You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
rocksdb/db/memtable.cc

818 lines
29 KiB

// Copyright (c) 2011-present, Facebook, Inc. All rights reserved.
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree. An additional grant
// of patent rights can be found in the PATENTS file in the same directory.
//
// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file. See the AUTHORS file for names of contributors.
#include "db/memtable.h"
#include <memory>
#include <algorithm>
Add a new mem-table representation based on cuckoo hash. Summary: = Major Changes = * Add a new mem-table representation, HashCuckooRep, which is based cuckoo hash. Cuckoo hash uses multiple hash functions. This allows each key to have multiple possible locations in the mem-table. - Put: When insert a key, it will try to find whether one of its possible locations is vacant and store the key. If none of its possible locations are available, then it will kick out a victim key and store at that location. The kicked-out victim key will then be stored at a vacant space of its possible locations or kick-out another victim. In this diff, the kick-out path (known as cuckoo-path) is found using BFS, which guarantees to be the shortest. - Get: Simply tries all possible locations of a key --- this guarantees worst-case constant time complexity. - Time complexity: O(1) for Get, and average O(1) for Put if the fullness of the mem-table is below 80%. - Default using two hash functions, the number of hash functions used by the cuckoo-hash may dynamically increase if it fails to find a short-enough kick-out path. - Currently, HashCuckooRep does not support iteration and snapshots, as our current main purpose of this is to optimize point access. = Minor Changes = * Add IsSnapshotSupported() to DB to indicate whether the current DB supports snapshots. If it returns false, then DB::GetSnapshot() will always return nullptr. Test Plan: Run existing tests. Will develop a test specifically for cuckoo hash in the next diff. Reviewers: sdong, haobo Reviewed By: sdong CC: leveldb, dhruba, igor Differential Revision: https://reviews.facebook.net/D16155
11 years ago
#include <limits>
#include "db/dbformat.h"
#include "db/merge_context.h"
#include "db/pinned_iterators_manager.h"
#include "db/writebuffer.h"
#include "rocksdb/comparator.h"
#include "rocksdb/env.h"
#include "rocksdb/iterator.h"
#include "rocksdb/merge_operator.h"
#include "rocksdb/slice_transform.h"
#include "table/internal_iterator.h"
#include "table/merger.h"
#include "util/arena.h"
#include "util/coding.h"
#include "util/murmurhash.h"
#include "util/mutexlock.h"
#include "util/perf_context_imp.h"
#include "util/statistics.h"
#include "util/stop_watch.h"
namespace rocksdb {
MemTableOptions::MemTableOptions(
const ImmutableCFOptions& ioptions,
const MutableCFOptions& mutable_cf_options)
: write_buffer_size(mutable_cf_options.write_buffer_size),
arena_block_size(mutable_cf_options.arena_block_size),
memtable_prefix_bloom_bits(mutable_cf_options.memtable_prefix_bloom_bits),
memtable_prefix_bloom_probes(
mutable_cf_options.memtable_prefix_bloom_probes),
memtable_prefix_bloom_huge_page_tlb_size(
mutable_cf_options.memtable_prefix_bloom_huge_page_tlb_size),
inplace_update_support(ioptions.inplace_update_support),
inplace_update_num_locks(mutable_cf_options.inplace_update_num_locks),
inplace_callback(ioptions.inplace_callback),
max_successive_merges(mutable_cf_options.max_successive_merges),
filter_deletes(mutable_cf_options.filter_deletes),
statistics(ioptions.statistics),
merge_operator(ioptions.merge_operator),
info_log(ioptions.info_log) {}
MemTable::MemTable(const InternalKeyComparator& cmp,
const ImmutableCFOptions& ioptions,
const MutableCFOptions& mutable_cf_options,
WriteBuffer* write_buffer, SequenceNumber earliest_seq)
: comparator_(cmp),
moptions_(ioptions, mutable_cf_options),
refs_(0),
kArenaBlockSize(OptimizeBlockSize(moptions_.arena_block_size)),
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
10 years ago
arena_(moptions_.arena_block_size, 0),
allocator_(&arena_, write_buffer),
table_(ioptions.memtable_factory->CreateMemTableRep(
comparator_, &allocator_, ioptions.prefix_extractor,
ioptions.info_log)),
data_size_(0),
num_entries_(0),
num_deletes_(0),
flush_in_progress_(false),
flush_completed_(false),
file_number_(0),
first_seqno_(0),
earliest_seqno_(earliest_seq),
mem_next_logfile_number_(0),
min_prep_log_referenced_(0),
locks_(moptions_.inplace_update_support
? moptions_.inplace_update_num_locks
: 0),
prefix_extractor_(ioptions.prefix_extractor),
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
10 years ago
flush_state_(FLUSH_NOT_REQUESTED),
env_(ioptions.env) {
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
10 years ago
UpdateFlushState();
// something went wrong if we need to flush before inserting anything
assert(!ShouldScheduleFlush());
if (prefix_extractor_ && moptions_.memtable_prefix_bloom_bits > 0) {
prefix_bloom_.reset(new DynamicBloom(
&allocator_,
moptions_.memtable_prefix_bloom_bits, ioptions.bloom_locality,
moptions_.memtable_prefix_bloom_probes, nullptr,
moptions_.memtable_prefix_bloom_huge_page_tlb_size,
ioptions.info_log));
}
}
MemTable::~MemTable() { assert(refs_ == 0); }
size_t MemTable::ApproximateMemoryUsage() {
