|
|
|
// 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.
|
|
|
|
// This source code is also licensed under the GPLv2 license found in the
|
|
|
|
// COPYING file in the root directory of this source tree.
|
|
|
|
//
|
|
|
|
// 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.
|
|
|
|
|
|
|
|
#pragma once
|
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
9 years ago
|
|
|
#include <vector>
|
|
|
|
#include "db/write_thread.h"
|
|
|
|
#include "rocksdb/types.h"
|
|
|
|
#include "rocksdb/write_batch.h"
|
|
|
|
#include "rocksdb/db.h"
|
|
|
|
#include "rocksdb/options.h"
|
|
|
|
#include "util/autovector.h"
|
|
|
|
|
|
|
|
namespace rocksdb {
|
|
|
|
|
|
|
|
class MemTable;
|
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
9 years ago
|
|
|
class FlushScheduler;
|
|
|
|
class ColumnFamilyData;
|
|
|
|
|
|
|
|
class ColumnFamilyMemTables {
|
|
|
|
public:
|
|
|
|
virtual ~ColumnFamilyMemTables() {}
|
|
|
|
virtual bool Seek(uint32_t column_family_id) = 0;
|
|
|
|
// returns true if the update to memtable should be ignored
|
|
|
|
// (useful when recovering from log whose updates have already
|
|
|
|
// been processed)
|
|
|
|
virtual uint64_t GetLogNumber() const = 0;
|
|
|
|
virtual MemTable* GetMemTable() const = 0;
|
|
|
|
virtual ColumnFamilyHandle* GetColumnFamilyHandle() = 0;
|
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
9 years ago
|
|
|
virtual ColumnFamilyData* current() { return nullptr; }
|
|
|
|
};
|
|
|
|
|
|
|
|
class ColumnFamilyMemTablesDefault : public ColumnFamilyMemTables {
|
|
|
|
public:
|
|
|
|
explicit ColumnFamilyMemTablesDefault(MemTable* mem)
|
|
|
|
: ok_(false), mem_(mem) {}
|
|
|
|
|
|
|
|
bool Seek(uint32_t column_family_id) override {
|
|
|
|
ok_ = (column_family_id == 0);
|
|
|
|
return ok_;
|
|
|
|
}
|
|
|
|
|
|
|
|
uint64_t GetLogNumber() const override { return 0; }
|
|
|
|
|
|
|
|
MemTable* GetMemTable() const override {
|
|
|
|
assert(ok_);
|
|
|
|
return mem_;
|
|
|
|
}
|
|
|
|
|
|
|
|
ColumnFamilyHandle* GetColumnFamilyHandle() override { return nullptr; }
|
|
|
|
|
|
|
|
private:
|
|
|
|
bool ok_;
|
|
|
|
MemTable* mem_;
|
|
|
|
};
|
|
|
|
|
|
|
|
// WriteBatchInternal provides static methods for manipulating a
|
|
|
|
// WriteBatch that we don't want in the public WriteBatch interface.
|
|
|
|
class WriteBatchInternal {
|
|
|
|
public:
|
|
|
|
|
|
|
|
// WriteBatch header has an 8-byte sequence number followed by a 4-byte count.
