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/db_impl.cc

6473 lines
225 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/db_impl.h"
#ifndef __STDC_FORMAT_MACROS
#define __STDC_FORMAT_MACROS
#endif
#include <inttypes.h>
#include <stdint.h>
#ifdef OS_SOLARIS
#include <alloca.h>
#endif
#ifdef ROCKSDB_JEMALLOC
#include "jemalloc/jemalloc.h"
#endif
#include <algorithm>
#include <climits>
#include <cstdio>
#include <map>
#include <set>
#include <stdexcept>
#include <string>
#include <unordered_map>
#include <unordered_set>
#include <utility>
#include <vector>
#include "db/auto_roll_logger.h"
#include "db/builder.h"
#include "db/compaction_job.h"
#include "db/db_info_dumper.h"
#include "db/db_iter.h"
#include "db/dbformat.h"
#include "db/event_helpers.h"
#include "db/external_sst_file_ingestion_job.h"
#include "db/filename.h"
#include "db/flush_job.h"
#include "db/forward_iterator.h"
#include "db/job_context.h"
#include "db/log_reader.h"
#include "db/log_writer.h"
#include "db/managed_iterator.h"
#include "db/memtable.h"
#include "db/memtable_list.h"
#include "db/merge_context.h"
#include "db/merge_helper.h"
#include "db/table_cache.h"
#include "db/table_properties_collector.h"
#include "db/transaction_log_impl.h"
#include "db/version_set.h"
#include "db/write_batch_internal.h"
#include "db/write_callback.h"
#include "db/xfunc_test_points.h"
#include "memtable/hash_linklist_rep.h"
#include "memtable/hash_skiplist_rep.h"
#include "port/likely.h"
#include "port/port.h"
#include "rocksdb/cache.h"
#include "rocksdb/compaction_filter.h"
#include "rocksdb/db.h"
#include "rocksdb/env.h"
#include "rocksdb/merge_operator.h"
#include "rocksdb/statistics.h"
#include "rocksdb/status.h"
#include "rocksdb/table.h"
#include "rocksdb/version.h"
#include "rocksdb/wal_filter.h"
#include "rocksdb/write_buffer_manager.h"
#include "table/block.h"
#include "table/block_based_table_factory.h"
#include "table/merger.h"
#include "table/table_builder.h"
#include "table/two_level_iterator.h"
#include "util/autovector.h"
#include "util/build_version.h"
#include "util/cf_options.h"
#include "util/coding.h"
#include "util/compression.h"
#include "util/crc32c.h"
#include "util/file_reader_writer.h"
#include "util/file_util.h"
#include "util/iostats_context_imp.h"
#include "util/log_buffer.h"
#include "util/logging.h"
#include "util/mutexlock.h"
#include "util/options_helper.h"
#include "util/options_parser.h"
#include "util/perf_context_imp.h"
#include "util/sst_file_manager_impl.h"
#include "util/stop_watch.h"
#include "util/string_util.h"
#include "util/sync_point.h"
#include "util/thread_status_updater.h"
#include "util/thread_status_util.h"
#include "util/xfunc.h"
namespace rocksdb {
const std::string kDefaultColumnFamilyName("default");
void DumpRocksDBBuildVersion(Logger * log);
struct DBImpl::WriteContext {
autovector<SuperVersion*> superversions_to_free_;
Support saving history in memtable_list Summary: For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit. After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much. This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit). However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions. Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here. Reviewers: sdong, rven, igor Reviewed By: igor Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D37443
9 years ago
autovector<MemTable*> memtables_to_free_;
~WriteContext() {
for (auto& sv : superversions_to_free_) {
delete sv;
}
Support saving history in memtable_list Summary: For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit. After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much. This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit). However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions. Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here. Reviewers: sdong, rven, igor Reviewed By: igor Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D37443
9 years ago
for (auto& m : memtables_to_free_) {
delete m;
}
}
};
Options SanitizeOptions(const std::string& dbname,
const Options& src) {
auto db_options = SanitizeOptions(dbname, DBOptions(src));
ImmutableDBOptions immutable_db_options(db_options);
auto cf_options =
SanitizeOptions(immutable_db_options, ColumnFamilyOptions(src));
return Options(db_options, cf_options);
}
DBOptions SanitizeOptions(const std::string& dbname, const DBOptions& src) {
DBOptions result(src);
// result.max_open_files means an "infinite" open files.
if (result.max_open_files != -1) {
int max_max_open_files = port::GetMaxOpenFiles();
if (max_max_open_files == -1) {
max_max_open_files = 1000000;
}
ClipToRange(&result.max_open_files, 20, max_max_open_files);
}
if (result.info_log == nullptr) {
Status s = CreateLoggerFromOptions(dbname, result, &result.info_log);
if (!s.ok()) {
// No place suitable for logging
result.info_log = nullptr;
}
}
if (!result.write_buffer_manager) {
result.write_buffer_manager.reset(
new WriteBufferManager(result.db_write_buffer_size));
}
if (result.base_background_compactions == -1) {
result.base_background_compactions = result.max_background_compactions;
}
if (result.base_background_compactions > result.max_background_compactions) {
result.base_background_compactions = result.max_background_compactions;
}
result.env->IncBackgroundThreadsIfNeeded(src.max_background_compactions,
Env::Priority::LOW);
result.env->IncBackgroundThreadsIfNeeded(src.max_background_flushes,
Env::Priority::HIGH);
if (result.rate_limiter.get() != nullptr) {
if (result.bytes_per_sync == 0) {
result.bytes_per_sync = 1024 * 1024;
}
}
if (result.WAL_ttl_seconds > 0 || result.WAL_size_limit_MB > 0) {
result.recycle_log_file_num = false;
}
if (result.recycle_log_file_num &&
(result.wal_recovery_mode == WALRecoveryMode::kPointInTimeRecovery ||
result.wal_recovery_mode == WALRecoveryMode::kAbsoluteConsistency)) {
// kPointInTimeRecovery is indistinguishable from
// kTolerateCorruptedTailRecords in recycle mode since we define
// the "end" of the log as the first corrupt record we encounter.
// kAbsoluteConsistency doesn't make sense because even a clean
// shutdown leaves old junk at the end of the log file.
result.wal_recovery_mode = WALRecoveryMode::kTolerateCorruptedTailRecords;
}
if (result.wal_dir.empty()) {
// Use dbname as default
result.wal_dir = dbname;
}
if (result.wal_dir.back() == '/') {
result.wal_dir = result.wal_dir.substr(0, result.wal_dir.size() - 1);
}
if (result.db_paths.size() == 0) {
result.db_paths.emplace_back(dbname, std::numeric_limits<uint64_t>::max());
}
if (result.compaction_readahead_size > 0) {
result.new_table_reader_for_compaction_inputs = true;
}
// Force flush on DB open if 2PC is enabled, since with 2PC we have no
// guarantee that consecutive log files have consecutive sequence id, which
// make recovery complicated.
if (result.allow_2pc) {
result.avoid_flush_during_recovery = false;
}
return result;
}
namespace {
Status SanitizeOptionsByTable(
const DBOptions& db_opts,
const std::vector<ColumnFamilyDescriptor>& column_families) {
Status s;
for (auto cf : column_families) {
s = cf.options.table_factory->SanitizeOptions(db_opts, cf.options);
if (!s.ok()) {
return s;
}
}
return Status::OK();
}
static Status ValidateOptions(
const DBOptions& db_options,
const std::vector<ColumnFamilyDescriptor>& column_families) {
Status s;
for (auto& cfd : column_families) {
s = CheckCompressionSupported(cfd.options);
if (s.ok() && db_options.allow_concurrent_memtable_write) {
s = CheckConcurrentWritesSupported(cfd.options);
}
if (!s.ok()) {
return s;
}
if (db_options.db_paths.size() > 1) {
if ((cfd.options.compaction_style != kCompactionStyleUniversal) &&
(cfd.options.compaction_style != kCompactionStyleLevel)) {
return Status::NotSupported(
"More than one DB paths are only supported in "
"universal and level compaction styles. ");
}
}
}
if (db_options.db_paths.size() > 4) {
return Status::NotSupported(
"More than four DB paths are not supported yet. ");
}
if (db_options.allow_mmap_reads && !db_options.allow_os_buffer) {
// Protect against assert in PosixMMapReadableFile constructor
return Status::NotSupported(
"If memory mapped reads (allow_mmap_reads) are enabled "
"then os caching (allow_os_buffer) must also be enabled. ");
}
return Status::OK();
}
CompressionType GetCompressionFlush(
const ImmutableCFOptions& ioptions,
const MutableCFOptions& mutable_cf_options) {
// Compressing memtable flushes might not help unless the sequential load
// optimization is used for leveled compaction. Otherwise the CPU and
// latency overhead is not offset by saving much space.
if (ioptions.compaction_style == kCompactionStyleUniversal) {
if (ioptions.compaction_options_universal.compression_size_percent < 0) {
return mutable_cf_options.compression;
} else {
return kNoCompression;
}
} else if (!ioptions.compression_per_level.empty()) {
// For leveled compress when min_level_to_compress != 0.
return ioptions.compression_per_level[0];
} else {
return mutable_cf_options.compression;
}
}
void DumpSupportInfo(Logger* logger) {
Header(logger, "Compression algorithms supported:");
Header(logger, "\tSnappy supported: %d", Snappy_Supported());
Header(logger, "\tZlib supported: %d", Zlib_Supported());
Header(logger, "\tBzip supported: %d", BZip2_Supported());
Header(logger, "\tLZ4 supported: %d", LZ4_Supported());
Header(logger, "\tZSTD supported: %d", ZSTD_Supported());
Header(logger, "Fast CRC32 supported: %d", crc32c::IsFastCrc32Supported());
}
} // namespace
DBImpl::DBImpl(const DBOptions& options, const std::string& dbname)
: env_(options.env),
dbname_(dbname),
initial_db_options_(SanitizeOptions(dbname, options)),
immutable_db_options_(initial_db_options_),
mutable_db_options_(initial_db_options_),
stats_(immutable_db_options_.statistics.get()),
db_lock_(nullptr),
mutex_(stats_, env_, DB_MUTEX_WAIT_MICROS,
immutable_db_options_.use_adaptive_mutex),
shutting_down_(false),
bg_cv_(&mutex_),
logfile_number_(0),
log_dir_synced_(false),
log_empty_(true),
default_cf_handle_(nullptr),
log_sync_cv_(&mutex_),
total_log_size_(0),
max_total_in_memory_state_(0),
is_snapshot_supported_(true),
write_buffer_manager_(immutable_db_options_.write_buffer_manager.get()),
write_thread_(immutable_db_options_.enable_write_thread_adaptive_yield
? immutable_db_options_.write_thread_max_yield_usec
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
: 0,
immutable_db_options_.write_thread_slow_yield_usec),
write_controller_(immutable_db_options_.delayed_write_rate),
last_batch_group_size_(0),
Rewritten system for scheduling background work Summary: When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue. The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction. Here are the performance results: Command: ./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000 --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333 Before the patch: fillrandom : 26.950 micros/op 37105 ops/sec; 4.1 MB/s After the patch: fillrandom : 17.404 micros/op 57456 ops/sec; 6.4 MB/s Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got: fillrandom : 7.590 micros/op 131758 ops/sec; 14.6 MB/s Test Plan: make check two stress tests: Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 max_background_flushes=0, to verify that this case also works correctly ./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: ljin, rven, yhchiang, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30123
10 years ago
unscheduled_flushes_(0),
unscheduled_compactions_(0),
bg_compaction_scheduled_(0),
num_running_compactions_(0),
bg_flush_scheduled_(0),
num_running_flushes_(0),
bg_purge_scheduled_(0),
disable_delete_obsolete_files_(0),
Speed up FindObsoleteFiles() Summary: There are two versions of FindObsoleteFiles(): * full scan, which is executed every 6 hours (and it's terribly slow) * no full scan, which is executed every time a background process finishes and iterator is deleted This diff is optimizing the second case (no full scan). Here's what we do before the diff: * Get the list of obsolete files (files with ref==0). Some files in obsolete_files set might actually be live. * Get the list of live files to avoid deleting files that are live. * Delete files that are in obsolete_files and not in live_files. After this diff: * The only files with ref==0 that are still live are files that have been part of move compaction. Don't include moved files in obsolete_files. * Get the list of obsolete files (which exclude moved files). * No need to get the list of live files, since all files in obsolete_files need to be deleted. I'll post the benchmark results, but you can get the feel of it here: https://reviews.facebook.net/D30123 This depends on D30123. P.S. We should do full scan only in failure scenarios, not every 6 hours. I'll do this in a follow-up diff. Test Plan: One new unit test. Made sure that unit test fails if we don't have a `if (!f->moved)` safeguard in ~Version. make check Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: yhchiang, rven, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30249
10 years ago
delete_obsolete_files_next_run_(
env_->NowMicros() +
immutable_db_options_.delete_obsolete_files_period_micros),
last_stats_dump_time_microsec_(0),
next_job_id_(1),
has_unpersisted_data_(false),
env_options_(BuildDBOptions(immutable_db_options_, mutable_db_options_)),
num_running_ingest_file_(0),
#ifndef ROCKSDB_LITE
wal_manager_(immutable_db_options_, env_options_),
#endif // ROCKSDB_LITE
event_logger_(immutable_db_options_.info_log.get()),
bg_work_paused_(0),
bg_compaction_paused_(0),
refitting_level_(false),
opened_successfully_(false) {
env_->GetAbsolutePath(dbname, &db_absolute_path_);
// Reserve ten files or so for other uses and give the rest to TableCache.
// Give a large number for setting of "infinite" open files.
const int table_cache_size = (immutable_db_options_.max_open_files == -1)
? 4194304
: immutable_db_options_.max_open_files - 10;
table_cache_ = NewLRUCache(table_cache_size,
immutable_db_options_.table_cache_numshardbits);
versions_.reset(new VersionSet(dbname_, &immutable_db_options_, env_options_,
table_cache_.get(), write_buffer_manager_,
&write_controller_));
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
column_family_memtables_.reset(
new ColumnFamilyMemTablesImpl(versions_->GetColumnFamilySet()));
DumpRocksDBBuildVersion(immutable_db_options_.info_log.get());
DumpDBFileSummary(immutable_db_options_, dbname_);
immutable_db_options_.Dump(immutable_db_options_.info_log.get());
mutable_db_options_.Dump(immutable_db_options_.info_log.get());
DumpSupportInfo(immutable_db_options_.info_log.get());
}
// Will lock the mutex_, will wait for completion if wait is true
void DBImpl::CancelAllBackgroundWork(bool wait) {
InstrumentedMutexLock l(&mutex_);
if (!shutting_down_.load(std::memory_order_acquire) &&
has_unpersisted_data_) {
for (auto cfd : *versions_->GetColumnFamilySet()) {
if (!cfd->IsDropped() && !cfd->mem()->IsEmpty()) {
cfd->Ref();
mutex_.Unlock();
FlushMemTable(cfd, FlushOptions());
mutex_.Lock();
cfd->Unref();
}
}
versions_->GetColumnFamilySet()->FreeDeadColumnFamilies();
}
Persist data during user initiated shutdown Summary: Move the manual memtable flush for databases containing data that has bypassed the WAL from DBImpl's destructor to CancleAllBackgroundWork(). CancelAllBackgroundWork() is a publicly exposed API which allows async operations performed by background threads to be disabled on a database. In effect, this places the database into a "shutdown" state in advance of calling the database object's destructor. No compactions or flushing of SST files can occur once a call to this API completes. When writes are issued to a database with WriteOptions::disableWAL set to true, DBImpl::has_unpersisted_data_ is set so that memtables can be flushed when the database object is destroyed. If CancelAllBackgroundWork() has been called prior to DBImpl's destructor, this flush operation is not possible and is skipped, causing unnecessary loss of data. Since CancelAllBackgroundWork() is already invoked by DBImpl's destructor in order to perform the thread join portion of its cleanup processing, moving the manual memtable flush to CancelAllBackgroundWork() ensures data is persisted regardless of client behavior. Test Plan: Write an amount of data that will not cause a memtable flush to a rocksdb database with all writes marked with WriteOptions::disableWAL. Properly "close" the database. Reopen database and verify that the data was persisted. Reviewers: IslamAbdelRahman, yiwu, yoshinorim, sdong Reviewed By: sdong Subscribers: andrewkr, dhruba Differential Revision: https://reviews.facebook.net/D62277
8 years ago
shutting_down_.store(true, std::memory_order_release);
bg_cv_.SignalAll();
if (!wait) {
return;
}
// Wait for background work to finish
while (bg_compaction_scheduled_ || bg_flush_scheduled_) {
bg_cv_.Wait();
}
}
DBImpl::~DBImpl() {
// CancelAllBackgroundWork called with false means we just set the shutdown
// marker. After this we do a variant of the waiting and unschedule work
// (to consider: moving all the waiting into CancelAllBackgroundWork(true))
CancelAllBackgroundWork(false);
int compactions_unscheduled = env_->UnSchedule(this, Env::Priority::LOW);
int flushes_unscheduled = env_->UnSchedule(this, Env::Priority::HIGH);
mutex_.Lock();
bg_compaction_scheduled_ -= compactions_unscheduled;
bg_flush_scheduled_ -= flushes_unscheduled;
// Wait for background work to finish
while (bg_compaction_scheduled_ || bg_flush_scheduled_ ||
bg_purge_scheduled_) {
TEST_SYNC_POINT("DBImpl::~DBImpl:WaitJob");
bg_cv_.Wait();
}
EraseThreadStatusDbInfo();
flush_scheduler_.Clear();
Rewritten system for scheduling background work Summary: When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue. The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction. Here are the performance results: Command: ./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000 --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333 Before the patch: fillrandom : 26.950 micros/op 37105 ops/sec; 4.1 MB/s After the patch: fillrandom : 17.404 micros/op 57456 ops/sec; 6.4 MB/s Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got: fillrandom : 7.590 micros/op 131758 ops/sec; 14.6 MB/s Test Plan: make check two stress tests: Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 max_background_flushes=0, to verify that this case also works correctly ./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: ljin, rven, yhchiang, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30123
10 years ago
while (!flush_queue_.empty()) {
auto cfd = PopFirstFromFlushQueue();
if (cfd->Unref()) {
delete cfd;
}
}
while (!compaction_queue_.empty()) {
auto cfd = PopFirstFromCompactionQueue();
if (cfd->Unref()) {
delete cfd;
}
}
if (default_cf_handle_ != nullptr) {
// we need to delete handle outside of lock because it does its own locking
mutex_.Unlock();
delete default_cf_handle_;
mutex_.Lock();
}
// Clean up obsolete files due to SuperVersion release.
// (1) Need to delete to obsolete files before closing because RepairDB()
// scans all existing files in the file system and builds manifest file.
// Keeping obsolete files confuses the repair process.
// (2) Need to check if we Open()/Recover() the DB successfully before
// deleting because if VersionSet recover fails (may be due to corrupted
// manifest file), it is not able to identify live files correctly. As a
// result, all "live" files can get deleted by accident. However, corrupted
// manifest is recoverable by RepairDB().
if (opened_successfully_) {
JobContext job_context(next_job_id_.fetch_add(1));
FindObsoleteFiles(&job_context, true);
mutex_.Unlock();
// manifest number starting from 2
job_context.manifest_file_number = 1;
if (job_context.HaveSomethingToDelete()) {
PurgeObsoleteFiles(job_context);
}
job_context.Clean();
mutex_.Lock();
}
for (auto l : logs_to_free_) {
delete l;
}
for (auto& log : logs_) {
log.ClearWriter();
}
logs_.clear();
// Table cache may have table handles holding blocks from the block cache.
// We need to release them before the block cache is destroyed. The block
// cache may be destroyed inside versions_.reset(), when column family data
// list is destroyed, so leaving handles in table cache after
// versions_.reset() may cause issues.
// Here we clean all unreferenced handles in table cache.
// Now we assume all user queries have finished, so only version set itself
// can possibly hold the blocks from block cache. After releasing unreferenced
// handles here, only handles held by version set left and inside
// versions_.reset(), we will release them. There, we need to make sure every
// time a handle is released, we erase it from the cache too. By doing that,
// we can guarantee that after versions_.reset(), table cache is empty
// so the cache can be safely destroyed.
table_cache_->EraseUnRefEntries();
for (auto& txn_entry : recovered_transactions_) {
delete txn_entry.second;
}
// versions need to be destroyed before table_cache since it can hold
// references to table_cache.
versions_.reset();
mutex_.Unlock();
if (db_lock_ != nullptr) {
env_->UnlockFile(db_lock_);
}
LogFlush(immutable_db_options_.info_log);
}
Status DBImpl::NewDB() {
VersionEdit new_db;
new_db.SetLogNumber(0);
new_db.SetNextFile(2);
new_db.SetLastSequence(0);
Status s;
Log(InfoLogLevel::INFO_LEVEL, immutable_db_options_.info_log,
"Creating manifest 1 \n");
const std::string manifest = DescriptorFileName(dbname_, 1);
{
unique_ptr<WritableFile> file;
EnvOptions env_options = env_->OptimizeForManifestWrite(env_options_);
s = NewWritableFile(env_, manifest, &file, env_options);
if (!s.ok()) {
return s;
}
file->SetPreallocationBlockSize(
immutable_db_options_.manifest_preallocation_size);
unique_ptr<WritableFileWriter> file_writer(
new WritableFileWriter(std::move(file), env_options));
log::Writer log(std::move(file_writer), 0, false);
std::string record;
new_db.EncodeTo(&record);
s = log.AddRecord(record);
if (s.ok()) {
s = SyncManifest(env_, &immutable_db_options_, log.file());
}
}
if (s.ok()) {
// Make "CURRENT" file that points to the new manifest file.
s = SetCurrentFile(env_, dbname_, 1, directories_.GetDbDir());
} else {
env_->DeleteFile(manifest);
}
return s;
}
void DBImpl::MaybeIgnoreError(Status* s) const {
if (s->ok() || immutable_db_options_.paranoid_checks) {
// No change needed
} else {
Log(InfoLogLevel::WARN_LEVEL, immutable_db_options_.info_log,
"Ignoring error %s", s->ToString().c_str());
*s = Status::OK();
}
}
const Status DBImpl::CreateArchivalDirectory() {
if (immutable_db_options_.wal_ttl_seconds > 0 ||
immutable_db_options_.wal_size_limit_mb > 0) {
std::string archivalPath = ArchivalDirectory(immutable_db_options_.wal_dir);
return env_->CreateDirIfMissing(archivalPath);
}
return Status::OK();
}
void DBImpl::PrintStatistics() {
auto dbstats = immutable_db_options_.statistics.get();
if (dbstats) {
Log(InfoLogLevel::WARN_LEVEL, immutable_db_options_.info_log,
"STATISTICS:\n %s", dbstats->ToString().c_str());
}
}
#ifndef ROCKSDB_LITE
#ifdef ROCKSDB_JEMALLOC
typedef struct {
char* cur;
char* end;
} MallocStatus;
static void GetJemallocStatus(void* mstat_arg, const char* status) {
MallocStatus* mstat = reinterpret_cast<MallocStatus*>(mstat_arg);
size_t status_len = status ? strlen(status) : 0;
size_t buf_size = (size_t)(mstat->end - mstat->cur);
if (!status_len || status_len > buf_size) {
return;
}
snprintf(mstat->cur, buf_size, "%s", status);
mstat->cur += status_len;
}
#endif // ROCKSDB_JEMALLOC
static void DumpMallocStats(std::string* stats) {
#ifdef ROCKSDB_JEMALLOC
MallocStatus mstat;
const uint kMallocStatusLen = 1000000;
std::unique_ptr<char> buf{new char[kMallocStatusLen + 1]};
mstat.cur = buf.get();
mstat.end = buf.get() + kMallocStatusLen;
malloc_stats_print(GetJemallocStatus, &mstat, "");
stats->append(buf.get());
#endif // ROCKSDB_JEMALLOC
}
#endif // !ROCKSDB_LITE
void DBImpl::MaybeDumpStats() {
if (immutable_db_options_.stats_dump_period_sec == 0) return;
const uint64_t now_micros = env_->NowMicros();
if (last_stats_dump_time_microsec_ +
immutable_db_options_.stats_dump_period_sec * 1000000 <=
now_micros) {
// Multiple threads could race in here simultaneously.
// However, the last one will update last_stats_dump_time_microsec_
// atomically. We could see more than one dump during one dump
// period in rare cases.