Add a new mem-table representation based on cuckoo hash. Summary: = Major Changes = * Add a new mem-table representation, HashCuckooRep, which is based cuckoo hash. Cuckoo hash uses multiple hash functions. This allows each key to have multiple possible locations in the mem-table. - Put: When insert a key, it will try to find whether one of its possible locations is vacant and store the key. If none of its possible locations are available, then it will kick out a victim key and store at that location. The kicked-out victim key will then be stored at a vacant space of its possible locations or kick-out another victim. In this diff, the kick-out path (known as cuckoo-path) is found using BFS, which guarantees to be the shortest. - Get: Simply tries all possible locations of a key --- this guarantees worst-case constant time complexity. - Time complexity: O(1) for Get, and average O(1) for Put if the fullness of the mem-table is below 80%. - Default using two hash functions, the number of hash functions used by the cuckoo-hash may dynamically increase if it fails to find a short-enough kick-out path. - Currently, HashCuckooRep does not support iteration and snapshots, as our current main purpose of this is to optimize point access. = Minor Changes = * Add IsSnapshotSupported() to DB to indicate whether the current DB supports snapshots. If it returns false, then DB::GetSnapshot() will always return nullptr. Test Plan: Run existing tests. Will develop a test specifically for cuckoo hash in the next diff. Reviewers: sdong, haobo Reviewed By: sdong CC: leveldb, dhruba, igor Differential Revision: https://reviews.facebook.net/D16155
11 years ago
size_t arena_usage = arena_.ApproximateMemoryUsage();
size_t table_usage = table_->ApproximateMemoryUsage();
// let MAX_USAGE = std::numeric_limits<size_t>::max()
// then if arena_usage + total_usage >= MAX_USAGE, return MAX_USAGE.
// the following variation is to avoid numeric overflow.
if (arena_usage >= std::numeric_limits<size_t>::max() - table_usage) {
return std::numeric_limits<size_t>::max();
}
// otherwise, return the actual usage
return arena_usage + table_usage;
}
bool MemTable::ShouldFlushNow() const {
// In a lot of times, we cannot allocate arena blocks that exactly matches the
// buffer size. Thus we have to decide if we should over-allocate or
// under-allocate.
// This constant variable can be interpreted as: if we still have more than
// "kAllowOverAllocationRatio * kArenaBlockSize" space left, we'd try to over
// allocate one more block.
const double kAllowOverAllocationRatio = 0.6;
// If arena still have room for new block allocation, we can safely say it
// shouldn't flush.
auto allocated_memory =
table_->ApproximateMemoryUsage() + arena_.MemoryAllocatedBytes();
// if we can still allocate one more block without exceeding the
// over-allocation ratio, then we should not flush.
if (allocated_memory + kArenaBlockSize <
moptions_.write_buffer_size +
kArenaBlockSize * kAllowOverAllocationRatio) {
return false;
}
// if user keeps adding entries that exceeds moptions.write_buffer_size,
// we need to flush earlier even though we still have much available
// memory left.
if (allocated_memory > moptions_.write_buffer_size +
kArenaBlockSize * kAllowOverAllocationRatio) {
return true;
}
// In this code path, Arena has already allocated its "last block", which
// means the total allocatedmemory size is either:
// (1) "moderately" over allocated the memory (no more than `0.6 * arena
// block size`. Or,
// (2) the allocated memory is less than write buffer size, but we'll stop
// here since if we allocate a new arena block, we'll over allocate too much
// more (half of the arena block size) memory.
//
// In either case, to avoid over-allocate, the last block will stop allocation
// when its usage reaches a certain ratio, which we carefully choose "0.75
// full" as the stop condition because it addresses the following issue with
// great simplicity: What if the next inserted entry's size is
// bigger than AllocatedAndUnused()?
//
// The answer is: if the entry size is also bigger than 0.25 *
// kArenaBlockSize, a dedicated block will be allocated for it; otherwise
// arena will anyway skip the AllocatedAndUnused() and allocate a new, empty
// and regular block. In either case, we *overly* over-allocated.
//
// Therefore, setting the last block to be at most "0.75 full" avoids both
// cases.
//
// NOTE: the average percentage of waste space of this approach can be counted
// as: "arena block size * 0.25 / write buffer size". User who specify a small
// write buffer size and/or big arena block size may suffer.
return arena_.AllocatedAndUnused() < kArenaBlockSize / 4;
}
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
10 years ago
void MemTable::UpdateFlushState() {
auto state = flush_state_.load(std::memory_order_relaxed);
if (state == FLUSH_NOT_REQUESTED && ShouldFlushNow()) {
// ignore CAS failure, because that means somebody else requested
// a flush
flush_state_.compare_exchange_strong(state, FLUSH_REQUESTED,
std::memory_order_relaxed,
std::memory_order_relaxed);
}
}
int MemTable::KeyComparator::operator()(const char* prefix_len_key1,
const char* prefix_len_key2) const {
// Internal keys are encoded as length-prefixed strings.
Slice k1 = GetLengthPrefixedSlice(prefix_len_key1);
Slice k2 = GetLengthPrefixedSlice(prefix_len_key2);
return comparator.Compare(k1, k2);
}
int MemTable::KeyComparator::operator()(const char* prefix_len_key,
const Slice& key)
const {
// Internal keys are encoded as length-prefixed strings.
