarily same version and settings.) The real workload is from production MySQL on RocksDB, filtering stacks containing "LRU", "ShardedCache" or "CacheShard." Lookup inclusive: 66% Insert inclusive: 17% Release inclusive: 15% An alternate simulated workload is MySQL running a LinkBench read test: Lookup inclusive: 54% Insert inclusive: 24% Release inclusive: 21% cache_bench default settings, prior to this change: Lookup inclusive: 35.8% Insert inclusive: 63.6% Release inclusive: 0% cache_bench after this change (intended as somewhat "tighter" workload than average production, more like LinkBench): Lookup inclusive: 52% Insert inclusive: 20% Release inclusive: 26% And top exclusive stacks (portion of stack samples as filtered above): Production MySQL: LRUHandleTable::FindPointer: 25.3% rocksdb::operator==: 15.1% <-- Slice == LRUCacheShard::LRU_Remove: 13.8% ShardedCache::Lookup: 8.9% __pthread_mutex_lock: 7.1% LRUCacheShard::LRU_Insert: 6.3% MurmurHash64A: 4.8% <-- Since upgraded to XXH3p ... Old cache_bench: LRUHandleTable::FindPointer: 23.6% __pthread_mutex_lock: 15.0% __pthread_mutex_unlock_usercnt: 11.7% __lll_lock_wait: 8.6% __lll_unlock_wake: 6.8% LRUCacheShard::LRU_Insert: 6.0% ShardedCache::Lookup: 4.4% LRUCacheShard::LRU_Remove: 2.8% ... rocksdb::operator==: 0.2% <-- Slice == ... New cache_bench: LRUHandleTable::FindPointer: 22.8% __pthread_mutex_unlock_usercnt: 14.3% rocksdb::operator==: 10.5% <-- Slice == LRUCacheShard::LRU_Insert: 9.0% __pthread_mutex_lock: 5.9% LRUCacheShard::LRU_Remove: 5.0% ... ShardedCache::Lookup: 2.9% ... So there's a bit more lock contention in the benchmark than in production, but otherwise looks similar enough to me. At least it's a big improvement over the existing code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6629 Test Plan: No production code changes, ran cache_bench with ASAN Reviewed By: ltamasi Differential Revision: D20824318 Pulled By: pdillinger fbshipit-source-id: 6f8dc5891ead0f87edbed3a615ecd5289d9abe12
2020-04-03 19:24:09 +02:00
const uint64_t max_key_;
// Cumulative thresholds in the space of a random uint64_t
const uint64_t lookup_insert_threshold_;
const uint64_t insert_threshold_;
const uint64_t lookup_threshold_;
const uint64_t erase_threshold_;
static void ThreadBody(void* v) {
Revamp cache_bench to resemble a real workload (#6629) Summary: I suspect LRUCache could use some optimization, and to support such an effort, a good benchmarking tool is needed. The existing cache_bench was heavily skewed toward insertion and lookup misses, and did not saturate memory with other work. This change should improve those things to better resemble a real workload. (All below using clang compiler, for some consistency, but not necessarily same version and settings.) The real workload is from production MySQL on RocksDB, filtering stacks containing "LRU", "ShardedCache" or "CacheShard." Lookup inclusive: 66% Insert inclusive: 17% Release inclusive: 15% An alternate simulated workload is MySQL running a LinkBench read test: Lookup inclusive: 54% Insert inclusive: 24% Release inclusive: 21% cache_bench default settings, prior to this change: Lookup inclusive: 35.8% Insert inclusive: 63.6% Release inclusive: 0% cache_bench after this change (intended as somewhat "tighter" workload than average production, more like LinkBench): Lookup inclusive: 52% Insert inclusive: 20% Release inclusive: 26% And top exclusive stacks (portion of stack samples as filtered above): Production MySQL: LRUHandleTable::FindPointer: 25.3% rocksdb::operator==: 15.1% <-- Slice == LRUCacheShard::LRU_Remove: 13.8% ShardedCache::Lookup: 8.9% __pthread_mutex_lock: 7.1% LRUCacheShard::LRU_Insert: 6.3% MurmurHash64A: 4.8% <-- Since upgraded to XXH3p ... Old cache_bench: LRUHandleTable::FindPointer: 23.6% __pthread_mutex_lock: 15.0% __pthread_mutex_unlock_usercnt: 11.7% __lll_lock_wait: 8.6% __lll_unlock_wake: 6.8% LRUCacheShard::LRU_Insert: 6.0% ShardedCache::Lookup: 4.4% LRUCacheShard::LRU_Remove: 2.8% ... rocksdb::operator==: 0.2% <-- Slice == ... New cache_bench: LRUHandleTable::FindPointer: 