rocksdb/db/write_thread.h

392 lines
14 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
#include <assert.h>
#include <stdint.h>
#include <atomic>
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
#include <chrono>
#include <condition_variable>
#include <mutex>
#include <type_traits>
#include <vector>
#include "db/write_callback.h"
#include "monitoring/instrumented_mutex.h"
#include "rocksdb/options.h"
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
#include "rocksdb/status.h"
#include "rocksdb/types.h"
#include "rocksdb/write_batch.h"
#include "util/autovector.h"
namespace rocksdb {
class WriteThread {
public:
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
enum State : uint8_t {
// The initial state of a writer. This is a Writer that is
// waiting in JoinBatchGroup. This state can be left when another
// thread informs the waiter that it has become a group leader
// (-> STATE_GROUP_LEADER), when a leader that has chosen to be
// non-parallel informs a follower that its writes have been committed
// (-> STATE_COMPLETED), or when a leader that has chosen to perform
// updates in parallel and needs this Writer to apply its batch (->
// STATE_PARALLEL_FOLLOWER).
STATE_INIT = 1,
// The state used to inform a waiting Writer that it has become the
// leader, and it should now build a write batch group. Tricky:
// this state is not used if newest_writer_ is empty when a writer
// enqueues itself, because there is no need to wait (or even to
// create the mutex and condvar used to wait) in that case. This is
// a terminal state unless the leader chooses to make this a parallel
// batch, in which case the last parallel worker to finish will move
// the leader to STATE_COMPLETED.
STATE_GROUP_LEADER = 2,
// The state used to inform a waiting writer that it has become the
// leader of memtable writer group. The leader will either write
// memtable for the whole group, or launch a parallel group write
// to memtable by calling LaunchParallelMemTableWrite.
STATE_MEMTABLE_WRITER_LEADER = 4,
// The state used to inform a waiting writer that it has become a
// parallel memtable writer. It can be the group leader who launch the
// praallel writer group, or one of the followers. The writer should then
// apply its batch to the memtable concurrently and call
// CompleteParallelMemTableWriter.
STATE_PARALLEL_MEMTABLE_WRITER = 8,
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
// A follower whose writes have been applied, or a parallel leader
// whose followers have all finished their work. This is a terminal
// state.
STATE_COMPLETED = 16,
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
// A state indicating that the thread may be waiting using StateMutex()
// and StateCondVar()
STATE_LOCKED_WAITING = 32,
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
};
struct Writer;
struct WriteGroup {
Writer* leader = nullptr;
Writer* last_writer = nullptr;
SequenceNumber last_sequence;
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
// before running goes to zero, status needs leader->StateMutex()
Status status;
std::atomic<size_t> running;
size_t size = 0;
struct Iterator {
Writer* writer;
Writer* last_writer;
explicit Iterator(Writer* w, Writer* last)
: writer(w), last_writer(last) {}
Writer* operator*() const { return writer; }
Iterator& operator++() {
assert(writer != nullptr);
if (writer == last_writer) {
writer = nullptr;
} else {
writer = writer->link_newer;
}
return *this;
}
bool operator!=(const Iterator& other) const {
return writer != other.writer;
}
};
Iterator begin() const { return Iterator(leader, last_writer); }
Iterator end() const { return Iterator(nullptr, nullptr); }
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
};
// Information kept for every waiting writer.
struct Writer {
WriteBatch* batch;
bool sync;
bool no_slowdown;
bool disable_wal;
bool disable_memtable;
uint64_t log_used; // log number that this batch was inserted into
uint64_t log_ref; // log number that memtable insert should reference
WriteCallback* callback;
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
bool made_waitable; // records lazy construction of mutex and cv
std::atomic<uint8_t> state; // write under StateMutex() or pre-link
WriteGroup* write_group;
SequenceNumber sequence; // the sequence number to use for the first key
Status status; // status of memtable inserter
Status callback_status; // status returned by callback->Callback()
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
std::aligned_storage<sizeof(std::mutex)>::type state_mutex_bytes;
std::aligned_storage<sizeof(std::condition_variable)>::type state_cv_bytes;
Writer* link_older; // read/write only before linking, or as leader
Writer* link_newer; // lazy, read/write only before linking, or as leader
Writer()
: batch(nullptr),
sync(false),
no_slowdown(false),
disable_wal(false),
disable_memtable(false),
log_used(0),
log_ref(0),
callback(nullptr),
made_waitable(false),
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
state(STATE_INIT),
write_group(nullptr),
link_older(nullptr),
link_newer(nullptr) {}
Writer(const WriteOptions& write_options, WriteBatch* _batch,
WriteCallback* _callback, uint64_t _log_ref, bool _disable_memtable)
: batch(_batch),
sync(write_options.sync),
no_slowdown(write_options.no_slowdown),
disable_wal(write_options.disableWAL),
disable_memtable(_disable_memtable),
log_used(0),
log_ref(_log_ref),
callback(_callback),
made_waitable(false),
state(STATE_INIT),
write_group(nullptr),
link_older(nullptr),
link_newer(nullptr) {}
~Writer() {
if (made_waitable) {
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
StateMutex().~mutex();
StateCV().~condition_variable();
}
}
bool CheckCallback(DB* db) {
if (callback != nullptr) {
callback_status = callback->Callback(db);
}
return callback_status.ok();
}
void CreateMutex() {
if (!made_waitable) {
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
// Note that made_waitable is tracked separately from state
// transitions, because we can't atomically create the mutex and
// link into the list.