Slice a = GetLengthPrefixedSlice(prefix_len_key);
return comparator.Compare(a, key);
}
Slice MemTableRep::UserKey(const char* key) const {
Slice slice = GetLengthPrefixedSlice(key);
return Slice(slice.data(), slice.size() - 8);
}
KeyHandle MemTableRep::Allocate(const size_t len, char** buf) {
*buf = allocator_->Allocate(len);
return static_cast<KeyHandle>(*buf);
}
// Encode a suitable internal key target for "target" and return it.
// Uses *scratch as scratch space, and the returned pointer will point
// into this scratch space.
const char* EncodeKey(std::string* scratch, const Slice& target) {
scratch->clear();
PutVarint32(scratch, static_cast<uint32_t>(target.size()));
scratch->append(target.data(), target.size());
return scratch->data();
}
class MemTableIterator : public InternalIterator {
public:
MemTableIterator(
const MemTable& mem, const ReadOptions& read_options, Arena* arena)
: bloom_(nullptr),
prefix_extractor_(mem.prefix_extractor_),
valid_(false),
arena_mode_(arena != nullptr) {
if (prefix_extractor_ != nullptr && !read_options.total_order_seek) {
bloom_ = mem.prefix_bloom_.get();
iter_ = mem.table_->GetDynamicPrefixIterator(arena);
} else {
iter_ = mem.table_->GetIterator(arena);
}
}
~MemTableIterator() {
#ifndef NDEBUG
// Assert that the MemTableIterator is never deleted while
// Pinning is Enabled.
assert(!pinned_iters_mgr_ ||
(pinned_iters_mgr_ && !pinned_iters_mgr_->PinningEnabled()));
#endif
if (arena_mode_) {
iter_->~Iterator();
} else {
delete iter_;
}
}
#ifndef NDEBUG
virtual void SetPinnedItersMgr(
PinnedIteratorsManager* pinned_iters_mgr) override {
pinned_iters_mgr_ = pinned_iters_mgr;
}
PinnedIteratorsManager* pinned_iters_mgr_ = nullptr;
#endif
virtual bool Valid() const override { return valid_; }
virtual void Seek(const Slice& k) override {
PERF_TIMER_GUARD(seek_on_memtable_time);
PERF_COUNTER_ADD(seek_on_memtable_count, 1);
if (bloom_ != nullptr) {
if (!bloom_->MayContain(
prefix_extractor_->Transform(ExtractUserKey(k)))) {
PERF_COUNTER_ADD(bloom_memtable_miss_count, 1);
valid_ = false;
return;
} else {
PERF_COUNTER_ADD(bloom_memtable_hit_count, 1);
}
}
iter_->Seek(k, nullptr);
valid_ = iter_->Valid();
}
virtual void SeekToFirst() override {
iter_->SeekToFirst();
valid_ = iter_->Valid();
}
virtual void SeekToLast() override {
iter_->SeekToLast();
valid_ = iter_->Valid();
}
virtual void Next() override {
assert(Valid());
iter_->Next();
valid_ = iter_->Valid();
}
virtual void Prev() override {
assert(Valid());
iter_->Prev();
valid_ = iter_->Valid();
}
virtual Slice key() const override {
assert(Valid());
return GetLengthPrefixedSlice(iter_->key());
}
virtual Slice value() const override {
assert(Valid());
Slice key_slice = GetLengthPrefixedSlice(iter_->key());
return GetLengthPrefixedSlice(key_slice.data() + key_slice.size());
}
virtual Status status() const override { return Status::OK(); }
Introduce ReadOptions::pin_data (support zero copy for keys) Summary: This patch update the Iterator API to introduce new functions that allow users to keep the Slices returned by key() valid as long as the Iterator is not deleted ReadOptions::pin_data : If true keep loaded blocks in memory as long as the iterator is not deleted Iterator::IsKeyPinned() : If true, this mean that the Slice returned by key() is valid as long as the iterator is not deleted Also add a new option BlockBasedTableOptions::use_delta_encoding to allow users to disable delta_encoding if needed. Benchmark results (using https://phabricator.fb.com/P20083553) ``` // $ du -h /home/tec/local/normal.4K.Snappy/db10077 // 6.1G /home/tec/local/normal.4K.Snappy/db10077 // $ du -h /home/tec/local/zero.8K.LZ4/db10077 // 6.4G /home/tec/local/zero.8K.LZ4/db10077 // Benchmarks for shard db10077 // _build/opt/rocks/benchmark/rocks_copy_benchmark \ // --normal_db_path="/home/tec/local/normal.4K.Snappy/db10077" \ // --zero_db_path="/home/tec/local/zero.8K.LZ4/db10077" // First run // ============================================================================ // rocks/benchmark/RocksCopyBenchmark.cpp relative time/iter iters/s // ============================================================================ // BM_StringCopy 1.73s 576.97m // BM_StringPiece 103.74% 1.67s 598.55m // ============================================================================ // Match rate : 1000000 / 1000000 // Second run // ============================================================================ // rocks/benchmark/RocksCopyBenchmark.cpp relative time/iter iters/s // ============================================================================ // BM_StringCopy 611.99ms 1.63 // BM_StringPiece 203.76% 300.35ms 3.33 // ============================================================================ // Match rate : 1000000 / 1000000 ``` Test Plan: Unit tests Reviewers: sdong, igor, anthony, yhchiang, rven Reviewed By: rven Subscribers: dhruba, lovro, adsharma Differential Revision: https://reviews.facebook.net/D48999
9 years ago
virtual bool IsKeyPinned() const override {
// memtable data is always pinned
return true;
}
private:
DynamicBloom* bloom_;
const SliceTransform* const prefix_extractor_;
MemTableRep::Iterator* iter_;
bool valid_;
bool arena_mode_;
// No copying allowed
MemTableIterator(const MemTableIterator&);
void operator=(const MemTableIterator&);
};
InternalIterator* MemTable::NewIterator(const ReadOptions& read_options,
Arena* arena) {
assert(arena != nullptr);
auto mem = arena->AllocateAligned(sizeof(MemTableIterator));
return new (mem) MemTableIterator(*this, read_options, arena);
}
port::RWMutex* MemTable::GetLock(const Slice& key) {
static murmur_hash hash;
return &locks_[hash(key) % locks_.size()];
}
uint64_t MemTable::ApproximateSize(const Slice& start_ikey,
const Slice& end_ikey) {
uint64_t entry_count = table_->ApproximateNumEntries(start_ikey, end_ikey);
if (entry_count == 0) {
return 0;
}
uint64_t n = num_entries_.load(std::memory_order_relaxed);
if (n == 0) {
return 0;
}
if (entry_count > n) {
// table_->ApproximateNumEntries() is just an estimate so it can be larger
// than actual entries we have. Cap it to entries we have to limit the
// inaccuracy.