22.8% __pthread_mutex_unlock_usercnt: 14.3% rocksdb::operator==: 10.5% <-- Slice == LRUCacheShard::LRU_Insert: 9.0% __pthread_mutex_lock: 5.9% LRUCacheShard::LRU_Remove: 5.0% ... ShardedCache::Lookup: 2.9% ... So there's a bit more lock contention in the benchmark than in production, but otherwise looks similar enough to me. At least it's a big improvement over the existing code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6629 Test Plan: No production code changes, ran cache_bench with ASAN Reviewed By: ltamasi Differential Revision: D20824318 Pulled By: pdillinger fbshipit-source-id: 6f8dc5891ead0f87edbed3a615ecd5289d9abe12
2020-04-03 19:24:09 +02:00
ThreadState* thread = static_cast<ThreadState*>(v);
SharedState* shared = thread->shared;
{
MutexLock l(shared->GetMutex());
shared->IncInitialized();
if (shared->AllInitialized()) {
shared->GetCondVar()->SignalAll();
}
while (!shared->Started()) {
shared->GetCondVar()->Wait();
}
}
thread->shared->GetCacheBench()->OperateCache(thread);
{
MutexLock l(shared->GetMutex());
shared->IncDone();
if (shared->AllDone()) {
shared->GetCondVar()->SignalAll();
}
}
}
void OperateCache(ThreadState* thread) {
Revamp cache_bench to resemble a real workload (#6629) Summary: I suspect LRUCache could use some optimization, and to support such an effort, a good benchmarking tool is needed. The existing cache_bench was heavily skewed toward insertion and lookup misses, and did not saturate memory with other work. This change should improve those things to better resemble a real workload. (All below using clang compiler, for some consistency, but not necessarily same version and settings.) The real workload is from production MySQL on RocksDB, filtering stacks containing "LRU", "ShardedCache" or "CacheShard." Lookup inclusive: 66% Insert inclusive: 17% Release inclusive: 15% An alternate simulated workload is MySQL running a LinkBench read test: Lookup inclusive: 54% Insert inclusive: 24% Release inclusive: 21% cache_bench default settings, prior to this change: Lookup inclusive: 35.8% Insert inclusive: 63.6% Release inclusive: 0% cache_bench after this change (intended as somewhat "tighter" workload than average production, more like LinkBench): Lookup inclusive: 52% Insert inclusive: 20% Release inclusive: 26% And top exclusive stacks (portion of stack samples as filtered above): Production MySQL: LRUHandleTable::FindPointer: 25.3% rocksdb::operator==: 15.1% <-- Slice == LRUCacheShard::LRU_Remove: 13.8% ShardedCache::Lookup: 8.9% __pthread_mutex_lock: 7.1% LRUCacheShard::LRU_Insert: 6.3% MurmurHash64A: 4.8% <-- Since upgraded to XXH3p ... Old cache_bench: LRUHandleTable::FindPointer: 23.6% __pthread_mutex_lock: 15.0% __pthread_mutex_unlock_usercnt: 11.7% __lll_lock_wait: 8.6% __lll_unlock_wake: 6.8% LRUCacheShard::LRU_Insert: 6.0% ShardedCache::Lookup: 4.4% LRUCacheShard::LRU_Remove: 2.8% ... rocksdb::operator==: 0.2% <-- Slice == ... New cache_bench: LRUHandleTable::FindPointer: 22.8% __pthread_mutex_unlock_usercnt: 14.3% rocksdb::operator==: 10.5% <-- Slice == LRUCacheShard::LRU_Insert: 9.0% __pthread_mutex_lock: 5.9% LRUCacheShard::LRU_Remove: 5.0% ... ShardedCache::Lookup: 2.9% ... So there's a bit more lock contention in the benchmark than in production, but otherwise looks similar enough to me. At least it's a big improvement over the existing code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6629 Test Plan: No production code changes, ran cache_bench with ASAN Reviewed By: ltamasi Differential Revision: D20824318 Pulled By: pdillinger fbshipit-source-id: 6f8dc5891ead0f87edbed3a615ecd5289d9abe12
2020-04-03 19:24:09 +02:00
// To use looked-up values
uint64_t result = 0;
// To hold handles for a non-trivial amount of time
Cache::Handle* handle = nullptr;
KeyGen gen;
for (uint64_t i = 0; i < FLAGS_ops_per_thread; i++) {