last_stats_dump_time_microsec_ = now_micros;
make internal stats independent of statistics Summary: also make it aware of column family output from db_bench ``` ** Compaction Stats [default] ** Level Files Size(MB) Score Read(GB) Rn(GB) Rnp1(GB) Write(GB) Wnew(GB) RW-Amp W-Amp Rd(MB/s) Wr(MB/s) Rn(cnt) Rnp1(cnt) Wnp1(cnt) Wnew(cnt) Comp(sec) Comp(cnt) Avg(sec) Stall(sec) Stall(cnt) Avg(ms) ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ L0 14 956 0.9 0.0 0.0 0.0 2.7 2.7 0.0 0.0 0.0 111.6 0 0 0 0 24 40 0.612 75.20 492387 0.15 L1 21 2001 2.0 5.7 2.0 3.7 5.3 1.6 5.4 2.6 71.2 65.7 31 43 55 12 82 2 41.242 43.72 41183 1.06 L2 217 18974 1.9 16.5 2.0 14.4 15.1 0.7 15.6 7.4 70.1 64.3 17 182 185 3 241 16 15.052 0.00 0 0.00 L3 1641 188245 1.8 9.1 1.1 8.0 8.5 0.5 15.4 7.4 61.3 57.2 9 75 76 1 152 9 16.887 0.00 0 0.00 L4 4447 449025 0.4 13.4 4.8 8.6 9.1 0.5 4.7 1.9 77.8 52.7 38 79 100 21 176 38 4.639 0.00 0 0.00 Sum 6340 659201 0.0 44.7 10.0 34.7 40.6 6.0 32.0 15.2 67.7 61.6 95 379 416 37 676 105 6.439 118.91 533570 0.22 Int 0 0 0.0 1.2 0.4 0.8 1.3 0.5 5.2 2.7 59.1 65.6 3 7 9 2 20 10 2.003 0.00 0 0.00 Stalls(secs): 75.197 level0_slowdown, 0.000 level0_numfiles, 0.000 memtable_compaction, 43.717 leveln_slowdown Stalls(count): 492387 level0_slowdown, 0 level0_numfiles, 0 memtable_compaction, 41183 leveln_slowdown ** DB Stats ** Uptime(secs): 202.1 total, 13.5 interval Cumulative writes: 6291456 writes, 6291456 batches, 1.0 writes per batch, 4.90 ingest GB Cumulative WAL: 6291456 writes, 6291456 syncs, 1.00 writes per sync, 4.90 GB written Interval writes: 1048576 writes, 1048576 batches, 1.0 writes per batch, 836.0 ingest MB Interval WAL: 1048576 writes, 1048576 syncs, 1.00 writes per sync, 0.82 MB written Test Plan: ran it Reviewers: sdong, yhchiang, igor Reviewed By: igor Subscribers: leveldb Differential Revision: https://reviews.facebook.net/D19917
10 years ago
#ifndef ROCKSDB_LITE
const DBPropertyInfo* cf_property_info =
GetPropertyInfo(DB::Properties::kCFStats);
assert(cf_property_info != nullptr);
const DBPropertyInfo* db_property_info =
GetPropertyInfo(DB::Properties::kDBStats);
assert(db_property_info != nullptr);
std::string stats;
make internal stats independent of statistics Summary: also make it aware of column family output from db_bench ``` ** Compaction Stats [default] ** Level Files Size(MB) Score Read(GB) Rn(GB) Rnp1(GB) Write(GB) Wnew(GB) RW-Amp W-Amp Rd(MB/s) Wr(MB/s) Rn(cnt) Rnp1(cnt) Wnp1(cnt) Wnew(cnt) Comp(sec) Comp(cnt) Avg(sec) Stall(sec) Stall(cnt) Avg(ms) ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ L0 14 956 0.9 0.0 0.0 0.0 2.7 2.7 0.0 0.0 0.0 111.6 0 0 0 0 24 40 0.612 75.20 492387 0.15 L1 21 2001 2.0 5.7 2.0 3.7 5.3 1.6 5.4 2.6 71.2 65.7 31 43 55 12 82 2 41.242 43.72 41183 1.06 L2 217 18974 1.9 16.5 2.0 14.4 15.1 0.7 15.6 7.4 70.1 64.3 17 182 185 3 241 16 15.052 0.00 0 0.00 L3 1641 188245 1.8 9.1 1.1 8.0 8.5 0.5 15.4 7.4 61.3 57.2 9 75 76 1 152 9 16.887 0.00 0 0.00 L4 4447 449025 0.4 13.4 4.8 8.6 9.1 0.5 4.7 1.9 77.8 52.7 38 79 100 21 176 38 4.639 0.00 0 0.00 Sum 6340 659201 0.0 44.7 10.0 34.7 40.6 6.0 32.0 15.2 67.7 61.6 95 379 416 37 676 105 6.439 118.91 533570 0.22 Int 0 0 0.0 1.2 0.4 0.8 1.3 0.5 5.2 2.7 59.1 65.6 3 7 9 2 20 10 2.003 0.00 0 0.00 Stalls(secs): 75.197 level0_slowdown, 0.000 level0_numfiles, 0.000 memtable_compaction, 43.717 leveln_slowdown Stalls(count): 492387 level0_slowdown, 0 level0_numfiles, 0 memtable_compaction, 41183 leveln_slowdown ** DB Stats ** Uptime(secs): 202.1 total, 13.5 interval Cumulative writes: 6291456 writes, 6291456 batches, 1.0 writes per batch, 4.90 ingest GB Cumulative WAL: 6291456 writes, 6291456 syncs, 1.00 writes per sync, 4.90 GB written Interval writes: 1048576 writes, 1048576 batches, 1.0 writes per batch, 836.0 ingest MB Interval WAL: 1048576 writes, 1048576 syncs, 1.00 writes per sync, 0.82 MB written Test Plan: ran it Reviewers: sdong, yhchiang, igor Reviewed By: igor Subscribers: leveldb Differential Revision: https://reviews.facebook.net/D19917
10 years ago
{
InstrumentedMutexLock l(&mutex_);
make internal stats independent of statistics Summary: also make it aware of column family output from db_bench ``` ** Compaction Stats [default] ** Level Files Size(MB) Score Read(GB) Rn(GB) Rnp1(GB) Write(GB) Wnew(GB) RW-Amp W-Amp Rd(MB/s) Wr(MB/s) Rn(cnt) Rnp1(cnt) Wnp1(cnt) Wnew(cnt) Comp(sec) Comp(cnt) Avg(sec) Stall(sec) Stall(cnt) Avg(ms) ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ L0 14 956 0.9 0.0 0.0 0.0 2.7 2.7 0.0 0.0 0.0 111.6 0 0 0 0 24 40 0.612 75.20 492387 0.15 L1 21 2001 2.0 5.7 2.0 3.7 5.3 1.6 5.4 2.6 71.2 65.7 31 43 55 12 82 2 41.242 43.72 41183 1.06 L2 217 18974 1.9 16.5 2.0 14.4 15.1 0.7 15.6 7.4 70.1 64.3 17 182 185 3 241 16 15.052 0.00 0 0.00 L3 1641 188245 1.8 9.1 1.1 8.0 8.5 0.5 15.4 7.4 61.3 57.2 9 75 76 1 152 9 16.887 0.00 0 0.00 L4 4447 449025 0.4 13.4 4.8 8.6 9.1 0.5 4.7 1.9 77.8 52.7 38 79 100 21 176 38 4.639 0.00 0 0.00 Sum 6340 659201 0.0 44.7 10.0 34.7 40.6 6.0 32.0 15.2 67.7 61.6 95 379 416 37 676 105 6.439 118.91 533570 0.22 Int 0 0 0.0 1.2 0.4 0.8 1.3 0.5 5.2 2.7 59.1 65.6 3 7 9 2 20 10 2.003 0.00 0 0.00 Stalls(secs): 75.197 level0_slowdown, 0.000 level0_numfiles, 0.000 memtable_compaction, 43.717 leveln_slowdown Stalls(count): 492387 level0_slowdown, 0 level0_numfiles, 0 memtable_compaction, 41183 leveln_slowdown ** DB Stats ** Uptime(secs): 202.1 total, 13.5 interval Cumulative writes: 6291456 writes, 6291456 batches, 1.0 writes per batch, 4.90 ingest GB Cumulative WAL: 6291456 writes, 6291456 syncs, 1.00 writes per sync, 4.90 GB written Interval writes: 1048576 writes, 1048576 batches, 1.0 writes per batch, 836.0 ingest MB Interval WAL: 1048576 writes, 1048576 syncs, 1.00 writes per sync, 0.82 MB written Test Plan: ran it Reviewers: sdong, yhchiang, igor Reviewed By: igor Subscribers: leveldb Differential Revision: https://reviews.facebook.net/D19917
10 years ago
for (auto cfd : *versions_->GetColumnFamilySet()) {
cfd->internal_stats()->GetStringProperty(
*cf_property_info, DB::Properties::kCFStats, &stats);
make internal stats independent of statistics Summary: also make it aware of column family output from db_bench ``` ** Compaction Stats [default] ** Level Files Size(MB) Score Read(GB) Rn(GB) Rnp1(GB) Write(GB) Wnew(GB) RW-Amp W-Amp Rd(MB/s) Wr(MB/s) Rn(cnt) Rnp1(cnt) Wnp1(cnt) Wnew(cnt) Comp(sec) Comp(cnt) Avg(sec) Stall(sec) Stall(cnt) Avg(ms) ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ L0 14 956 0.9 0.0 0.0 0.0 2.7 2.7 0.0 0.0 0.0 111.6 0 0 0 0 24 40 0.612 75.20 492387 0.15 L1 21 2001 2.0 5.7 2.0 3.7 5.3 1.6 5.4 2.6 71.2 65.7 31 43 55 12 82 2 41.242 43.72 41183 1.06 L2 217 18974 1.9 16.5 2.0 14.4 15.1 0.7 15.6 7.4 70.1 64.3 17 182 185 3 241 16 15.052 0.00 0 0.00 L3 1641 188245 1.8 9.1 1.1 8.0 8.5 0.5 15.4 7.4 61.3 57.2 9 75 76 1 152 9 16.887 0.00 0 0.00 L4 4447 449025 0.4 13.4 4.8 8.6 9.1 0.5 4.7 1.9 77.8 52.7 38 79 100 21 176 38 4.639 0.00 0 0.00 Sum 6340 659201 0.0 44.7 10.0 34.7 40.6 6.0 32.0 15.2 67.7 61.6 95 379 416 37 676 105 6.439 118.91 533570 0.22 Int 0 0 0.0 1.2 0.4 0.8 1.3 0.5 5.2 2.7 59.1 65.6 3 7 9 2 20 10 2.003 0.00 0 0.00 Stalls(secs): 75.197 level0_slowdown, 0.000 level0_numfiles, 0.000 memtable_compaction, 43.717 leveln_slowdown Stalls(count): 492387 level0_slowdown, 0 level0_numfiles, 0 memtable_compaction, 41183 leveln_slowdown ** DB Stats ** Uptime(secs): 202.1 total, 13.5 interval Cumulative writes: 6291456 writes, 6291456 batches, 1.0 writes per batch, 4.90 ingest GB Cumulative WAL: 6291456 writes, 6291456 syncs, 1.00 writes per sync, 4.90 GB written Interval writes: 1048576 writes, 1048576 batches, 1.0 writes per batch, 836.0 ingest MB Interval WAL: 1048576 writes, 1048576 syncs, 1.00 writes per sync, 0.82 MB written Test Plan: ran it Reviewers: sdong, yhchiang, igor Reviewed By: igor Subscribers: leveldb Differential Revision: https://reviews.facebook.net/D19917
10 years ago
}
default_cf_internal_stats_->GetStringProperty(
*db_property_info, DB::Properties::kDBStats, &stats);
make internal stats independent of statistics Summary: also make it aware of column family output from db_bench ``` ** Compaction Stats [default] ** Level Files Size(MB) Score Read(GB) Rn(GB) Rnp1(GB) Write(GB) Wnew(GB) RW-Amp W-Amp Rd(MB/s) Wr(MB/s) Rn(cnt) Rnp1(cnt) Wnp1(cnt) Wnew(cnt) Comp(sec) Comp(cnt) Avg(sec) Stall(sec) Stall(cnt) Avg(ms) ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ L0 14 956 0.9 0.0 0.0 0.0 2.7 2.7 0.0 0.0 0.0 111.6 0 0 0 0 24 40 0.612 75.20 492387 0.15 L1 21 2001 2.0 5.7 2.0 3.7 5.3 1.6 5.4 2.6 71.2 65.7 31 43 55 12 82 2 41.242 43.72 41183 1.06 L2 217 18974 1.9 16.5 2.0 14.4 15.1 0.7 15.6 7.4 70.1 64.3 17 182 185 3 241 16 15.052 0.00 0 0.00 L3 1641 188245 1.8 9.1 1.1 8.0 8.5 0.5 15.4 7.4 61.3 57.2 9 75 76 1 152 9 16.887 0.00 0 0.00 L4 4447 449025 0.4 13.4 4.8 8.6 9.1 0.5 4.7 1.9 77.8 52.7 38 79 100 21 176 38 4.639 0.00 0 0.00 Sum 6340 659201 0.0 44.7 10.0 34.7 40.6 6.0 32.0 15.2 67.7 61.6 95 379 416 37 676 105 6.439 118.91 533570 0.22 Int 0 0 0.0 1.2 0.4 0.8 1.3 0.5 5.2 2.7 59.1 65.6 3 7 9 2 20 10 2.003 0.00 0 0.00 Stalls(secs): 75.197 level0_slowdown, 0.000 level0_numfiles, 0.000 memtable_compaction, 43.717 leveln_slowdown Stalls(count): 492387 level0_slowdown, 0 level0_numfiles, 0 memtable_compaction, 41183 leveln_slowdown ** DB Stats ** Uptime(secs): 202.1 total, 13.5 interval Cumulative writes: 6291456 writes, 6291456 batches, 1.0 writes per batch, 4.90 ingest GB Cumulative WAL: 6291456 writes, 6291456 syncs, 1.00 writes per sync, 4.90 GB written Interval writes: 1048576 writes, 1048576 batches, 1.0 writes per batch, 836.0 ingest MB Interval WAL: 1048576 writes, 1048576 syncs, 1.00 writes per sync, 0.82 MB written Test Plan: ran it Reviewers: sdong, yhchiang, igor Reviewed By: igor Subscribers: leveldb Differential Revision: https://reviews.facebook.net/D19917
10 years ago
}
Log(InfoLogLevel::WARN_LEVEL, immutable_db_options_.info_log,
"------- DUMPING STATS -------");
Log(InfoLogLevel::WARN_LEVEL, immutable_db_options_.info_log, "%s",
stats.c_str());
if (immutable_db_options_.dump_malloc_stats) {
stats.clear();
DumpMallocStats(&stats);
if (!stats.empty()) {
Log(InfoLogLevel::WARN_LEVEL, immutable_db_options_.info_log,
"------- Malloc STATS -------");
Log(InfoLogLevel::WARN_LEVEL, immutable_db_options_.info_log, "%s",
stats.c_str());
}
}
#endif // !ROCKSDB_LITE
make internal stats independent of statistics Summary: also make it aware of column family output from db_bench ``` ** Compaction Stats [default] ** Level Files Size(MB) Score Read(GB) Rn(GB) Rnp1(GB) Write(GB) Wnew(GB) RW-Amp W-Amp Rd(MB/s) Wr(MB/s) Rn(cnt) Rnp1(cnt) Wnp1(cnt) Wnew(cnt) Comp(sec) Comp(cnt) Avg(sec) Stall(sec) Stall(cnt) Avg(ms) ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ L0 14 956 0.9 0.0 0.0 0.0 2.7 2.7 0.0 0.0 0.0 111.6 0 0 0 0 24 40 0.612 75.20 492387 0.15 L1 21 2001 2.0 5.7 2.0 3.7 5.3 1.6 5.4 2.6 71.2 65.7 31 43 55 12 82 2 41.242 43.72 41183 1.06 L2 217 18974 1.9 16.5 2.0 14.4 15.1 0.7 15.6 7.4 70.1 64.3 17 182 185 3 241 16 15.052 0.00 0 0.00 L3 1641 188245 1.8 9.1 1.1 8.0 8.5 0.5 15.4 7.4 61.3 57.2 9 75 76 1 152 9 16.887 0.00 0 0.00 L4 4447 449025 0.4 13.4 4.8 8.6 9.1 0.5 4.7 1.9 77.8 52.7 38 79 100 21 176 38 4.639 0.00 0 0.00 Sum 6340 659201 0.0 44.7 10.0 34.7 40.6 6.0 32.0 15.2 67.7 61.6 95 379 416 37 676 105 6.439 118.91 533570 0.22 Int 0 0 0.0 1.2 0.4 0.8 1.3 0.5 5.2 2.7 59.1 65.6 3 7 9 2 20 10 2.003 0.00 0 0.00 Stalls(secs): 75.197 level0_slowdown, 0.000 level0_numfiles, 0.000 memtable_compaction, 43.717 leveln_slowdown Stalls(count): 492387 level0_slowdown, 0 level0_numfiles, 0 memtable_compaction, 41183 leveln_slowdown ** DB Stats ** Uptime(secs): 202.1 total, 13.5 interval Cumulative writes: 6291456 writes, 6291456 batches, 1.0 writes per batch, 4.90 ingest GB Cumulative WAL: 6291456 writes, 6291456 syncs, 1.00 writes per sync, 4.90 GB written Interval writes: 1048576 writes, 1048576 batches, 1.0 writes per batch, 836.0 ingest MB Interval WAL: 1048576 writes, 1048576 syncs, 1.00 writes per sync, 0.82 MB written Test Plan: ran it Reviewers: sdong, yhchiang, igor Reviewed By: igor Subscribers: leveldb Differential Revision: https://reviews.facebook.net/D19917
10 years ago
PrintStatistics();
}
}
uint64_t DBImpl::FindMinPrepLogReferencedByMemTable() {
uint64_t min_log = 0;
// we must look through the memtables for two phase transactions
// that have been committed but not yet flushed
for (auto loop_cfd : *versions_->GetColumnFamilySet()) {
if (loop_cfd->IsDropped()) {
continue;
}
auto log = loop_cfd->imm()->GetMinLogContainingPrepSection();
if (log > 0 && (min_log == 0 || log < min_log)) {
min_log = log;
}
log = loop_cfd->mem()->GetMinLogContainingPrepSection();
if (log > 0 && (min_log == 0 || log < min_log)) {
min_log = log;
}
}
return min_log;
}
void DBImpl::MarkLogAsHavingPrepSectionFlushed(uint64_t log) {
assert(log != 0);
std::lock_guard<std::mutex> lock(prep_heap_mutex_);
auto it = prepared_section_completed_.find(log);
assert(it != prepared_section_completed_.end());
it->second += 1;
}
void DBImpl::MarkLogAsContainingPrepSection(uint64_t log) {
assert(log != 0);
std::lock_guard<std::mutex> lock(prep_heap_mutex_);
min_log_with_prep_.push(log);
auto it = prepared_section_completed_.find(log);
if (it == prepared_section_completed_.end()) {
prepared_section_completed_[log] = 0;
}
}
uint64_t DBImpl::FindMinLogContainingOutstandingPrep() {
8 years ago
std::lock_guard<std::mutex> lock(prep_heap_mutex_);
uint64_t min_log = 0;
// first we look in the prepared heap where we keep
// track of transactions that have been prepared (written to WAL)
// but not yet committed.
while (!min_log_with_prep_.empty()) {
min_log = min_log_with_prep_.top();
auto it = prepared_section_completed_.find(min_log);
// value was marked as 'deleted' from heap
if (it != prepared_section_completed_.end() && it->second > 0) {
it->second -= 1;
min_log_with_prep_.pop();
// back to squere one...
min_log = 0;
continue;
} else {
// found a valid value
break;
}
}
return min_log;
}
void DBImpl::ScheduleBgLogWriterClose(JobContext* job_context) {
if (!job_context->logs_to_free.empty()) {
for (auto l : job_context->logs_to_free) {
AddToLogsToFreeQueue(l);
}
job_context->logs_to_free.clear();
SchedulePurge();
}
}
// * Returns the list of live files in 'sst_live'
Speed up FindObsoleteFiles() Summary: There are two versions of FindObsoleteFiles(): * full scan, which is executed every 6 hours (and it's terribly slow) * no full scan, which is executed every time a background process finishes and iterator is deleted This diff is optimizing the second case (no full scan). Here's what we do before the diff: * Get the list of obsolete files (files with ref==0). Some files in obsolete_files set might actually be live. * Get the list of live files to avoid deleting files that are live. * Delete files that are in obsolete_files and not in live_files. After this diff: * The only files with ref==0 that are still live are files that have been part of move compaction. Don't include moved files in obsolete_files. * Get the list of obsolete files (which exclude moved files). * No need to get the list of live files, since all files in obsolete_files need to be deleted. I'll post the benchmark results, but you can get the feel of it here: https://reviews.facebook.net/D30123 This depends on D30123. P.S. We should do full scan only in failure scenarios, not every 6 hours. I'll do this in a follow-up diff. Test Plan: One new unit test. Made sure that unit test fails if we don't have a `if (!f->moved)` safeguard in ~Version. make check Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: yhchiang, rven, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30249
10 years ago
// If it's doing full scan:
// * Returns the list of all files in the filesystem in
// 'full_scan_candidate_files'.
Speed up FindObsoleteFiles() Summary: There are two versions of FindObsoleteFiles(): * full scan, which is executed every 6 hours (and it's terribly slow) * no full scan, which is executed every time a background process finishes and iterator is deleted This diff is optimizing the second case (no full scan). Here's what we do before the diff: * Get the list of obsolete files (files with ref==0). Some files in obsolete_files set might actually be live. * Get the list of live files to avoid deleting files that are live. * Delete files that are in obsolete_files and not in live_files. After this diff: * The only files with ref==0 that are still live are files that have been part of move compaction. Don't include moved files in obsolete_files. * Get the list of obsolete files (which exclude moved files). * No need to get the list of live files, since all files in obsolete_files need to be deleted. I'll post the benchmark results, but you can get the feel of it here: https://reviews.facebook.net/D30123 This depends on D30123. P.S. We should do full scan only in failure scenarios, not every 6 hours. I'll do this in a follow-up diff. Test Plan: One new unit test. Made sure that unit test fails if we don't have a `if (!f->moved)` safeguard in ~Version. make check Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: yhchiang, rven, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30249
10 years ago
// Otherwise, gets obsolete files from VersionSet.
// no_full_scan = true -- never do the full scan using GetChildren()
// force = false -- don't force the full scan, except every
// immutable_db_options_.delete_obsolete_files_period_micros
// force = true -- force the full scan
void DBImpl::FindObsoleteFiles(JobContext* job_context, bool force,
bool no_full_scan) {
mutex_.AssertHeld();
// if deletion is disabled, do nothing
if (disable_delete_obsolete_files_ > 0) {
return;
}
bool doing_the_full_scan = false;
// logic for figurint out if we're doing the full scan
if (no_full_scan) {
doing_the_full_scan = false;
} else if (force ||
immutable_db_options_.delete_obsolete_files_period_micros == 0) {
doing_the_full_scan = true;
} else {
const uint64_t now_micros = env_->NowMicros();
Speed up FindObsoleteFiles() Summary: There are two versions of FindObsoleteFiles(): * full scan, which is executed every 6 hours (and it's terribly slow) * no full scan, which is executed every time a background process finishes and iterator is deleted This diff is optimizing the second case (no full scan). Here's what we do before the diff: * Get the list of obsolete files (files with ref==0). Some files in obsolete_files set might actually be live. * Get the list of live files to avoid deleting files that are live. * Delete files that are in obsolete_files and not in live_files. After this diff: * The only files with ref==0 that are still live are files that have been part of move compaction. Don't include moved files in obsolete_files. * Get the list of obsolete files (which exclude moved files). * No need to get the list of live files, since all files in obsolete_files need to be deleted. I'll post the benchmark results, but you can get the feel of it here: https://reviews.facebook.net/D30123 This depends on D30123. P.S. We should do full scan only in failure scenarios, not every 6 hours. I'll do this in a follow-up diff. Test Plan: One new unit test. Made sure that unit test fails if we don't have a `if (!f->moved)` safeguard in ~Version. make check Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: yhchiang, rven, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30249
10 years ago
if (delete_obsolete_files_next_run_ < now_micros) {
doing_the_full_scan = true;
Speed up FindObsoleteFiles() Summary: There are two versions of FindObsoleteFiles(): * full scan, which is executed every 6 hours (and it's terribly slow) * no full scan, which is executed every time a background process finishes and iterator is deleted This diff is optimizing the second case (no full scan). Here's what we do before the diff: * Get the list of obsolete files (files with ref==0). Some files in obsolete_files set might actually be live. * Get the list of live files to avoid deleting files that are live. * Delete files that are in obsolete_files and not in live_files. After this diff: * The only files with ref==0 that are still live are files that have been part of move compaction. Don't include moved files in obsolete_files. * Get the list of obsolete files (which exclude moved files). * No need to get the list of live files, since all files in obsolete_files need to be deleted. I'll post the benchmark results, but you can get the feel of it here: https://reviews.facebook.net/D30123 This depends on D30123. P.S. We should do full scan only in failure scenarios, not every 6 hours. I'll do this in a follow-up diff. Test Plan: One new unit test. Made sure that unit test fails if we don't have a `if (!f->moved)` safeguard in ~Version. make check Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: yhchiang, rven, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30249
10 years ago
delete_obsolete_files_next_run_ =
now_micros +
immutable_db_options_.delete_obsolete_files_period_micros;
}
}
// don't delete files that might be currently written to from compaction
// threads
// Since job_context->min_pending_output is set, until file scan finishes,
// mutex_ cannot be released. Otherwise, we might see no min_pending_output
// here but later find newer generated unfinalized files while scannint.
if (!pending_outputs_.empty()) {
job_context->min_pending_output = *pending_outputs_.begin();
} else {
// delete all of them
job_context->min_pending_output = std::numeric_limits<uint64_t>::max();
}
// Get obsolete files. This function will also update the list of
// pending files in VersionSet().
versions_->GetObsoleteFiles(&job_context->sst_delete_files,
&job_context->manifest_delete_files,
job_context->min_pending_output);
// store the current filenum, lognum, etc
job_context->manifest_file_number = versions_->manifest_file_number();
job_context->pending_manifest_file_number =
versions_->pending_manifest_file_number();
job_context->log_number = versions_->MinLogNumber();
if (allow_2pc()) {
// if are 2pc we must consider logs containing prepared
// sections of outstanding transactions.
//
// We must check min logs with outstanding prep before we check
// logs referneces by memtables because a log referenced by the
// first data structure could transition to the second under us.
//
// TODO(horuff): iterating over all column families under db mutex.
// should find more optimial solution
auto min_log_in_prep_heap = FindMinLogContainingOutstandingPrep();
if (min_log_in_prep_heap != 0 &&
min_log_in_prep_heap < job_context->log_number) {
job_context->log_number = min_log_in_prep_heap;
}
auto min_log_refed_by_mem = FindMinPrepLogReferencedByMemTable();
if (min_log_refed_by_mem != 0 &&
min_log_refed_by_mem < job_context->log_number) {
job_context->log_number = min_log_refed_by_mem;
}
}
job_context->prev_log_number = versions_->prev_log_number();
versions_->AddLiveFiles(&job_context->sst_live);
if (doing_the_full_scan) {
for (size_t path_id = 0; path_id < immutable_db_options_.db_paths.size();
path_id++) {
// set of all files in the directory. We'll exclude files that are still
// alive in the subsequent processings.
std::vector<std::string> files;
env_->GetChildren(immutable_db_options_.db_paths[path_id].path,
&files); // Ignore errors
for (std::string file : files) {
// TODO(icanadi) clean up this mess to avoid having one-off "/" prefixes
job_context->full_scan_candidate_files.emplace_back(
"/" + file, static_cast<uint32_t>(path_id));
}
}
// Add log files in wal_dir
if (immutable_db_options_.wal_dir != dbname_) {
std::vector<std::string> log_files;
env_->GetChildren(immutable_db_options_.wal_dir,
&log_files); // Ignore errors
for (std::string log_file : log_files) {
job_context->full_scan_candidate_files.emplace_back(log_file, 0);
}
}
// Add info log files in db_log_dir
if (!immutable_db_options_.db_log_dir.empty() &&
immutable_db_options_.db_log_dir != dbname_) {
std::vector<std::string> info_log_files;
// Ignore errors
env_->GetChildren(immutable_db_options_.db_log_dir, &info_log_files);
for (std::string log_file : info_log_files) {
job_context->full_scan_candidate_files.emplace_back(log_file, 0);
}
}
}
// logs_ is empty when called during recovery, in which case there can't yet
// be any tracked obsolete logs
if (!alive_log_files_.empty() && !logs_.empty()) {
uint64_t min_log_number = job_context->log_number;
size_t num_alive_log_files = alive_log_files_.size();
// find newly obsoleted log files
while (alive_log_files_.begin()->number < min_log_number) {
auto& earliest = *alive_log_files_.begin();
if (immutable_db_options_.recycle_log_file_num >
log_recycle_files.size()) {
Log(InfoLogLevel::INFO_LEVEL, immutable_db_options_.info_log,
"adding log %" PRIu64 " to recycle list\n", earliest.number);
log_recycle_files.push_back(earliest.number);
} else {
job_context->log_delete_files.push_back(earliest.number);
}
if (job_context->size_log_to_delete == 0) {
job_context->prev_total_log_size = total_log_size_;
job_context->num_alive_log_files = num_alive_log_files;
}
job_context->size_log_to_delete += earliest.size;
total_log_size_ -= earliest.size;
alive_log_files_.pop_front();
// Current log should always stay alive since it can't have
// number < MinLogNumber().
assert(alive_log_files_.size());
}
while (!logs_.empty() && logs_.front().number < min_log_number) {
auto& log = logs_.front();
if (log.getting_synced) {
log_sync_cv_.Wait();
// logs_ could have changed while we were waiting.
continue;
}
logs_to_free_.push_back(log.ReleaseWriter());
logs_.pop_front();
}
// Current log cannot be obsolete.
assert(!logs_.empty());
}
// We're just cleaning up for DB::Write().
assert(job_context->logs_to_free.empty());
job_context->logs_to_free = logs_to_free_;
job_context->log_recycle_files.assign(log_recycle_files.begin(),
log_recycle_files.end());
logs_to_free_.clear();
}
namespace {
bool CompareCandidateFile(const JobContext::CandidateFileInfo& first,
const JobContext::CandidateFileInfo& second) {
if (first.file_name > second.file_name) {
return true;
} else if (first.file_name < second.file_name) {
return false;
} else {
return (first.path_id > second.path_id);
}
}
}; // namespace
// Delete obsolete files and log status and information of file deletion
void DBImpl::DeleteObsoleteFileImpl(Status file_deletion_status, int job_id,
const std::string& fname, FileType type,
uint64_t number, uint32_t path_id) {
if (type == kTableFile) {
file_deletion_status =
DeleteSSTFile(&immutable_db_options_, fname, path_id);
} else {
file_deletion_status = env_->DeleteFile(fname);
}
if (file_deletion_status.ok()) {
Log(InfoLogLevel::DEBUG_LEVEL, immutable_db_options_.info_log,
"[JOB %d] Delete %s type=%d #%" PRIu64 " -- %s\n", job_id,
fname.c_str(), type, number, file_deletion_status.ToString().c_str());
} else if (env_->FileExists(fname).IsNotFound()) {
Log(InfoLogLevel::INFO_LEVEL, immutable_db_options_.info_log,
"[JOB %d] Tried to delete a non-existing file %s type=%d #%" PRIu64
" -- %s\n",
job_id, fname.c_str(), type, number,
file_deletion_status.ToString().c_str());
} else {
Log(InfoLogLevel::ERROR_LEVEL, immutable_db_options_.info_log,
"[JOB %d] Failed to delete %s type=%d #%" PRIu64 " -- %s\n", job_id,
fname.c_str(), type, number, file_deletion_status.ToString().c_str());
}
if (type == kTableFile) {
EventHelpers::LogAndNotifyTableFileDeletion(
&event_logger_, job_id, number, fname, file_deletion_status, GetName(),
immutable_db_options_.listeners);
}
}
// Diffs the files listed in filenames and those that do not
// belong to live files are posibly removed. Also, removes all the
// files in sst_delete_files and log_delete_files.
// It is not necessary to hold the mutex when invoking this method.
void DBImpl::PurgeObsoleteFiles(const JobContext& state, bool schedule_only) {
// we'd better have sth to delete
assert(state.HaveSomethingToDelete());
// this checks if FindObsoleteFiles() was run before. If not, don't do
// PurgeObsoleteFiles(). If FindObsoleteFiles() was run, we need to also
// run PurgeObsoleteFiles(), even if disable_delete_obsolete_files_ is true
if (state.manifest_file_number == 0) {
return;
}
// Now, convert live list to an unordered map, WITHOUT mutex held;
// set is slow.
std::unordered_map<uint64_t, const FileDescriptor*> sst_live_map;
for (const FileDescriptor& fd : state.sst_live) {
sst_live_map[fd.GetNumber()] = &fd;
}
std::unordered_set<uint64_t> log_recycle_files_set(
state.log_recycle_files.begin(), state.log_recycle_files.end());
Speed up FindObsoleteFiles() Summary: There are two versions of FindObsoleteFiles(): * full scan, which is executed every 6 hours (and it's terribly slow) * no full scan, which is executed every time a background process finishes and iterator is deleted This diff is optimizing the second case (no full scan). Here's what we do before the diff: * Get the list of obsolete files (files with ref==0). Some files in obsolete_files set might actually be live. * Get the list of live files to avoid deleting files that are live. * Delete files that are in obsolete_files and not in live_files. After this diff: * The only files with ref==0 that are still live are files that have been part of move compaction. Don't include moved files in obsolete_files. * Get the list of obsolete files (which exclude moved files). * No need to get the list of live files, since all files in obsolete_files need to be deleted. I'll post the benchmark results, but you can get the feel of it here: https://reviews.facebook.net/D30123 This depends on D30123. P.S. We should do full scan only in failure scenarios, not every 6 hours. I'll do this in a follow-up diff. Test Plan: One new unit test. Made sure that unit test fails if we don't have a `if (!f->moved)` safeguard in ~Version. make check Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: yhchiang, rven, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30249
10 years ago
auto candidate_files = state.full_scan_candidate_files;
candidate_files.reserve(
candidate_files.size() + state.sst_delete_files.size() +
state.log_delete_files.size() + state.manifest_delete_files.size());
// We may ignore the dbname when generating the file names.
const char* kDumbDbName = "";
for (auto file : state.sst_delete_files) {
candidate_files.emplace_back(
MakeTableFileName(kDumbDbName, file->fd.GetNumber()),
file->fd.GetPathId());
delete file;
}
for (auto file_num : state.log_delete_files) {
if (file_num > 0) {
candidate_files.emplace_back(LogFileName(kDumbDbName, file_num), 0);
}
}
for (const auto& filename : state.manifest_delete_files) {
candidate_files.emplace_back(filename, 0);
}
// dedup state.candidate_files so we don't try to delete the same
// file twice
std::sort(candidate_files.begin(), candidate_files.end(),
CompareCandidateFile);
candidate_files.erase(
std::unique(candidate_files.begin(), candidate_files.end()),
candidate_files.end());
if (state.prev_total_log_size > 0) {
Log(InfoLogLevel::INFO_LEVEL, immutable_db_options_.info_log,
"[JOB %d] Try to delete WAL files size %" PRIu64
", prev total WAL file size %" PRIu64
", number of live WAL files %" ROCKSDB_PRIszt ".\n",
state.job_id, state.size_log_to_delete, state.prev_total_log_size,
state.num_alive_log_files);
}
std::vector<std::string> old_info_log_files;
InfoLogPrefix info_log_prefix(!immutable_db_options_.db_log_dir.empty(),
dbname_);
for (const auto& candidate_file : candidate_files) {
std::string to_delete = candidate_file.file_name;
uint32_t path_id = candidate_file.path_id;
uint64_t number;
FileType type;
// Ignore file if we cannot recognize it.
if (!ParseFileName(to_delete, &number, info_log_prefix.prefix, &type)) {
continue;
}
bool keep = true;
switch (type) {
case kLogFile:
keep = ((number >= state.log_number) ||
(number == state.prev_log_number) ||
(log_recycle_files_set.find(number) !=
log_recycle_files_set.end()));
break;
case kDescriptorFile:
// Keep my manifest file, and any newer incarnations'
// (can happen during manifest roll)
keep = (number >= state.manifest_file_number);
break;
case kTableFile:
// If the second condition is not there, this makes
// DontDeletePendingOutputs fail
keep = (sst_live_map.find(number) != sst_live_map.end()) ||
number >= state.min_pending_output;
break;
case kTempFile:
// Any temp files that are currently being written to must
// be recorded in pending_outputs_, which is inserted into "live".