entry_count = n;
}
uint64_t data_size = data_size_.load(std::memory_order_relaxed);
return entry_count * (data_size / n);
}
void MemTable::Add(SequenceNumber s, ValueType type,
const Slice& key, /* user key */
support for concurrent adds to memtable Summary: This diff adds support for concurrent adds to the skiplist memtable implementations. Memory allocation is made thread-safe by the addition of a spinlock, with small per-core buffers to avoid contention. Concurrent memtable writes are made via an additional method and don't impose a performance overhead on the non-concurrent case, so parallelism can be selected on a per-batch basis. Write thread synchronization is an increasing bottleneck for higher levels of concurrency, so this diff adds --enable_write_thread_adaptive_yield (default off). This feature causes threads joining a write batch group to spin for a short time (default 100 usec) using sched_yield, rather than going to sleep on a mutex. If the timing of the yield calls indicates that another thread has actually run during the yield then spinning is avoided. This option improves performance for concurrent situations even without parallel adds, although it has the potential to increase CPU usage (and the heuristic adaptation is not yet mature). Parallel writes are not currently compatible with inplace updates, update callbacks, or delete filtering. Enable it with --allow_concurrent_memtable_write (and --enable_write_thread_adaptive_yield). Parallel memtable writes are performance neutral when there is no actual parallelism, and in my experiments (SSD server-class Linux and varying contention and key sizes for fillrandom) they are always a performance win when there is more than one thread. Statistics are updated earlier in the write path, dropping the number of DB mutex acquisitions from 2 to 1 for almost all cases. This diff was motivated and inspired by Yahoo's cLSM work. It is more conservative than cLSM: RocksDB's write batch group leader role is preserved (along with all of the existing flush and write throttling logic) and concurrent writers are blocked until all memtable insertions have completed and the sequence number has been advanced, to preserve linearizability. My test config is "db_bench -benchmarks=fillrandom -threads=$T -batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T -level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999 -disable_auto_compactions --max_write_buffer_number=8 -max_background_flushes=8 --disable_wal --write_buffer_size=160000000 --block_size=16384 --allow_concurrent_memtable_write" on a two-socket Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1 thread I get ~440Kops/sec. Peak performance for 1 socket (numactl -N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance across both sockets happens at 30 threads, and is ~900Kops/sec, although with fewer threads there is less performance loss when the system has background work. Test Plan: 1. concurrent stress tests for InlineSkipList and DynamicBloom 2. make clean; make check 3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench 4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench 5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench 6. make clean; OPT=-DROCKSDB_LITE make check 7. verify no perf regressions when disabled Reviewers: igor, sdong Reviewed By: sdong Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba Differential Revision: https://reviews.facebook.net/D50589
10 years ago
const Slice& value, bool allow_concurrent) {
// Format of an entry is concatenation of:
// key_size : varint32 of internal_key.size()
// key bytes : char[internal_key.size()]
// value_size : varint32 of value.size()
// value bytes : char[value.size()]
uint32_t key_size = static_cast<uint32_t>(key.size());
uint32_t val_size = static_cast<uint32_t>(value.size());
uint32_t internal_key_size = key_size + 8;
const uint32_t encoded_len = VarintLength(internal_key_size) +
internal_key_size + VarintLength(val_size) +
val_size;
char* buf = nullptr;
KeyHandle handle = table_->Allocate(encoded_len, &buf);
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
10 years ago
char* p = EncodeVarint32(buf, internal_key_size);
memcpy(p, key.data(), key_size);
p += key_size;
uint64_t packed = PackSequenceAndType(s, type);
EncodeFixed64(p, packed);
p += 8;
p = EncodeVarint32(p, val_size);
memcpy(p, value.data(), val_size);
assert((unsigned)(p + val_size - buf) == (unsigned)encoded_len);
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
10 years ago
if (!allow_concurrent) {
table_->Insert(handle);
// this is a bit ugly, but is the way to avoid locked instructions
// when incrementing an atomic
num_entries_.store(num_entries_.load(std::memory_order_relaxed) + 1,
std::memory_order_relaxed);
data_size_.store(data_size_.load(std::memory_order_relaxed) + encoded_len,
std::memory_order_relaxed);
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
10 years ago
if (type == kTypeDeletion) {
num_deletes_.store(num_deletes_.load(std::memory_order_relaxed) + 1,
std::memory_order_relaxed);
}
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
10 years ago
if (prefix_bloom_) {
assert(prefix_extractor_);
prefix_bloom_->Add(prefix_extractor_->Transform(key));
}