|
|
|
|
static const size_t kHeader = 12;
|
|
|
|
|
|
|
|
// WriteBatch methods with column_family_id instead of ColumnFamilyHandle*
|
|
|
|
static Status Put(WriteBatch* batch, uint32_t column_family_id,
|
|
|
|
const Slice& key, const Slice& value);
|
|
|
|
|
|
|
|
static Status Put(WriteBatch* batch, uint32_t column_family_id,
|
|
|
|
const SliceParts& key, const SliceParts& value);
|
|
|
|
|
|
|
|
static Status Delete(WriteBatch* batch, uint32_t column_family_id,
|
|
|
|
const SliceParts& key);
|
|
|
|
|
|
|
|
static Status Delete(WriteBatch* batch, uint32_t column_family_id,
|
|
|
|
const Slice& key);
|
|
|
|
|
|
|
|
static Status SingleDelete(WriteBatch* batch, uint32_t column_family_id,
|
|
|
|
const SliceParts& key);
|
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
|
|
|
|
|
|
|
static Status SingleDelete(WriteBatch* batch, uint32_t column_family_id,
|
|
|
|
const Slice& key);
|
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
|
|
|
|
|
|
|
static Status DeleteRange(WriteBatch* b, uint32_t column_family_id,
|
|
|
|
const Slice& begin_key, const Slice& end_key);
|
|
|
|
|
|
|
|
static Status DeleteRange(WriteBatch* b, uint32_t column_family_id,
|
|
|
|
const SliceParts& begin_key,
|
|
|
|
const SliceParts& end_key);
|
|
|
|
|
|
|
|
static Status Merge(WriteBatch* batch, uint32_t column_family_id,
|
|
|
|
const Slice& key, const Slice& value);
|
|
|
|
|
|
|
|
static Status Merge(WriteBatch* batch, uint32_t column_family_id,
|
|
|
|
const SliceParts& key, const SliceParts& value);
|
|
|
|
|
|
|
|
static Status MarkEndPrepare(WriteBatch* batch, const Slice& xid);
|
Modification of WriteBatch to support two phase commit
Summary: Adds three new WriteBatch data types: Prepare(xid), Commit(xid), Rollback(xid). Prepare(xid) should precede the (single) operation to which is applies. There can obviously be multiple Prepare(xid) markers. There should only be one Rollback(xid) or Commit(xid) marker yet not both. None of this logic is currently enforced and will most likely be implemented further up such as in the memtableinserter. All three markers are similar to PutLogData in that they are writebatch meta-data, ie stored but not counted. All three markers differ from PutLogData in that they will actually be written to disk. As for WriteBatchWithIndex, Prepare, Commit, Rollback are all implemented just as PutLogData and none are tested just as PutLogData.
Test Plan: single unit test in write_batch_test.
Reviewers: hermanlee4, sdong, anthony
Subscribers: leveldb, dhruba, vasilep, andrewkr
Differential Revision: https://reviews.facebook.net/D57867
9 years ago
|
|
|
|
|
|
|
static Status MarkRollback(WriteBatch* batch, const Slice& xid);
|
Modification of WriteBatch to support two phase commit
Summary: Adds three new WriteBatch data types: Prepare(xid), Commit(xid), Rollback(xid). Prepare(xid) should precede the (single) operation to which is applies. There can obviously be multiple Prepare(xid) markers. There should only be one Rollback(xid) or Commit(xid) marker yet not both. None of this logic is currently enforced and will most likely be implemented further up such as in the memtableinserter. All three markers are similar to PutLogData in that they are writebatch meta-data, ie stored but not counted. All three markers differ from PutLogData in that they will actually be written to disk. As for WriteBatchWithIndex, Prepare, Commit, Rollback are all implemented just as PutLogData and none are tested just as PutLogData.
Test Plan: single unit test in write_batch_test.
Reviewers: hermanlee4, sdong, anthony
Subscribers: leveldb, dhruba, vasilep, andrewkr
Differential Revision: https://reviews.facebook.net/D57867
9 years ago
|
|
|
|
|
|
|
static Status MarkCommit(WriteBatch* batch, const Slice& xid);
|
Modification of WriteBatch to support two phase commit
Summary: Adds three new WriteBatch data types: Prepare(xid), Commit(xid), Rollback(xid). Prepare(xid) should precede the (single) operation to which is applies. There can obviously be multiple Prepare(xid) markers. There should only be one Rollback(xid) or Commit(xid) marker yet not both. None of this logic is currently enforced and will most likely be implemented further up such as in the memtableinserter. All three markers are similar to PutLogData in that they are writebatch meta-data, ie stored but not counted. All three markers differ from PutLogData in that they will actually be written to disk. As for WriteBatchWithIndex, Prepare, Commit, Rollback are all implemented just as PutLogData and none are tested just as PutLogData.