// Also, SetCurrentFile creates a temp file when writing out new
// manifest, which is equal to state.pending_manifest_file_number. We
// should not delete that file
//
// TODO(yhchiang): carefully modify the third condition to safely
// remove the temp options files.
keep = (sst_live_map.find(number) != sst_live_map.end()) ||
(number == state.pending_manifest_file_number) ||
(to_delete.find(kOptionsFileNamePrefix) != std::string::npos);
break;
case kInfoLogFile:
keep = true;
if (number != 0) {
old_info_log_files.push_back(to_delete);
}
break;
case kCurrentFile:
case kDBLockFile:
case kIdentityFile:
case kMetaDatabase:
case kOptionsFile:
keep = true;
break;
}
if (keep) {
continue;
}
std::string fname;
if (type == kTableFile) {
// evict from cache
TableCache::Evict(table_cache_.get(), number);
fname = TableFileName(immutable_db_options_.db_paths, number, path_id);
} else {
fname = ((type == kLogFile) ? immutable_db_options_.wal_dir : dbname_) +
"/" + to_delete;
}
#ifndef ROCKSDB_LITE
if (type == kLogFile && (immutable_db_options_.wal_ttl_seconds > 0 ||
immutable_db_options_.wal_size_limit_mb > 0)) {
wal_manager_.ArchiveWALFile(fname, number);
continue;
}
#endif // !ROCKSDB_LITE
Status file_deletion_status;
if (schedule_only) {
InstrumentedMutexLock guard_lock(&mutex_);
SchedulePendingPurge(fname, type, number, path_id, state.job_id);
} else {
DeleteObsoleteFileImpl(file_deletion_status, state.job_id, fname, type,
number, path_id);
}
}
// Delete old info log files.
size_t old_info_log_file_count = old_info_log_files.size();
if (old_info_log_file_count != 0 &&
old_info_log_file_count >= immutable_db_options_.keep_log_file_num) {
std::sort(old_info_log_files.begin(), old_info_log_files.end());
size_t end =
old_info_log_file_count - immutable_db_options_.keep_log_file_num;
for (unsigned int i = 0; i <= end; i++) {
std::string& to_delete = old_info_log_files.at(i);
std::string full_path_to_delete =
(immutable_db_options_.db_log_dir.empty()
? dbname_
: immutable_db_options_.db_log_dir) +
"/" + to_delete;
Log(InfoLogLevel::INFO_LEVEL, immutable_db_options_.info_log,
"[JOB %d] Delete info log file %s\n", state.job_id,
full_path_to_delete.c_str());
Status s = env_->DeleteFile(full_path_to_delete);
if (!s.ok()) {
if (env_->FileExists(full_path_to_delete).IsNotFound()) {
Log(InfoLogLevel::INFO_LEVEL, immutable_db_options_.info_log,
"[JOB %d] Tried to delete non-existing info log file %s FAILED "
"-- %s\n",
state.job_id, to_delete.c_str(), s.ToString().c_str());
} else {
Log(InfoLogLevel::ERROR_LEVEL, immutable_db_options_.info_log,
"[JOB %d] Delete info log file %s FAILED -- %s\n", state.job_id,
to_delete.c_str(), s.ToString().c_str());
}
}
}
}
#ifndef ROCKSDB_LITE
wal_manager_.PurgeObsoleteWALFiles();
#endif // ROCKSDB_LITE
LogFlush(immutable_db_options_.info_log);
}
void DBImpl::DeleteObsoleteFiles() {
mutex_.AssertHeld();
JobContext job_context(next_job_id_.fetch_add(1));
FindObsoleteFiles(&job_context, true);
mutex_.Unlock();
if (job_context.HaveSomethingToDelete()) {
PurgeObsoleteFiles(job_context);
}
job_context.Clean();
mutex_.Lock();
}
Status DBImpl::Directories::CreateAndNewDirectory(
Env* env, const std::string& dirname,
std::unique_ptr<Directory>* directory) const {
// We call CreateDirIfMissing() as the directory may already exist (if we
// are reopening a DB), when this happens we don't want creating the
// directory to cause an error. However, we need to check if creating the
// directory fails or else we may get an obscure message about the lock
// file not existing. One real-world example of this occurring is if
// env->CreateDirIfMissing() doesn't create intermediate directories, e.g.
// when dbname_ is "dir/db" but when "dir" doesn't exist.
Status s = env->CreateDirIfMissing(dirname);
if (!s.ok()) {
return s;
}
return env->NewDirectory(dirname, directory);
}
Status DBImpl::Directories::SetDirectories(
Env* env, const std::string& dbname, const std::string& wal_dir,
const std::vector<DbPath>& data_paths) {
Status s = CreateAndNewDirectory(env, dbname, &db_dir_);
if (!s.ok()) {
return s;
}
if (!wal_dir.empty() && dbname != wal_dir) {
s = CreateAndNewDirectory(env, wal_dir, &wal_dir_);
if (!s.ok()) {
return s;
}
}
data_dirs_.clear();
for (auto& p : data_paths) {
const std::string db_path = p.path;
if (db_path == dbname) {
data_dirs_.emplace_back(nullptr);
} else {
std::unique_ptr<Directory> path_directory;
s = CreateAndNewDirectory(env, db_path, &path_directory);
if (!s.ok()) {
return s;
}
data_dirs_.emplace_back(path_directory.release());
}
}
assert(data_dirs_.size() == data_paths.size());
return Status::OK();
}
Directory* DBImpl::Directories::GetDataDir(size_t path_id) {
assert(path_id < data_dirs_.size());
Directory* ret_dir = data_dirs_[path_id].get();
if (ret_dir == nullptr) {
// Should use db_dir_
return db_dir_.get();
}
return ret_dir;
}
Status DBImpl::Recover(
const std::vector<ColumnFamilyDescriptor>& column_families, bool read_only,
bool error_if_log_file_exist, bool error_if_data_exists_in_logs) {
mutex_.AssertHeld();
bool is_new_db = false;
assert(db_lock_ == nullptr);
if (!read_only) {
Status s = directories_.SetDirectories(env_, dbname_,
immutable_db_options_.wal_dir,
immutable_db_options_.db_paths);
if (!s.ok()) {
return s;
}
s = env_->LockFile(LockFileName(dbname_), &db_lock_);
if (!s.ok()) {
return s;
}
s = env_->FileExists(CurrentFileName(dbname_));
if (s.IsNotFound()) {
if (immutable_db_options_.create_if_missing) {
s = NewDB();
is_new_db = true;
if (!s.ok()) {
return s;
}
} else {
return Status::InvalidArgument(
dbname_, "does not exist (create_if_missing is false)");
}
} else if (s.ok()) {
if (immutable_db_options_.error_if_exists) {
return Status::InvalidArgument(
dbname_, "exists (error_if_exists is true)");
}
} else {
// Unexpected error reading file
assert(s.IsIOError());
return s;
}
Dbid feature Summary: Create a new type of file on startup if it doesn't already exist called DBID. This will store a unique number generated from boost library's uuid header file. The use-case is to identify the case of a db losing all its data and coming back up either empty or from an image(backup/live replica's recovery) the key point to note is that DBID is not stored in a backup or db snapshot It's preferable to use Boost for uuid because: 1) A non-standard way of generating uuid is not good 2) /proc/sys/kernel/random/uuid generates a uuid but only on linux environments and the solution would not be clean 3) c++ doesn't have any direct way to get a uuid 4) Boost is a very good library that was already having linkage in rocksdb from third-party Note: I had to update the TOOLCHAIN_REV in build files to get latest verison of boost from third-party as the older version had a bug. I had to put Wno-uninitialized in Makefile because boost-1.51 has an unitialized variable and rocksdb would not comiple otherwise. Latet open-source for boost is 1.54 but is not there in third-party. I have notified the concerned people in fbcode about it. @kailiu : While releasing to third-party, an additional dependency will need to be created for boost in TARGETS file. I can help identify. Test Plan: Expand db_test to test 2 cases 1) Restarting db with Id file present - verify that no change to Id 2)Restarting db with Id file deleted - verify that a different Id is there after reopen Also run make all check Reviewers: dhruba, haobo, kailiu, sdong Reviewed By: dhruba CC: leveldb Differential Revision: https://reviews.facebook.net/D13587
11 years ago
// Check for the IDENTITY file and create it if not there
s = env_->FileExists(IdentityFileName(dbname_));
if (s.IsNotFound()) {
Dbid feature Summary: Create a new type of file on startup if it doesn't already exist called DBID. This will store a unique number generated from boost library's uuid header file. The use-case is to identify the case of a db losing all its data and coming back up either empty or from an image(backup/live replica's recovery) the key point to note is that DBID is not stored in a backup or db snapshot It's preferable to use Boost for uuid because: 1) A non-standard way of generating uuid is not good 2) /proc/sys/kernel/random/uuid generates a uuid but only on linux environments and the solution would not be clean 3) c++ doesn't have any direct way to get a uuid 4) Boost is a very good library that was already having linkage in rocksdb from third-party Note: I had to update the TOOLCHAIN_REV in build files to get latest verison of boost from third-party as the older version had a bug. I had to put Wno-uninitialized in Makefile because boost-1.51 has an unitialized variable and rocksdb would not comiple otherwise. Latet open-source for boost is 1.54 but is not there in third-party. I have notified the concerned people in fbcode about it. @kailiu : While releasing to third-party, an additional dependency will need to be created for boost in TARGETS file. I can help identify. Test Plan: Expand db_test to test 2 cases 1) Restarting db with Id file present - verify that no change to Id 2)Restarting db with Id file deleted - verify that a different Id is there after reopen Also run make all check Reviewers: dhruba, haobo, kailiu, sdong Reviewed By: dhruba CC: leveldb Differential Revision: https://reviews.facebook.net/D13587
11 years ago
s = SetIdentityFile(env_, dbname_);
if (!s.ok()) {
return s;
}
} else if (!s.ok()) {
assert(s.IsIOError());
return s;
Dbid feature Summary: Create a new type of file on startup if it doesn't already exist called DBID. This will store a unique number generated from boost library's uuid header file. The use-case is to identify the case of a db losing all its data and coming back up either empty or from an image(backup/live replica's recovery) the key point to note is that DBID is not stored in a backup or db snapshot It's preferable to use Boost for uuid because: 1) A non-standard way of generating uuid is not good 2) /proc/sys/kernel/random/uuid generates a uuid but only on linux environments and the solution would not be clean 3) c++ doesn't have any direct way to get a uuid 4) Boost is a very good library that was already having linkage in rocksdb from third-party Note: I had to update the TOOLCHAIN_REV in build files to get latest verison of boost from third-party as the older version had a bug. I had to put Wno-uninitialized in Makefile because boost-1.51 has an unitialized variable and rocksdb would not comiple otherwise. Latet open-source for boost is 1.54 but is not there in third-party. I have notified the concerned people in fbcode about it. @kailiu : While releasing to third-party, an additional dependency will need to be created for boost in TARGETS file. I can help identify. Test Plan: Expand db_test to test 2 cases 1) Restarting db with Id file present - verify that no change to Id 2)Restarting db with Id file deleted - verify that a different Id is there after reopen Also run make all check Reviewers: dhruba, haobo, kailiu, sdong Reviewed By: dhruba CC: leveldb Differential Revision: https://reviews.facebook.net/D13587
11 years ago
}
}
Status s = versions_->Recover(column_families, read_only);
if (immutable_db_options_.paranoid_checks && s.ok()) {
s = CheckConsistency();
}
if (s.ok()) {
SequenceNumber next_sequence(kMaxSequenceNumber);
default_cf_handle_ = new ColumnFamilyHandleImpl(
versions_->GetColumnFamilySet()->GetDefault(), this, &mutex_);
make internal stats independent of statistics Summary: also make it aware of column family output from db_bench ``` ** Compaction Stats [default] ** Level Files Size(MB) Score Read(GB) Rn(GB) Rnp1(GB) Write(GB) Wnew(GB) RW-Amp W-Amp Rd(MB/s) Wr(MB/s) Rn(cnt) Rnp1(cnt) Wnp1(cnt) Wnew(cnt) Comp(sec) Comp(cnt) Avg(sec) Stall(sec) Stall(cnt) Avg(ms) ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ L0 14 956 0.9 0.0 0.0 0.0 2.7 2.7 0.0 0.0 0.0 111.6 0 0 0 0 24 40 0.612 75.20 492387 0.15 L1 21 2001 2.0 5.7 2.0 3.7 5.3 1.6 5.4 2.6 71.2 65.7 31 43 55 12 82 2 41.242 43.72 41183 1.06 L2 217 18974 1.9 16.5 2.0 14.4 15.1 0.7 15.6 7.4 70.1 64.3 17 182 185 3 241 16 15.052 0.00 0 0.00 L3 1641 188245 1.8 9.1 1.1 8.0 8.5 0.5 15.4 7.4 61.3 57.2 9 75 76 1 152 9 16.887 0.00 0 0.00 L4 4447 449025 0.4 13.4 4.8 8.6 9.1 0.5 4.7 1.9 77.8 52.7 38 79 100 21 176 38 4.639 0.00 0 0.00 Sum 6340 659201 0.0 44.7 10.0 34.7 40.6 6.0 32.0 15.2 67.7 61.6 95 379 416 37 676 105 6.439 118.91 533570 0.22 Int 0 0 0.0 1.2 0.4 0.8 1.3 0.5 5.2 2.7 59.1 65.6 3 7 9 2 20 10 2.003 0.00 0 0.00 Stalls(secs): 75.197 level0_slowdown, 0.000 level0_numfiles, 0.000 memtable_compaction, 43.717 leveln_slowdown Stalls(count): 492387 level0_slowdown, 0 level0_numfiles, 0 memtable_compaction, 41183 leveln_slowdown ** DB Stats ** Uptime(secs): 202.1 total, 13.5 interval Cumulative writes: 6291456 writes, 6291456 batches, 1.0 writes per batch, 4.90 ingest GB Cumulative WAL: 6291456 writes, 6291456 syncs, 1.00 writes per sync, 4.90 GB written Interval writes: 1048576 writes, 1048576 batches, 1.0 writes per batch, 836.0 ingest MB Interval WAL: 1048576 writes, 1048576 syncs, 1.00 writes per sync, 0.82 MB written Test Plan: ran it Reviewers: sdong, yhchiang, igor Reviewed By: igor Subscribers: leveldb Differential Revision: https://reviews.facebook.net/D19917
10 years ago
default_cf_internal_stats_ = default_cf_handle_->cfd()->internal_stats();
single_column_family_mode_ =
versions_->GetColumnFamilySet()->NumberOfColumnFamilies() == 1;
// Recover from all newer log files than the ones named in the
// descriptor (new log files may have been added by the previous
// incarnation without registering them in the descriptor).
//
// Note that prev_log_number() is no longer used, but we pay
// attention to it in case we are recovering a database
// produced by an older version of rocksdb.
std::vector<std::string> filenames;
s = env_->GetChildren(immutable_db_options_.wal_dir, &filenames);
if (!s.ok()) {
return s;
}
std::vector<uint64_t> logs;
for (size_t i = 0; i < filenames.size(); i++) {
uint64_t number;
FileType type;
if (ParseFileName(filenames[i], &number, &type) && type == kLogFile) {
if (is_new_db) {
return Status::Corruption(
"While creating a new Db, wal_dir contains "
"existing log file: ",
filenames[i]);
} else {
logs.push_back(number);
}
}
}
if (logs.size() > 0) {
if (error_if_log_file_exist) {
return Status::Corruption(
"The db was opened in readonly mode with error_if_log_file_exist"
"flag but a log file already exists");
} else if (error_if_data_exists_in_logs) {
for (auto& log : logs) {
std::string fname = LogFileName(immutable_db_options_.wal_dir, log);
uint64_t bytes;
s = env_->GetFileSize(fname, &bytes);
if (s.ok()) {
if (bytes > 0) {
return Status::Corruption(
"error_if_data_exists_in_logs is set but there are data "
" in log files.");
}
}
}
}
}
if (!logs.empty()) {
// Recover in the order in which the logs were generated
std::sort(logs.begin(), logs.end());
s = RecoverLogFiles(logs, &next_sequence, read_only);
if (!s.ok()) {
// Clear memtables if recovery failed
for (auto cfd : *versions_->GetColumnFamilySet()) {
cfd->CreateNewMemtable(*cfd->GetLatestMutableCFOptions(),
kMaxSequenceNumber);
}
}
}
SetTickerCount(stats_, SEQUENCE_NUMBER, versions_->LastSequence());
}
// Initial value
max_total_in_memory_state_ = 0;
for (auto cfd : *versions_->GetColumnFamilySet()) {
auto* mutable_cf_options = cfd->GetLatestMutableCFOptions();
max_total_in_memory_state_ += mutable_cf_options->write_buffer_size *
mutable_cf_options->max_write_buffer_number;
}
return s;
}
// REQUIRES: log_numbers are sorted in ascending order
Status DBImpl::RecoverLogFiles(const std::vector<uint64_t>& log_numbers,
SequenceNumber* next_sequence, bool read_only) {
struct LogReporter : public log::Reader::Reporter {
Env* env;
Logger* info_log;
const char* fname;
Status* status; // nullptr if immutable_db_options_.paranoid_checks==false
virtual void Corruption(size_t bytes, const Status& s) override {
Log(InfoLogLevel::WARN_LEVEL,
info_log, "%s%s: dropping %d bytes; %s",
(this->status == nullptr ? "(ignoring error) " : ""),
fname, static_cast<int>(bytes), s.ToString().c_str());
if (this->status != nullptr && this->status->ok()) {
*this->status = s;
}
}
};
mutex_.AssertHeld();
Status status;
std::unordered_map<int, VersionEdit> version_edits;
// no need to refcount because iteration is under mutex
for (auto cfd : *versions_->GetColumnFamilySet()) {
VersionEdit edit;
edit.SetColumnFamily(cfd->GetID());
version_edits.insert({cfd->GetID(), edit});
}
int job_id = next_job_id_.fetch_add(1);
{
auto stream = event_logger_.Log();
stream << "job" << job_id << "event"
<< "recovery_started";
stream << "log_files";
stream.StartArray();
for (auto log_number : log_numbers) {
stream << log_number;
}
stream.EndArray();
}
Refactor Recover() code Summary: This diff does two things: * Rethinks how we call Recover() with read_only option. Before, we call it with pointer to memtable where we'd like to apply those changes to. This memtable is set in db_impl_readonly.cc and it's actually DBImpl::mem_. Why don't we just apply updates to mem_ right away? It seems more intuitive. * Changes when we apply updates to manifest. Before, the process is to recover all the logs, flush it to sst files and then do one giant commit that atomically adds all recovered sst files and sets the next log number. This works good enough, but causes some small troubles for my column family approach, since I can't have one VersionEdit apply to more than single column family[1]. The change here is to commit the files recovered from logs right away. Here is the state of the world before the change: 1. Recover log 5, add new sst files to edit 2. Recover log 7, add new sst files to edit 3. Recover log 8, add new sst files to edit 4. Commit all added sst files to manifest and mark log files 5, 7 and 8 as recoverd (via SetLogNumber(9) function) After the change, we'll do: 1. Recover log 5, commit the new sst files and set log 5 as recovered 2. Recover log 7, commit the new sst files and set log 7 as recovered 3. Recover log 8, commit the new sst files and set log 8 as recovered The added (small) benefit is that if we fail after (2), the new recovery will only have to recover log 8. In previous case, we'll have to restart the recovery from the beginning. The bigger benefit will be to enable easier integration of multiple column families in Recovery code path. [1] I'm happy to dicuss this decison, but I believe this is the cleanest way to go. It also makes backward compatibility much easier. We don't have a requirement of adding multiple column families atomically. Test Plan: make check Reviewers: dhruba, haobo, kailiu, sdong Reviewed By: kailiu CC: leveldb Differential Revision: https://reviews.facebook.net/D15237
11 years ago
#ifndef ROCKSDB_LITE
if (immutable_db_options_.wal_filter != nullptr) {
std::map<std::string, uint32_t> cf_name_id_map;
std::map<uint32_t, uint64_t> cf_lognumber_map;
for (auto cfd : *versions_->GetColumnFamilySet()) {
cf_name_id_map.insert(
std::make_pair(cfd->GetName(), cfd->GetID()));
cf_lognumber_map.insert(
std::make_pair(cfd->GetID(), cfd->GetLogNumber()));
}
immutable_db_options_.wal_filter->ColumnFamilyLogNumberMap(cf_lognumber_map,
cf_name_id_map);
}
#endif
bool stop_replay_by_wal_filter = false;
bool stop_replay_for_corruption = false;
bool flushed = false;
for (auto log_number : log_numbers) {
// The previous incarnation may not have written any MANIFEST
// records after allocating this log number. So we manually
// update the file number allocation counter in VersionSet.
versions_->MarkFileNumberUsedDuringRecovery(log_number);
// Open the log file
std::string fname = LogFileName(immutable_db_options_.wal_dir, log_number);
Log(InfoLogLevel::INFO_LEVEL, immutable_db_options_.info_log,
"Recovering log #%" PRIu64 " mode %d", log_number,
immutable_db_options_.wal_recovery_mode);
auto logFileDropped = [this, &fname]() {
uint64_t bytes;
if (env_->GetFileSize(fname, &bytes).ok()) {
auto info_log = immutable_db_options_.info_log.get();
Log(InfoLogLevel::WARN_LEVEL, info_log, "%s: dropping %d bytes",
fname.c_str(), static_cast<int>(bytes));
}
};
if (stop_replay_by_wal_filter) {
logFileDropped();
continue;
}
unique_ptr<SequentialFileReader> file_reader;
{
unique_ptr<SequentialFile> file;
status = env_->NewSequentialFile(fname, &file, env_options_);
if (!status.ok()) {
MaybeIgnoreError(&status);
if (!status.ok()) {
return status;
} else {
// Fail with one log file, but that's ok.
// Try next one.
continue;
}
}
file_reader.reset(new SequentialFileReader(std::move(file)));
}
// Create the log reader.
LogReporter reporter;
reporter.env = env_;
reporter.info_log = immutable_db_options_.info_log.get();
reporter.fname = fname.c_str();
if (!immutable_db_options_.paranoid_checks ||
immutable_db_options_.wal_recovery_mode ==
WALRecoveryMode::kSkipAnyCorruptedRecords) {
reporter.status = nullptr;
} else {
reporter.status = &status;
}
// We intentially make log::Reader do checksumming even if
// paranoid_checks==false so that corruptions cause entire commits
// to be skipped instead of propagating bad information (like overly
// large sequence numbers).
log::Reader reader(immutable_db_options_.info_log, std::move(file_reader),
&reporter, true /*checksum*/, 0 /*initial_offset*/,
log_number);
// Determine if we should tolerate incomplete records at the tail end of the
// Read all the records and add to a memtable
std::string scratch;
Slice record;
WriteBatch batch;
while (!stop_replay_by_wal_filter &&
reader.ReadRecord(&record, &scratch,
immutable_db_options_.wal_recovery_mode) &&
status.ok()) {
if (record.size() < WriteBatchInternal::kHeader) {
reporter.Corruption(record.size(),
Status::Corruption("log record too small"));
continue;
}
WriteBatchInternal::SetContents(&batch, record);
SequenceNumber sequence = WriteBatchInternal::Sequence(&batch);
if (immutable_db_options_.wal_recovery_mode ==
WALRecoveryMode::kPointInTimeRecovery) {
// In point-in-time recovery mode, if sequence id of log files are
// consecutive, we continue recovery despite corruption. This could
// happen when we open and write to a corrupted DB, where sequence id
// will start from the last sequence id we recovered.
if (sequence == *next_sequence) {
stop_replay_for_corruption = false;
}
if (stop_replay_for_corruption) {
logFileDropped();
break;
}
}
bool no_prev_seq = true;
if (!immutable_db_options_.allow_2pc) {
*next_sequence = sequence;
} else {
if (*next_sequence == kMaxSequenceNumber) {
*next_sequence = sequence;
} else {
no_prev_seq = false;
WriteBatchInternal::SetSequence(&batch, *next_sequence);
}
}
#ifndef ROCKSDB_LITE
if (immutable_db_options_.wal_filter != nullptr) {
WriteBatch new_batch;
bool batch_changed = false;
WalFilter::WalProcessingOption wal_processing_option =
immutable_db_options_.wal_filter->LogRecordFound(
log_number, fname, batch, &new_batch, &batch_changed);
switch (wal_processing_option) {
case WalFilter::WalProcessingOption::kContinueProcessing:
// do nothing, proceeed normally
break;
case WalFilter::WalProcessingOption::kIgnoreCurrentRecord:
// skip current record
continue;
case WalFilter::WalProcessingOption::kStopReplay:
// skip current record and stop replay
stop_replay_by_wal_filter = true;
continue;
case WalFilter::WalProcessingOption::kCorruptedRecord: {
status =
Status::Corruption("Corruption reported by Wal Filter ",
immutable_db_options_.wal_filter->Name());
MaybeIgnoreError(&status);
if (!status.ok()) {
reporter.Corruption(record.size(), status);
continue;
}
break;
}
default: {
assert(false); // unhandled case
status = Status::NotSupported(
"Unknown WalProcessingOption returned"
" by Wal Filter ",
immutable_db_options_.wal_filter->Name());
MaybeIgnoreError(&status);
if (!status.ok()) {
return status;
} else {
// Ignore the error with current record processing.
continue;
}
}
}
if (batch_changed) {
// Make sure that the count in the new batch is
// within the orignal count.
int new_count = WriteBatchInternal::Count(&new_batch);
int original_count = WriteBatchInternal::Count(&batch);
if (new_count > original_count) {
Log(InfoLogLevel::FATAL_LEVEL, immutable_db_options_.info_log,
"Recovering log #%" PRIu64
" mode %d log filter %s returned "
"more records (%d) than original (%d) which is not allowed. "
"Aborting recovery.",
log_number, immutable_db_options_.wal_recovery_mode,
immutable_db_options_.wal_filter->Name(), new_count,
original_count);
status = Status::NotSupported(
"More than original # of records "
"returned by Wal Filter ",
immutable_db_options_.wal_filter->Name());
return status;
}
// Set the same sequence number in the new_batch
// as the original batch.
WriteBatchInternal::SetSequence(&new_batch,
WriteBatchInternal::Sequence(&batch));
batch = new_batch;
}
}
#endif // ROCKSDB_LITE
// If column family was not found, it might mean that the WAL write
// batch references to the column family that was dropped after the
// insert. We don't want to fail the whole write batch in that case --
// we just ignore the update.