Revamp cache_bench to resemble a real workload (#6629) Summary: I suspect LRUCache could use some optimization, and to support such an effort, a good benchmarking tool is needed. The existing cache_bench was heavily skewed toward insertion and lookup misses, and did not saturate memory with other work. This change should improve those things to better resemble a real workload. (All below using clang compiler, for some consistency, but not necessarily same version and settings.) The real workload is from production MySQL on RocksDB, filtering stacks containing "LRU", "ShardedCache" or "CacheShard." Lookup inclusive: 66% Insert inclusive: 17% Release inclusive: 15% An alternate simulated workload is MySQL running a LinkBench read test: Lookup inclusive: 54% Insert inclusive: 24% Release inclusive: 21% cache_bench default settings, prior to this change: Lookup inclusive: 35.8% Insert inclusive: 63.6% Release inclusive: 0% cache_bench after this change (intended as somewhat "tighter" workload than average production, more like LinkBench): Lookup inclusive: 52% Insert inclusive: 20% Release inclusive: 26% And top exclusive stacks (portion of stack samples as filtered above): Production MySQL: LRUHandleTable::FindPointer: 25.3% rocksdb::operator==: 15.1% <-- Slice == LRUCacheShard::LRU_Remove: 13.8% ShardedCache::Lookup: 8.9% __pthread_mutex_lock: 7.1% LRUCacheShard::LRU_Insert: 6.3% MurmurHash64A: 4.8% <-- Since upgraded to XXH3p ... Old cache_bench: LRUHandleTable::FindPointer: 23.6% __pthread_mutex_lock: 15.0% __pthread_mutex_unlock_usercnt: 11.7% __lll_lock_wait: 8.6% __lll_unlock_wake: 6.8% LRUCacheShard::LRU_Insert: 6.0% ShardedCache::Lookup: 4.4% LRUCacheShard::LRU_Remove: 2.8% ... rocksdb::operator==: 0.2% <-- Slice == ... New cache_bench: LRUHandleTable::FindPointer: 22.8% __pthread_mutex_unlock_usercnt: 14.3% rocksdb::operator==: 10.5% <-- Slice == LRUCacheShard::LRU_Insert: 9.0% __pthread_mutex_lock: 5.9% LRUCacheShard::LRU_Remove: 5.0% ... ShardedCache::Lookup: 2.9% ... So there's a bit more lock contention in the benchmark than in production, but otherwise looks similar enough to me. At least it's a big improvement over the existing code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6629 Test Plan: No production code changes, ran cache_bench with ASAN Reviewed By: ltamasi Differential Revision: D20824318 Pulled By: pdillinger fbshipit-source-id: 6f8dc5891ead0f87edbed3a615ecd5289d9abe12
2020-04-03 19:24:09 +02:00
Slice key = gen.GetRand(thread->rnd, max_key_);
uint64_t random_op = thread->rnd.Next();
if (random_op < lookup_insert_threshold_) {
if (handle) {
cache_->Release(handle);
handle = nullptr;
}
// do lookup
Revamp cache_bench to resemble a real workload (#6629) Summary: I suspect LRUCache could use some optimization, and to support such an effort, a good benchmarking tool is needed. The existing cache_bench was heavily skewed toward insertion and lookup misses, and did not saturate memory with other work. This change should improve those things to better resemble a real workload. (All below using clang compiler, for some consistency, but not necessarily same version and settings.) The real workload is from production MySQL on RocksDB, filtering stacks containing "LRU", "ShardedCache" or "CacheShard." Lookup inclusive: 66% Insert inclusive: 17% Release inclusive: 15% An alternate simulated workload is MySQL running a LinkBench read test: Lookup inclusive: 54% Insert inclusive: 24% Release inclusive: 21% cache_bench default settings, prior to this change: Lookup inclusive: 35.8% Insert inclusive: 63.6% Release inclusive: 0% cache_bench after this change (intended as somewhat "tighter" workload than average production, more like LinkBench): Lookup inclusive: 52% Insert inclusive: 20% Release inclusive: 26% And top exclusive stacks (portion of stack samples as filtered above): Production MySQL: LRUHandleTable::FindPointer: 25.3% rocksdb::operator==: 15.1% <-- Slice == LRUCacheShard::LRU_Remove: 13.8% ShardedCache::Lookup: 8.9% __pthread_mutex_lock: 7.1% LRUCacheShard::LRU_Insert: 6.3% MurmurHash64A: 4.8% <-- Since upgraded to XXH3p ... Old cache_bench: LRUHandleTable::FindPointer: 23.6% __pthread_mutex_lock: 15.0% __pthread_mutex_unlock_usercnt: 11.7% __lll_lock_wait: 8.6% __lll_unlock_wake: 