made_waitable = true;
support for concurrent adds to memtable Summary: This diff adds support for concurrent adds to the skiplist memtable implementations. Memory allocation is made thread-safe by the addition of a spinlock, with small per-core buffers to avoid contention. Concurrent memtable writes are made via an additional method and don't impose a performance overhead on the non-concurrent case, so parallelism can be selected on a per-batch basis. Write thread synchronization is an increasing bottleneck for higher levels of concurrency, so this diff adds --enable_write_thread_adaptive_yield (default off). This feature causes threads joining a write batch group to spin for a short time (default 100 usec) using sched_yield, rather than going to sleep on a mutex. If the timing of the yield calls indicates that another thread has actually run during the yield then spinning is avoided. This option improves performance for concurrent situations even without parallel adds, although it has the potential to increase CPU usage (and the heuristic adaptation is not yet mature). Parallel writes are not currently compatible with inplace updates, update callbacks, or delete filtering. Enable it with --allow_concurrent_memtable_write (and --enable_write_thread_adaptive_yield). Parallel memtable writes are performance neutral when there is no actual parallelism, and in my experiments (SSD server-class Linux and varying contention and key sizes for fillrandom) they are always a performance win when there is more than one thread. Statistics are updated earlier in the write path, dropping the number of DB mutex acquisitions from 2 to 1 for almost all cases. This diff was motivated and inspired by Yahoo's cLSM work. It is more conservative than cLSM: RocksDB's write batch group leader role is preserved (along with all of the existing flush and write throttling logic) and concurrent writers are blocked until all memtable insertions have completed and the sequence number has been advanced, to preserve linearizability. My test config is "db_bench -benchmarks=fillrandom -threads=$T -batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T -level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999 -disable_auto_compactions --max_write_buffer_number=8 -max_background_flushes=8 --disable_wal --write_buffer_size=160000000 --block_size=16384 --allow_concurrent_memtable_write" on a two-socket Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1 thread I get ~440Kops/sec. Peak performance for 1 socket (numactl -N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance across both sockets happens at 30 threads, and is ~900Kops/sec, although with fewer threads there is less performance loss when the system has background work. Test Plan: 1. concurrent stress tests for InlineSkipList and DynamicBloom 2. make clean; make check 3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench 4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench 5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench 6. make clean; OPT=-DROCKSDB_LITE make check 7. verify no perf regressions when disabled Reviewers: igor, sdong Reviewed By: sdong Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba Differential Revision: https://reviews.facebook.net/D50589
2015-08-15 01:59:07 +02:00
new (&state_mutex_bytes) std::mutex;
new (&state_cv_bytes) std::condition_variable;
}
}
// returns the aggregate status of this Writer
Status FinalStatus() {
if (!status.ok()) {
// a non-ok memtable write status takes presidence
assert(callback == nullptr || callback_status.ok());
return status;
} else if (!callback_status.ok()) {
// if the callback failed then that is the status we want
// because a memtable insert should not have been attempted
assert(callback != nullptr);
assert(status.ok());
return callback_status;
} else {
// if there is no callback then we only care about
// the memtable insert status
assert(callback == nullptr || callback_status.ok());
return status;
}
}
bool CallbackFailed() {
return (callback != nullptr) && !callback_status.ok();
}
bool ShouldWriteToMemtable() {
return status.ok() && !CallbackFailed() && !disable_memtable;
}
bool ShouldWriteToWAL() {
return status.ok() && !CallbackFailed() && !disable_wal;
}
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
// No other mutexes may be acquired while holding StateMutex(), it is
// always last in the order
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
std::mutex& StateMutex() {
assert(made_waitable);
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
return *static_cast<std::mutex*>(static_cast<void*>(&state_mutex_bytes));
}
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
std::condition_variable& StateCV() {
assert(made_waitable);
return *static_cast<std::condition_variable*>(
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
static_cast<void*>(&state_cv_bytes));
}
};
struct AdaptationContext {
const char* name;
std::atomic<int32_t> value;
explicit AdaptationContext(const char* name0) : name(name0), value(0) {}
};
explicit WriteThread(const ImmutableDBOptions& db_options);
virtual ~WriteThread() = default;
// IMPORTANT: None of the methods in this class rely on the db mutex
// for correctness. All of the methods except JoinBatchGroup and
// EnterUnbatched may be called either with or without the db mutex held.