// That's why we set ignore missing column families to true
bool has_valid_writes = false;
status = WriteBatchInternal::InsertInto(
&batch, column_family_memtables_.get(), &flush_scheduler_, true,
log_number, this, false /* concurrent_memtable_writes */,
next_sequence, &has_valid_writes);
// If it is the first log file and there is no column family updated
// after replaying the file, this file may be a stale file. We ignore
// sequence IDs from the file. Otherwise, if a newer stale log file that
// has been deleted, the sequenceID may be wrong.
if (immutable_db_options_.allow_2pc) {
if (no_prev_seq && !has_valid_writes) {
*next_sequence = kMaxSequenceNumber;
}
}
MaybeIgnoreError(&status);
if (!status.ok()) {
// We are treating this as a failure while reading since we read valid
// blocks that do not form coherent data
reporter.Corruption(record.size(), status);
continue;
}
if (has_valid_writes && !read_only) {
// we can do this because this is called before client has access to the
// DB and there is only a single thread operating on DB
ColumnFamilyData* cfd;
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
while ((cfd = flush_scheduler_.TakeNextColumnFamily()) != nullptr) {
cfd->Unref();
// If this asserts, it means that InsertInto failed in
// filtering updates to already-flushed column families
assert(cfd->GetLogNumber() <= log_number);
auto iter = version_edits.find(cfd->GetID());
assert(iter != version_edits.end());
VersionEdit* edit = &iter->second;
status = WriteLevel0TableForRecovery(job_id, cfd, cfd->mem(), edit);
if (!status.ok()) {
// Reflect errors immediately so that conditions like full
// file-systems cause the DB::Open() to fail.
return status;
}
flushed = true;
cfd->CreateNewMemtable(*cfd->GetLatestMutableCFOptions(),
*next_sequence);
}
}
}
if (!status.ok()) {
if (immutable_db_options_.wal_recovery_mode ==
WALRecoveryMode::kSkipAnyCorruptedRecords) {
// We should ignore all errors unconditionally
status = Status::OK();
} else if (immutable_db_options_.wal_recovery_mode ==
WALRecoveryMode::kPointInTimeRecovery) {
// We should ignore the error but not continue replaying
status = Status::OK();
stop_replay_for_corruption = true;
Log(InfoLogLevel::INFO_LEVEL, immutable_db_options_.info_log,
"Point in time recovered to log #%" PRIu64 " seq #%" PRIu64,
log_number, *next_sequence);
} else {
assert(immutable_db_options_.wal_recovery_mode ==
WALRecoveryMode::kTolerateCorruptedTailRecords ||
immutable_db_options_.wal_recovery_mode ==
WALRecoveryMode::kAbsoluteConsistency);
return status;
}
}
flush_scheduler_.Clear();
auto last_sequence = *next_sequence - 1;
if ((*next_sequence != kMaxSequenceNumber) &&
(versions_->LastSequence() <= last_sequence)) {
versions_->SetLastSequence(last_sequence);
}
}
if (!read_only) {
// no need to refcount since client still doesn't have access
// to the DB and can not drop column families while we iterate
auto max_log_number = log_numbers.back();
for (auto cfd : *versions_->GetColumnFamilySet()) {
auto iter = version_edits.find(cfd->GetID());
assert(iter != version_edits.end());
VersionEdit* edit = &iter->second;
if (cfd->GetLogNumber() > max_log_number) {
// Column family cfd has already flushed the data
// from all logs. Memtable has to be empty because
// we filter the updates based on log_number
// (in WriteBatch::InsertInto)
assert(cfd->mem()->GetFirstSequenceNumber() == 0);
assert(edit->NumEntries() == 0);
continue;
}
// flush the final memtable (if non-empty)
if (cfd->mem()->GetFirstSequenceNumber() != 0) {
// If flush happened in the middle of recovery (e.g. due to memtable
// being full), we flush at the end. Otherwise we'll need to record
// where we were on last flush, which make the logic complicated.
if (flushed || !immutable_db_options_.avoid_flush_during_recovery) {
status = WriteLevel0TableForRecovery(job_id, cfd, cfd->mem(), edit);
if (!status.ok()) {
// Recovery failed
break;
}
flushed = true;
cfd->CreateNewMemtable(*cfd->GetLatestMutableCFOptions(),
*next_sequence);
}
}
// write MANIFEST with update
// writing log_number in the manifest means that any log file
// with number strongly less than (log_number + 1) is already
// recovered and should be ignored on next reincarnation.
// Since we already recovered max_log_number, we want all logs
// with numbers `<= max_log_number` (includes this one) to be ignored
if (flushed || cfd->mem()->GetFirstSequenceNumber() == 0) {
edit->SetLogNumber(max_log_number + 1);
}
// we must mark the next log number as used, even though it's
// not actually used. that is because VersionSet assumes
// VersionSet::next_file_number_ always to be strictly greater than any
// log number
versions_->MarkFileNumberUsedDuringRecovery(max_log_number + 1);
status = versions_->LogAndApply(
cfd, *cfd->GetLatestMutableCFOptions(), edit, &mutex_);
if (!status.ok()) {
// Recovery failed
break;
}
}
}
Refactor Recover() code Summary: This diff does two things: * Rethinks how we call Recover() with read_only option. Before, we call it with pointer to memtable where we'd like to apply those changes to. This memtable is set in db_impl_readonly.cc and it's actually DBImpl::mem_. Why don't we just apply updates to mem_ right away? It seems more intuitive. * Changes when we apply updates to manifest. Before, the process is to recover all the logs, flush it to sst files and then do one giant commit that atomically adds all recovered sst files and sets the next log number. This works good enough, but causes some small troubles for my column family approach, since I can't have one VersionEdit apply to more than single column family[1]. The change here is to commit the files recovered from logs right away. Here is the state of the world before the change: 1. Recover log 5, add new sst files to edit 2. Recover log 7, add new sst files to edit 3. Recover log 8, add new sst files to edit 4. Commit all added sst files to manifest and mark log files 5, 7 and 8 as recoverd (via SetLogNumber(9) function) After the change, we'll do: 1. Recover log 5, commit the new sst files and set log 5 as recovered 2. Recover log 7, commit the new sst files and set log 7 as recovered 3. Recover log 8, commit the new sst files and set log 8 as recovered The added (small) benefit is that if we fail after (2), the new recovery will only have to recover log 8. In previous case, we'll have to restart the recovery from the beginning. The bigger benefit will be to enable easier integration of multiple column families in Recovery code path. [1] I'm happy to dicuss this decison, but I believe this is the cleanest way to go. It also makes backward compatibility much easier. We don't have a requirement of adding multiple column families atomically. Test Plan: make check Reviewers: dhruba, haobo, kailiu, sdong Reviewed By: kailiu CC: leveldb Differential Revision: https://reviews.facebook.net/D15237
11 years ago
if (!flushed) {
// Mark these as alive so they'll be considered for deletion later by
// FindObsoleteFiles()
for (auto log_number : log_numbers) {
alive_log_files_.push_back(LogFileNumberSize(log_number));
}
}
event_logger_.Log() << "job" << job_id << "event"
<< "recovery_finished";
return status;
}
Status DBImpl::WriteLevel0TableForRecovery(int job_id, ColumnFamilyData* cfd,
MemTable* mem, VersionEdit* edit) {
mutex_.AssertHeld();
const uint64_t start_micros = env_->NowMicros();
FileMetaData meta;
auto pending_outputs_inserted_elem =
CaptureCurrentFileNumberInPendingOutputs();
meta.fd = FileDescriptor(versions_->NewFileNumber(), 0, 0);
ReadOptions ro;
ro.total_order_seek = true;
Arena arena;
Status s;
TableProperties table_properties;
{
ScopedArenaIterator iter(mem->NewIterator(ro, &arena));
Log(InfoLogLevel::DEBUG_LEVEL, immutable_db_options_.info_log,
"[%s] [WriteLevel0TableForRecovery]"
" Level-0 table #%" PRIu64 ": started",
cfd->GetName().c_str(), meta.fd.GetNumber());
// Get the latest mutable cf options while the mutex is still locked
const MutableCFOptions mutable_cf_options =
*cfd->GetLatestMutableCFOptions();
bool paranoid_file_checks =
cfd->GetLatestMutableCFOptions()->paranoid_file_checks;
{
mutex_.Unlock();
SequenceNumber earliest_write_conflict_snapshot;
std::vector<SequenceNumber> snapshot_seqs =
snapshots_.GetAll(&earliest_write_conflict_snapshot);
s = BuildTable(
dbname_, env_, *cfd->ioptions(), mutable_cf_options, env_options_,
cfd->table_cache(), iter.get(),
ScopedArenaIterator(mem->NewRangeTombstoneIterator(ro, &arena)),
&meta, cfd->internal_comparator(),
cfd->int_tbl_prop_collector_factories(), cfd->GetID(), cfd->GetName(),
snapshot_seqs, earliest_write_conflict_snapshot,
GetCompressionFlush(*cfd->ioptions(), mutable_cf_options),
cfd->ioptions()->compression_opts, paranoid_file_checks,
cfd->internal_stats(), TableFileCreationReason::kRecovery,
&event_logger_, job_id);
LogFlush(immutable_db_options_.info_log);
Log(InfoLogLevel::DEBUG_LEVEL, immutable_db_options_.info_log,
"[%s] [WriteLevel0TableForRecovery]"
" Level-0 table #%" PRIu64 ": %" PRIu64 " bytes %s",
cfd->GetName().c_str(), meta.fd.GetNumber(), meta.fd.GetFileSize(),
s.ToString().c_str());
mutex_.Lock();
}
}
ReleaseFileNumberFromPendingOutputs(pending_outputs_inserted_elem);
// Note that if file_size is zero, the file has been deleted and
// should not be added to the manifest.
int level = 0;
if (s.ok() && meta.fd.GetFileSize() > 0) {
edit->AddFile(level, meta.fd.GetNumber(), meta.fd.GetPathId(),
meta.fd.GetFileSize(), meta.smallest, meta.largest,
meta.smallest_seqno, meta.largest_seqno,
meta.marked_for_compaction);
}
InternalStats::CompactionStats stats(1);
stats.micros = env_->NowMicros() - start_micros;
stats.bytes_written = meta.fd.GetFileSize();
stats.num_output_files = 1;
cfd->internal_stats()->AddCompactionStats(level, stats);
make internal stats independent of statistics Summary: also make it aware of column family output from db_bench ``` ** Compaction Stats [default] ** Level Files Size(MB) Score Read(GB) Rn(GB) Rnp1(GB) Write(GB) Wnew(GB) RW-Amp W-Amp Rd(MB/s) Wr(MB/s) Rn(cnt) Rnp1(cnt) Wnp1(cnt) Wnew(cnt) Comp(sec) Comp(cnt) Avg(sec) Stall(sec) Stall(cnt) Avg(ms) ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ L0 14 956 0.9 0.0 0.0 0.0 2.7 2.7 0.0 0.0 0.0 111.6 0 0 0 0 24 40 0.612 75.20 492387 0.15 L1 21 2001 2.0 5.7 2.0 3.7 5.3 1.6 5.4 2.6 71.2 65.7 31 43 55 12 82 2 41.242 43.72 41183 1.06 L2 217 18974 1.9 16.5 2.0 14.4 15.1 0.7 15.6 7.4 70.1 64.3 17 182 185 3 241 16 15.052 0.00 0 0.00 L3 1641 188245 1.8 9.1 1.1 8.0 8.5 0.5 15.4 7.4 61.3 57.2 9 75 76 1 152 9 16.887 0.00 0 0.00 L4 4447 449025 0.4 13.4 4.8 8.6 9.1 0.5 4.7 1.9 77.8 52.7 38 79 100 21 176 38 4.639 0.00 0 0.00 Sum 6340 659201 0.0 44.7 10.0 34.7 40.6 6.0 32.0 15.2 67.7 61.6 95 379 416 37 676 105 6.439 118.91 533570 0.22 Int 0 0 0.0 1.2 0.4 0.8 1.3 0.5 5.2 2.7 59.1 65.6 3 7 9 2 20 10 2.003 0.00 0 0.00 Stalls(secs): 75.197 level0_slowdown, 0.000 level0_numfiles, 0.000 memtable_compaction, 43.717 leveln_slowdown Stalls(count): 492387 level0_slowdown, 0 level0_numfiles, 0 memtable_compaction, 41183 leveln_slowdown ** DB Stats ** Uptime(secs): 202.1 total, 13.5 interval Cumulative writes: 6291456 writes, 6291456 batches, 1.0 writes per batch, 4.90 ingest GB Cumulative WAL: 6291456 writes, 6291456 syncs, 1.00 writes per sync, 4.90 GB written Interval writes: 1048576 writes, 1048576 batches, 1.0 writes per batch, 836.0 ingest MB Interval WAL: 1048576 writes, 1048576 syncs, 1.00 writes per sync, 0.82 MB written Test Plan: ran it Reviewers: sdong, yhchiang, igor Reviewed By: igor Subscribers: leveldb Differential Revision: https://reviews.facebook.net/D19917
10 years ago
cfd->internal_stats()->AddCFStats(
InternalStats::BYTES_FLUSHED, meta.fd.GetFileSize());
RecordTick(stats_, COMPACT_WRITE_BYTES, meta.fd.GetFileSize());
return s;
}
Status DBImpl::SyncClosedLogs(JobContext* job_context) {
mutex_.AssertHeld();
autovector<log::Writer*, 1> logs_to_sync;
uint64_t current_log_number = logfile_number_;
while (logs_.front().number < current_log_number &&
logs_.front().getting_synced) {
log_sync_cv_.Wait();
}
for (auto it = logs_.begin();
it != logs_.end() && it->number < current_log_number; ++it) {
auto& log = *it;
assert(!log.getting_synced);
log.getting_synced = true;
logs_to_sync.push_back(log.writer);
}
Status s;
if (!logs_to_sync.empty()) {
mutex_.Unlock();
for (log::Writer* log : logs_to_sync) {
Log(InfoLogLevel::INFO_LEVEL, immutable_db_options_.info_log,
"[JOB %d] Syncing log #%" PRIu64, job_context->job_id,
log->get_log_number());
s = log->file()->Sync(immutable_db_options_.use_fsync);
}
if (s.ok()) {
s = directories_.GetWalDir()->Fsync();
}
mutex_.Lock();
// "number <= current_log_number - 1" is equivalent to
// "number < current_log_number".
MarkLogsSynced(current_log_number - 1, true, s);
if (!s.ok()) {
bg_error_ = s;
return s;
}
}
return s;
}
Status DBImpl::FlushMemTableToOutputFile(
ColumnFamilyData* cfd, const MutableCFOptions& mutable_cf_options,
bool* made_progress, JobContext* job_context, LogBuffer* log_buffer) {
mutex_.AssertHeld();
Support saving history in memtable_list Summary: For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit. After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much. This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit). However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions. Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here. Reviewers: sdong, rven, igor Reviewed By: igor Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D37443
9 years ago
assert(cfd->imm()->NumNotFlushed() != 0);
assert(cfd->imm()->IsFlushPending());
SequenceNumber earliest_write_conflict_snapshot;
std::vector<SequenceNumber> snapshot_seqs =
snapshots_.GetAll(&earliest_write_conflict_snapshot);
FlushJob flush_job(
dbname_, cfd, immutable_db_options_, mutable_cf_options, env_options_,
versions_.get(), &mutex_, &shutting_down_, snapshot_seqs,
earliest_write_conflict_snapshot, job_context, log_buffer,
directories_.GetDbDir(), directories_.GetDataDir(0U),
GetCompressionFlush(*cfd->ioptions(), mutable_cf_options), stats_,
&event_logger_, mutable_cf_options.report_bg_io_stats);
FileMetaData file_meta;
flush_job.PickMemTable();
Status s;
if (logfile_number_ > 0 &&
versions_->GetColumnFamilySet()->NumberOfColumnFamilies() > 0 &&
!immutable_db_options_.disable_data_sync) {
// If there are more than one column families, we need to make sure that
// all the log files except the most recent one are synced. Otherwise if
// the host crashes after flushing and before WAL is persistent, the
// flushed SST may contain data from write batches whose updates to
// other column families are missing.
// SyncClosedLogs() may unlock and re-lock the db_mutex.
s = SyncClosedLogs(job_context);
}
// Within flush_job.Run, rocksdb may call event listener to notify
// file creation and deletion.
//
// Note that flush_job.Run will unlock and lock the db_mutex,
// and EventListener callback will be called when the db_mutex
// is unlocked by the current thread.
if (s.ok()) {
s = flush_job.Run(&file_meta);
}
if (s.ok()) {
InstallSuperVersionAndScheduleWorkWrapper(cfd, job_context,
mutable_cf_options);
if (made_progress) {
*made_progress = 1;
}
VersionStorageInfo::LevelSummaryStorage tmp;
LogToBuffer(log_buffer, "[%s] Level summary: %s\n", cfd->GetName().c_str(),
cfd->current()->storage_info()->LevelSummary(&tmp));
}
if (!s.ok() && !s.IsShutdownInProgress() &&
immutable_db_options_.paranoid_checks && bg_error_.ok()) {
// if a bad error happened (not ShutdownInProgress) and paranoid_checks is
// true, mark DB read-only
bg_error_ = s;
}
if (s.ok()) {
#ifndef ROCKSDB_LITE
// may temporarily unlock and lock the mutex.
NotifyOnFlushCompleted(cfd, &file_meta, mutable_cf_options,
job_context->job_id, flush_job.GetTableProperties());
#endif // ROCKSDB_LITE
auto sfm = static_cast<SstFileManagerImpl*>(
immutable_db_options_.sst_file_manager.get());
if (sfm) {
// Notify sst_file_manager that a new file was added
std::string file_path = MakeTableFileName(
immutable_db_options_.db_paths[0].path, file_meta.fd.GetNumber());
sfm->OnAddFile(file_path);
if (sfm->IsMaxAllowedSpaceReached() && bg_error_.ok()) {
bg_error_ = Status::IOError("Max allowed space was reached");
TEST_SYNC_POINT(
"DBImpl::FlushMemTableToOutputFile:MaxAllowedSpaceReached");
}
}
}
return s;
}
void DBImpl::NotifyOnFlushCompleted(ColumnFamilyData* cfd,
FileMetaData* file_meta,
const MutableCFOptions& mutable_cf_options,
int job_id, TableProperties prop) {
#ifndef ROCKSDB_LITE
if (immutable_db_options_.listeners.size() == 0U) {
return;
}
mutex_.AssertHeld();
if (shutting_down_.load(std::memory_order_acquire)) {
return;
}
bool triggered_writes_slowdown =
(cfd->current()->storage_info()->NumLevelFiles(0) >=
mutable_cf_options.level0_slowdown_writes_trigger);
bool triggered_writes_stop =
(cfd->current()->storage_info()->NumLevelFiles(0) >=
mutable_cf_options.level0_stop_writes_trigger);
// release lock while notifying events
mutex_.Unlock();
{
FlushJobInfo info;
info.cf_name = cfd->GetName();
// TODO(yhchiang): make db_paths dynamic in case flush does not
// go to L0 in the future.
info.file_path = MakeTableFileName(immutable_db_options_.db_paths[0].path,
file_meta->fd.GetNumber());
info.thread_id = env_->GetThreadID();
info.job_id = job_id;
info.triggered_writes_slowdown = triggered_writes_slowdown;
info.triggered_writes_stop = triggered_writes_stop;
info.smallest_seqno = file_meta->smallest_seqno;
info.largest_seqno = file_meta->largest_seqno;
info.table_properties = prop;
for (auto listener : immutable_db_options_.listeners) {
listener->OnFlushCompleted(this, info);
}
}
mutex_.Lock();
// no need to signal bg_cv_ as it will be signaled at the end of the
// flush process.
#endif // ROCKSDB_LITE
}
Status DBImpl::CompactRange(const CompactRangeOptions& options,
ColumnFamilyHandle* column_family,
const Slice* begin, const Slice* end) {
if (options.target_path_id >= immutable_db_options_.db_paths.size()) {
return Status::InvalidArgument("Invalid target path ID");
}
auto cfh = reinterpret_cast<ColumnFamilyHandleImpl*>(column_family);
auto cfd = cfh->cfd();
bool exclusive = options.exclusive_manual_compaction;
Status s = FlushMemTable(cfd, FlushOptions());
if (!s.ok()) {
LogFlush(immutable_db_options_.info_log);
return s;
}
int max_level_with_files = 0;
{
InstrumentedMutexLock l(&mutex_);
Version* base = cfd->current();
for (int level = 1; level < base->storage_info()->num_non_empty_levels();
level++) {
if (base->storage_info()->OverlapInLevel(level, begin, end)) {
max_level_with_files = level;
}
}
}
int final_output_level = 0;
if (cfd->ioptions()->compaction_style == kCompactionStyleUniversal &&
cfd->NumberLevels() > 1) {
// Always compact all files together.
s = RunManualCompaction(cfd, ColumnFamilyData::kCompactAllLevels,
cfd->NumberLevels() - 1, options.target_path_id,
begin, end, exclusive);
final_output_level = cfd->NumberLevels() - 1;
} else {
for (int level = 0; level <= max_level_with_files; level++) {
int output_level;
// in case the compaction is universal or if we're compacting the
// bottom-most level, the output level will be the same as input one.
// level 0 can never be the bottommost level (i.e. if all files are in
// level 0, we will compact to level 1)
if (cfd->ioptions()->compaction_style == kCompactionStyleUniversal ||
cfd->ioptions()->compaction_style == kCompactionStyleFIFO) {
output_level = level;
} else if (level == max_level_with_files && level > 0) {
if (options.bottommost_level_compaction ==
BottommostLevelCompaction::kSkip) {
// Skip bottommost level compaction
continue;
} else if (options.bottommost_level_compaction ==
BottommostLevelCompaction::kIfHaveCompactionFilter &&
cfd->ioptions()->compaction_filter == nullptr &&
cfd->ioptions()->compaction_filter_factory == nullptr) {
// Skip bottommost level compaction since we don't have a compaction
// filter
continue;
}
output_level = level;
} else {
output_level = level + 1;
if (cfd->ioptions()->compaction_style == kCompactionStyleLevel &&
cfd->ioptions()->level_compaction_dynamic_level_bytes &&
level == 0) {
output_level = ColumnFamilyData::kCompactToBaseLevel;
}
}
s = RunManualCompaction(cfd, level, output_level, options.target_path_id,
begin, end, exclusive);
if (!s.ok()) {
break;
}
if (output_level == ColumnFamilyData::kCompactToBaseLevel) {
final_output_level = cfd->NumberLevels() - 1;
} else if (output_level > final_output_level) {
final_output_level = output_level;
}
TEST_SYNC_POINT("DBImpl::RunManualCompaction()::1");
TEST_SYNC_POINT("DBImpl::RunManualCompaction()::2");
}
}
if (!s.ok()) {
LogFlush(immutable_db_options_.info_log);
return s;
}
if (options.change_level) {
Log(InfoLogLevel::INFO_LEVEL, immutable_db_options_.info_log,
"[RefitLevel] waiting for background threads to stop");
s = PauseBackgroundWork();
if (s.ok()) {
s = ReFitLevel(cfd, final_output_level, options.target_level);
}
ContinueBackgroundWork();
}
LogFlush(immutable_db_options_.info_log);
{
InstrumentedMutexLock l(&mutex_);
// an automatic compaction that has been scheduled might have been
// preempted by the manual compactions. Need to schedule it back.
MaybeScheduleFlushOrCompaction();
}
return s;
}
Status DBImpl::CompactFiles(
const CompactionOptions& compact_options,
ColumnFamilyHandle* column_family,
const std::vector<std::string>& input_file_names,
const int output_level, const int output_path_id) {
#ifdef ROCKSDB_LITE
// not supported in lite version
return Status::NotSupported("Not supported in ROCKSDB LITE");
#else
if (column_family == nullptr) {
return Status::InvalidArgument("ColumnFamilyHandle must be non-null.");
}
auto cfd = reinterpret_cast<ColumnFamilyHandleImpl*>(column_family)->cfd();
assert(cfd);
Status s;
JobContext job_context(0, true);
LogBuffer log_buffer(InfoLogLevel::INFO_LEVEL,
immutable_db_options_.info_log.get());
// Perform CompactFiles
SuperVersion* sv = GetAndRefSuperVersion(cfd);
{
InstrumentedMutexLock l(&mutex_);
// This call will unlock/lock the mutex to wait for current running
// IngestExternalFile() calls to finish.
WaitForIngestFile();
s = CompactFilesImpl(compact_options, cfd, sv->current,
input_file_names, output_level,
output_path_id, &job_context, &log_buffer);
}
ReturnAndCleanupSuperVersion(cfd, sv);
// Find and delete obsolete files
{
InstrumentedMutexLock l(&mutex_);
// If !s.ok(), this means that Compaction failed. In that case, we want
// to delete all obsolete files we might have created and we force
// FindObsoleteFiles(). This is because job_context does not
// catch all created files if compaction failed.
FindObsoleteFiles(&job_context, !s.ok());
} // release the mutex
// delete unnecessary files if any, this is done outside the mutex
if (job_context.HaveSomethingToDelete() || !log_buffer.IsEmpty()) {
// Have to flush the info logs before bg_compaction_scheduled_--
// because if bg_flush_scheduled_ becomes 0 and the lock is
// released, the deconstructor of DB can kick in and destroy all the
// states of DB so info_log might not be available after that point.
// It also applies to access other states that DB owns.
log_buffer.FlushBufferToLog();
if (job_context.HaveSomethingToDelete()) {
// no mutex is locked here. No need to Unlock() and Lock() here.
PurgeObsoleteFiles(job_context);
}
job_context.Clean();
}
return s;
#endif // ROCKSDB_LITE
}
#ifndef ROCKSDB_LITE
Status DBImpl::CompactFilesImpl(
const CompactionOptions& compact_options, ColumnFamilyData* cfd,
Version* version, const std::vector<std::string>& input_file_names,
const int output_level, int output_path_id, JobContext* job_context,
LogBuffer* log_buffer) {
mutex_.AssertHeld();
if (shutting_down_.load(std::memory_order_acquire)) {
return Status::ShutdownInProgress();
}
std::unordered_set<uint64_t> input_set;
for (auto file_name : input_file_names) {
input_set.insert(TableFileNameToNumber(file_name));
}
ColumnFamilyMetaData cf_meta;
// TODO(yhchiang): can directly use version here if none of the
// following functions call is pluggable to external developers.
version->GetColumnFamilyMetaData(&cf_meta);
if (output_path_id < 0) {
if (immutable_db_options_.db_paths.size() == 1U) {
output_path_id = 0;
} else {
return Status::NotSupported(
"Automatic output path selection is not "
"yet supported in CompactFiles()");
}
}
Status s = cfd->compaction_picker()->SanitizeCompactionInputFiles(
&input_set, cf_meta, output_level);
if (!s.ok()) {
return s;
}
Make Compaction class easier to use Summary: The goal of this diff is to make Compaction class easier to use. This should also make new compaction algorithms easier to write (like CompactFiles from @yhchiang and dynamic leveled and multi-leveled universal from @sdong). Here are couple of things demonstrating that Compaction class is hard to use: 1. we have two constructors of Compaction class 2. there's this thing called grandparents_, but it appears to only be setup for leveled compaction and not compactfiles 3. it's easy to introduce a subtle and dangerous bug like this: D36225 4. SetupBottomMostLevel() is hard to understand and it shouldn't be. See this comment: https://github.com/facebook/rocksdb/blob/afbafeaeaebfd27a0f3e992fee8e0c57d07658fa/db/compaction.cc#L236-L241. It also made it harder for @yhchiang to write CompactFiles, as evidenced by this: https://github.com/facebook/rocksdb/blob/afbafeaeaebfd27a0f3e992fee8e0c57d07658fa/db/compaction_picker.cc#L204-L210 The problem is that we create Compaction object, which holds a lot of state, and then pass it around to some functions. After those functions are done mutating, then we call couple of functions on Compaction object, like SetupBottommostLevel() and MarkFilesBeingCompacted(). It is very hard to see what's happening with all that Compaction's state while it's travelling across different functions. If you're writing a new PickCompaction() function you need to try really hard to understand what are all the functions you need to run on Compaction object and what state you need to setup. My proposed solution is to make important parts of Compaction immutable after construction. PickCompaction() should calculate compaction inputs and then pass them onto Compaction object once they are finalized. That makes it easy to create a new compaction -- just provide all the parameters to the constructor and you're done. No need to call confusing functions after you created your object. This diff doesn't fully achieve that goal, but it comes pretty close. Here are some of the changes: * have one Compaction constructor instead of two. * inputs_ is constant after construction * MarkFilesBeingCompacted() is now private to Compaction class and automatically called on construction/destruction. * SetupBottommostLevel() is gone. Compaction figures it out on its own based on the input. * CompactionPicker's functions are not passing around Compaction object anymore. They are only passing around the state that they need. Test Plan: make check make asan_check make valgrind_check Reviewers: rven, anthony, sdong, yhchiang Reviewed By: yhchiang Subscribers: sdong, yhchiang, dhruba, leveldb Differential Revision: https://reviews.facebook.net/D36687
9 years ago
std::vector<CompactionInputFiles> input_files;
s = cfd->compaction_picker()->GetCompactionInputsFromFileNumbers(
&input_files, &input_set, version->storage_info(), compact_options);
if (!s.ok()) {
return s;
}
for (auto inputs : input_files) {
if (cfd->compaction_picker()->FilesInCompaction(inputs.files)) {
return Status::Aborted(
"Some of the necessary compaction input "
"files are already being compacted");
}
}
// At this point, CompactFiles will be run.
bg_compaction_scheduled_++;
unique_ptr<Compaction> c;
assert(cfd->compaction_picker());
c.reset(cfd->compaction_picker()->FormCompaction(
Make Compaction class easier to use Summary: The goal of this diff is to make Compaction class easier to use. This should also make new compaction algorithms easier to write (like CompactFiles from @yhchiang and dynamic leveled and multi-leveled universal from @sdong). Here are couple of things demonstrating that Compaction class is hard to use: 1. we have two constructors of Compaction class 2. there's this thing called grandparents_, but it appears to only be setup for leveled compaction and not compactfiles 3. it's easy to introduce a subtle and dangerous bug like this: D36225 4. SetupBottomMostLevel() is hard to understand and it shouldn't be. See this comment: https://github.com/facebook/rocksdb/blob/afbafeaeaebfd27a0f3e992fee8e0c57d07658fa/db/compaction.cc#L236-L241. It also made it harder for @yhchiang to write CompactFiles, as evidenced by this: https://github.com/facebook/rocksdb/blob/afbafeaeaebfd27a0f3e992fee8e0c57d07658fa/db/compaction_picker.cc#L204-L210 The problem is that we create Compaction object, which holds a lot of state, and then pass it around to some functions. After those functions are done mutating, then we call couple of functions on Compaction object, like SetupBottommostLevel() and MarkFilesBeingCompacted(). It is very hard to see what's happening with all that Compaction's state while it's travelling across different functions. If you're writing a new PickCompaction() function you need to try really hard to understand what are all the functions you need to run on Compaction object and what state you need to setup. My proposed solution is to make important parts of Compaction immutable after construction. PickCompaction() should calculate compaction inputs and then pass them onto Compaction object once they are finalized. That makes it easy to create a new compaction -- just provide all the parameters to the constructor and you're done. No need to call confusing functions after you created your object. This diff doesn't fully achieve that goal, but it comes pretty close. Here are some of the changes: * have one Compaction constructor instead of two. * inputs_ is constant after construction * MarkFilesBeingCompacted() is now private to Compaction class and automatically called on construction/destruction. * SetupBottommostLevel() is gone. Compaction figures it out on its own based on the input. * CompactionPicker's functions are not passing around Compaction object anymore. They are only passing around the state that they need. Test Plan: make check make asan_check make valgrind_check Reviewers: rven, anthony, sdong, yhchiang Reviewed By: yhchiang Subscribers: sdong, yhchiang, dhruba, leveldb Differential Revision: https://reviews.facebook.net/D36687
9 years ago
compact_options, input_files, output_level, version->storage_info(),
*cfd->GetLatestMutableCFOptions(), output_path_id));
if (!c) {
return Status::Aborted("Another Level 0 compaction is running");
}
c->SetInputVersion(version);
// deletion compaction currently not allowed in CompactFiles.
assert(!c->deletion_compaction());
SequenceNumber earliest_write_conflict_snapshot;
std::vector<SequenceNumber> snapshot_seqs =
snapshots_.GetAll(&earliest_write_conflict_snapshot);
auto pending_outputs_inserted_elem =
CaptureCurrentFileNumberInPendingOutputs();
assert(is_snapshot_supported_ || snapshots_.empty());
CompactionJob compaction_job(
job_context->job_id, c.get(), immutable_db_options_, env_options_,
versions_.get(), &shutting_down_, log_buffer, directories_.GetDbDir(),
directories_.GetDataDir(c->output_path_id()), stats_, &mutex_, &bg_error_,
snapshot_seqs, earliest_write_conflict_snapshot, table_cache_,
&event_logger_, c->mutable_cf_options()->paranoid_file_checks,
c->mutable_cf_options()->report_bg_io_stats, dbname_,
nullptr); // Here we pass a nullptr for CompactionJobStats because
// CompactFiles does not trigger OnCompactionCompleted(),
// which is the only place where CompactionJobStats is
// returned. The idea of not triggering OnCompationCompleted()
// is that CompactFiles runs in the caller thread, so the user
// should always know when it completes. As a result, it makes
// less sense to notify the users something they should already
// know.