support for concurrent adds to memtable Summary: This diff adds support for concurrent adds to the skiplist memtable implementations. Memory allocation is made thread-safe by the addition of a spinlock, with small per-core buffers to avoid contention. Concurrent memtable writes are made via an additional method and don't impose a performance overhead on the non-concurrent case, so parallelism can be selected on a per-batch basis. Write thread synchronization is an increasing bottleneck for higher levels of concurrency, so this diff adds --enable_write_thread_adaptive_yield (default off). This feature causes threads joining a write batch group to spin for a short time (default 100 usec) using sched_yield, rather than going to sleep on a mutex. If the timing of the yield calls indicates that another thread has actually run during the yield then spinning is avoided. This option improves performance for concurrent situations even without parallel adds, although it has the potential to increase CPU usage (and the heuristic adaptation is not yet mature). Parallel writes are not currently compatible with inplace updates, update callbacks, or delete filtering. Enable it with --allow_concurrent_memtable_write (and --enable_write_thread_adaptive_yield). Parallel memtable writes are performance neutral when there is no actual parallelism, and in my experiments (SSD server-class Linux and varying contention and key sizes for fillrandom) they are always a performance win when there is more than one thread. Statistics are updated earlier in the write path, dropping the number of DB mutex acquisitions from 2 to 1 for almost all cases. This diff was motivated and inspired by Yahoo's cLSM work. It is more conservative than cLSM: RocksDB's write batch group leader role is preserved (along with all of the existing flush and write throttling logic) and concurrent writers are blocked until all memtable insertions have completed and the sequence number has been advanced, to preserve linearizability. My test config is "db_bench -benchmarks=fillrandom -threads=$T -batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T -level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999 -disable_auto_compactions --max_write_buffer_number=8 -max_background_flushes=8 --disable_wal --write_buffer_size=160000000 --block_size=16384 --allow_concurrent_memtable_write" on a two-socket Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1 thread I get ~440Kops/sec. Peak performance for 1 socket (numactl -N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance across both sockets happens at 30 threads, and is ~900Kops/sec, although with fewer threads there is less performance loss when the system has background work. Test Plan: 1. concurrent stress tests for InlineSkipList and DynamicBloom 2. make clean; make check 3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench 4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench 5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench 6. make clean; OPT=-DROCKSDB_LITE make check 7. verify no perf regressions when disabled Reviewers: igor, sdong Reviewed By: sdong Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba Differential Revision: https://reviews.facebook.net/D50589
10 years ago
// The first sequence number inserted into the memtable
assert(first_seqno_ == 0 || s > first_seqno_);
if (first_seqno_ == 0) {
first_seqno_.store(s, std::memory_order_relaxed);
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
10 years ago
if (earliest_seqno_ == kMaxSequenceNumber) {
earliest_seqno_.store(GetFirstSequenceNumber(),
std::memory_order_relaxed);
}
assert(first_seqno_.load() >= earliest_seqno_.load());
}
} else {
table_->InsertConcurrently(handle);
num_entries_.fetch_add(1, std::memory_order_relaxed);
data_size_.fetch_add(encoded_len, std::memory_order_relaxed);
if (type == kTypeDeletion) {
num_deletes_.fetch_add(1, std::memory_order_relaxed);
}
if (prefix_bloom_) {
assert(prefix_extractor_);
prefix_bloom_->AddConcurrently(prefix_extractor_->Transform(key));
}
// atomically update first_seqno_ and earliest_seqno_.
uint64_t cur_seq_num = first_seqno_.load(std::memory_order_relaxed);
while ((cur_seq_num == 0 || s < cur_seq_num) &&
!first_seqno_.compare_exchange_weak(cur_seq_num, s)) {
}
uint64_t cur_earliest_seqno =
earliest_seqno_.load(std::memory_order_relaxed);
while (
(cur_earliest_seqno == kMaxSequenceNumber || s < cur_earliest_seqno) &&
!first_seqno_.compare_exchange_weak(cur_earliest_seqno, s)) {
}
}
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
10 years ago
UpdateFlushState();
}
// Callback from MemTable::Get()
namespace {
struct Saver {
Status* status;
const LookupKey* key;
bool* found_final_value; // Is value set correctly? Used by KeyMayExist
bool* merge_in_progress;
std::string* value;
SequenceNumber seq;
const MergeOperator* merge_operator;
// the merge operations encountered;
MergeContext* merge_context;
MemTable* mem;
Logger* logger;
Statistics* statistics;
bool inplace_update_support;
Env* env_;
};
} // namespace
static bool SaveValue(void* arg, const char* entry) {
Saver* s = reinterpret_cast<Saver*>(arg);
MergeContext* merge_context = s->merge_context;
const MergeOperator* merge_operator = s->merge_operator;
assert(s != nullptr && merge_context != nullptr);
// entry format is:
// klength varint32
// userkey char[klength-8]
// tag uint64
// vlength varint32
// value char[vlength]
// Check that it belongs to same user key. We do not check the
// sequence number since the Seek() call above should have skipped
// all entries with overly large sequence numbers.