// Correctness is maintained by ensuring that only a single thread is
// a leader at a time.
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
// Registers w as ready to become part of a batch group, waits until the
// caller should perform some work, and returns the current state of the
// writer. If w has become the leader of a write batch group, returns
// STATE_GROUP_LEADER. If w has been made part of a sequential batch
// group and the leader has performed the write, returns STATE_DONE.
2016-06-25 09:29:40 +02:00
// If w has been made part of a parallel batch group and is responsible
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
// for updating the memtable, returns STATE_PARALLEL_FOLLOWER.
//
// The db mutex SHOULD NOT be held when calling this function, because
// it will block.
//
// Writer* w: Writer to be executed as part of a batch group
void JoinBatchGroup(Writer* w);
// Constructs a write batch group led by leader, which should be a
// Writer passed to JoinBatchGroup on the current thread.
//
// Writer* leader: Writer that is STATE_GROUP_LEADER
// WriteGroup* write_group: Out-param of group members
// returns: Total batch group byte size
size_t EnterAsBatchGroupLeader(Writer* leader, WriteGroup* write_group);
// Unlinks the Writer-s in a batch group, wakes up the non-leaders,
// and wakes up the next leader (if any).
//
// WriteGroup* write_group: the write group
// Status status: Status of write operation
void ExitAsBatchGroupLeader(WriteGroup& write_group, Status status);
// Exit batch group on behalf of batch group leader.
void ExitAsBatchGroupFollower(Writer* w);
// Constructs a write batch group led by leader from newest_memtable_writers_
// list. The leader should either write memtable for the whole group and
// call ExitAsMemTableWriter, or launch parallel memtable write through
// LaunchParallelMemTableWriters.
void EnterAsMemTableWriter(Writer* leader, WriteGroup* write_grup);
// Memtable writer group leader, or the last finished writer in a parallel
// write group, exit from the newest_memtable_writers_ list, and wake up
// the next leader if needed.
void ExitAsMemTableWriter(Writer* self, WriteGroup& write_group);
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
// Causes JoinBatchGroup to return STATE_PARALLEL_FOLLOWER for all of the
// non-leader members of this write batch group. Sets Writer::sequence
// before waking them up.
//
// WriteGroup* write_group: Extra state used to coordinate the parallel add
void LaunchParallelMemTableWriters(WriteGroup* write_group);
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
// Reports the completion of w's batch to the parallel group leader, and
// waits for the rest of the parallel batch to complete. Returns true
// if this thread is the last to complete, and hence should advance
// the sequence number and then call EarlyExitParallelGroup, false if
// someone else has already taken responsibility for that.
bool CompleteParallelMemTableWriter(Writer* w);
// Waits for all preceding writers (unlocking mu while waiting), then
// registers w as the currently proceeding writer.
//
// Writer* w: A Writer not eligible for batching
// InstrumentedMutex* mu: The db mutex, to unlock while waiting
// REQUIRES: db mutex held
void EnterUnbatched(Writer* w, InstrumentedMutex* mu);
// Completes a Writer begun with EnterUnbatched, unblocking subsequent
// writers.
void ExitUnbatched(Writer* w);
// Wait for all parallel memtable writers to finish, in case pipelined
// write is enabled.
void WaitForMemTableWriters();
SequenceNumber UpdateLastSequence(SequenceNumber sequence) {
if (sequence > last_sequence_) {
last_sequence_ = sequence;
}
return last_sequence_;
}
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
private:
// See AwaitState.
const uint64_t max_yield_usec_;
const uint64_t slow_yield_usec_;
// Allow multiple writers write to memtable concurrently.
const bool allow_concurrent_memtable_write_;
// Enable pipelined write to WAL and memtable.