//
// In the future, if we would like to add CompactionJobStats
// support for CompactFiles, we should have CompactFiles API
// pass a pointer of CompactionJobStats as the out-value
// instead of using EventListener.
// Creating a compaction influences the compaction score because the score
// takes running compactions into account (by skipping files that are already
// being compacted). Since we just changed compaction score, we recalculate it
// here.
version->storage_info()->ComputeCompactionScore(*cfd->ioptions(),
*c->mutable_cf_options());
compaction_job.Prepare();
mutex_.Unlock();
TEST_SYNC_POINT("CompactFilesImpl:0");
TEST_SYNC_POINT("CompactFilesImpl:1");
compaction_job.Run();
TEST_SYNC_POINT("CompactFilesImpl:2");
TEST_SYNC_POINT("CompactFilesImpl:3");
mutex_.Lock();
Status status = compaction_job.Install(*c->mutable_cf_options());
if (status.ok()) {
InstallSuperVersionAndScheduleWorkWrapper(
c->column_family_data(), job_context, *c->mutable_cf_options());
}
c->ReleaseCompactionFiles(s);
ReleaseFileNumberFromPendingOutputs(pending_outputs_inserted_elem);
if (status.ok()) {
// Done
} else if (status.IsShutdownInProgress()) {
// Ignore compaction errors found during shutting down
} else {
Log(InfoLogLevel::WARN_LEVEL, immutable_db_options_.info_log,
"[%s] [JOB %d] Compaction error: %s",
c->column_family_data()->GetName().c_str(), job_context->job_id,
status.ToString().c_str());
if (immutable_db_options_.paranoid_checks && bg_error_.ok()) {
bg_error_ = status;
}
}
c.reset();
bg_compaction_scheduled_--;
if (bg_compaction_scheduled_ == 0) {
bg_cv_.SignalAll();
}
return status;
}
#endif // ROCKSDB_LITE
Status DBImpl::PauseBackgroundWork() {
InstrumentedMutexLock guard_lock(&mutex_);
bg_compaction_paused_++;
while (bg_compaction_scheduled_ > 0 || bg_flush_scheduled_ > 0) {
bg_cv_.Wait();
}
bg_work_paused_++;
return Status::OK();
}
Status DBImpl::ContinueBackgroundWork() {
InstrumentedMutexLock guard_lock(&mutex_);
if (bg_work_paused_ == 0) {
return Status::InvalidArgument();
}
assert(bg_work_paused_ > 0);
assert(bg_compaction_paused_ > 0);
bg_compaction_paused_--;
bg_work_paused_--;
// It's sufficient to check just bg_work_paused_ here since
// bg_work_paused_ is always no greater than bg_compaction_paused_
if (bg_work_paused_ == 0) {
MaybeScheduleFlushOrCompaction();
}
return Status::OK();
}
void DBImpl::NotifyOnCompactionCompleted(
ColumnFamilyData* cfd, Compaction *c, const Status &st,
const CompactionJobStats& compaction_job_stats,
const int job_id) {
#ifndef ROCKSDB_LITE
if (immutable_db_options_.listeners.size() == 0U) {
return;
}
mutex_.AssertHeld();
if (shutting_down_.load(std::memory_order_acquire)) {
return;
}
// release lock while notifying events
mutex_.Unlock();
TEST_SYNC_POINT("DBImpl::NotifyOnCompactionCompleted::UnlockMutex");
{
CompactionJobInfo info;
info.cf_name = cfd->GetName();
info.status = st;
info.thread_id = env_->GetThreadID();
info.job_id = job_id;
info.base_input_level = c->start_level();
info.output_level = c->output_level();
info.stats = compaction_job_stats;
info.table_properties = c->GetOutputTableProperties();
info.compaction_reason = c->compaction_reason();
info.compression = c->output_compression();
for (size_t i = 0; i < c->num_input_levels(); ++i) {
for (const auto fmd : *c->inputs(i)) {
auto fn = TableFileName(immutable_db_options_.db_paths,
fmd->fd.GetNumber(), fmd->fd.GetPathId());
info.input_files.push_back(fn);
if (info.table_properties.count(fn) == 0) {
std::shared_ptr<const TableProperties> tp;
auto s = cfd->current()->GetTableProperties(&tp, fmd, &fn);
if (s.ok()) {
info.table_properties[fn] = tp;
}
}
}
}
for (const auto newf : c->edit()->GetNewFiles()) {
info.output_files.push_back(TableFileName(immutable_db_options_.db_paths,
newf.second.fd.GetNumber(),
newf.second.fd.GetPathId()));
}
for (auto listener : immutable_db_options_.listeners) {
listener->OnCompactionCompleted(this, info);
}
}
mutex_.Lock();
// no need to signal bg_cv_ as it will be signaled at the end of the
// flush process.
#endif // ROCKSDB_LITE
}
Status DBImpl::SetOptions(ColumnFamilyHandle* column_family,
const std::unordered_map<std::string, std::string>& options_map) {
#ifdef ROCKSDB_LITE
return Status::NotSupported("Not supported in ROCKSDB LITE");
#else
auto* cfd = reinterpret_cast<ColumnFamilyHandleImpl*>(column_family)->cfd();
if (options_map.empty()) {
Log(InfoLogLevel::WARN_LEVEL, immutable_db_options_.info_log,
"SetOptions() on column family [%s], empty input",
cfd->GetName().c_str());
return Status::InvalidArgument("empty input");
}
MutableCFOptions new_options;
Status s;
Status persist_options_status;
{
InstrumentedMutexLock l(&mutex_);
s = cfd->SetOptions(options_map);
if (s.ok()) {
new_options = *cfd->GetLatestMutableCFOptions();
// Append new version to recompute compaction score.
VersionEdit dummy_edit;
versions_->LogAndApply(cfd, new_options, &dummy_edit, &mutex_,
directories_.GetDbDir());
// Trigger possible flush/compactions. This has to be before we persist
// options to file, otherwise there will be a deadlock with writer
// thread.
auto* old_sv =
InstallSuperVersionAndScheduleWork(cfd, nullptr, new_options);
delete old_sv;
persist_options_status = PersistOptions();
}
}
Log(InfoLogLevel::INFO_LEVEL, immutable_db_options_.info_log,
"SetOptions() on column family [%s], inputs:", cfd->GetName().c_str());
for (const auto& o : options_map) {
Log(InfoLogLevel::INFO_LEVEL, immutable_db_options_.info_log, "%s: %s\n",
o.first.c_str(), o.second.c_str());
}
if (s.ok()) {
Log(InfoLogLevel::INFO_LEVEL, immutable_db_options_.info_log,
"[%s] SetOptions() succeeded", cfd->GetName().c_str());
new_options.Dump(immutable_db_options_.info_log.get());
if (!persist_options_status.ok()) {
if (immutable_db_options_.fail_if_options_file_error) {
s = Status::IOError(
"SetOptions() succeeded, but unable to persist options",
persist_options_status.ToString());
}
Warn(immutable_db_options_.info_log,
"Unable to persist options in SetOptions() -- %s",
persist_options_status.ToString().c_str());
}
} else {
Log(InfoLogLevel::WARN_LEVEL, immutable_db_options_.info_log,
"[%s] SetOptions() failed", cfd->GetName().c_str());
}
LogFlush(immutable_db_options_.info_log);
return s;
#endif // ROCKSDB_LITE
}
Status DBImpl::SetDBOptions(
const std::unordered_map<std::string, std::string>& options_map) {
#ifdef ROCKSDB_LITE
return Status::NotSupported("Not supported in ROCKSDB LITE");
#else
if (options_map.empty()) {
Log(InfoLogLevel::WARN_LEVEL, immutable_db_options_.info_log,
"SetDBOptions(), empty input.");
return Status::InvalidArgument("empty input");
}
MutableDBOptions new_options;
Status s;
Status persist_options_status;
{
InstrumentedMutexLock l(&mutex_);
s = GetMutableDBOptionsFromStrings(mutable_db_options_, options_map,
&new_options);
if (s.ok()) {
if (new_options.max_background_compactions >
mutable_db_options_.max_background_compactions) {
env_->IncBackgroundThreadsIfNeeded(
new_options.max_background_compactions, Env::Priority::LOW);
MaybeScheduleFlushOrCompaction();
}
mutable_db_options_ = new_options;
persist_options_status = PersistOptions();
}
}
Log(InfoLogLevel::INFO_LEVEL, immutable_db_options_.info_log,
"SetDBOptions(), inputs:");
for (const auto& o : options_map) {
Log(InfoLogLevel::INFO_LEVEL, immutable_db_options_.info_log, "%s: %s\n",
o.first.c_str(), o.second.c_str());
}
if (s.ok()) {
Log(InfoLogLevel::INFO_LEVEL, immutable_db_options_.info_log,
"SetDBOptions() succeeded");
new_options.Dump(immutable_db_options_.info_log.get());
if (!persist_options_status.ok()) {
if (immutable_db_options_.fail_if_options_file_error) {
s = Status::IOError(
"SetDBOptions() succeeded, but unable to persist options",
persist_options_status.ToString());
}
Warn(immutable_db_options_.info_log,
"Unable to persist options in SetDBOptions() -- %s",
persist_options_status.ToString().c_str());
}
} else {
Log(InfoLogLevel::WARN_LEVEL, immutable_db_options_.info_log,
"SetDBOptions failed");
}
LogFlush(immutable_db_options_.info_log);
return s;
#endif // ROCKSDB_LITE
}
Status DBImpl::PersistOptions() {
mutex_.AssertHeld();
WriteThread::Writer w;
write_thread_.EnterUnbatched(&w, &mutex_);
Status s = WriteOptionsFile();
write_thread_.ExitUnbatched(&w);
return s;
}
// return the same level if it cannot be moved
int DBImpl::FindMinimumEmptyLevelFitting(ColumnFamilyData* cfd,
const MutableCFOptions& mutable_cf_options, int level) {
mutex_.AssertHeld();
const auto* vstorage = cfd->current()->storage_info();
int minimum_level = level;
for (int i = level - 1; i > 0; --i) {
// stop if level i is not empty
if (vstorage->NumLevelFiles(i) > 0) break;
// stop if level i is too small (cannot fit the level files)
if (vstorage->MaxBytesForLevel(i) < vstorage->NumLevelBytes(level)) {
break;
}
minimum_level = i;
}
return minimum_level;
}
// REQUIREMENT: block all background work by calling PauseBackgroundWork()
// before calling this function
Status DBImpl::ReFitLevel(ColumnFamilyData* cfd, int level, int target_level) {
assert(level < cfd->NumberLevels());
if (target_level >= cfd->NumberLevels()) {
return Status::InvalidArgument("Target level exceeds number of levels");
}
std::unique_ptr<SuperVersion> superversion_to_free;
std::unique_ptr<SuperVersion> new_superversion(new SuperVersion());
Status status;
InstrumentedMutexLock guard_lock(&mutex_);
// only allow one thread refitting
if (refitting_level_) {
Log(InfoLogLevel::INFO_LEVEL, immutable_db_options_.info_log,
"[ReFitLevel] another thread is refitting");
return Status::NotSupported("another thread is refitting");
}
refitting_level_ = true;
const MutableCFOptions mutable_cf_options = *cfd->GetLatestMutableCFOptions();
// move to a smaller level
int to_level = target_level;
if (target_level < 0) {
to_level = FindMinimumEmptyLevelFitting(cfd, mutable_cf_options, level);
}
auto* vstorage = cfd->current()->storage_info();
if (to_level > level) {
if (level == 0) {
return Status::NotSupported(
"Cannot change from level 0 to other levels.");
}
// Check levels are empty for a trivial move
for (int l = level + 1; l <= to_level; l++) {
if (vstorage->NumLevelFiles(l) > 0) {
return Status::NotSupported(
"Levels between source and target are not empty for a move.");
}
}
}
if (to_level != level) {
Log(InfoLogLevel::DEBUG_LEVEL, immutable_db_options_.info_log,
"[%s] Before refitting:\n%s", cfd->GetName().c_str(),
cfd->current()->DebugString().data());
VersionEdit edit;
edit.SetColumnFamily(cfd->GetID());
for (const auto& f : vstorage->LevelFiles(level)) {
edit.DeleteFile(level, f->fd.GetNumber());
edit.AddFile(to_level, f->fd.GetNumber(), f->fd.GetPathId(),
f->fd.GetFileSize(), f->smallest, f->largest,
f->smallest_seqno, f->largest_seqno,
f->marked_for_compaction);
}
Log(InfoLogLevel::DEBUG_LEVEL, immutable_db_options_.info_log,
"[%s] Apply version edit:\n%s", cfd->GetName().c_str(),
edit.DebugString().data());
status = versions_->LogAndApply(cfd, mutable_cf_options, &edit, &mutex_,
directories_.GetDbDir());
superversion_to_free.reset(InstallSuperVersionAndScheduleWork(
cfd, new_superversion.release(), mutable_cf_options));
Log(InfoLogLevel::DEBUG_LEVEL, immutable_db_options_.info_log,
"[%s] LogAndApply: %s\n", cfd->GetName().c_str(),
status.ToString().data());
if (status.ok()) {
Log(InfoLogLevel::DEBUG_LEVEL, immutable_db_options_.info_log,
"[%s] After refitting:\n%s", cfd->GetName().c_str(),
cfd->current()->DebugString().data());
}
}
refitting_level_ = false;
return status;
}
int DBImpl::NumberLevels(ColumnFamilyHandle* column_family) {
auto cfh = reinterpret_cast<ColumnFamilyHandleImpl*>(column_family);
return cfh->cfd()->NumberLevels();
}
int DBImpl::MaxMemCompactionLevel(ColumnFamilyHandle* column_family) {
return 0;
}
int DBImpl::Level0StopWriteTrigger(ColumnFamilyHandle* column_family) {
auto cfh = reinterpret_cast<ColumnFamilyHandleImpl*>(column_family);
InstrumentedMutexLock l(&mutex_);
return cfh->cfd()->GetSuperVersion()->
mutable_cf_options.level0_stop_writes_trigger;
}
Status DBImpl::Flush(const FlushOptions& flush_options,
ColumnFamilyHandle* column_family) {
auto cfh = reinterpret_cast<ColumnFamilyHandleImpl*>(column_family);
return FlushMemTable(cfh->cfd(), flush_options);
}
Status DBImpl::SyncWAL() {
autovector<log::Writer*, 1> logs_to_sync;
bool need_log_dir_sync;
uint64_t current_log_number;
{
InstrumentedMutexLock l(&mutex_);
assert(!logs_.empty());
// This SyncWAL() call only cares about logs up to this number.
current_log_number = logfile_number_;
while (logs_.front().number <= current_log_number &&
logs_.front().getting_synced) {
log_sync_cv_.Wait();
}
// First check that logs are safe to sync in background.
for (auto it = logs_.begin();
it != logs_.end() && it->number <= current_log_number; ++it) {
if (!it->writer->file()->writable_file()->IsSyncThreadSafe()) {
return Status::NotSupported(
"SyncWAL() is not supported for this implementation of WAL file",
immutable_db_options_.allow_mmap_writes
? "try setting Options::allow_mmap_writes to false"
: Slice());
}
}
for (auto it = logs_.begin();
it != logs_.end() && it->number <= current_log_number; ++it) {
auto& log = *it;
assert(!log.getting_synced);
log.getting_synced = true;
logs_to_sync.push_back(log.writer);
}
need_log_dir_sync = !log_dir_synced_;
}
RecordTick(stats_, WAL_FILE_SYNCED);
Status status;
for (log::Writer* log : logs_to_sync) {
status = log->file()->SyncWithoutFlush(immutable_db_options_.use_fsync);
if (!status.ok()) {
break;
}
}
if (status.ok() && need_log_dir_sync) {
status = directories_.GetWalDir()->Fsync();
}
TEST_SYNC_POINT("DBImpl::SyncWAL:BeforeMarkLogsSynced:1");
{
InstrumentedMutexLock l(&mutex_);
MarkLogsSynced(current_log_number, need_log_dir_sync, status);
}
TEST_SYNC_POINT("DBImpl::SyncWAL:BeforeMarkLogsSynced:2");
return status;
}
void DBImpl::MarkLogsSynced(
uint64_t up_to, bool synced_dir, const Status& status) {
mutex_.AssertHeld();
if (synced_dir &&
logfile_number_ == up_to &&
status.ok()) {
log_dir_synced_ = true;
}
for (auto it = logs_.begin(); it != logs_.end() && it->number <= up_to;) {
auto& log = *it;
assert(log.getting_synced);
if (status.ok() && logs_.size() > 1) {
logs_to_free_.push_back(log.ReleaseWriter());
it = logs_.erase(it);
} else {
log.getting_synced = false;
++it;
}
}
assert(logs_.empty() || logs_[0].number > up_to ||
(logs_.size() == 1 && !logs_[0].getting_synced));
log_sync_cv_.SignalAll();
}
SequenceNumber DBImpl::GetLatestSequenceNumber() const {
return versions_->LastSequence();
}
Status DBImpl::RunManualCompaction(ColumnFamilyData* cfd, int input_level,
int output_level, uint32_t output_path_id,
Allowing L0 -> L1 trivial move on sorted data Summary: This diff updates the logic of how we do trivial move, now trivial move can run on any number of files in input level as long as they are not overlapping The conditions for trivial move have been updated Introduced conditions: - Trivial move cannot happen if we have a compaction filter (except if the compaction is not manual) - Input level files cannot be overlapping Removed conditions: - Trivial move only run when the compaction is not manual - Input level should can contain only 1 file More context on what tests failed because of Trivial move ``` DBTest.CompactionsGenerateMultipleFiles This test is expecting compaction on a file in L0 to generate multiple files in L1, this test will fail with trivial move because we end up with one file in L1 ``` ``` DBTest.NoSpaceCompactRange This test expect compaction to fail when we force environment to report running out of space, of course this is not valid in trivial move situation because trivial move does not need any extra space, and did not check for that ``` ``` DBTest.DropWrites Similar to DBTest.NoSpaceCompactRange ``` ``` DBTest.DeleteObsoleteFilesPendingOutputs This test expect that a file in L2 is deleted after it's moved to L3, this is not valid with trivial move because although the file was moved it is now used by L3 ``` ``` CuckooTableDBTest.CompactionIntoMultipleFiles Same as DBTest.CompactionsGenerateMultipleFiles ``` This diff is based on a work by @sdong https://reviews.facebook.net/D34149 Test Plan: make -j64 check Reviewers: rven, sdong, igor Reviewed By: igor Subscribers: yhchiang, ott, march, dhruba, sdong Differential Revision: https://reviews.facebook.net/D34797
9 years ago
const Slice* begin, const Slice* end,
bool exclusive, bool disallow_trivial_move) {
assert(input_level == ColumnFamilyData::kCompactAllLevels ||
input_level >= 0);
InternalKey begin_storage, end_storage;
CompactionArg* ca;
bool scheduled = false;
bool manual_conflict = false;
ManualCompaction manual;
manual.cfd = cfd;
manual.input_level = input_level;
manual.output_level = output_level;
manual.output_path_id = output_path_id;
manual.done = false;
manual.in_progress = false;
manual.incomplete = false;
manual.exclusive = exclusive;
Allowing L0 -> L1 trivial move on sorted data Summary: This diff updates the logic of how we do trivial move, now trivial move can run on any number of files in input level as long as they are not overlapping The conditions for trivial move have been updated Introduced conditions: - Trivial move cannot happen if we have a compaction filter (except if the compaction is not manual) - Input level files cannot be overlapping Removed conditions: - Trivial move only run when the compaction is not manual - Input level should can contain only 1 file More context on what tests failed because of Trivial move ``` DBTest.CompactionsGenerateMultipleFiles This test is expecting compaction on a file in L0 to generate multiple files in L1, this test will fail with trivial move because we end up with one file in L1 ``` ``` DBTest.NoSpaceCompactRange This test expect compaction to fail when we force environment to report running out of space, of course this is not valid in trivial move situation because trivial move does not need any extra space, and did not check for that ``` ``` DBTest.DropWrites Similar to DBTest.NoSpaceCompactRange ``` ``` DBTest.DeleteObsoleteFilesPendingOutputs This test expect that a file in L2 is deleted after it's moved to L3, this is not valid with trivial move because although the file was moved it is now used by L3 ``` ``` CuckooTableDBTest.CompactionIntoMultipleFiles Same as DBTest.CompactionsGenerateMultipleFiles ``` This diff is based on a work by @sdong https://reviews.facebook.net/D34149 Test Plan: make -j64 check Reviewers: rven, sdong, igor Reviewed By: igor Subscribers: yhchiang, ott, march, dhruba, sdong Differential Revision: https://reviews.facebook.net/D34797
9 years ago
manual.disallow_trivial_move = disallow_trivial_move;
// For universal compaction, we enforce every manual compaction to compact
// all files.
if (begin == nullptr ||
cfd->ioptions()->compaction_style == kCompactionStyleUniversal ||
cfd->ioptions()->compaction_style == kCompactionStyleFIFO) {
manual.begin = nullptr;
} else {
begin_storage.SetMaxPossibleForUserKey(*begin);
manual.begin = &begin_storage;
}
if (end == nullptr ||
cfd->ioptions()->compaction_style == kCompactionStyleUniversal ||
cfd->ioptions()->compaction_style == kCompactionStyleFIFO) {
manual.end = nullptr;
} else {
end_storage.SetMinPossibleForUserKey(*end);
manual.end = &end_storage;
}
TEST_SYNC_POINT("DBImpl::RunManualCompaction:0");
TEST_SYNC_POINT("DBImpl::RunManualCompaction:1");
InstrumentedMutexLock l(&mutex_);
Fix a deadlock in CompactRange() Summary: The way DBImpl::TEST_CompactRange() throttles down the number of bg compactions can cause it to deadlock when CompactRange() is called concurrently from multiple threads. Imagine a following scenario with only two threads (max_background_compactions is 10 and bg_compaction_scheduled_ is initially 0): 1. Thread #1 increments bg_compaction_scheduled_ (to LargeNumber), sets bg_compaction_scheduled_ to 9 (newvalue), schedules the compaction (bg_compaction_scheduled_ is now 10) and waits for it to complete. 2. Thread #2 calls TEST_CompactRange(), increments bg_compaction_scheduled_ (now LargeNumber + 10) and waits on a cv for bg_compaction_scheduled_ to drop to LargeNumber. 3. BG thread completes the first manual compaction, decrements bg_compaction_scheduled_ and wakes up all threads waiting on bg_cv_. Thread #1 runs, increments bg_compaction_scheduled_ by LargeNumber again (now 2*LargeNumber + 9). Since that's more than LargeNumber + newvalue, thread #2 also goes to sleep (waiting on bg_cv_), without resetting bg_compaction_scheduled_. This diff attempts to address the problem by introducing a new counter bg_manual_only_ (when positive, MaybeScheduleFlushOrCompaction() will only schedule manual compactions). Test Plan: I could pretty much consistently reproduce the deadlock with a program that calls CompactRange(nullptr, nullptr) immediately after Write() from multiple threads. This no longer happens with this patch. Tests (make check) pass. Reviewers: dhruba, igor, sdong, haobo Reviewed By: igor CC: leveldb Differential Revision: https://reviews.facebook.net/D14799
11 years ago
// When a manual compaction arrives, temporarily disable scheduling of
// non-manual compactions and wait until the number of scheduled compaction
// jobs drops to zero. This is needed to ensure that this manual compaction
// can compact any range of keys/files.
//
// HasPendingManualCompaction() is true when at least one thread is inside
// RunManualCompaction(), i.e. during that time no other compaction will
Fix a deadlock in CompactRange() Summary: The way DBImpl::TEST_CompactRange() throttles down the number of bg compactions can cause it to deadlock when CompactRange() is called concurrently from multiple threads. Imagine a following scenario with only two threads (max_background_compactions is 10 and bg_compaction_scheduled_ is initially 0): 1. Thread #1 increments bg_compaction_scheduled_ (to LargeNumber), sets bg_compaction_scheduled_ to 9 (newvalue), schedules the compaction (bg_compaction_scheduled_ is now 10) and waits for it to complete. 2. Thread #2 calls TEST_CompactRange(), increments bg_compaction_scheduled_ (now LargeNumber + 10) and waits on a cv for bg_compaction_scheduled_ to drop to LargeNumber. 3. BG thread completes the first manual compaction, decrements bg_compaction_scheduled_ and wakes up all threads waiting on bg_cv_. Thread #1 runs, increments bg_compaction_scheduled_ by LargeNumber again (now 2*LargeNumber + 9). Since that's more than LargeNumber + newvalue, thread #2 also goes to sleep (waiting on bg_cv_), without resetting bg_compaction_scheduled_. This diff attempts to address the problem by introducing a new counter bg_manual_only_ (when positive, MaybeScheduleFlushOrCompaction() will only schedule manual compactions). Test Plan: I could pretty much consistently reproduce the deadlock with a program that calls CompactRange(nullptr, nullptr) immediately after Write() from multiple threads. This no longer happens with this patch. Tests (make check) pass. Reviewers: dhruba, igor, sdong, haobo Reviewed By: igor CC: leveldb Differential Revision: https://reviews.facebook.net/D14799
11 years ago
// get scheduled (see MaybeScheduleFlushOrCompaction).
//
// Note that the following loop doesn't stop more that one thread calling
// RunManualCompaction() from getting to the second while loop below.
Fix a deadlock in CompactRange() Summary: The way DBImpl::TEST_CompactRange() throttles down the number of bg compactions can cause it to deadlock when CompactRange() is called concurrently from multiple threads. Imagine a following scenario with only two threads (max_background_compactions is 10 and bg_compaction_scheduled_ is initially 0): 1. Thread #1 increments bg_compaction_scheduled_ (to LargeNumber), sets bg_compaction_scheduled_ to 9 (newvalue), schedules the compaction (bg_compaction_scheduled_ is now 10) and waits for it to complete. 2. Thread #2 calls TEST_CompactRange(), increments bg_compaction_scheduled_ (now LargeNumber + 10) and waits on a cv for bg_compaction_scheduled_ to drop to LargeNumber. 3. BG thread completes the first manual compaction, decrements bg_compaction_scheduled_ and wakes up all threads waiting on bg_cv_. Thread #1 runs, increments bg_compaction_scheduled_ by LargeNumber again (now 2*LargeNumber + 9). Since that's more than LargeNumber + newvalue, thread #2 also goes to sleep (waiting on bg_cv_), without resetting bg_compaction_scheduled_. This diff attempts to address the problem by introducing a new counter bg_manual_only_ (when positive, MaybeScheduleFlushOrCompaction() will only schedule manual compactions). Test Plan: I could pretty much consistently reproduce the deadlock with a program that calls CompactRange(nullptr, nullptr) immediately after Write() from multiple threads. This no longer happens with this patch. Tests (make check) pass. Reviewers: dhruba, igor, sdong, haobo Reviewed By: igor CC: leveldb Differential Revision: https://reviews.facebook.net/D14799
11 years ago
// However, only one of them will actually schedule compaction, while
// others will wait on a condition variable until it completes.