6.8% LRUCacheShard::LRU_Insert: 6.0% ShardedCache::Lookup: 4.4% LRUCacheShard::LRU_Remove: 2.8% ... rocksdb::operator==: 0.2% <-- Slice == ... New cache_bench: LRUHandleTable::FindPointer: 22.8% __pthread_mutex_unlock_usercnt: 14.3% rocksdb::operator==: 10.5% <-- Slice == LRUCacheShard::LRU_Insert: 9.0% __pthread_mutex_lock: 5.9% LRUCacheShard::LRU_Remove: 5.0% ... ShardedCache::Lookup: 2.9% ... So there's a bit more lock contention in the benchmark than in production, but otherwise looks similar enough to me. At least it's a big improvement over the existing code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6629 Test Plan: No production code changes, ran cache_bench with ASAN Reviewed By: ltamasi Differential Revision: D20824318 Pulled By: pdillinger fbshipit-source-id: 6f8dc5891ead0f87edbed3a615ecd5289d9abe12
2020-04-03 19:24:09 +02:00
handle = cache_->Lookup(key);
if (handle) {
// do something with the data
result += NPHash64(static_cast<char*>(cache_->Value(handle)),
FLAGS_value_bytes);
} else {
// do insert
cache_->Insert(key, createValue(thread->rnd), FLAGS_value_bytes,
&deleter, &handle);
}
} else if (random_op < insert_threshold_) {
if (handle) {
cache_->Release(handle);
Revamp cache_bench to resemble a real workload (#6629) Summary: I suspect LRUCache could use some optimization, and to support such an effort, a good benchmarking tool is needed. The existing cache_bench was heavily skewed toward insertion and lookup misses, and did not saturate memory with other work. This change should improve those things to better resemble a real workload. (All below using clang compiler, for some consistency, but not necessarily same version and settings.) The real workload is from production MySQL on RocksDB, filtering stacks containing "LRU", "ShardedCache" or "CacheShard." Lookup inclusive: 66% Insert inclusive: 17% Release inclusive: 15% An alternate simulated workload is MySQL running a LinkBench read test: Lookup inclusive: 54% Insert inclusive: 24% Release inclusive: 21% cache_bench default settings, prior to this change: Lookup inclusive: 35.8% Insert inclusive: 63.6% Release inclusive: 0% cache_bench after this change (intended as somewhat "tighter" workload than average production, more like LinkBench): Lookup inclusive: 52% Insert inclusive: 20% Release inclusive: 26% And top exclusive stacks (portion of stack samples as filtered above): Production MySQL: LRUHandleTable::FindPointer: 25.3% rocksdb::operator==: 15.1% <-- Slice == LRUCacheShard::LRU_Remove: 13.8% ShardedCache::Lookup: 8.9% __pthread_mutex_lock: 7.1% LRUCacheShard::LRU_Insert: 6.3% MurmurHash64A: 4.8% <-- Since upgraded to XXH3p ... Old cache_bench: LRUHandleTable::FindPointer: 23.6% __pthread_mutex_lock: 15.0% __pthread_mutex_unlock_usercnt: 11.7% __lll_lock_wait: 8.6% __lll_unlock_wake: 6.8% LRUCacheShard::LRU_Insert: 6.0% ShardedCache::Lookup: 4.4% LRUCacheShard::LRU_Remove: 2.8% ... rocksdb::operator==: 0.2% <-- Slice == ... New cache_bench: LRUHandleTable::FindPointer: 22.8% __pthread_mutex_unlock_usercnt: 14.3% rocksdb::operator==: 10.5% <-- Slice == LRUCacheShard::LRU_Insert: 9.0% __pthread_mutex_lock: 5.9% LRUCacheShard::LRU_Remove: 5.0% ... ShardedCache::Lookup: 2.9% ... So there's a bit more lock contention in the benchmark than in production, but otherwise looks similar enough to me. At least it's a big improvement over the existing code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6629 Test Plan: No production code changes, ran cache_bench with ASAN Reviewed By: ltamasi Differential Revision: D20824318 Pulled By: pdillinger fbshipit-source-id: 6f8dc5891ead0f87edbed3a615ecd5289d9abe12
2020-04-03 19:24:09 +02:00
handle = nullptr;
}