uint32_t key_length;
const char* key_ptr = GetVarint32Ptr(entry, entry + 5, &key_length);
if (s->mem->GetInternalKeyComparator().user_comparator()->Equal(
Slice(key_ptr, key_length - 8), s->key->user_key())) {
// Correct user key
const uint64_t tag = DecodeFixed64(key_ptr + key_length - 8);
ValueType type;
UnPackSequenceAndType(tag, &s->seq, &type);
switch (type) {
case kTypeValue: {
if (s->inplace_update_support) {
s->mem->GetLock(s->key->user_key())->ReadLock();
}
Slice v = GetLengthPrefixedSlice(key_ptr + key_length);
*(s->status) = Status::OK();
if (*(s->merge_in_progress)) {
assert(merge_operator);
bool merge_success = false;
{
StopWatchNano timer(s->env_, s->statistics != nullptr);
PERF_TIMER_GUARD(merge_operator_time_nanos);
merge_success = merge_operator->FullMerge(
s->key->user_key(), &v, merge_context->GetOperands(), s->value,
s->logger);
RecordTick(s->statistics, MERGE_OPERATION_TOTAL_TIME,
timer.ElapsedNanos());
}
if (!merge_success) {
RecordTick(s->statistics, NUMBER_MERGE_FAILURES);
*(s->status) =
Status::Corruption("Error: Could not perform merge.");
}
} else if (s->value != nullptr) {
s->value->assign(v.data(), v.size());
}
if (s->inplace_update_support) {
s->mem->GetLock(s->key->user_key())->ReadUnlock();
}
*(s->found_final_value) = true;
return false;
}
Support for SingleDelete() Summary: This patch fixes #7460559. It introduces SingleDelete as a new database operation. This operation can be used to delete keys that were never overwritten (no put following another put of the same key). If an overwritten key is single deleted the behavior is undefined. Single deletion of a non-existent key has no effect but multiple consecutive single deletions are not allowed (see limitations). In contrast to the conventional Delete() operation, the deletion entry is removed along with the value when the two are lined up in a compaction. Note: The semantics are similar to @igor's prototype that allowed to have this behavior on the granularity of a column family ( https://reviews.facebook.net/D42093 ). This new patch, however, is more aggressive when it comes to removing tombstones: It removes the SingleDelete together with the value whenever there is no snapshot between them while the older patch only did this when the sequence number of the deletion was older than the earliest snapshot. Most of the complex additions are in the Compaction Iterator, all other changes should be relatively straightforward. The patch also includes basic support for single deletions in db_stress and db_bench. Limitations: - Not compatible with cuckoo hash tables - Single deletions cannot be used in combination with merges and normal deletions on the same key (other keys are not affected by this) - Consecutive single deletions are currently not allowed (and older version of this patch supported this so it could be resurrected if needed) Test Plan: make all check Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor Reviewed By: igor Subscribers: maykov, dhruba, leveldb Differential Revision: https://reviews.facebook.net/D43179
9 years ago
case kTypeDeletion:
case kTypeSingleDeletion: {
if (*(s->merge_in_progress)) {
Support for SingleDelete() Summary: This patch fixes #7460559. It introduces SingleDelete as a new database operation. This operation can be used to delete keys that were never overwritten (no put following another put of the same key). If an overwritten key is single deleted the behavior is undefined. Single deletion of a non-existent key has no effect but multiple consecutive single deletions are not allowed (see limitations). In contrast to the conventional Delete() operation, the deletion entry is removed along with the value when the two are lined up in a compaction. Note: The semantics are similar to @igor's prototype that allowed to have this behavior on the granularity of a column family ( https://reviews.facebook.net/D42093 ). This new patch, however, is more aggressive when it comes to removing tombstones: It removes the SingleDelete together with the value whenever there is no snapshot between them while the older patch only did this when the sequence number of the deletion was older than the earliest snapshot. Most of the complex additions are in the Compaction Iterator, all other changes should be relatively straightforward. The patch also includes basic support for single deletions in db_stress and db_bench. Limitations: - Not compatible with cuckoo hash tables - Single deletions cannot be used in combination with merges and normal deletions on the same key (other keys are not affected by this) - Consecutive single deletions are currently not allowed (and older version of this patch supported this so it could be resurrected if needed) Test Plan: make all check Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor Reviewed By: igor Subscribers: maykov, dhruba, leveldb Differential Revision: https://reviews.facebook.net/D43179
9 years ago
assert(merge_operator != nullptr);
*(s->status) = Status::OK();
bool merge_success = false;
{
StopWatchNano timer(s->env_, s->statistics != nullptr);
PERF_TIMER_GUARD(merge_operator_time_nanos);
merge_success = merge_operator->FullMerge(
s->key->user_key(), nullptr, merge_context->GetOperands(),
s->value, s->logger);
RecordTick(s->statistics, MERGE_OPERATION_TOTAL_TIME,
timer.ElapsedNanos());
}
if (!merge_success) {
RecordTick(s->statistics, NUMBER_MERGE_FAILURES);
*(s->status) =
Status::Corruption("Error: Could not perform merge.");
}
} else {
*(s->status) = Status::NotFound();
}
*(s->found_final_value) = true;
return false;
}
case kTypeMerge: {
if (!merge_operator) {
*(s->status) = Status::InvalidArgument(
"merge_operator is not properly initialized.");
// Normally we continue the loop (return true) when we see a merge
// operand. But in case of an error, we should stop the loop
// immediately and pretend we have found the value to stop further
// seek. Otherwise, the later call will override this error status.