Test Plan: single unit test in write_batch_test.
Reviewers: hermanlee4, sdong, anthony
Subscribers: leveldb, dhruba, vasilep, andrewkr
Differential Revision: https://reviews.facebook.net/D57867
9 years ago
|
|
|
|
|
|
|
static Status InsertNoop(WriteBatch* batch);
|
Modification of WriteBatch to support two phase commit
Summary: Adds three new WriteBatch data types: Prepare(xid), Commit(xid), Rollback(xid). Prepare(xid) should precede the (single) operation to which is applies. There can obviously be multiple Prepare(xid) markers. There should only be one Rollback(xid) or Commit(xid) marker yet not both. None of this logic is currently enforced and will most likely be implemented further up such as in the memtableinserter. All three markers are similar to PutLogData in that they are writebatch meta-data, ie stored but not counted. All three markers differ from PutLogData in that they will actually be written to disk. As for WriteBatchWithIndex, Prepare, Commit, Rollback are all implemented just as PutLogData and none are tested just as PutLogData.
Test Plan: single unit test in write_batch_test.
Reviewers: hermanlee4, sdong, anthony
Subscribers: leveldb, dhruba, vasilep, andrewkr
Differential Revision: https://reviews.facebook.net/D57867
9 years ago
|
|
|
|
|
|
|
// Return the number of entries in the batch.
|
|
|
|
static int Count(const WriteBatch* batch);
|
|
|
|
|
|
|
|
// Set the count for the number of entries in the batch.
|
|
|
|
static void SetCount(WriteBatch* batch, int n);
|
|
|
|
|
|
|
|
// Return the seqeunce number for the start of this batch.
|
|
|
|
static SequenceNumber Sequence(const WriteBatch* batch);
|
|
|
|
|
|
|
|
// Store the specified number as the seqeunce number for the start of
|
|
|
|
// this batch.
|
|
|
|
static void SetSequence(WriteBatch* batch, SequenceNumber seq);
|
|
|
|
|
|
|
|
// Returns the offset of the first entry in the batch.
|
|
|
|
// This offset is only valid if the batch is not empty.
|
|
|
|
static size_t GetFirstOffset(WriteBatch* batch);
|
|
|
|
|
|
|
|
static Slice Contents(const WriteBatch* batch) {
|
|
|
|
return Slice(batch->rep_);
|
|
|
|
}
|
|
|
|
|
|
|
|
static size_t ByteSize(const WriteBatch* batch) {
|
|
|
|
return batch->rep_.size();
|
|
|
|
}
|
|
|
|
|
|
|
|
static Status SetContents(WriteBatch* batch, const Slice& contents);
|
|
|
|
|
|
|
|
// Inserts batches[i] into memtable, for i in 0..num_batches-1 inclusive.
|
|
|
|
//
|
|
|
|
// If ignore_missing_column_families == true. WriteBatch
|
|
|
|
// referencing non-existing column family will be ignored.
|
|
|
|
// If ignore_missing_column_families == false, processing of the
|
|
|
|
// batches will be stopped if a reference is found to a non-existing
|
|
|
|
// column family and InvalidArgument() will be returned. The writes
|
|
|
|
// in batches may be only partially applied at that point.
|
|
|
|
//
|
|
|
|
// If log_number is non-zero, the memtable will be updated only if
|
|
|
|
// memtables->GetLogNumber() >= log_number.
|
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
9 years ago
|
|
|
//
|
|
|
|
// If flush_scheduler is non-null, it will be invoked if the memtable
|
|
|
|
// should be flushed.
|
|
|
|
//
|
|
|
|
// Under concurrent use, the caller is responsible for making sure that
|
|
|
|
// the memtables object itself is thread-local.
|
|
|
|
static Status InsertInto(WriteThread::WriteGroup& write_group,
|
|
|
|
SequenceNumber sequence,
|
|
|
|
ColumnFamilyMemTables* memtables,
|
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
9 years ago
|
|
|
FlushScheduler* flush_scheduler,
|
|
|
|
bool ignore_missing_column_families = false,
|
|
|
|
uint64_t log_number = 0, DB* db = 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
9 years ago
|
|
|
bool concurrent_memtable_writes = false);
|
|
|
|
|
|
|
|
// Convenience form of InsertInto when you have only one batch
|
[rocksdb] Recovery path sequence miscount fix
Summary:
Consider the following WAL with 4 batch entries prefixed with their sequence at time of memtable insert.