Revamp cache_bench to resemble a real workload (#6629) Summary: I suspect LRUCache could use some optimization, and to support such an effort, a good benchmarking tool is needed. The existing cache_bench was heavily skewed toward insertion and lookup misses, and did not saturate memory with other work. This change should improve those things to better resemble a real workload. (All below using clang compiler, for some consistency, but not necessarily same version and settings.) The real workload is from production MySQL on RocksDB, filtering stacks containing "LRU", "ShardedCache" or "CacheShard." Lookup inclusive: 66% Insert inclusive: 17% Release inclusive: 15% An alternate simulated workload is MySQL running a LinkBench read test: Lookup inclusive: 54% Insert inclusive: 24% Release inclusive: 21% cache_bench default settings, prior to this change: Lookup inclusive: 35.8% Insert inclusive: 63.6% Release inclusive: 0% cache_bench after this change (intended as somewhat "tighter" workload than average production, more like LinkBench): Lookup inclusive: 52% Insert inclusive: 20% Release inclusive: 26% And top exclusive stacks (portion of stack samples as filtered above): Production MySQL: LRUHandleTable::FindPointer: 25.3% rocksdb::operator==: 15.1% <-- Slice == LRUCacheShard::LRU_Remove: 13.8% ShardedCache::Lookup: 8.9% __pthread_mutex_lock: 7.1% LRUCacheShard::LRU_Insert: 6.3% MurmurHash64A: 4.8% <-- Since upgraded to XXH3p ... Old cache_bench: LRUHandleTable::FindPointer: 23.6% __pthread_mutex_lock: 15.0% __pthread_mutex_unlock_usercnt: 11.7% __lll_lock_wait: 8.6% __lll_unlock_wake: 6.8% LRUCacheShard::LRU_Insert: 6.0% ShardedCache::Lookup: 4.4% LRUCacheShard::LRU_Remove: 2.8% ... rocksdb::operator==: 0.2% <-- Slice == ... New cache_bench: LRUHandleTable::FindPointer: 22.8% __pthread_mutex_unlock_usercnt: 14.3% rocksdb::operator==: 10.5% <-- Slice == LRUCacheShard::LRU_Insert: 9.0% __pthread_mutex_lock: 5.9% LRUCacheShard::LRU_Remove: 5.0% ... ShardedCache::Lookup: 2.9% ... So there's a bit more lock contention in the benchmark than in production, but otherwise looks similar enough to me. At least it's a big improvement over the existing code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6629 Test Plan: No production code changes, ran cache_bench with ASAN Reviewed By: ltamasi Differential Revision: D20824318 Pulled By: pdillinger fbshipit-source-id: 6f8dc5891ead0f87edbed3a615ecd5289d9abe12
2020-04-03 19:24:09 +02:00
// do insert
cache_->Insert(key, createValue(thread->rnd), FLAGS_value_bytes,
&deleter, &handle);
} else if (random_op < lookup_threshold_) {
if (handle) {
cache_->Release(handle);
handle = nullptr;
}
// do lookup
handle = cache_->Lookup(key);
if (handle) {
// do something with the data
result += NPHash64(static_cast<char*>(cache_->Value(handle)),
FLAGS_value_bytes);
}
} else if (random_op < erase_threshold_) {
// do erase
cache_->Erase(key);
Revamp cache_bench to resemble a real workload (#6629) Summary: I suspect LRUCache could use some optimization, and to support such an effort, a good benchmarking tool is needed. The existing cache_bench was heavily skewed toward insertion and lookup misses, and did not saturate memory with other work. This change should improve those things to better resemble a real workload. (All below using clang compiler, for some consistency, but not necessarily same version and settings.) The real workload is from production MySQL on RocksDB, filtering stacks containing "LRU", "ShardedCache" or "CacheShard." Lookup inclusive: 66% Insert inclusive: 17% Release inclusive: 15% An alternate simulated workload is MySQL running a LinkBench read test: Lookup inclusive: 54% Insert inclusive: 24% Release inclusive: 21% cache_bench default settings, prior to this change: Lookup inclusive: 35.8% Insert inclusive: 63.6% Release inclusive: 0% cache_bench after this change (intended as somewhat "tighter" workload than average production, more like LinkBench): Lookup inclusive: 52% Insert inclusive: 20% Release inclusive: 26% And top exclusive stacks (portion of stack samples as filtered above): Production MySQL: LRUHandleTable::FindPointer: 25.3% rocksdb::operator==: 15.1% <-- Slice == LRUCacheShard::LRU_Remove: 13.8% ShardedCache::Lookup: 8.9% __pthread_mutex_lock: 7.1% LRUCacheShard::LRU_Insert: 6.3% MurmurHash64A: 4.8% <-- Since upgraded to XXH3p ... Old cache_bench: LRUHandleTable::FindPointer: 23.6% __pthread_mutex_lock: 