const bool enable_pipelined_write_;
support for concurrent adds to memtable Summary: This diff adds support for concurrent adds to the skiplist memtable implementations. Memory allocation is made thread-safe by the addition of a spinlock, with small per-core buffers to avoid contention. Concurrent memtable writes are made via an additional method and don't impose a performance overhead on the non-concurrent case, so parallelism can be selected on a per-batch basis. Write thread synchronization is an increasing bottleneck for higher levels of concurrency, so this diff adds --enable_write_thread_adaptive_yield (default off). This feature causes threads joining a write batch group to spin for a short time (default 100 usec) using sched_yield, rather than going to sleep on a mutex. If the timing of the yield calls indicates that another thread has actually run during the yield then spinning is avoided. This option improves performance for concurrent situations even without parallel adds, although it has the potential to increase CPU usage (and the heuristic adaptation is not yet mature). Parallel writes are not currently compatible with inplace updates, update callbacks, or delete filtering. Enable it with --allow_concurrent_memtable_write (and --enable_write_thread_adaptive_yield). Parallel memtable writes are performance neutral when there is no actual parallelism, and in my experiments (SSD server-class Linux and varying contention and key sizes for fillrandom) they are always a performance win when there is more than one thread. Statistics are updated earlier in the write path, dropping the number of DB mutex acquisitions from 2 to 1 for almost all cases. This diff was motivated and inspired by Yahoo's cLSM work. It is more conservative than cLSM: RocksDB's write batch group leader role is preserved (along with all of the existing flush and write throttling logic) and concurrent writers are blocked until all memtable insertions have completed and the sequence number has been advanced, to preserve linearizability. My test config is "db_bench -benchmarks=fillrandom -threads=$T -batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T -level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999 -disable_auto_compactions --max_write_buffer_number=8 -max_background_flushes=8 --disable_wal --write_buffer_size=160000000 --block_size=16384 --allow_concurrent_memtable_write" on a two-socket Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1 thread I get ~440Kops/sec. Peak performance for 1 socket (numactl -N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance across both sockets happens at 30 threads, and is ~900Kops/sec, although with fewer threads there is less performance loss when the system has background work. Test Plan: 1. concurrent stress tests for InlineSkipList and DynamicBloom 2. make clean; make check 3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench 4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench 5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench 6. make clean; OPT=-DROCKSDB_LITE make check 7. verify no perf regressions when disabled Reviewers: igor, sdong Reviewed By: sdong Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba Differential Revision: https://reviews.facebook.net/D50589
2015-08-15 01:59:07 +02:00
// Points to the newest pending writer. Only leader can remove
// elements, adding can be done lock-free by anybody.
std::atomic<Writer*> newest_writer_;
// Points to the newest pending memtable writer. Used only when pipelined
// write is enabled.
std::atomic<Writer*> newest_memtable_writer_;
// The last sequence that have been consumed by a writer. The sequence
// is not necessary visible to reads because the writer can be ongoing.
SequenceNumber last_sequence_;
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
// Waits for w->state & goal_mask using w->StateMutex(). Returns
// the state that satisfies goal_mask.
uint8_t BlockingAwaitState(Writer* w, uint8_t goal_mask);
// Blocks until w->state & goal_mask, returning the state value
// that satisfied the predicate. Uses ctx to adaptively use
// std::this_thread::yield() to avoid mutex overheads. ctx should be
// a context-dependent static.
uint8_t AwaitState(Writer* w, uint8_t goal_mask, AdaptationContext* ctx);
// Set writer state and wake the writer up if it is waiting.
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
void SetState(Writer* w, uint8_t new_state);
// Links w into the newest_writer list. Return true if w was linked directly
// into the leader position. Safe to call from multiple threads without
// external locking.
bool LinkOne(Writer* w, std::atomic<Writer*>* newest_writer);
// Link write group into the newest_writer list as a whole, while keeping the
// order of the writers unchanged. Return true if the group was linked
// directly into the leader position.
bool LinkGroup(WriteGroup& write_group, std::atomic<Writer*>* newest_writer);
// Computes any missing link_newer links. Should not be called
// concurrently with itself.
void CreateMissingNewerLinks(Writer* head);
// Set the leader in write_group to completed state and remove it from the
// write group.
void CompleteLeader(WriteGroup& write_group);
// Set a follower in write_group to completed state and remove it from the
// write group.
void CompleteFollower(Writer* w, WriteGroup& write_group);
};
} // namespace rocksdb