*(s->found_final_value) = true;
return false;
}
Slice v = GetLengthPrefixedSlice(key_ptr + key_length);
*(s->merge_in_progress) = true;
merge_context->PushOperand(v);
return true;
}
default:
assert(false);
return true;
}
}
// s->state could be Corrupt, merge or notfound
return false;
}
bool MemTable::Get(const LookupKey& key, std::string* value, Status* s,
MergeContext* merge_context, SequenceNumber* seq) {
// The sequence number is updated synchronously in version_set.h
if (IsEmpty()) {
// Avoiding recording stats for speed.
return false;
}
PERF_TIMER_GUARD(get_from_memtable_time);
Slice user_key = key.user_key();
bool found_final_value = false;
bool merge_in_progress = s->IsMergeInProgress();
bool const may_contain =
nullptr == prefix_bloom_
? false
: prefix_bloom_->MayContain(prefix_extractor_->Transform(user_key));
if (prefix_bloom_ && !may_contain) {
// iter is null if prefix bloom says the key does not exist
PERF_COUNTER_ADD(bloom_memtable_miss_count, 1);
*seq = kMaxSequenceNumber;
} else {
if (prefix_bloom_) {
PERF_COUNTER_ADD(bloom_memtable_hit_count, 1);
}
Saver saver;
saver.status = s;
saver.found_final_value = &found_final_value;
saver.merge_in_progress = &merge_in_progress;
saver.key = &key;
saver.value = value;
saver.seq = kMaxSequenceNumber;
saver.mem = this;
saver.merge_context = merge_context;
saver.merge_operator = moptions_.merge_operator;
saver.logger = moptions_.info_log;
saver.inplace_update_support = moptions_.inplace_update_support;
saver.statistics = moptions_.statistics;
saver.env_ = env_;
table_->Get(key, &saver, SaveValue);
*seq = saver.seq;
}
[RocksDB] [MergeOperator] The new Merge Interface! Uses merge sequences. Summary: Here are the major changes to the Merge Interface. It has been expanded to handle cases where the MergeOperator is not associative. It does so by stacking up merge operations while scanning through the key history (i.e.: during Get() or Compaction), until a valid Put/Delete/end-of-history is encountered; it then applies all of the merge operations in the correct sequence starting with the base/sentinel value. I have also introduced an "AssociativeMerge" function which allows the user to take advantage of associative merge operations (such as in the case of counters). The implementation will always attempt to merge the operations/operands themselves together when they are encountered, and will resort to the "stacking" method if and only if the "associative-merge" fails. This implementation is conjectured to allow MergeOperator to handle the general case, while still providing the user with the ability to take advantage of certain efficiencies in their own merge-operator / data-structure. NOTE: This is a preliminary diff. This must still go through a lot of review, revision, and testing. Feedback welcome! Test Plan: -This is a preliminary diff. I have only just begun testing/debugging it. -I will be testing this with the existing MergeOperator use-cases and unit-tests (counters, string-append, and redis-lists) -I will be "desk-checking" and walking through the code with the help gdb. -I will find a way of stress-testing the new interface / implementation using db_bench, db_test, merge_test, and/or db_stress. -I will ensure that my tests cover all cases: Get-Memtable, Get-Immutable-Memtable, Get-from-Disk, Iterator-Range-Scan, Flush-Memtable-to-L0, Compaction-L0-L1, Compaction-Ln-L(n+1), Put/Delete found, Put/Delete not-found, end-of-history, end-of-file, etc. -A lot of feedback from the reviewers. Reviewers: haobo, dhruba, zshao, emayanke Reviewed By: haobo CC: leveldb Differential Revision: https://reviews.facebook.net/D11499
12 years ago
// No change to value, since we have not yet found a Put/Delete
if (!found_final_value && merge_in_progress) {
*s = Status::MergeInProgress();
}
PERF_COUNTER_ADD(get_from_memtable_count, 1);
return found_final_value;
}
void MemTable::Update(SequenceNumber seq,
const Slice& key,
const Slice& value) {
LookupKey lkey(key, seq);
Slice mem_key = lkey.memtable_key();
std::unique_ptr<MemTableRep::Iterator> iter(
table_->GetDynamicPrefixIterator());
iter->Seek(lkey.internal_key(), mem_key.data());
if (iter->Valid()) {
// entry format is:
// key_length varint32
// userkey char[klength-8]
// tag uint64
// vlength varint32
// value char[vlength]
// Check that it belongs to same user key. We do not check the
// sequence number since the Seek() call above should have skipped
// all entries with overly large sequence numbers.