15.0% __pthread_mutex_unlock_usercnt: 11.7% __lll_lock_wait: 8.6% __lll_unlock_wake: 6.8% LRUCacheShard::LRU_Insert: 6.0% ShardedCache::Lookup: 4.4% LRUCacheShard::LRU_Remove: 2.8% ... rocksdb::operator==: 0.2% <-- Slice == ... New cache_bench: LRUHandleTable::FindPointer: 22.8% __pthread_mutex_unlock_usercnt: 14.3% rocksdb::operator==: 10.5% <-- Slice == LRUCacheShard::LRU_Insert: 9.0% __pthread_mutex_lock: 5.9% LRUCacheShard::LRU_Remove: 5.0% ... ShardedCache::Lookup: 2.9% ... So there's a bit more lock contention in the benchmark than in production, but otherwise looks similar enough to me. At least it's a big improvement over the existing code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6629 Test Plan: No production code changes, ran cache_bench with ASAN Reviewed By: ltamasi Differential Revision: D20824318 Pulled By: pdillinger fbshipit-source-id: 6f8dc5891ead0f87edbed3a615ecd5289d9abe12
2020-04-03 19:24:09 +02:00
} else {
// Should be extremely unlikely (noop)
assert(random_op >= kHundredthUint64 * 100U);
}
}
Revamp cache_bench to resemble a real workload (#6629) Summary: I suspect LRUCache could use some optimization, and to support such an effort, a good benchmarking tool is needed. The existing cache_bench was heavily skewed toward insertion and lookup misses, and did not saturate memory with other work. This change should improve those things to better resemble a real workload. (All below using clang compiler, for some consistency, but not necessarily same version and settings.) The real workload is from production MySQL on RocksDB, filtering stacks containing "LRU", "ShardedCache" or "CacheShard." Lookup inclusive: 66% Insert inclusive: 17% Release inclusive: 15% An alternate simulated workload is MySQL running a LinkBench read test: Lookup inclusive: 54% Insert inclusive: 24% Release inclusive: 21% cache_bench default settings, prior to this change: Lookup inclusive: 35.8% Insert inclusive: 63.6% Release inclusive: 0% cache_bench after this change (intended as somewhat "tighter" workload than average production, more like LinkBench): Lookup inclusive: 52% Insert inclusive: 20% Release inclusive: 26% And top exclusive stacks (portion of stack samples as filtered above): Production MySQL: LRUHandleTable::FindPointer: 25.3% rocksdb::operator==: 15.1% <-- Slice == LRUCacheShard::LRU_Remove: 13.8% ShardedCache::Lookup: 8.9% __pthread_mutex_lock: 7.1% LRUCacheShard::LRU_Insert: 6.3% MurmurHash64A: 4.8% <-- Since upgraded to XXH3p ... Old cache_bench: LRUHandleTable::FindPointer: 23.6% __pthread_mutex_lock: 15.0% __pthread_mutex_unlock_usercnt: 11.7% __lll_lock_wait: 8.6% __lll_unlock_wake: 6.8% LRUCacheShard::LRU_Insert: 6.0% ShardedCache::Lookup: 4.4% LRUCacheShard::LRU_Remove: 2.8% ... rocksdb::operator==: 0.2% <-- Slice == ... New cache_bench: LRUHandleTable::FindPointer: 22.8% __pthread_mutex_unlock_usercnt: 14.3% rocksdb::operator==: 10.5% <-- Slice == LRUCacheShard::LRU_Insert: 9.0% __pthread_mutex_lock: 5.9% LRUCacheShard::LRU_Remove: 5.0% ... ShardedCache::Lookup: 2.9% ... So there's a bit more lock contention in the benchmark than in production, but otherwise looks similar enough to me. At least it's a big improvement over the existing code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6629 Test Plan: No production code changes, ran cache_bench with ASAN Reviewed By: ltamasi Differential Revision: D20824318 Pulled By: pdillinger fbshipit-source-id: 6f8dc5891ead0f87edbed3a615ecd5289d9abe12
2020-04-03 19:24:09 +02:00
if (handle) {
cache_->Release(handle);
handle = nullptr;
}
}
void PrintEnv() const {
printf("RocksDB version : %d.%d\n", kMajorVersion, kMinorVersion);