AddManualCompaction(&manual);
TEST_SYNC_POINT_CALLBACK("DBImpl::RunManualCompaction:NotScheduled", &mutex_);
if (exclusive) {
while (bg_compaction_scheduled_ > 0) {
Log(InfoLogLevel::INFO_LEVEL, immutable_db_options_.info_log,
"[%s] Manual compaction waiting for all other scheduled background "
"compactions to finish",
cfd->GetName().c_str());
bg_cv_.Wait();
}
}
Fix a deadlock in CompactRange() Summary: The way DBImpl::TEST_CompactRange() throttles down the number of bg compactions can cause it to deadlock when CompactRange() is called concurrently from multiple threads. Imagine a following scenario with only two threads (max_background_compactions is 10 and bg_compaction_scheduled_ is initially 0): 1. Thread #1 increments bg_compaction_scheduled_ (to LargeNumber), sets bg_compaction_scheduled_ to 9 (newvalue), schedules the compaction (bg_compaction_scheduled_ is now 10) and waits for it to complete. 2. Thread #2 calls TEST_CompactRange(), increments bg_compaction_scheduled_ (now LargeNumber + 10) and waits on a cv for bg_compaction_scheduled_ to drop to LargeNumber. 3. BG thread completes the first manual compaction, decrements bg_compaction_scheduled_ and wakes up all threads waiting on bg_cv_. Thread #1 runs, increments bg_compaction_scheduled_ by LargeNumber again (now 2*LargeNumber + 9). Since that's more than LargeNumber + newvalue, thread #2 also goes to sleep (waiting on bg_cv_), without resetting bg_compaction_scheduled_. This diff attempts to address the problem by introducing a new counter bg_manual_only_ (when positive, MaybeScheduleFlushOrCompaction() will only schedule manual compactions). Test Plan: I could pretty much consistently reproduce the deadlock with a program that calls CompactRange(nullptr, nullptr) immediately after Write() from multiple threads. This no longer happens with this patch. Tests (make check) pass. Reviewers: dhruba, igor, sdong, haobo Reviewed By: igor CC: leveldb Differential Revision: https://reviews.facebook.net/D14799
11 years ago
Log(InfoLogLevel::INFO_LEVEL, immutable_db_options_.info_log,
"[%s] Manual compaction starting", cfd->GetName().c_str());
// We don't check bg_error_ here, because if we get the error in compaction,
// the compaction will set manual.status to bg_error_ and set manual.done to
// true.
while (!manual.done) {
assert(HasPendingManualCompaction());
manual_conflict = false;
if (ShouldntRunManualCompaction(&manual) || (manual.in_progress == true) ||
scheduled ||
((manual.manual_end = &manual.tmp_storage1)&&(
(manual.compaction = manual.cfd->CompactRange(
*manual.cfd->GetLatestMutableCFOptions(), manual.input_level,
manual.output_level, manual.output_path_id, manual.begin,
manual.end, &manual.manual_end, &manual_conflict)) ==
nullptr) &&
manual_conflict)) {
// exclusive manual compactions should not see a conflict during
// CompactRange
assert(!exclusive || !manual_conflict);
Fix a deadlock in CompactRange() Summary: The way DBImpl::TEST_CompactRange() throttles down the number of bg compactions can cause it to deadlock when CompactRange() is called concurrently from multiple threads. Imagine a following scenario with only two threads (max_background_compactions is 10 and bg_compaction_scheduled_ is initially 0): 1. Thread #1 increments bg_compaction_scheduled_ (to LargeNumber), sets bg_compaction_scheduled_ to 9 (newvalue), schedules the compaction (bg_compaction_scheduled_ is now 10) and waits for it to complete. 2. Thread #2 calls TEST_CompactRange(), increments bg_compaction_scheduled_ (now LargeNumber + 10) and waits on a cv for bg_compaction_scheduled_ to drop to LargeNumber. 3. BG thread completes the first manual compaction, decrements bg_compaction_scheduled_ and wakes up all threads waiting on bg_cv_. Thread #1 runs, increments bg_compaction_scheduled_ by LargeNumber again (now 2*LargeNumber + 9). Since that's more than LargeNumber + newvalue, thread #2 also goes to sleep (waiting on bg_cv_), without resetting bg_compaction_scheduled_. This diff attempts to address the problem by introducing a new counter bg_manual_only_ (when positive, MaybeScheduleFlushOrCompaction() will only schedule manual compactions). Test Plan: I could pretty much consistently reproduce the deadlock with a program that calls CompactRange(nullptr, nullptr) immediately after Write() from multiple threads. This no longer happens with this patch. Tests (make check) pass. Reviewers: dhruba, igor, sdong, haobo Reviewed By: igor CC: leveldb Differential Revision: https://reviews.facebook.net/D14799
11 years ago
// Running either this or some other manual compaction
bg_cv_.Wait();
if (scheduled && manual.incomplete == true) {
assert(!manual.in_progress);
scheduled = false;
manual.incomplete = false;
}
} else if (!scheduled) {
if (manual.compaction == nullptr) {
manual.done = true;
bg_cv_.SignalAll();
continue;
}
ca = new CompactionArg;
ca->db = this;
ca->m = &manual;
manual.incomplete = false;
bg_compaction_scheduled_++;
env_->Schedule(&DBImpl::BGWorkCompaction, ca, Env::Priority::LOW, this,
&DBImpl::UnscheduleCallback);
scheduled = true;
}
}
Fix a deadlock in CompactRange() Summary: The way DBImpl::TEST_CompactRange() throttles down the number of bg compactions can cause it to deadlock when CompactRange() is called concurrently from multiple threads. Imagine a following scenario with only two threads (max_background_compactions is 10 and bg_compaction_scheduled_ is initially 0): 1. Thread #1 increments bg_compaction_scheduled_ (to LargeNumber), sets bg_compaction_scheduled_ to 9 (newvalue), schedules the compaction (bg_compaction_scheduled_ is now 10) and waits for it to complete. 2. Thread #2 calls TEST_CompactRange(), increments bg_compaction_scheduled_ (now LargeNumber + 10) and waits on a cv for bg_compaction_scheduled_ to drop to LargeNumber. 3. BG thread completes the first manual compaction, decrements bg_compaction_scheduled_ and wakes up all threads waiting on bg_cv_. Thread #1 runs, increments bg_compaction_scheduled_ by LargeNumber again (now 2*LargeNumber + 9). Since that's more than LargeNumber + newvalue, thread #2 also goes to sleep (waiting on bg_cv_), without resetting bg_compaction_scheduled_. This diff attempts to address the problem by introducing a new counter bg_manual_only_ (when positive, MaybeScheduleFlushOrCompaction() will only schedule manual compactions). Test Plan: I could pretty much consistently reproduce the deadlock with a program that calls CompactRange(nullptr, nullptr) immediately after Write() from multiple threads. This no longer happens with this patch. Tests (make check) pass. Reviewers: dhruba, igor, sdong, haobo Reviewed By: igor CC: leveldb Differential Revision: https://reviews.facebook.net/D14799
11 years ago
assert(!manual.in_progress);
assert(HasPendingManualCompaction());
RemoveManualCompaction(&manual);
bg_cv_.SignalAll();
return manual.status;
}
InternalIterator* DBImpl::NewInternalIterator(
Arena* arena, ColumnFamilyHandle* column_family) {
ColumnFamilyData* cfd;
if (column_family == nullptr) {
cfd = default_cf_handle_->cfd();
} else {
auto cfh = reinterpret_cast<ColumnFamilyHandleImpl*>(column_family);
cfd = cfh->cfd();
}
mutex_.Lock();
SuperVersion* super_version = cfd->GetSuperVersion()->Ref();
mutex_.Unlock();
ReadOptions roptions;
return NewInternalIterator(roptions, cfd, super_version, arena);
}
Status DBImpl::FlushMemTable(ColumnFamilyData* cfd,
const FlushOptions& flush_options,
bool writes_stopped) {
Status s;
{
WriteContext context;
InstrumentedMutexLock guard_lock(&mutex_);
Support saving history in memtable_list Summary: For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit. After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much. This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit). However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions. Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here. Reviewers: sdong, rven, igor Reviewed By: igor Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D37443
9 years ago
if (cfd->imm()->NumNotFlushed() == 0 && cfd->mem()->IsEmpty()) {
// Nothing to flush
return Status::OK();
}
WriteThread::Writer w;
if (!writes_stopped) {
write_thread_.EnterUnbatched(&w, &mutex_);
}
// SwitchMemtable() will release and reacquire mutex
// during execution
s = SwitchMemtable(cfd, &context);
if (!writes_stopped) {
write_thread_.ExitUnbatched(&w);
}
Rewritten system for scheduling background work Summary: When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue. The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction. Here are the performance results: Command: ./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000 --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333 Before the patch: fillrandom : 26.950 micros/op 37105 ops/sec; 4.1 MB/s After the patch: fillrandom : 17.404 micros/op 57456 ops/sec; 6.4 MB/s Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got: fillrandom : 7.590 micros/op 131758 ops/sec; 14.6 MB/s Test Plan: make check two stress tests: Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 max_background_flushes=0, to verify that this case also works correctly ./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: ljin, rven, yhchiang, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30123
10 years ago
cfd->imm()->FlushRequested();
Rewritten system for scheduling background work Summary: When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue. The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction. Here are the performance results: Command: ./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000 --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333 Before the patch: fillrandom : 26.950 micros/op 37105 ops/sec; 4.1 MB/s After the patch: fillrandom : 17.404 micros/op 57456 ops/sec; 6.4 MB/s Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got: fillrandom : 7.590 micros/op 131758 ops/sec; 14.6 MB/s Test Plan: make check two stress tests: Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 max_background_flushes=0, to verify that this case also works correctly ./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: ljin, rven, yhchiang, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30123
10 years ago
// schedule flush
SchedulePendingFlush(cfd);
MaybeScheduleFlushOrCompaction();
}
if (s.ok() && flush_options.wait) {
// Wait until the compaction completes
s = WaitForFlushMemTable(cfd);
}
return s;
}
Status DBImpl::WaitForFlushMemTable(ColumnFamilyData* cfd) {
Status s;
// Wait until the compaction completes
InstrumentedMutexLock l(&mutex_);
Support saving history in memtable_list Summary: For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit. After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much. This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit). However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions. Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here. Reviewers: sdong, rven, igor Reviewed By: igor Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D37443
9 years ago
while (cfd->imm()->NumNotFlushed() > 0 && bg_error_.ok()) {
if (shutting_down_.load(std::memory_order_acquire)) {
return Status::ShutdownInProgress();
}
bg_cv_.Wait();
}
if (!bg_error_.ok()) {
s = bg_error_;
}
return s;
}
Status DBImpl::EnableAutoCompaction(
const std::vector<ColumnFamilyHandle*>& column_family_handles) {
Status s;
for (auto cf_ptr : column_family_handles) {
add call to install superversion and schedule work in enableautocompactions Summary: This patch fixes https://github.com/facebook/mysql-5.6/issues/121 There is a recent change in rocksdb to disable auto compactions on startup: https://reviews.facebook.net/D51147. However, there is a small timing window where a column family needs to be compacted and schedules a compaction, but the scheduled compaction fails when it checks the disable_auto_compactions setting. The expectation is once the application is ready, it will call EnableAutoCompactions() to allow new compactions to go through. However, if the Column family is stalled because L0 is full, and no writes can go through, it is possible the column family may never have a new compaction request get scheduled. EnableAutoCompaction() should probably schedule an new flush and compaction event when it resets disable_auto_compaction. Using InstallSuperVersionAndScheduleWork, we call SchedulePendingFlush, SchedulePendingCompaction, as well as MaybeScheduleFlushOrcompaction on all the column families to avoid the situation above. This is still a first pass for feedback. Could also just call SchedePendingFlush and SchedulePendingCompaction directly. Test Plan: Run on Asan build cd _build-5.6-ASan/ && ./mysql-test/mtr --mem --big --testcase-timeout=36000 --suite-timeout=12000 --parallel=16 --suite=rocksdb,rocksdb_rpl,rocksdb_sys_vars --mysqld=--default-storage-engine=rocksdb --mysqld=--skip-innodb --mysqld=--default-tmp-storage-engine=MyISAM --mysqld=--rocksdb rocksdb_rpl.rpl_rocksdb_stress_crash --repeat=1000 Ensure that it no longer hangs during the test. Reviewers: hermanlee4, yhchiang, anthony Reviewed By: anthony Subscribers: leveldb, yhchiang, dhruba Differential Revision: https://reviews.facebook.net/D51747
9 years ago
Status status =
this->SetOptions(cf_ptr, {{"disable_auto_compactions", "false"}});
if (!status.ok()) {
add call to install superversion and schedule work in enableautocompactions Summary: This patch fixes https://github.com/facebook/mysql-5.6/issues/121 There is a recent change in rocksdb to disable auto compactions on startup: https://reviews.facebook.net/D51147. However, there is a small timing window where a column family needs to be compacted and schedules a compaction, but the scheduled compaction fails when it checks the disable_auto_compactions setting. The expectation is once the application is ready, it will call EnableAutoCompactions() to allow new compactions to go through. However, if the Column family is stalled because L0 is full, and no writes can go through, it is possible the column family may never have a new compaction request get scheduled. EnableAutoCompaction() should probably schedule an new flush and compaction event when it resets disable_auto_compaction. Using InstallSuperVersionAndScheduleWork, we call SchedulePendingFlush, SchedulePendingCompaction, as well as MaybeScheduleFlushOrcompaction on all the column families to avoid the situation above. This is still a first pass for feedback. Could also just call SchedePendingFlush and SchedulePendingCompaction directly. Test Plan: Run on Asan build cd _build-5.6-ASan/ && ./mysql-test/mtr --mem --big --testcase-timeout=36000 --suite-timeout=12000 --parallel=16 --suite=rocksdb,rocksdb_rpl,rocksdb_sys_vars --mysqld=--default-storage-engine=rocksdb --mysqld=--skip-innodb --mysqld=--default-tmp-storage-engine=MyISAM --mysqld=--rocksdb rocksdb_rpl.rpl_rocksdb_stress_crash --repeat=1000 Ensure that it no longer hangs during the test. Reviewers: hermanlee4, yhchiang, anthony Reviewed By: anthony Subscribers: leveldb, yhchiang, dhruba Differential Revision: https://reviews.facebook.net/D51747
9 years ago
s = status;
}
}
return s;
}
void DBImpl::MaybeScheduleFlushOrCompaction() {
mutex_.AssertHeld();
if (!opened_successfully_) {
// Compaction may introduce data race to DB open
return;
}
if (bg_work_paused_ > 0) {
// we paused the background work
Rewritten system for scheduling background work Summary: When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue. The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction. Here are the performance results: Command: ./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000 --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333 Before the patch: fillrandom : 26.950 micros/op 37105 ops/sec; 4.1 MB/s After the patch: fillrandom : 17.404 micros/op 57456 ops/sec; 6.4 MB/s Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got: fillrandom : 7.590 micros/op 131758 ops/sec; 14.6 MB/s Test Plan: make check two stress tests: Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 max_background_flushes=0, to verify that this case also works correctly ./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: ljin, rven, yhchiang, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30123
10 years ago
return;
} else if (shutting_down_.load(std::memory_order_acquire)) {
// DB is being deleted; no more background compactions
Rewritten system for scheduling background work Summary: When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue. The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction. Here are the performance results: Command: ./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000 --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333 Before the patch: fillrandom : 26.950 micros/op 37105 ops/sec; 4.1 MB/s After the patch: fillrandom : 17.404 micros/op 57456 ops/sec; 6.4 MB/s Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got: fillrandom : 7.590 micros/op 131758 ops/sec; 14.6 MB/s Test Plan: make check two stress tests: Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 max_background_flushes=0, to verify that this case also works correctly ./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: ljin, rven, yhchiang, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30123
10 years ago
return;
}
Rewritten system for scheduling background work Summary: When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue. The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction. Here are the performance results: Command: ./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000 --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333 Before the patch: fillrandom : 26.950 micros/op 37105 ops/sec; 4.1 MB/s After the patch: fillrandom : 17.404 micros/op 57456 ops/sec; 6.4 MB/s Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got: fillrandom : 7.590 micros/op 131758 ops/sec; 14.6 MB/s Test Plan: make check two stress tests: Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 max_background_flushes=0, to verify that this case also works correctly ./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: ljin, rven, yhchiang, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30123
10 years ago
while (unscheduled_flushes_ > 0 &&
bg_flush_scheduled_ < immutable_db_options_.max_background_flushes) {
Rewritten system for scheduling background work Summary: When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue. The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction. Here are the performance results: Command: ./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000 --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333 Before the patch: fillrandom : 26.950 micros/op 37105 ops/sec; 4.1 MB/s After the patch: fillrandom : 17.404 micros/op 57456 ops/sec; 6.4 MB/s Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got: fillrandom : 7.590 micros/op 131758 ops/sec; 14.6 MB/s Test Plan: make check two stress tests: Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 max_background_flushes=0, to verify that this case also works correctly ./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: ljin, rven, yhchiang, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30123
10 years ago
unscheduled_flushes_--;
bg_flush_scheduled_++;
env_->Schedule(&DBImpl::BGWorkFlush, this, Env::Priority::HIGH, this);
Rewritten system for scheduling background work Summary: When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue. The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction. Here are the performance results: Command: ./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000 --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333 Before the patch: fillrandom : 26.950 micros/op 37105 ops/sec; 4.1 MB/s After the patch: fillrandom : 17.404 micros/op 57456 ops/sec; 6.4 MB/s Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got: fillrandom : 7.590 micros/op 131758 ops/sec; 14.6 MB/s Test Plan: make check two stress tests: Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 max_background_flushes=0, to verify that this case also works correctly ./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: ljin, rven, yhchiang, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30123
10 years ago
}
auto bg_compactions_allowed = BGCompactionsAllowed();
// special case -- if max_background_flushes == 0, then schedule flush on a
// compaction thread
if (immutable_db_options_.max_background_flushes == 0) {
while (unscheduled_flushes_ > 0 &&
bg_flush_scheduled_ + bg_compaction_scheduled_ <
bg_compactions_allowed) {
unscheduled_flushes_--;
bg_flush_scheduled_++;
env_->Schedule(&DBImpl::BGWorkFlush, this, Env::Priority::LOW, this);
}
}
if (bg_compaction_paused_ > 0) {
// we paused the background compaction
return;
}
if (HasExclusiveManualCompaction()) {
// only manual compactions are allowed to run. don't schedule automatic
// compactions
return;
}
while (bg_compaction_scheduled_ < bg_compactions_allowed &&
Rewritten system for scheduling background work Summary: When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue. The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction. Here are the performance results: Command: ./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000 --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333 Before the patch: fillrandom : 26.950 micros/op 37105 ops/sec; 4.1 MB/s After the patch: fillrandom : 17.404 micros/op 57456 ops/sec; 6.4 MB/s Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got: fillrandom : 7.590 micros/op 131758 ops/sec; 14.6 MB/s Test Plan: make check two stress tests: Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 max_background_flushes=0, to verify that this case also works correctly ./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: ljin, rven, yhchiang, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30123
10 years ago
unscheduled_compactions_ > 0) {
CompactionArg* ca = new CompactionArg;
ca->db = this;
ca->m = nullptr;
Rewritten system for scheduling background work Summary: When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue. The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction. Here are the performance results: Command: ./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000 --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333 Before the patch: fillrandom : 26.950 micros/op 37105 ops/sec; 4.1 MB/s After the patch: fillrandom : 17.404 micros/op 57456 ops/sec; 6.4 MB/s Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got: fillrandom : 7.590 micros/op 131758 ops/sec; 14.6 MB/s Test Plan: make check two stress tests: Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 max_background_flushes=0, to verify that this case also works correctly ./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: ljin, rven, yhchiang, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30123
10 years ago
bg_compaction_scheduled_++;
unscheduled_compactions_--;
env_->Schedule(&DBImpl::BGWorkCompaction, ca, Env::Priority::LOW, this,
&DBImpl::UnscheduleCallback);
Rewritten system for scheduling background work Summary: When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue. The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction. Here are the performance results: Command: ./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000 --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333 Before the patch: fillrandom : 26.950 micros/op 37105 ops/sec; 4.1 MB/s After the patch: fillrandom : 17.404 micros/op 57456 ops/sec; 6.4 MB/s Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got: fillrandom : 7.590 micros/op 131758 ops/sec; 14.6 MB/s Test Plan: make check two stress tests: Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 max_background_flushes=0, to verify that this case also works correctly ./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: ljin, rven, yhchiang, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30123
10 years ago
}
}
void DBImpl::SchedulePurge() {
mutex_.AssertHeld();
assert(opened_successfully_);
// Purge operations are put into High priority queue
bg_purge_scheduled_++;
env_->Schedule(&DBImpl::BGWorkPurge, this, Env::Priority::HIGH, nullptr);
}
int DBImpl::BGCompactionsAllowed() const {
mutex_.AssertHeld();
if (write_controller_.NeedSpeedupCompaction()) {
return mutable_db_options_.max_background_compactions;
} else {
return mutable_db_options_.base_background_compactions;
}
}
Rewritten system for scheduling background work Summary: When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue. The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction. Here are the performance results: Command: ./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000 --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333 Before the patch: fillrandom : 26.950 micros/op 37105 ops/sec; 4.1 MB/s After the patch: fillrandom : 17.404 micros/op 57456 ops/sec; 6.4 MB/s Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got: fillrandom : 7.590 micros/op 131758 ops/sec; 14.6 MB/s Test Plan: make check two stress tests: Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 max_background_flushes=0, to verify that this case also works correctly ./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: ljin, rven, yhchiang, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30123
10 years ago
void DBImpl::AddToCompactionQueue(ColumnFamilyData* cfd) {
assert(!cfd->pending_compaction());
cfd->Ref();
compaction_queue_.push_back(cfd);
cfd->set_pending_compaction(true);
}
ColumnFamilyData* DBImpl::PopFirstFromCompactionQueue() {
assert(!compaction_queue_.empty());
auto cfd = *compaction_queue_.begin();
compaction_queue_.pop_front();
assert(cfd->pending_compaction());
cfd->set_pending_compaction(false);
return cfd;
}
void DBImpl::AddToFlushQueue(ColumnFamilyData* cfd) {
assert(!cfd->pending_flush());
cfd->Ref();
flush_queue_.push_back(cfd);
cfd->set_pending_flush(true);
}
ColumnFamilyData* DBImpl::PopFirstFromFlushQueue() {
assert(!flush_queue_.empty());
auto cfd = *flush_queue_.begin();
flush_queue_.pop_front();
assert(cfd->pending_flush());
cfd->set_pending_flush(false);
return cfd;
}
void DBImpl::SchedulePendingFlush(ColumnFamilyData* cfd) {
if (!cfd->pending_flush() && cfd->imm()->IsFlushPending()) {
AddToFlushQueue(cfd);
++unscheduled_flushes_;
}
}
void DBImpl::SchedulePendingCompaction(ColumnFamilyData* cfd) {
if (!cfd->pending_compaction() && cfd->NeedsCompaction()) {
AddToCompactionQueue(cfd);
++unscheduled_compactions_;
}
}
void DBImpl::SchedulePendingPurge(std::string fname, FileType type,
uint64_t number, uint32_t path_id,
int job_id) {
mutex_.AssertHeld();
PurgeFileInfo file_info(fname, type, number, path_id, job_id);
purge_queue_.push_back(std::move(file_info));
}
void DBImpl::BGWorkFlush(void* db) {
IOSTATS_SET_THREAD_POOL_ID(Env::Priority::HIGH);
TEST_SYNC_POINT("DBImpl::BGWorkFlush");
reinterpret_cast<DBImpl*>(db)->BackgroundCallFlush();
TEST_SYNC_POINT("DBImpl::BGWorkFlush:done");
}
void DBImpl::BGWorkCompaction(void* arg) {
CompactionArg ca = *(reinterpret_cast<CompactionArg*>(arg));
delete reinterpret_cast<CompactionArg*>(arg);
IOSTATS_SET_THREAD_POOL_ID(Env::Priority::LOW);
TEST_SYNC_POINT("DBImpl::BGWorkCompaction");
reinterpret_cast<DBImpl*>(ca.db)->BackgroundCallCompaction(ca.m);
}
void DBImpl::BGWorkPurge(void* db) {
IOSTATS_SET_THREAD_POOL_ID(Env::Priority::HIGH);
TEST_SYNC_POINT("DBImpl::BGWorkPurge:start");
reinterpret_cast<DBImpl*>(db)->BackgroundCallPurge();
TEST_SYNC_POINT("DBImpl::BGWorkPurge:end");
}
void DBImpl::UnscheduleCallback(void* arg) {
CompactionArg ca = *(reinterpret_cast<CompactionArg*>(arg));
delete reinterpret_cast<CompactionArg*>(arg);
if ((ca.m != nullptr) && (ca.m->compaction != nullptr)) {
delete ca.m->compaction;
}
TEST_SYNC_POINT("DBImpl::UnscheduleCallback");
}
void DBImpl::BackgroundCallPurge() {
mutex_.Lock();
// We use one single loop to clear both queues so that after existing the loop
// both queues are empty. This is stricter than what is needed, but can make
// it easier for us to reason the correctness.
while (!purge_queue_.empty() || !logs_to_free_queue_.empty()) {
if (!purge_queue_.empty()) {
auto purge_file = purge_queue_.begin();
auto fname = purge_file->fname;
auto type = purge_file->type;
auto number = purge_file->number;
auto path_id = purge_file->path_id;
auto job_id = purge_file->job_id;
purge_queue_.pop_front();
mutex_.Unlock();
Status file_deletion_status;
DeleteObsoleteFileImpl(file_deletion_status, job_id, fname, type, number,
path_id);
mutex_.Lock();
} else {
assert(!logs_to_free_queue_.empty());
log::Writer* log_writer = *(logs_to_free_queue_.begin());
logs_to_free_queue_.pop_front();
mutex_.Unlock();
delete log_writer;
mutex_.Lock();
}
}
bg_purge_scheduled_--;
bg_cv_.SignalAll();
// IMPORTANT:there should be no code after calling SignalAll. This call may
// signal the DB destructor that it's OK to proceed with destruction. In
// that case, all DB variables will be dealloacated and referencing them
// will cause trouble.
mutex_.Unlock();
}
Status DBImpl::BackgroundFlush(bool* made_progress, JobContext* job_context,
LogBuffer* log_buffer) {
mutex_.AssertHeld();
Status status = bg_error_;
if (status.ok() && shutting_down_.load(std::memory_order_acquire)) {
status = Status::ShutdownInProgress();
}
if (!status.ok()) {
return status;
}
Rewritten system for scheduling background work Summary: When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue. The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction. Here are the performance results: Command: ./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000 --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333 Before the patch: fillrandom : 26.950 micros/op 37105 ops/sec; 4.1 MB/s After the patch: fillrandom : 17.404 micros/op 57456 ops/sec; 6.4 MB/s Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got: fillrandom : 7.590 micros/op 131758 ops/sec; 14.6 MB/s Test Plan: make check two stress tests: Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 max_background_flushes=0, to verify that this case also works correctly ./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: ljin, rven, yhchiang, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30123
10 years ago
ColumnFamilyData* cfd = nullptr;
while (!flush_queue_.empty()) {
// This cfd is already referenced
auto first_cfd = PopFirstFromFlushQueue();
Rewritten system for scheduling background work Summary: When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue. The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction. Here are the performance results: Command: ./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000 --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333 Before the patch: fillrandom : 26.950 micros/op 37105 ops/sec; 4.1 MB/s After the patch: fillrandom : 17.404 micros/op 57456 ops/sec; 6.4 MB/s Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got: fillrandom : 7.590 micros/op 131758 ops/sec; 14.6 MB/s Test Plan: make check two stress tests: Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 max_background_flushes=0, to verify that this case also works correctly ./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: ljin, rven, yhchiang, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30123
10 years ago
if (first_cfd->IsDropped() || !first_cfd->imm()->IsFlushPending()) {
Rewritten system for scheduling background work Summary: When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue. The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction. Here are the performance results: Command: ./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000 --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333 Before the patch: fillrandom : 26.950 micros/op 37105 ops/sec; 4.1 MB/s After the patch: fillrandom : 17.404 micros/op 57456 ops/sec; 6.4 MB/s Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got: fillrandom : 7.590 micros/op 131758 ops/sec; 14.6 MB/s Test Plan: make check two stress tests: Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 max_background_flushes=0, to verify that this case also works correctly ./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: ljin, rven, yhchiang, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30123
10 years ago
// can't flush this CF, try next one
if (first_cfd->Unref()) {
delete first_cfd;
Rewritten system for scheduling background work Summary: When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue. The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction. Here are the performance results: Command: ./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000 --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333 Before the patch: fillrandom : 26.950 micros/op 37105 ops/sec; 4.1 MB/s After the patch: fillrandom : 17.404 micros/op 57456 ops/sec; 6.4 MB/s Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got: fillrandom : 7.590 micros/op 131758 ops/sec; 14.6 MB/s Test Plan: make check two stress tests: Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 max_background_flushes=0, to verify that this case also works correctly ./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: ljin, rven, yhchiang, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30123
10 years ago
}
continue;
}
Rewritten system for scheduling background work Summary: When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue. The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction. Here are the performance results: Command: ./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000 --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333 Before the patch: fillrandom : 26.950 micros/op 37105 ops/sec; 4.1 MB/s After the patch: fillrandom : 17.404 micros/op 57456 ops/sec; 6.4 MB/s Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got: fillrandom : 7.590 micros/op 131758 ops/sec; 14.6 MB/s Test Plan: make check two stress tests: Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 max_background_flushes=0, to verify that this case also works correctly ./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: ljin, rven, yhchiang, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30123
10 years ago
// found a flush!
cfd = first_cfd;
Rewritten system for scheduling background work Summary: When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue. The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction. Here are the performance results: Command: ./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000 --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333 Before the patch: fillrandom : 26.950 micros/op 37105 ops/sec; 4.1 MB/s After the patch: fillrandom : 17.404 micros/op 57456 ops/sec; 6.4 MB/s Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got: fillrandom : 7.590 micros/op 131758 ops/sec; 14.6 MB/s Test Plan: make check two stress tests: Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 max_background_flushes=0, to verify that this case also works correctly ./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: ljin, rven, yhchiang, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30123
10 years ago
break;
}
if (cfd != nullptr) {
const MutableCFOptions mutable_cf_options =
*cfd->GetLatestMutableCFOptions();
LogToBuffer(
log_buffer,
"Calling FlushMemTableToOutputFile with column "
"family [%s], flush slots available %d, compaction slots allowed %d, "
"compaction slots scheduled %d",
cfd->GetName().c_str(), immutable_db_options_.max_background_flushes,
bg_flush_scheduled_, BGCompactionsAllowed() - bg_compaction_scheduled_);
status = FlushMemTableToOutputFile(cfd, mutable_cf_options, made_progress,
Rewritten system for scheduling background work Summary: When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue. The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction. Here are the performance results: Command: ./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000 --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333 Before the patch: fillrandom : 26.950 micros/op 37105 ops/sec; 4.1 MB/s After the patch: fillrandom : 17.404 micros/op 57456 ops/sec; 6.4 MB/s Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got: fillrandom : 7.590 micros/op 131758 ops/sec; 14.6 MB/s Test Plan: make check two stress tests: Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 max_background_flushes=0, to verify that this case also works correctly ./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: ljin, rven, yhchiang, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30123
10 years ago
job_context, log_buffer);
if (cfd->Unref()) {
delete cfd;
}
}
Rewritten system for scheduling background work Summary: When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue. The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction. Here are the performance results: Command: ./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000 --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333 Before the patch: fillrandom : 26.950 micros/op 37105 ops/sec; 4.1 MB/s After the patch: fillrandom : 17.404 micros/op 57456 ops/sec; 6.4 MB/s Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got: fillrandom : 7.590 micros/op 131758 ops/sec; 14.6 MB/s Test Plan: make check two stress tests: Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 max_background_flushes=0, to verify that this case also works correctly ./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: ljin, rven, yhchiang, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30123
10 years ago
return status;
}
void DBImpl::BackgroundCallFlush() {
bool made_progress = false;
JobContext job_context(next_job_id_.fetch_add(1), true);
assert(bg_flush_scheduled_);
TEST_SYNC_POINT("DBImpl::BackgroundCallFlush:start");
LogBuffer log_buffer(InfoLogLevel::INFO_LEVEL,
immutable_db_options_.info_log.get());
{
InstrumentedMutexLock l(&mutex_);
num_running_flushes_++;
auto pending_outputs_inserted_elem =
CaptureCurrentFileNumberInPendingOutputs();
Status s = BackgroundFlush(&made_progress, &job_context, &log_buffer);
if (!s.ok() && !s.IsShutdownInProgress()) {
// Wait a little bit before retrying background flush in
// case this is an environmental problem and we do not want to
// chew up resources for failed flushes for the duration of
// the problem.
uint64_t error_cnt =
default_cf_internal_stats_->BumpAndGetBackgroundErrorCount();
bg_cv_.SignalAll(); // In case a waiter can proceed despite the error
mutex_.Unlock();
Log(InfoLogLevel::ERROR_LEVEL, immutable_db_options_.info_log,
"Waiting after background flush error: %s"
"Accumulated background error counts: %" PRIu64,
s.ToString().c_str(), error_cnt);
log_buffer.FlushBufferToLog();
LogFlush(immutable_db_options_.info_log);
env_->SleepForMicroseconds(1000000);
mutex_.Lock();
}
ReleaseFileNumberFromPendingOutputs(pending_outputs_inserted_elem);
// If flush failed, we want to delete all temporary files that we might have
// created. Thus, we force full scan in FindObsoleteFiles()
FindObsoleteFiles(&job_context, !s.ok() && !s.IsShutdownInProgress());
// delete unnecessary files if any, this is done outside the mutex
if (job_context.HaveSomethingToDelete() || !log_buffer.IsEmpty()) {
mutex_.Unlock();
// Have to flush the info logs before bg_flush_scheduled_--
// because if bg_flush_scheduled_ becomes 0 and the lock is
// released, the deconstructor of DB can kick in and destroy all the
// states of DB so info_log might not be available after that point.