const char* entry = iter->key();
uint32_t key_length = 0;
const char* key_ptr = GetVarint32Ptr(entry, entry + 5, &key_length);
if (comparator_.comparator.user_comparator()->Equal(
Slice(key_ptr, key_length - 8), lkey.user_key())) {
// Correct user key
const uint64_t tag = DecodeFixed64(key_ptr + key_length - 8);
ValueType type;
SequenceNumber unused;
UnPackSequenceAndType(tag, &unused, &type);
switch (type) {
case kTypeValue: {
Slice prev_value = GetLengthPrefixedSlice(key_ptr + key_length);
uint32_t prev_size = static_cast<uint32_t>(prev_value.size());
uint32_t new_size = static_cast<uint32_t>(value.size());
// Update value, if new value size <= previous value size
if (new_size <= prev_size ) {
char* p = EncodeVarint32(const_cast<char*>(key_ptr) + key_length,
new_size);
WriteLock wl(GetLock(lkey.user_key()));
memcpy(p, value.data(), value.size());
assert((unsigned)((p + value.size()) - entry) ==
(unsigned)(VarintLength(key_length) + key_length +
VarintLength(value.size()) + value.size()));
return;
}
}
default:
// If the latest value is kTypeDeletion, kTypeMerge or kTypeLogData
// we don't have enough space for update inplace
Add(seq, kTypeValue, key, value);
return;
}
}
}
// key doesn't exist
Add(seq, kTypeValue, key, value);
}
bool MemTable::UpdateCallback(SequenceNumber seq,
const Slice& key,
const Slice& delta) {
LookupKey lkey(key, seq);
Slice memkey = lkey.memtable_key();
std::unique_ptr<MemTableRep::Iterator> iter(
table_->GetDynamicPrefixIterator());
iter->Seek(lkey.internal_key(), memkey.data());
if (iter->Valid()) {
// entry format is:
// key_length varint32
// userkey char[klength-8]
// tag uint64
// vlength varint32
// value char[vlength]
// Check that it belongs to same user key. We do not check the
// sequence number since the Seek() call above should have skipped
// all entries with overly large sequence numbers.
const char* entry = iter->key();
uint32_t key_length = 0;
const char* key_ptr = GetVarint32Ptr(entry, entry + 5, &key_length);
if (comparator_.comparator.user_comparator()->Equal(
Slice(key_ptr, key_length - 8), lkey.user_key())) {
// Correct user key
const uint64_t tag = DecodeFixed64(key_ptr + key_length - 8);
ValueType type;
uint64_t unused;
UnPackSequenceAndType(tag, &unused, &type);
switch (type) {
case kTypeValue: {
Slice prev_value = GetLengthPrefixedSlice(key_ptr + key_length);
uint32_t prev_size = static_cast<uint32_t>(prev_value.size());
char* prev_buffer = const_cast<char*>(prev_value.data());
uint32_t new_prev_size = prev_size;
std::string str_value;
WriteLock wl(GetLock(lkey.user_key()));
auto status = moptions_.inplace_callback(prev_buffer, &new_prev_size,
delta, &str_value);
if (status == UpdateStatus::UPDATED_INPLACE) {
// Value already updated by callback.
assert(new_prev_size <= prev_size);
if (new_prev_size < prev_size) {
// overwrite the new prev_size
char* p = EncodeVarint32(const_cast<char*>(key_ptr) + key_length,
new_prev_size);
if (VarintLength(new_prev_size) < VarintLength(prev_size)) {
// shift the value buffer as well.
memcpy(p, prev_buffer, new_prev_size);
}
}
RecordTick(moptions_.statistics, NUMBER_KEYS_UPDATED);
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
10 years ago
UpdateFlushState();
return true;
} else if (status == UpdateStatus::UPDATED) {
Add(seq, kTypeValue, key, Slice(str_value));
RecordTick(moptions_.statistics, NUMBER_KEYS_WRITTEN);
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
10 years ago
UpdateFlushState();
return true;
} else if (status == UpdateStatus::UPDATE_FAILED) {
// No action required. Return.
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
10 years ago
UpdateFlushState();
return true;
}
}
default:
break;
}
}
}
// If the latest value is not kTypeValue
// or key doesn't exist
return false;
}
size_t MemTable::CountSuccessiveMergeEntries(const LookupKey& key) {
Slice memkey = key.memtable_key();
// A total ordered iterator is costly for some memtablerep (prefix aware
// reps). By passing in the user key, we allow efficient iterator creation.
// The iterator only needs to be ordered within the same user key.
std::unique_ptr<MemTableRep::Iterator> iter(
table_->GetDynamicPrefixIterator());
iter->Seek(key.internal_key(), memkey.data());
size_t num_successive_merges = 0;
for (; iter->Valid(); iter->Next()) {
const char* entry = iter->key();
uint32_t key_length = 0;
const char* iter_key_ptr = GetVarint32Ptr(entry, entry + 5, &key_length);
if (!comparator_.comparator.user_comparator()->Equal(
Slice(iter_key_ptr, key_length - 8), key.user_key())) {
break;
}
const uint64_t tag = DecodeFixed64(iter_key_ptr + key_length - 8);
ValueType type;
uint64_t unused;
UnPackSequenceAndType(tag, &unused, &type);
if (type != kTypeMerge) {
break;
}
++num_successive_merges;
}
return num_successive_merges;
}
void MemTableRep::Get(const LookupKey& k, void* callback_args,
bool (*callback_func)(void* arg, const char* entry)) {
auto iter = GetDynamicPrefixIterator();
for (iter->Seek(k.internal_key(), k.memtable_key().data());
iter->Valid() && callback_func(callback_args, iter->key());
iter->Next()) {
}
}
void MemTable::RefLogContainingPrepSection(uint64_t log) {
assert(log > 0);
auto cur = min_prep_log_referenced_.load();
while ((log < cur || cur == 0) &&
!min_prep_log_referenced_.compare_exchange_strong(cur, log)) {
cur = min_prep_log_referenced_.load();
}
}
uint64_t MemTable::GetMinLogContainingPrepSection() {
return min_prep_log_referenced_.load();
}
} // namespace rocksdb