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(a)]
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(b)]
[4: COMMIT(a)]
[7: COMMIT(b)]
The first two batches do not consume any sequence numbers so are both prefixed with seq=1.
For 2pc commit, memtable insertion takes place before COMMIT batch is written to WAL.
We can see that sequence number consumption takes place between WAL entries giving us the seemingly sparse sequence prefix for WAL entries.
This is a valid WAL.
Because with 2PC markers one WriteBatch points to another batch containing its inserts a writebatch can consume more or less sequence numbers than the number of sequence consuming entries that it contains.
We can see that, given the entries in the WAL, 6 sequence ids were consumed. Yet on recovery the maximum sequence consumed would be 7 + 3 (the number of sequence numbers consumed by COMMIT(b))
So, now upon recovery we must track the actual consumption of sequence numbers.
In the provided scenario there will be no sequence gaps, but it is possible to produce a sequence gap. This should not be a problem though. correct?
Test Plan: provided test.
Reviewers: sdong
Subscribers: andrewkr, leveldb, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D57645
9 years ago
|
|
|
// last_seq_used returns the last sequnce number used in a MemTable insert
|
|
|
|
static Status InsertInto(const WriteBatch* batch,
|
|
|
|
ColumnFamilyMemTables* memtables,
|
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
9 years ago
|
|
|
FlushScheduler* flush_scheduler,
|
|
|
|
bool ignore_missing_column_families = false,
|
|
|
|
uint64_t log_number = 0, DB* db = nullptr,
|
[rocksdb] Recovery path sequence miscount fix
Summary:
Consider the following WAL with 4 batch entries prefixed with their sequence at time of memtable insert.
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(a)]
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(b)]
[4: COMMIT(a)]
[7: COMMIT(b)]
The first two batches do not consume any sequence numbers so are both prefixed with seq=1.
For 2pc commit, memtable insertion takes place before COMMIT batch is written to WAL.
We can see that sequence number consumption takes place between WAL entries giving us the seemingly sparse sequence prefix for WAL entries.
This is a valid WAL.
Because with 2PC markers one WriteBatch points to another batch containing its inserts a writebatch can consume more or less sequence numbers than the number of sequence consuming entries that it contains.
We can see that, given the entries in the WAL, 6 sequence ids were consumed. Yet on recovery the maximum sequence consumed would be 7 + 3 (the number of sequence numbers consumed by COMMIT(b))
So, now upon recovery we must track the actual consumption of sequence numbers.
In the provided scenario there will be no sequence gaps, but it is possible to produce a sequence gap. This should not be a problem though. correct?
Test Plan: provided test.
Reviewers: sdong
Subscribers: andrewkr, leveldb, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D57645
9 years ago
|
|
|
bool concurrent_memtable_writes = false,
|
Ignore stale logs while restarting DBs
Summary:
Stale log files can be deleted out of order. This can happen for various reasons. One of the reason is that no data is ever inserted to a column family and we have an optimization to update its log number, but not all the old log files are cleaned up (the case shown in the unit tests added). It can also happen when we simply delete multiple log files out of order.
This causes data corruption because we simply increase seqID after processing the next row and we may end up with writing data with smaller seqID than what is already flushed to memtables.
In DB recovery, for the oldest files we are replaying, if there it contains no data for any column family, we ignore the sequence IDs in the file.
Test Plan: Add two unit tests that fail without the fix.