Revamp cache_bench to resemble a real workload (#6629) Summary: I suspect LRUCache could use some optimization, and to support such an effort, a good benchmarking tool is needed. The existing cache_bench was heavily skewed toward insertion and lookup misses, and did not saturate memory with other work. This change should improve those things to better resemble a real workload. (All below using clang compiler, for some consistency, but not necessarily same version and settings.) The real workload is from production MySQL on RocksDB, filtering stacks containing "LRU", "ShardedCache" or "CacheShard." Lookup inclusive: 66% Insert inclusive: 17% Release inclusive: 15% An alternate simulated workload is MySQL running a LinkBench read test: Lookup inclusive: 54% Insert inclusive: 24% Release inclusive: 21% cache_bench default settings, prior to this change: Lookup inclusive: 35.8% Insert inclusive: 63.6% Release inclusive: 0% cache_bench after this change (intended as somewhat "tighter" workload than average production, more like LinkBench): Lookup inclusive: 52% Insert inclusive: 20% Release inclusive: 26% And top exclusive stacks (portion of stack samples as filtered above): Production MySQL: LRUHandleTable::FindPointer: 25.3% rocksdb::operator==: 15.1% <-- Slice == LRUCacheShard::LRU_Remove: 13.8% ShardedCache::Lookup: 8.9% __pthread_mutex_lock: 7.1% LRUCacheShard::LRU_Insert: 6.3% MurmurHash64A: 4.8% <-- Since upgraded to XXH3p ... Old cache_bench: LRUHandleTable::FindPointer: 23.6% __pthread_mutex_lock: 15.0% __pthread_mutex_unlock_usercnt: 11.7% __lll_lock_wait: 8.6% __lll_unlock_wake: 6.8% LRUCacheShard::LRU_Insert: 6.0% ShardedCache::Lookup: 4.4% LRUCacheShard::LRU_Remove: 2.8% ... rocksdb::operator==: 0.2% <-- Slice == ... New cache_bench: LRUHandleTable::FindPointer: 22.8% __pthread_mutex_unlock_usercnt: 14.3% rocksdb::operator==: 10.5% <-- Slice == LRUCacheShard::LRU_Insert: 9.0% __pthread_mutex_lock: 5.9% LRUCacheShard::LRU_Remove: 5.0% ... ShardedCache::Lookup: 2.9% ... So there's a bit more lock contention in the benchmark than in production, but otherwise looks similar enough to me. At least it's a big improvement over the existing code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6629 Test Plan: No production code changes, ran cache_bench with ASAN Reviewed By: ltamasi Differential Revision: D20824318 Pulled By: pdillinger fbshipit-source-id: 6f8dc5891ead0f87edbed3a615ecd5289d9abe12
2020-04-03 19:24:09 +02:00
printf("Number of threads : %u\n", FLAGS_threads);
printf("Ops per thread : %" PRIu64 "\n", FLAGS_ops_per_thread);
printf("Cache size : %" PRIu64 "\n", FLAGS_cache_size);
Revamp cache_bench to resemble a real workload (#6629) Summary: I suspect LRUCache could use some optimization, and to support such an effort, a good benchmarking tool is needed. The existing cache_bench was heavily skewed toward insertion and lookup misses, and did not saturate memory with other work. This change should improve those things to better resemble a real workload. (All below using clang compiler, for some consistency, but not necessarily same version and settings.) The real workload is from production MySQL on RocksDB, filtering stacks containing "LRU", "ShardedCache" or "CacheShard." Lookup inclusive: 66% Insert inclusive: 17% Release inclusive: 15% An alternate simulated workload is MySQL running a LinkBench read test: Lookup inclusive: 54% Insert inclusive: 24% Release inclusive: 21% cache_bench default settings, prior to this change: Lookup inclusive: 35.8% Insert inclusive: 63.6% Release inclusive: 0% cache_bench after this change (intended as somewhat "tighter" workload than average production, more like LinkBench): Lookup inclusive: 52% Insert inclusive: 20% Release inclusive: 26% And top exclusive stacks (portion of stack samples as filtered above): Production MySQL: LRUHandleTable::FindPointer: 25.3% rocksdb::operator==: 15.1% <-- Slice == LRUCacheShard::LRU_Remove: 13.8% ShardedCache::Lookup: 8.9% __pthread_mutex_lock: 7.1% LRUCacheShard::LRU_Insert: 6.3% MurmurHash64A: 4.8% <-- Since upgraded to XXH3p ... Old cache_bench: LRUHandleTable::FindPointer: 23.6% __pthread_mutex_lock: 15.0% __pthread_mutex_unlock_usercnt: 11.7% __lll_lock_wait: 8.6% __lll_unlock_wake: 6.8% LRUCacheShard::LRU_Insert: 6.0% ShardedCache::Lookup: 4.4% LRUCacheShard::LRU_Remove: 2.8% ... rocksdb::operator==: 0.2% <-- Slice == ... New cache_bench: LRUHandleTable::FindPointer: 22.8% __pthread_mutex_unlock_usercnt: 14.3% rocksdb::operator==: 10.5% <-- Slice == LRUCacheShard::LRU_Insert: 9.0% __pthread_mutex_lock: 5.9% LRUCacheShard::LRU_Remove: 5.0% ... ShardedCache::Lookup: 2.9% ... So there's a bit more lock contention in the benchmark than in production, but otherwise looks similar enough to me. At least it's a big improvement over the existing code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6629 Test Plan: No production code changes, ran cache_bench with ASAN Reviewed By: ltamasi Differential Revision: D20824318 Pulled By: pdillinger fbshipit-source-id: 6f8dc5891ead0f87edbed3a615ecd5289d9abe12