// It also applies to access other states that DB owns.
log_buffer.FlushBufferToLog();
if (job_context.HaveSomethingToDelete()) {
PurgeObsoleteFiles(job_context);
}
job_context.Clean();
mutex_.Lock();
}
assert(num_running_flushes_ > 0);
num_running_flushes_--;
bg_flush_scheduled_--;
Rewritten system for scheduling background work Summary: When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue. The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction. Here are the performance results: Command: ./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000 --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333 Before the patch: fillrandom : 26.950 micros/op 37105 ops/sec; 4.1 MB/s After the patch: fillrandom : 17.404 micros/op 57456 ops/sec; 6.4 MB/s Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got: fillrandom : 7.590 micros/op 131758 ops/sec; 14.6 MB/s Test Plan: make check two stress tests: Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 max_background_flushes=0, to verify that this case also works correctly ./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: ljin, rven, yhchiang, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30123
10 years ago
// See if there's more work to be done
MaybeScheduleFlushOrCompaction();
bg_cv_.SignalAll();
// IMPORTANT: there should be no code after calling SignalAll. This call may
// signal the DB destructor that it's OK to proceed with destruction. In
// that case, all DB variables will be dealloacated and referencing them
// will cause trouble.
}
}
void DBImpl::BackgroundCallCompaction(void* arg) {
bool made_progress = false;
ManualCompaction* m = reinterpret_cast<ManualCompaction*>(arg);
JobContext job_context(next_job_id_.fetch_add(1), true);
TEST_SYNC_POINT("BackgroundCallCompaction:0");
MaybeDumpStats();
LogBuffer log_buffer(InfoLogLevel::INFO_LEVEL,
immutable_db_options_.info_log.get());
{
InstrumentedMutexLock l(&mutex_);
// This call will unlock/lock the mutex to wait for current running
// IngestExternalFile() calls to finish.
WaitForIngestFile();
num_running_compactions_++;
auto pending_outputs_inserted_elem =
CaptureCurrentFileNumberInPendingOutputs();
assert(bg_compaction_scheduled_);
Status s =
BackgroundCompaction(&made_progress, &job_context, &log_buffer, m);
TEST_SYNC_POINT("BackgroundCallCompaction:1");
if (!s.ok() && !s.IsShutdownInProgress()) {
// Wait a little bit before retrying background compaction in
// case this is an environmental problem and we do not want to
// chew up resources for failed compactions for the duration of
// the problem.
uint64_t error_cnt =
make internal stats independent of statistics Summary: also make it aware of column family output from db_bench ``` ** Compaction Stats [default] ** Level Files Size(MB) Score Read(GB) Rn(GB) Rnp1(GB) Write(GB) Wnew(GB) RW-Amp W-Amp Rd(MB/s) Wr(MB/s) Rn(cnt) Rnp1(cnt) Wnp1(cnt) Wnew(cnt) Comp(sec) Comp(cnt) Avg(sec) Stall(sec) Stall(cnt) Avg(ms) ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ L0 14 956 0.9 0.0 0.0 0.0 2.7 2.7 0.0 0.0 0.0 111.6 0 0 0 0 24 40 0.612 75.20 492387 0.15 L1 21 2001 2.0 5.7 2.0 3.7 5.3 1.6 5.4 2.6 71.2 65.7 31 43 55 12 82 2 41.242 43.72 41183 1.06 L2 217 18974 1.9 16.5 2.0 14.4 15.1 0.7 15.6 7.4 70.1 64.3 17 182 185 3 241 16 15.052 0.00 0 0.00 L3 1641 188245 1.8 9.1 1.1 8.0 8.5 0.5 15.4 7.4 61.3 57.2 9 75 76 1 152 9 16.887 0.00 0 0.00 L4 4447 449025 0.4 13.4 4.8 8.6 9.1 0.5 4.7 1.9 77.8 52.7 38 79 100 21 176 38 4.639 0.00 0 0.00 Sum 6340 659201 0.0 44.7 10.0 34.7 40.6 6.0 32.0 15.2 67.7 61.6 95 379 416 37 676 105 6.439 118.91 533570 0.22 Int 0 0 0.0 1.2 0.4 0.8 1.3 0.5 5.2 2.7 59.1 65.6 3 7 9 2 20 10 2.003 0.00 0 0.00 Stalls(secs): 75.197 level0_slowdown, 0.000 level0_numfiles, 0.000 memtable_compaction, 43.717 leveln_slowdown Stalls(count): 492387 level0_slowdown, 0 level0_numfiles, 0 memtable_compaction, 41183 leveln_slowdown ** DB Stats ** Uptime(secs): 202.1 total, 13.5 interval Cumulative writes: 6291456 writes, 6291456 batches, 1.0 writes per batch, 4.90 ingest GB Cumulative WAL: 6291456 writes, 6291456 syncs, 1.00 writes per sync, 4.90 GB written Interval writes: 1048576 writes, 1048576 batches, 1.0 writes per batch, 836.0 ingest MB Interval WAL: 1048576 writes, 1048576 syncs, 1.00 writes per sync, 0.82 MB written Test Plan: ran it Reviewers: sdong, yhchiang, igor Reviewed By: igor Subscribers: leveldb Differential Revision: https://reviews.facebook.net/D19917
10 years ago
default_cf_internal_stats_->BumpAndGetBackgroundErrorCount();
bg_cv_.SignalAll(); // In case a waiter can proceed despite the error
mutex_.Unlock();
log_buffer.FlushBufferToLog();
Log(InfoLogLevel::ERROR_LEVEL, immutable_db_options_.info_log,
"Waiting after background compaction error: %s, "
"Accumulated background error counts: %" PRIu64,
s.ToString().c_str(), error_cnt);
LogFlush(immutable_db_options_.info_log);
env_->SleepForMicroseconds(1000000);
mutex_.Lock();
}
ReleaseFileNumberFromPendingOutputs(pending_outputs_inserted_elem);
// If compaction failed, we want to delete all temporary files that we might
// have created (they might not be all recorded in job_context in case of a
// failure). Thus, we force full scan in FindObsoleteFiles()
FindObsoleteFiles(&job_context, !s.ok() && !s.IsShutdownInProgress());
// delete unnecessary files if any, this is done outside the mutex
if (job_context.HaveSomethingToDelete() || !log_buffer.IsEmpty()) {
mutex_.Unlock();
// Have to flush the info logs before bg_compaction_scheduled_--
// because if bg_flush_scheduled_ becomes 0 and the lock is
// released, the deconstructor of DB can kick in and destroy all the
// states of DB so info_log might not be available after that point.
// It also applies to access other states that DB owns.
log_buffer.FlushBufferToLog();
if (job_context.HaveSomethingToDelete()) {
PurgeObsoleteFiles(job_context);
}
job_context.Clean();
mutex_.Lock();
}
assert(num_running_compactions_ > 0);
num_running_compactions_--;
bg_compaction_scheduled_--;
versions_->GetColumnFamilySet()->FreeDeadColumnFamilies();
Rewritten system for scheduling background work Summary: When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue. The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction. Here are the performance results: Command: ./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000 --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333 Before the patch: fillrandom : 26.950 micros/op 37105 ops/sec; 4.1 MB/s After the patch: fillrandom : 17.404 micros/op 57456 ops/sec; 6.4 MB/s Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got: fillrandom : 7.590 micros/op 131758 ops/sec; 14.6 MB/s Test Plan: make check two stress tests: Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 max_background_flushes=0, to verify that this case also works correctly ./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: ljin, rven, yhchiang, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30123
10 years ago
// See if there's more work to be done
MaybeScheduleFlushOrCompaction();
if (made_progress || bg_compaction_scheduled_ == 0 ||
HasPendingManualCompaction()) {
// signal if
// * made_progress -- need to wakeup DelayWrite
// * bg_compaction_scheduled_ == 0 -- need to wakeup ~DBImpl
// * HasPendingManualCompaction -- need to wakeup RunManualCompaction
// If none of this is true, there is no need to signal since nobody is
// waiting for it
bg_cv_.SignalAll();
}
// IMPORTANT: there should be no code after calling SignalAll. This call may
// signal the DB destructor that it's OK to proceed with destruction. In
// that case, all DB variables will be dealloacated and referencing them
// will cause trouble.
}
}
Status DBImpl::BackgroundCompaction(bool* made_progress,
JobContext* job_context,
LogBuffer* log_buffer, void* arg) {
ManualCompaction* manual_compaction =
reinterpret_cast<ManualCompaction*>(arg);
*made_progress = false;
mutex_.AssertHeld();
TEST_SYNC_POINT("DBImpl::BackgroundCompaction:Start");
bool is_manual = (manual_compaction != nullptr);
// (manual_compaction->in_progress == false);
bool trivial_move_disallowed =
is_manual && manual_compaction->disallow_trivial_move;
CompactionJobStats compaction_job_stats;
Status status = bg_error_;
if (status.ok() && shutting_down_.load(std::memory_order_acquire)) {
status = Status::ShutdownInProgress();
}
if (!status.ok()) {
if (is_manual) {
manual_compaction->status = status;
manual_compaction->done = true;
manual_compaction->in_progress = false;
delete manual_compaction->compaction;
manual_compaction = nullptr;
}
return status;
}
if (is_manual) {
// another thread cannot pick up the same work
manual_compaction->in_progress = true;
}
unique_ptr<Compaction> c;
// InternalKey manual_end_storage;
// InternalKey* manual_end = &manual_end_storage;
if (is_manual) {
ManualCompaction* m = manual_compaction;
assert(m->in_progress);
c.reset(std::move(m->compaction));
if (!c) {
m->done = true;
m->manual_end = nullptr;
LogToBuffer(log_buffer,
"[%s] Manual compaction from level-%d from %s .. "
"%s; nothing to do\n",
m->cfd->GetName().c_str(), m->input_level,
(m->begin ? m->begin->DebugString().c_str() : "(begin)"),
(m->end ? m->end->DebugString().c_str() : "(end)"));
} else {
LogToBuffer(log_buffer,
"[%s] Manual compaction from level-%d to level-%d from %s .. "
"%s; will stop at %s\n",
m->cfd->GetName().c_str(), m->input_level, c->output_level(),
(m->begin ? m->begin->DebugString().c_str() : "(begin)"),
(m->end ? m->end->DebugString().c_str() : "(end)"),
((m->done || m->manual_end == nullptr)
? "(end)"
: m->manual_end->DebugString().c_str()));
}
Rewritten system for scheduling background work Summary: When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue. The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction. Here are the performance results: Command: ./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000 --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333 Before the patch: fillrandom : 26.950 micros/op 37105 ops/sec; 4.1 MB/s After the patch: fillrandom : 17.404 micros/op 57456 ops/sec; 6.4 MB/s Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got: fillrandom : 7.590 micros/op 131758 ops/sec; 14.6 MB/s Test Plan: make check two stress tests: Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 max_background_flushes=0, to verify that this case also works correctly ./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: ljin, rven, yhchiang, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30123
10 years ago
} else if (!compaction_queue_.empty()) {
// cfd is referenced here
auto cfd = PopFirstFromCompactionQueue();
// We unreference here because the following code will take a Ref() on
// this cfd if it is going to use it (Compaction class holds a
// reference).
// This will all happen under a mutex so we don't have to be afraid of
// somebody else deleting it.
if (cfd->Unref()) {
delete cfd;
// This was the last reference of the column family, so no need to
// compact.
return Status::OK();
}
if (HaveManualCompaction(cfd)) {
// Can't compact right now, but try again later
TEST_SYNC_POINT("DBImpl::BackgroundCompaction()::Conflict");
return Status::OK();
}
Rewritten system for scheduling background work Summary: When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue. The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction. Here are the performance results: Command: ./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000 --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333 Before the patch: fillrandom : 26.950 micros/op 37105 ops/sec; 4.1 MB/s After the patch: fillrandom : 17.404 micros/op 57456 ops/sec; 6.4 MB/s Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got: fillrandom : 7.590 micros/op 131758 ops/sec; 14.6 MB/s Test Plan: make check two stress tests: Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 max_background_flushes=0, to verify that this case also works correctly ./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: ljin, rven, yhchiang, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30123
10 years ago
// Pick up latest mutable CF Options and use it throughout the
// compaction job
// Compaction makes a copy of the latest MutableCFOptions. It should be used
// throughout the compaction procedure to make sure consistency. It will
// eventually be installed into SuperVersion
auto* mutable_cf_options = cfd->GetLatestMutableCFOptions();
if (!mutable_cf_options->disable_auto_compactions && !cfd->IsDropped()) {
// NOTE: try to avoid unnecessary copy of MutableCFOptions if
// compaction is not necessary. Need to make sure mutex is held
// until we make a copy in the following code
TEST_SYNC_POINT("DBImpl::BackgroundCompaction():BeforePickCompaction");
Rewritten system for scheduling background work Summary: When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue. The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction. Here are the performance results: Command: ./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000 --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333 Before the patch: fillrandom : 26.950 micros/op 37105 ops/sec; 4.1 MB/s After the patch: fillrandom : 17.404 micros/op 57456 ops/sec; 6.4 MB/s Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got: fillrandom : 7.590 micros/op 131758 ops/sec; 14.6 MB/s Test Plan: make check two stress tests: Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 max_background_flushes=0, to verify that this case also works correctly ./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: ljin, rven, yhchiang, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30123
10 years ago
c.reset(cfd->PickCompaction(*mutable_cf_options, log_buffer));
TEST_SYNC_POINT("DBImpl::BackgroundCompaction():AfterPickCompaction");
Rewritten system for scheduling background work Summary: When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue. The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction. Here are the performance results: Command: ./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000 --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333 Before the patch: fillrandom : 26.950 micros/op 37105 ops/sec; 4.1 MB/s After the patch: fillrandom : 17.404 micros/op 57456 ops/sec; 6.4 MB/s Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got: fillrandom : 7.590 micros/op 131758 ops/sec; 14.6 MB/s Test Plan: make check two stress tests: Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 max_background_flushes=0, to verify that this case also works correctly ./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: ljin, rven, yhchiang, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30123
10 years ago
if (c != nullptr) {
// update statistics
MeasureTime(stats_, NUM_FILES_IN_SINGLE_COMPACTION,
c->inputs(0)->size());
// There are three things that can change compaction score:
// 1) When flush or compaction finish. This case is covered by
// InstallSuperVersionAndScheduleWork
Rewritten system for scheduling background work Summary: When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue. The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction. Here are the performance results: Command: ./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000 --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333 Before the patch: fillrandom : 26.950 micros/op 37105 ops/sec; 4.1 MB/s After the patch: fillrandom : 17.404 micros/op 57456 ops/sec; 6.4 MB/s Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got: fillrandom : 7.590 micros/op 131758 ops/sec; 14.6 MB/s Test Plan: make check two stress tests: Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 max_background_flushes=0, to verify that this case also works correctly ./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: ljin, rven, yhchiang, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30123
10 years ago
// 2) When MutableCFOptions changes. This case is also covered by
// InstallSuperVersionAndScheduleWork, because this is when the new
// options take effect.
Rewritten system for scheduling background work Summary: When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue. The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction. Here are the performance results: Command: ./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000 --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333 Before the patch: fillrandom : 26.950 micros/op 37105 ops/sec; 4.1 MB/s After the patch: fillrandom : 17.404 micros/op 57456 ops/sec; 6.4 MB/s Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got: fillrandom : 7.590 micros/op 131758 ops/sec; 14.6 MB/s Test Plan: make check two stress tests: Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 max_background_flushes=0, to verify that this case also works correctly ./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: ljin, rven, yhchiang, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30123
10 years ago
// 3) When we Pick a new compaction, we "remove" those files being
// compacted from the calculation, which then influences compaction
// score. Here we check if we need the new compaction even without the
// files that are currently being compacted. If we need another
// compaction, we might be able to execute it in parallel, so we add it
// to the queue and schedule a new thread.
if (cfd->NeedsCompaction()) {
// Yes, we need more compactions!
AddToCompactionQueue(cfd);
++unscheduled_compactions_;
MaybeScheduleFlushOrCompaction();
}
}
}
}
if (!c) {
// Nothing to do
LogToBuffer(log_buffer, "Compaction nothing to do");
} else if (c->deletion_compaction()) {
// TODO(icanadi) Do we want to honor snapshots here? i.e. not delete old
// file if there is alive snapshot pointing to it
assert(c->num_input_files(1) == 0);
assert(c->level() == 0);
assert(c->column_family_data()->ioptions()->compaction_style ==
kCompactionStyleFIFO);
compaction_job_stats.num_input_files = c->num_input_files(0);
for (const auto& f : *c->inputs(0)) {
c->edit()->DeleteFile(c->level(), f->fd.GetNumber());
}
status = versions_->LogAndApply(c->column_family_data(),
*c->mutable_cf_options(), c->edit(),
&mutex_, directories_.GetDbDir());
InstallSuperVersionAndScheduleWorkWrapper(
c->column_family_data(), job_context, *c->mutable_cf_options());
LogToBuffer(log_buffer, "[%s] Deleted %d files\n",
c->column_family_data()->GetName().c_str(),
c->num_input_files(0));
*made_progress = true;
Allowing L0 -> L1 trivial move on sorted data Summary: This diff updates the logic of how we do trivial move, now trivial move can run on any number of files in input level as long as they are not overlapping The conditions for trivial move have been updated Introduced conditions: - Trivial move cannot happen if we have a compaction filter (except if the compaction is not manual) - Input level files cannot be overlapping Removed conditions: - Trivial move only run when the compaction is not manual - Input level should can contain only 1 file More context on what tests failed because of Trivial move ``` DBTest.CompactionsGenerateMultipleFiles This test is expecting compaction on a file in L0 to generate multiple files in L1, this test will fail with trivial move because we end up with one file in L1 ``` ``` DBTest.NoSpaceCompactRange This test expect compaction to fail when we force environment to report running out of space, of course this is not valid in trivial move situation because trivial move does not need any extra space, and did not check for that ``` ``` DBTest.DropWrites Similar to DBTest.NoSpaceCompactRange ``` ``` DBTest.DeleteObsoleteFilesPendingOutputs This test expect that a file in L2 is deleted after it's moved to L3, this is not valid with trivial move because although the file was moved it is now used by L3 ``` ``` CuckooTableDBTest.CompactionIntoMultipleFiles Same as DBTest.CompactionsGenerateMultipleFiles ``` This diff is based on a work by @sdong https://reviews.facebook.net/D34149 Test Plan: make -j64 check Reviewers: rven, sdong, igor Reviewed By: igor Subscribers: yhchiang, ott, march, dhruba, sdong Differential Revision: https://reviews.facebook.net/D34797
9 years ago
} else if (!trivial_move_disallowed && c->IsTrivialMove()) {
TEST_SYNC_POINT("DBImpl::BackgroundCompaction:TrivialMove");
// Instrument for event update
// TODO(yhchiang): add op details for showing trivial-move.
ThreadStatusUtil::SetColumnFamily(
c->column_family_data(), c->column_family_data()->ioptions()->env,
immutable_db_options_.enable_thread_tracking);
ThreadStatusUtil::SetThreadOperation(ThreadStatus::OP_COMPACTION);
compaction_job_stats.num_input_files = c->num_input_files(0);
Allowing L0 -> L1 trivial move on sorted data Summary: This diff updates the logic of how we do trivial move, now trivial move can run on any number of files in input level as long as they are not overlapping The conditions for trivial move have been updated Introduced conditions: - Trivial move cannot happen if we have a compaction filter (except if the compaction is not manual) - Input level files cannot be overlapping Removed conditions: - Trivial move only run when the compaction is not manual - Input level should can contain only 1 file More context on what tests failed because of Trivial move ``` DBTest.CompactionsGenerateMultipleFiles This test is expecting compaction on a file in L0 to generate multiple files in L1, this test will fail with trivial move because we end up with one file in L1 ``` ``` DBTest.NoSpaceCompactRange This test expect compaction to fail when we force environment to report running out of space, of course this is not valid in trivial move situation because trivial move does not need any extra space, and did not check for that ``` ``` DBTest.DropWrites Similar to DBTest.NoSpaceCompactRange ``` ``` DBTest.DeleteObsoleteFilesPendingOutputs This test expect that a file in L2 is deleted after it's moved to L3, this is not valid with trivial move because although the file was moved it is now used by L3 ``` ``` CuckooTableDBTest.CompactionIntoMultipleFiles Same as DBTest.CompactionsGenerateMultipleFiles ``` This diff is based on a work by @sdong https://reviews.facebook.net/D34149 Test Plan: make -j64 check Reviewers: rven, sdong, igor Reviewed By: igor Subscribers: yhchiang, ott, march, dhruba, sdong Differential Revision: https://reviews.facebook.net/D34797
9 years ago
// Move files to next level
int32_t moved_files = 0;
int64_t moved_bytes = 0;
for (unsigned int l = 0; l < c->num_input_levels(); l++) {
if (c->level(l) == c->output_level()) {
continue;
}
for (size_t i = 0; i < c->num_input_files(l); i++) {
FileMetaData* f = c->input(l, i);
c->edit()->DeleteFile(c->level(l), f->fd.GetNumber());
c->edit()->AddFile(c->output_level(), f->fd.GetNumber(),
f->fd.GetPathId(), f->fd.GetFileSize(), f->smallest,
f->largest, f->smallest_seqno, f->largest_seqno,
f->marked_for_compaction);
LogToBuffer(log_buffer,
"[%s] Moving #%" PRIu64 " to level-%d %" PRIu64 " bytes\n",
c->column_family_data()->GetName().c_str(),
f->fd.GetNumber(), c->output_level(), f->fd.GetFileSize());
++moved_files;
moved_bytes += f->fd.GetFileSize();
}
Allowing L0 -> L1 trivial move on sorted data Summary: This diff updates the logic of how we do trivial move, now trivial move can run on any number of files in input level as long as they are not overlapping The conditions for trivial move have been updated Introduced conditions: - Trivial move cannot happen if we have a compaction filter (except if the compaction is not manual) - Input level files cannot be overlapping Removed conditions: - Trivial move only run when the compaction is not manual - Input level should can contain only 1 file More context on what tests failed because of Trivial move ``` DBTest.CompactionsGenerateMultipleFiles This test is expecting compaction on a file in L0 to generate multiple files in L1, this test will fail with trivial move because we end up with one file in L1 ``` ``` DBTest.NoSpaceCompactRange This test expect compaction to fail when we force environment to report running out of space, of course this is not valid in trivial move situation because trivial move does not need any extra space, and did not check for that ``` ``` DBTest.DropWrites Similar to DBTest.NoSpaceCompactRange ``` ``` DBTest.DeleteObsoleteFilesPendingOutputs This test expect that a file in L2 is deleted after it's moved to L3, this is not valid with trivial move because although the file was moved it is now used by L3 ``` ``` CuckooTableDBTest.CompactionIntoMultipleFiles Same as DBTest.CompactionsGenerateMultipleFiles ``` This diff is based on a work by @sdong https://reviews.facebook.net/D34149 Test Plan: make -j64 check Reviewers: rven, sdong, igor Reviewed By: igor Subscribers: yhchiang, ott, march, dhruba, sdong Differential Revision: https://reviews.facebook.net/D34797
9 years ago
}
status = versions_->LogAndApply(c->column_family_data(),
*c->mutable_cf_options(), c->edit(),
&mutex_, directories_.GetDbDir());
// Use latest MutableCFOptions
InstallSuperVersionAndScheduleWorkWrapper(
c->column_family_data(), job_context, *c->mutable_cf_options());
VersionStorageInfo::LevelSummaryStorage tmp;
c->column_family_data()->internal_stats()->IncBytesMoved(c->output_level(),
Allowing L0 -> L1 trivial move on sorted data Summary: This diff updates the logic of how we do trivial move, now trivial move can run on any number of files in input level as long as they are not overlapping The conditions for trivial move have been updated Introduced conditions: - Trivial move cannot happen if we have a compaction filter (except if the compaction is not manual) - Input level files cannot be overlapping Removed conditions: - Trivial move only run when the compaction is not manual - Input level should can contain only 1 file More context on what tests failed because of Trivial move ``` DBTest.CompactionsGenerateMultipleFiles This test is expecting compaction on a file in L0 to generate multiple files in L1, this test will fail with trivial move because we end up with one file in L1 ``` ``` DBTest.NoSpaceCompactRange This test expect compaction to fail when we force environment to report running out of space, of course this is not valid in trivial move situation because trivial move does not need any extra space, and did not check for that ``` ``` DBTest.DropWrites Similar to DBTest.NoSpaceCompactRange ``` ``` DBTest.DeleteObsoleteFilesPendingOutputs This test expect that a file in L2 is deleted after it's moved to L3, this is not valid with trivial move because although the file was moved it is now used by L3 ``` ``` CuckooTableDBTest.CompactionIntoMultipleFiles Same as DBTest.CompactionsGenerateMultipleFiles ``` This diff is based on a work by @sdong https://reviews.facebook.net/D34149 Test Plan: make -j64 check Reviewers: rven, sdong, igor Reviewed By: igor Subscribers: yhchiang, ott, march, dhruba, sdong Differential Revision: https://reviews.facebook.net/D34797
9 years ago
moved_bytes);
{
event_logger_.LogToBuffer(log_buffer)
<< "job" << job_context->job_id << "event"
<< "trivial_move"
<< "destination_level" << c->output_level() << "files" << moved_files
Allowing L0 -> L1 trivial move on sorted data Summary: This diff updates the logic of how we do trivial move, now trivial move can run on any number of files in input level as long as they are not overlapping The conditions for trivial move have been updated Introduced conditions: - Trivial move cannot happen if we have a compaction filter (except if the compaction is not manual) - Input level files cannot be overlapping Removed conditions: - Trivial move only run when the compaction is not manual - Input level should can contain only 1 file More context on what tests failed because of Trivial move ``` DBTest.CompactionsGenerateMultipleFiles This test is expecting compaction on a file in L0 to generate multiple files in L1, this test will fail with trivial move because we end up with one file in L1 ``` ``` DBTest.NoSpaceCompactRange This test expect compaction to fail when we force environment to report running out of space, of course this is not valid in trivial move situation because trivial move does not need any extra space, and did not check for that ``` ``` DBTest.DropWrites Similar to DBTest.NoSpaceCompactRange ``` ``` DBTest.DeleteObsoleteFilesPendingOutputs This test expect that a file in L2 is deleted after it's moved to L3, this is not valid with trivial move because although the file was moved it is now used by L3 ``` ``` CuckooTableDBTest.CompactionIntoMultipleFiles Same as DBTest.CompactionsGenerateMultipleFiles ``` This diff is based on a work by @sdong https://reviews.facebook.net/D34149 Test Plan: make -j64 check Reviewers: rven, sdong, igor Reviewed By: igor Subscribers: yhchiang, ott, march, dhruba, sdong Differential Revision: https://reviews.facebook.net/D34797
9 years ago
<< "total_files_size" << moved_bytes;
}
LogToBuffer(
log_buffer,
Allowing L0 -> L1 trivial move on sorted data Summary: This diff updates the logic of how we do trivial move, now trivial move can run on any number of files in input level as long as they are not overlapping The conditions for trivial move have been updated Introduced conditions: - Trivial move cannot happen if we have a compaction filter (except if the compaction is not manual) - Input level files cannot be overlapping Removed conditions: - Trivial move only run when the compaction is not manual - Input level should can contain only 1 file More context on what tests failed because of Trivial move ``` DBTest.CompactionsGenerateMultipleFiles This test is expecting compaction on a file in L0 to generate multiple files in L1, this test will fail with trivial move because we end up with one file in L1 ``` ``` DBTest.NoSpaceCompactRange This test expect compaction to fail when we force environment to report running out of space, of course this is not valid in trivial move situation because trivial move does not need any extra space, and did not check for that ``` ``` DBTest.DropWrites Similar to DBTest.NoSpaceCompactRange ``` ``` DBTest.DeleteObsoleteFilesPendingOutputs This test expect that a file in L2 is deleted after it's moved to L3, this is not valid with trivial move because although the file was moved it is now used by L3 ``` ``` CuckooTableDBTest.CompactionIntoMultipleFiles Same as DBTest.CompactionsGenerateMultipleFiles ``` This diff is based on a work by @sdong https://reviews.facebook.net/D34149 Test Plan: make -j64 check Reviewers: rven, sdong, igor Reviewed By: igor Subscribers: yhchiang, ott, march, dhruba, sdong Differential Revision: https://reviews.facebook.net/D34797
9 years ago
"[%s] Moved #%d files to level-%d %" PRIu64 " bytes %s: %s\n",
c->column_family_data()->GetName().c_str(), moved_files,
c->output_level(), moved_bytes, status.ToString().c_str(),
c->column_family_data()->current()->storage_info()->LevelSummary(&tmp));
*made_progress = true;
// Clear Instrument
ThreadStatusUtil::ResetThreadStatus();
} else {
int output_level __attribute__((unused)) = c->output_level();
TEST_SYNC_POINT_CALLBACK("DBImpl::BackgroundCompaction:NonTrivial",
&output_level);
SequenceNumber earliest_write_conflict_snapshot;
std::vector<SequenceNumber> snapshot_seqs =
snapshots_.GetAll(&earliest_write_conflict_snapshot);
assert(is_snapshot_supported_ || snapshots_.empty());
CompactionJob compaction_job(
job_context->job_id, c.get(), immutable_db_options_, env_options_,
versions_.get(), &shutting_down_, log_buffer, directories_.GetDbDir(),
directories_.GetDataDir(c->output_path_id()), stats_, &mutex_,
&bg_error_, snapshot_seqs, earliest_write_conflict_snapshot,
table_cache_, &event_logger_,
c->mutable_cf_options()->paranoid_file_checks,
c->mutable_cf_options()->report_bg_io_stats, dbname_,
&compaction_job_stats);
compaction_job.Prepare();
mutex_.Unlock();
compaction_job.Run();
TEST_SYNC_POINT("DBImpl::BackgroundCompaction:NonTrivial:AfterRun");
mutex_.Lock();
status = compaction_job.Install(*c->mutable_cf_options());
if (status.ok()) {
InstallSuperVersionAndScheduleWorkWrapper(
c->column_family_data(), job_context, *c->mutable_cf_options());
}
*made_progress = true;
}
if (c != nullptr) {
c->ReleaseCompactionFiles(status);
*made_progress = true;
NotifyOnCompactionCompleted(
c->column_family_data(), c.get(), status,
compaction_job_stats, job_context->job_id);
}
// this will unref its input_version and column_family_data
c.reset();
if (status.ok()) {
// Done
} else if (status.IsShutdownInProgress()) {
// Ignore compaction errors found during shutting down
} else {
Log(InfoLogLevel::WARN_LEVEL, immutable_db_options_.info_log,
"Compaction error: %s", status.ToString().c_str());
if (immutable_db_options_.paranoid_checks && bg_error_.ok()) {
bg_error_ = status;
}
}
if (is_manual) {
ManualCompaction* m = manual_compaction;
if (!status.ok()) {
m->status = status;
m->done = true;
}
// For universal compaction:
// Because universal compaction always happens at level 0, so one
// compaction will pick up all overlapped files. No files will be
// filtered out due to size limit and left for a successive compaction.