Reviewers: IslamAbdelRahman, igor, yiwu
Reviewed By: yiwu
Subscribers: hermanlee4, yoshinorim, leveldb, andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D60891
8 years ago
|
|
|
SequenceNumber* last_seq_used = nullptr,
|
|
|
|
bool* has_valid_writes = nullptr);
|
[rocksdb] Recovery path sequence miscount fix
Summary:
Consider the following WAL with 4 batch entries prefixed with their sequence at time of memtable insert.
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(a)]
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(b)]
[4: COMMIT(a)]
[7: COMMIT(b)]
The first two batches do not consume any sequence numbers so are both prefixed with seq=1.
For 2pc commit, memtable insertion takes place before COMMIT batch is written to WAL.
We can see that sequence number consumption takes place between WAL entries giving us the seemingly sparse sequence prefix for WAL entries.
This is a valid WAL.
Because with 2PC markers one WriteBatch points to another batch containing its inserts a writebatch can consume more or less sequence numbers than the number of sequence consuming entries that it contains.
We can see that, given the entries in the WAL, 6 sequence ids were consumed. Yet on recovery the maximum sequence consumed would be 7 + 3 (the number of sequence numbers consumed by COMMIT(b))
So, now upon recovery we must track the actual consumption of sequence numbers.
In the provided scenario there will be no sequence gaps, but it is possible to produce a sequence gap. This should not be a problem though. correct?
Test Plan: provided test.
Reviewers: sdong
Subscribers: andrewkr, leveldb, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D57645
9 years ago
|
|
|
|
|
|
|
static Status InsertInto(WriteThread::Writer* writer,
|
|
|
|
ColumnFamilyMemTables* memtables,
|
|
|
|
FlushScheduler* flush_scheduler,
|
|
|
|
bool ignore_missing_column_families = false,
|
|
|
|
uint64_t log_number = 0, DB* db = nullptr,
|
|
|
|
bool concurrent_memtable_writes = false);
|
|
|
|
|
|
|
|
static Status Append(WriteBatch* dst, const WriteBatch* src,
|
|
|
|
const bool WAL_only = false);
|
|
|
|
|
|
|
|
// Returns the byte size of appending a WriteBatch with ByteSize
|
|
|
|
// leftByteSize and a WriteBatch with ByteSize rightByteSize
|
|
|
|
static size_t AppendedByteSize(size_t leftByteSize, size_t rightByteSize);
|
|
|
|
};
|
|
|
|
|
|
|
|
// LocalSavePoint is similar to a scope guard
|
|
|
|
class LocalSavePoint {
|
|
|
|
public:
|
|
|
|
explicit LocalSavePoint(WriteBatch* batch)
|
|
|
|
: batch_(batch),
|
|
|
|
savepoint_(batch->GetDataSize(), batch->Count(),
|
|
|
|
batch->content_flags_.load(std::memory_order_relaxed))
|
|
|
|
#ifndef NDEBUG
|
|
|
|
,
|
|
|
|
committed_(false)
|
|
|
|
#endif
|
|
|
|
{
|
|
|
|
}
|
|
|
|
|
|
|
|
#ifndef NDEBUG
|
|
|
|
~LocalSavePoint() { assert(committed_); }
|
|
|
|
#endif
|
|
|
|
Status commit() {
|
|
|
|
#ifndef NDEBUG
|
|
|
|
committed_ = true;
|
|
|
|
#endif
|
|
|
|
if (batch_->max_bytes_ && batch_->rep_.size() > batch_->max_bytes_) {
|
|
|
|
batch_->rep_.resize(savepoint_.size);
|
|
|
|
WriteBatchInternal::SetCount(batch_, savepoint_.count);
|
|
|
|
batch_->content_flags_.store(savepoint_.content_flags,
|
|
|
|
std::memory_order_relaxed);
|
|
|
|
return Status::MemoryLimit();
|
|
|
|
}
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
private:
|
|
|
|
WriteBatch* batch_;
|
|
|
|
SavePoint savepoint_;
|
|
|
|
#ifndef NDEBUG
|
|
|
|
bool committed_;
|
|
|
|
#endif
|
|
|
|
};
|
|
|
|
|
|
|
|
} // namespace rocksdb
|