2020-04-03 19:24:09 +02:00
printf("Num shard bits : %u\n", FLAGS_num_shard_bits);
printf("Max key : %" PRIu64 "\n", max_key_);
printf("Resident ratio : %g\n", FLAGS_resident_ratio);
printf("Skew degree : %u\n", FLAGS_skew);
printf("Populate cache : %d\n", int{FLAGS_populate_cache});
printf("Lookup+Insert pct : %u%%\n", FLAGS_lookup_insert_percent);
printf("Insert percentage : %u%%\n", FLAGS_insert_percent);
printf("Lookup percentage : %u%%\n", FLAGS_lookup_percent);
printf("Erase percentage : %u%%\n", FLAGS_erase_percent);
printf("----------------------------\n");
}
};
} // namespace ROCKSDB_NAMESPACE
int main(int argc, char** argv) {
ParseCommandLineFlags(&argc, &argv, true);
if (FLAGS_threads <= 0) {
fprintf(stderr, "threads number <= 0\n");
exit(1);
}
ROCKSDB_NAMESPACE::CacheBench bench;
if (FLAGS_populate_cache) {
bench.PopulateCache();
Revamp cache_bench to resemble a real workload (#6629) Summary: I suspect LRUCache could use some optimization, and to support such an effort, a good benchmarking tool is needed. The existing cache_bench was heavily skewed toward insertion and lookup misses, and did not saturate memory with other work. This change should improve those things to better resemble a real workload. (All below using clang compiler, for some consistency, but not necessarily same version and settings.) The real workload is from production MySQL on RocksDB, filtering stacks containing "LRU", "ShardedCache" or "CacheShard." Lookup inclusive: 66% Insert inclusive: 17% Release inclusive: 15% An alternate simulated workload is MySQL running a LinkBench read test: Lookup inclusive: 54% Insert inclusive: 24% Release inclusive: 21% cache_bench default settings, prior to this change: Lookup inclusive: 35.8% Insert inclusive: 63.6% Release inclusive: 0% cache_bench after this change (intended as somewhat "tighter" workload than average production, more like LinkBench): Lookup inclusive: 52% Insert inclusive: 20% Release inclusive: 26% And top exclusive stacks (portion of stack samples as filtered above): Production MySQL: LRUHandleTable::FindPointer: 25.3% rocksdb::operator==: 15.1% <-- Slice == LRUCacheShard::LRU_Remove: 13.8% ShardedCache::Lookup: 8.9% __pthread_mutex_lock: 7.1% LRUCacheShard::LRU_Insert: 6.3% MurmurHash64A: 4.8% <-- Since upgraded to XXH3p ... Old cache_bench: LRUHandleTable::FindPointer: 23.6% __pthread_mutex_lock: 15.0% __pthread_mutex_unlock_usercnt: 11.7% __lll_lock_wait: 8.6% __lll_unlock_wake: 6.8% LRUCacheShard::LRU_Insert: 6.0% ShardedCache::Lookup: 4.4% LRUCacheShard::LRU_Remove: 2.8% ... rocksdb::operator==: 0.2% <-- Slice == ... New cache_bench: LRUHandleTable::FindPointer: 22.8% __pthread_mutex_unlock_usercnt: 14.3% rocksdb::operator==: 10.5% <-- Slice == LRUCacheShard::LRU_Insert: 9.0% __pthread_mutex_lock: 5.9% LRUCacheShard::LRU_Remove: 5.0% ... ShardedCache::Lookup: 2.9% ... So there's a bit more lock contention in the benchmark than in production, but otherwise looks similar enough to me. At least it's a big improvement over the existing code. Pull Request resolved: https://github.com/facebook/rocksdb/pull/6629 Test Plan: No production code changes, ran cache_bench with ASAN Reviewed By: ltamasi Differential Revision: D20824318 Pulled By: pdillinger fbshipit-source-id: 6f8dc5891ead0f87edbed3a615ecd5289d9abe12
2020-04-03 19:24:09 +02:00
printf("Population complete\n");
printf("----------------------------\n");
}
if (bench.Run()) {
return 0;
} else {
return 1;
}
}
#endif // GFLAGS