// So we can safely conclude the current compaction.
//
// Also note that, if we don't stop here, then the current compaction
// writes a new file back to level 0, which will be used in successive
// compaction. Hence the manual compaction will never finish.
//
// Stop the compaction if manual_end points to nullptr -- this means
// that we compacted the whole range. manual_end should always point
// to nullptr in case of universal compaction
if (m->manual_end == nullptr) {
m->done = true;
}
if (!m->done) {
// We only compacted part of the requested range. Update *m
// to the range that is left to be compacted.
// Universal and FIFO compactions should always compact the whole range
assert(m->cfd->ioptions()->compaction_style !=
kCompactionStyleUniversal ||
m->cfd->ioptions()->num_levels > 1);
assert(m->cfd->ioptions()->compaction_style != kCompactionStyleFIFO);
m->tmp_storage = *m->manual_end;
m->begin = &m->tmp_storage;
m->incomplete = true;
}
m->in_progress = false; // not being processed anymore
}
TEST_SYNC_POINT("DBImpl::BackgroundCompaction:Finish");
return status;
}
bool DBImpl::HasPendingManualCompaction() {
return (!manual_compaction_dequeue_.empty());
}
void DBImpl::AddManualCompaction(DBImpl::ManualCompaction* m) {
manual_compaction_dequeue_.push_back(m);
}
void DBImpl::RemoveManualCompaction(DBImpl::ManualCompaction* m) {
// Remove from queue
std::deque<ManualCompaction*>::iterator it =
manual_compaction_dequeue_.begin();
while (it != manual_compaction_dequeue_.end()) {
if (m == (*it)) {
it = manual_compaction_dequeue_.erase(it);
return;
}
it++;
}
assert(false);
return;
}
bool DBImpl::ShouldntRunManualCompaction(ManualCompaction* m) {
if (num_running_ingest_file_ > 0) {
// We need to wait for other IngestExternalFile() calls to finish
// before running a manual compaction.
return true;
}
if (m->exclusive) {
return (bg_compaction_scheduled_ > 0);
}
std::deque<ManualCompaction*>::iterator it =
manual_compaction_dequeue_.begin();
bool seen = false;
while (it != manual_compaction_dequeue_.end()) {
if (m == (*it)) {
it++;
seen = true;
continue;
} else if (MCOverlap(m, (*it)) && (!seen && !(*it)->in_progress)) {
// Consider the other manual compaction *it, conflicts if:
// overlaps with m
// and (*it) is ahead in the queue and is not yet in progress
return true;
}
it++;
}
return false;
}
bool DBImpl::HaveManualCompaction(ColumnFamilyData* cfd) {
// Remove from priority queue
std::deque<ManualCompaction*>::iterator it =
manual_compaction_dequeue_.begin();
while (it != manual_compaction_dequeue_.end()) {
if ((*it)->exclusive) {
return true;
}
if ((cfd == (*it)->cfd) && (!((*it)->in_progress || (*it)->done))) {
// Allow automatic compaction if manual compaction is
// is in progress
return true;
}
it++;
}
return false;
}
bool DBImpl::HasExclusiveManualCompaction() {
// Remove from priority queue
std::deque<ManualCompaction*>::iterator it =
manual_compaction_dequeue_.begin();
while (it != manual_compaction_dequeue_.end()) {
if ((*it)->exclusive) {
return true;
}
it++;
}
return false;
}
bool DBImpl::MCOverlap(ManualCompaction* m, ManualCompaction* m1) {
if ((m->exclusive) || (m1->exclusive)) {
return true;
}
if (m->cfd != m1->cfd) {
return false;
}
return true;
}
size_t DBImpl::GetWalPreallocateBlockSize(uint64_t write_buffer_size) const {
size_t bsize = write_buffer_size / 10 + write_buffer_size;
// Some users might set very high write_buffer_size and rely on
// max_total_wal_size or other parameters to control the WAL size.
if (immutable_db_options_.max_total_wal_size > 0) {
bsize = std::min<size_t>(bsize, immutable_db_options_.max_total_wal_size);
}
if (immutable_db_options_.db_write_buffer_size > 0) {
bsize = std::min<size_t>(bsize, immutable_db_options_.db_write_buffer_size);
}
if (immutable_db_options_.write_buffer_manager &&
immutable_db_options_.write_buffer_manager->enabled()) {
bsize = std::min<size_t>(
bsize, immutable_db_options_.write_buffer_manager->buffer_size());
}
return bsize;
}
namespace {
struct IterState {
IterState(DBImpl* _db, InstrumentedMutex* _mu, SuperVersion* _super_version,
bool _background_purge)
: db(_db),
mu(_mu),
super_version(_super_version),
background_purge(_background_purge) {}
DBImpl* db;
InstrumentedMutex* mu;
SuperVersion* super_version;
bool background_purge;
};
static void CleanupIteratorState(void* arg1, void* arg2) {
IterState* state = reinterpret_cast<IterState*>(arg1);
if (state->super_version->Unref()) {
// Job id == 0 means that this is not our background process, but rather
// user thread
JobContext job_context(0);
state->mu->Lock();
state->super_version->Cleanup();
state->db->FindObsoleteFiles(&job_context, false, true);
if (state->background_purge) {
state->db->ScheduleBgLogWriterClose(&job_context);
}
state->mu->Unlock();
delete state->super_version;
if (job_context.HaveSomethingToDelete()) {
if (state->background_purge) {
// PurgeObsoleteFiles here does not delete files. Instead, it adds the
// files to be deleted to a job queue, and deletes it in a separate
// background thread.
state->db->PurgeObsoleteFiles(job_context, true /* schedule only */);
state->mu->Lock();
state->db->SchedulePurge();
state->mu->Unlock();
} else {
state->db->PurgeObsoleteFiles(job_context);
}
}
job_context.Clean();
MemTableListVersion Summary: MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful. This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team. I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet()) Test Plan: `make check` works. I will also do `make valgrind_check` before commit Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak Reviewed By: kailiu CC: leveldb Differential Revision: https://reviews.facebook.net/D15255
11 years ago
}
delete state;
}
} // namespace
InternalIterator* DBImpl::NewInternalIterator(const ReadOptions& read_options,
ColumnFamilyData* cfd,
SuperVersion* super_version,
Arena* arena) {
InternalIterator* internal_iter;
assert(arena != nullptr);
// Need to create internal iterator from the arena.
MergeIteratorBuilder merge_iter_builder(
&cfd->internal_comparator(), arena,
!read_options.total_order_seek &&
cfd->ioptions()->prefix_extractor != nullptr);
// Collect iterator for mutable mem
merge_iter_builder.AddIterator(
super_version->mem->NewIterator(read_options, arena));
// Collect all needed child iterators for immutable memtables
super_version->imm->AddIterators(read_options, &merge_iter_builder);
// Collect iterators for files in L0 - Ln
super_version->current->AddIterators(read_options, env_options_,
&merge_iter_builder);
internal_iter = merge_iter_builder.Finish();
IterState* cleanup =
new IterState(this, &mutex_, super_version,
read_options.background_purge_on_iterator_cleanup);
internal_iter->RegisterCleanup(CleanupIteratorState, cleanup, nullptr);
return internal_iter;
}
ColumnFamilyHandle* DBImpl::DefaultColumnFamily() const {
return default_cf_handle_;
}
Status DBImpl::Get(const ReadOptions& read_options,
ColumnFamilyHandle* column_family, const Slice& key,
std::string* value) {
return GetImpl(read_options, column_family, key, value);
}
// JobContext gets created and destructed outside of the lock --
// we
// use this convinently to:
// * malloc one SuperVersion() outside of the lock -- new_superversion
// * delete SuperVersion()s outside of the lock -- superversions_to_free
//
// However, if InstallSuperVersionAndScheduleWork() gets called twice with the
// same job_context, we can't reuse the SuperVersion() that got
// malloced because
// first call already used it. In that rare case, we take a hit and create a
// new SuperVersion() inside of the mutex. We do similar thing
// for superversion_to_free
void DBImpl::InstallSuperVersionAndScheduleWorkWrapper(
ColumnFamilyData* cfd, JobContext* job_context,
const MutableCFOptions& mutable_cf_options) {
mutex_.AssertHeld();
SuperVersion* old_superversion = InstallSuperVersionAndScheduleWork(
cfd, job_context->new_superversion, mutable_cf_options);
job_context->new_superversion = nullptr;
job_context->superversions_to_free.push_back(old_superversion);
}
SuperVersion* DBImpl::InstallSuperVersionAndScheduleWork(
ColumnFamilyData* cfd, SuperVersion* new_sv,
const MutableCFOptions& mutable_cf_options) {
mutex_.AssertHeld();
// Update max_total_in_memory_state_
size_t old_memtable_size = 0;
auto* old_sv = cfd->GetSuperVersion();
if (old_sv) {
old_memtable_size = old_sv->mutable_cf_options.write_buffer_size *
old_sv->mutable_cf_options.max_write_buffer_number;
}
auto* old = cfd->InstallSuperVersion(
new_sv ? new_sv : new SuperVersion(), &mutex_, mutable_cf_options);
Rewritten system for scheduling background work Summary: When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue. The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction. Here are the performance results: Command: ./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000 --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333 Before the patch: fillrandom : 26.950 micros/op 37105 ops/sec; 4.1 MB/s After the patch: fillrandom : 17.404 micros/op 57456 ops/sec; 6.4 MB/s Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got: fillrandom : 7.590 micros/op 131758 ops/sec; 14.6 MB/s Test Plan: make check two stress tests: Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 max_background_flushes=0, to verify that this case also works correctly ./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: ljin, rven, yhchiang, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30123
10 years ago
// Whenever we install new SuperVersion, we might need to issue new flushes or
// compactions.
SchedulePendingFlush(cfd);
SchedulePendingCompaction(cfd);
MaybeScheduleFlushOrCompaction();
// Update max_total_in_memory_state_
max_total_in_memory_state_ =
max_total_in_memory_state_ - old_memtable_size +
mutable_cf_options.write_buffer_size *
mutable_cf_options.max_write_buffer_number;
return old;
}
Status DBImpl::GetImpl(const ReadOptions& read_options,
ColumnFamilyHandle* column_family, const Slice& key,
std::string* value, bool* value_found) {
StopWatch sw(env_, stats_, DB_GET);
PERF_TIMER_GUARD(get_snapshot_time);
auto cfh = reinterpret_cast<ColumnFamilyHandleImpl*>(column_family);
auto cfd = cfh->cfd();
SequenceNumber snapshot;
if (read_options.snapshot != nullptr) {
snapshot = reinterpret_cast<const SnapshotImpl*>(
read_options.snapshot)->number_;
} else {
snapshot = versions_->LastSequence();
}
// Acquire SuperVersion
SuperVersion* sv = GetAndRefSuperVersion(cfd);
[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
11 years ago
// Prepare to store a list of merge operations if merge occurs.
MergeContext merge_context;
[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
11 years ago
Status s;
// First look in the memtable, then in the immutable memtable (if any).
// s is both in/out. When in, s could either be OK or MergeInProgress.
[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
11 years ago
// merge_operands will contain the sequence of merges in the latter case.
LookupKey lkey(key, snapshot);
PERF_TIMER_STOP(get_snapshot_time);
bool skip_memtable =
(read_options.read_tier == kPersistedTier && has_unpersisted_data_);
bool done = false;
if (!skip_memtable) {
if (sv->mem->Get(lkey, value, &s, &merge_context)) {
done = true;
RecordTick(stats_, MEMTABLE_HIT);
} else if (sv->imm->Get(lkey, value, &s, &merge_context)) {
done = true;
RecordTick(stats_, MEMTABLE_HIT);
}
}
if (!done) {
PERF_TIMER_GUARD(get_from_output_files_time);
sv->current->Get(read_options, lkey, value, &s, &merge_context,
value_found);
RecordTick(stats_, MEMTABLE_MISS);
}
{
PERF_TIMER_GUARD(get_post_process_time);
ReturnAndCleanupSuperVersion(cfd, sv);
RecordTick(stats_, NUMBER_KEYS_READ);
RecordTick(stats_, BYTES_READ, value->size());
MeasureTime(stats_, BYTES_PER_READ, value->size());
}
return s;
}
std::vector<Status> DBImpl::MultiGet(
const ReadOptions& read_options,
const std::vector<ColumnFamilyHandle*>& column_family,
const std::vector<Slice>& keys, std::vector<std::string>* values) {
StopWatch sw(env_, stats_, DB_MULTIGET);
PERF_TIMER_GUARD(get_snapshot_time);
SequenceNumber snapshot;
struct MultiGetColumnFamilyData {
ColumnFamilyData* cfd;
SuperVersion* super_version;
};
std::unordered_map<uint32_t, MultiGetColumnFamilyData*> multiget_cf_data;
// fill up and allocate outside of mutex
for (auto cf : column_family) {
auto cfh = reinterpret_cast<ColumnFamilyHandleImpl*>(cf);
auto cfd = cfh->cfd();
if (multiget_cf_data.find(cfd->GetID()) == multiget_cf_data.end()) {
auto mgcfd = new MultiGetColumnFamilyData();
mgcfd->cfd = cfd;
multiget_cf_data.insert({cfd->GetID(), mgcfd});
}
}
mutex_.Lock();
if (read_options.snapshot != nullptr) {
snapshot = reinterpret_cast<const SnapshotImpl*>(
read_options.snapshot)->number_;
} else {
snapshot = versions_->LastSequence();
}
for (auto mgd_iter : multiget_cf_data) {
mgd_iter.second->super_version =
mgd_iter.second->cfd->GetSuperVersion()->Ref();
}
mutex_.Unlock();
// Contain a list of merge operations if merge occurs.
MergeContext merge_context;
[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
11 years ago
// Note: this always resizes the values array
size_t num_keys = keys.size();
std::vector<Status> stat_list(num_keys);
values->resize(num_keys);
// Keep track of bytes that we read for statistics-recording later
uint64_t bytes_read = 0;
PERF_TIMER_STOP(get_snapshot_time);
// For each of the given keys, apply the entire "get" process as follows:
// First look in the memtable, then in the immutable memtable (if any).
// s is both in/out. When in, s could either be OK or MergeInProgress.
[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
11 years ago
// merge_operands will contain the sequence of merges in the latter case.
for (size_t i = 0; i < num_keys; ++i) {
merge_context.Clear();
Status& s = stat_list[i];
std::string* value = &(*values)[i];
LookupKey lkey(keys[i], snapshot);
auto cfh = reinterpret_cast<ColumnFamilyHandleImpl*>(column_family[i]);
auto mgd_iter = multiget_cf_data.find(cfh->cfd()->GetID());
assert(mgd_iter != multiget_cf_data.end());
auto mgd = mgd_iter->second;
auto super_version = mgd->super_version;
bool skip_memtable =
(read_options.read_tier == kPersistedTier && has_unpersisted_data_);
bool done = false;
if (!skip_memtable) {
if (super_version->mem->Get(lkey, value, &s, &merge_context)) {
done = true;
// TODO(?): RecordTick(stats_, MEMTABLE_HIT)?
} else if (super_version->imm->Get(lkey, value, &s, &merge_context)) {
done = true;
// TODO(?): RecordTick(stats_, MEMTABLE_HIT)?
}
}
if (!done) {
PERF_TIMER_GUARD(get_from_output_files_time);
super_version->current->Get(read_options, lkey, value, &s,
&merge_context);
// TODO(?): RecordTick(stats_, MEMTABLE_MISS)?
}
if (s.ok()) {
bytes_read += value->size();
}
}
// Post processing (decrement reference counts and record statistics)
PERF_TIMER_GUARD(get_post_process_time);
autovector<SuperVersion*> superversions_to_delete;
// TODO(icanadi) do we need lock here or just around Cleanup()?
mutex_.Lock();
for (auto mgd_iter : multiget_cf_data) {
auto mgd = mgd_iter.second;
if (mgd->super_version->Unref()) {
mgd->super_version->Cleanup();
superversions_to_delete.push_back(mgd->super_version);
}
}
mutex_.Unlock();
for (auto td : superversions_to_delete) {
delete td;
}
for (auto mgd : multiget_cf_data) {
delete mgd.second;
}
RecordTick(stats_, NUMBER_MULTIGET_CALLS);
RecordTick(stats_, NUMBER_MULTIGET_KEYS_READ, num_keys);
RecordTick(stats_, NUMBER_MULTIGET_BYTES_READ, bytes_read);
MeasureTime(stats_, BYTES_PER_MULTIGET, bytes_read);
PERF_TIMER_STOP(get_post_process_time);
return stat_list;
}
Status DBImpl::CreateColumnFamily(const ColumnFamilyOptions& cf_options,
const std::string& column_family_name,
ColumnFamilyHandle** handle) {
Status s;
Status persist_options_status;
*handle = nullptr;
s = CheckCompressionSupported(cf_options);
if (s.ok() && immutable_db_options_.allow_concurrent_memtable_write) {
support for concurrent adds to memtable Summary: This diff adds support for concurrent adds to the skiplist memtable implementations. Memory allocation is made thread-safe by the addition of a spinlock, with small per-core buffers to avoid contention. Concurrent memtable writes are made via an additional method and don't impose a performance overhead on the non-concurrent case, so parallelism can be selected on a per-batch basis. Write thread synchronization is an increasing bottleneck for higher levels of concurrency, so this diff adds --enable_write_thread_adaptive_yield (default off). This feature causes threads joining a write batch group to spin for a short time (default 100 usec) using sched_yield, rather than going to sleep on a mutex. If the timing of the yield calls indicates that another thread has actually run during the yield then spinning is avoided. This option improves performance for concurrent situations even without parallel adds, although it has the potential to increase CPU usage (and the heuristic adaptation is not yet mature). Parallel writes are not currently compatible with inplace updates, update callbacks, or delete filtering. Enable it with --allow_concurrent_memtable_write (and --enable_write_thread_adaptive_yield). Parallel memtable writes are performance neutral when there is no actual parallelism, and in my experiments (SSD server-class Linux and varying contention and key sizes for fillrandom) they are always a performance win when there is more than one thread. Statistics are updated earlier in the write path, dropping the number of DB mutex acquisitions from 2 to 1 for almost all cases. This diff was motivated and inspired by Yahoo's cLSM work. It is more conservative than cLSM: RocksDB's write batch group leader role is preserved (along with all of the existing flush and write throttling logic) and concurrent writers are blocked until all memtable insertions have completed and the sequence number has been advanced, to preserve linearizability. My test config is "db_bench -benchmarks=fillrandom -threads=$T -batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T -level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999 -disable_auto_compactions --max_write_buffer_number=8 -max_background_flushes=8 --disable_wal --write_buffer_size=160000000 --block_size=16384 --allow_concurrent_memtable_write" on a two-socket Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1 thread I get ~440Kops/sec. Peak performance for 1 socket (numactl -N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance across both sockets happens at 30 threads, and is ~900Kops/sec, although with fewer threads there is less performance loss when the system has background work. Test Plan: 1. concurrent stress tests for InlineSkipList and DynamicBloom 2. make clean; make check 3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench 4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench 5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench 6. make clean; OPT=-DROCKSDB_LITE make check 7. verify no perf regressions when disabled Reviewers: igor, sdong Reviewed By: sdong Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba Differential Revision: https://reviews.facebook.net/D50589
9 years ago
s = CheckConcurrentWritesSupported(cf_options);
}
if (!s.ok()) {
return s;
}
{
InstrumentedMutexLock l(&mutex_);
if (versions_->GetColumnFamilySet()->GetColumnFamily(column_family_name) !=
nullptr) {
return Status::InvalidArgument("Column family already exists");
}
VersionEdit edit;
edit.AddColumnFamily(column_family_name);
uint32_t new_id = versions_->GetColumnFamilySet()->GetNextColumnFamilyID();
edit.SetColumnFamily(new_id);
edit.SetLogNumber(logfile_number_);
edit.SetComparatorName(cf_options.comparator->Name());
// LogAndApply will both write the creation in MANIFEST and create
// ColumnFamilyData object
{ // write thread
WriteThread::Writer w;
write_thread_.EnterUnbatched(&w, &mutex_);
// LogAndApply will both write the creation in MANIFEST and create
// ColumnFamilyData object
s = versions_->LogAndApply(nullptr, MutableCFOptions(cf_options), &edit,
&mutex_, directories_.GetDbDir(), false,
&cf_options);
if (s.ok()) {
// If the column family was created successfully, we then persist
// the updated RocksDB options under the same single write thread
persist_options_status = WriteOptionsFile();
}
write_thread_.ExitUnbatched(&w);
}
if (s.ok()) {
single_column_family_mode_ = false;
auto* cfd =
versions_->GetColumnFamilySet()->GetColumnFamily(column_family_name);
assert(cfd != nullptr);
delete InstallSuperVersionAndScheduleWork(
cfd, nullptr, *cfd->GetLatestMutableCFOptions());
if (!cfd->mem()->IsSnapshotSupported()) {
is_snapshot_supported_ = false;
}
*handle = new ColumnFamilyHandleImpl(cfd, this, &mutex_);
Log(InfoLogLevel::INFO_LEVEL, immutable_db_options_.info_log,
"Created column family [%s] (ID %u)", column_family_name.c_str(),
(unsigned)cfd->GetID());
} else {
Log(InfoLogLevel::ERROR_LEVEL, immutable_db_options_.info_log,
"Creating column family [%s] FAILED -- %s",
column_family_name.c_str(), s.ToString().c_str());
}
} // InstrumentedMutexLock l(&mutex_)
// this is outside the mutex
if (s.ok()) {
NewThreadStatusCfInfo(
reinterpret_cast<ColumnFamilyHandleImpl*>(*handle)->cfd());
if (!persist_options_status.ok()) {
if (immutable_db_options_.fail_if_options_file_error) {
s = Status::IOError(
"ColumnFamily has been created, but unable to persist"
"options in CreateColumnFamily()",
persist_options_status.ToString().c_str());
}
Warn(immutable_db_options_.info_log,
"Unable to persist options in CreateColumnFamily() -- %s",
persist_options_status.ToString().c_str());
}
}
return s;
}
Status DBImpl::DropColumnFamily(ColumnFamilyHandle* column_family) {
auto cfh = reinterpret_cast<ColumnFamilyHandleImpl*>(column_family);
auto cfd = cfh->cfd();
if (cfd->GetID() == 0) {
return Status::InvalidArgument("Can't drop default column family");
}
bool cf_support_snapshot = cfd->mem()->IsSnapshotSupported();
VersionEdit edit;
edit.DropColumnFamily();
edit.SetColumnFamily(cfd->GetID());
Status s;
Status options_persist_status;
{
InstrumentedMutexLock l(&mutex_);
if (cfd->IsDropped()) {
s = Status::InvalidArgument("Column family already dropped!\n");
}
if (s.ok()) {
// we drop column family from a single write thread
WriteThread::Writer w;
write_thread_.EnterUnbatched(&w, &mutex_);
s = versions_->LogAndApply(cfd, *cfd->GetLatestMutableCFOptions(),
&edit, &mutex_);
if (s.ok()) {
// If the column family was dropped successfully, we then persist
// the updated RocksDB options under the same single write thread
options_persist_status = WriteOptionsFile();
}
write_thread_.ExitUnbatched(&w);
}
if (!cf_support_snapshot) {
// Dropped Column Family doesn't support snapshot. Need to recalculate
// is_snapshot_supported_.
bool new_is_snapshot_supported = true;
for (auto c : *versions_->GetColumnFamilySet()) {
if (!c->IsDropped() && !c->mem()->IsSnapshotSupported()) {
new_is_snapshot_supported = false;
break;
}
}
is_snapshot_supported_ = new_is_snapshot_supported;
}
}
if (s.ok()) {
// Note that here we erase the associated cf_info of the to-be-dropped
// cfd before its ref-count goes to zero to avoid having to erase cf_info
// later inside db_mutex.
EraseThreadStatusCfInfo(cfd);
assert(cfd->IsDropped());
auto* mutable_cf_options = cfd->GetLatestMutableCFOptions();
max_total_in_memory_state_ -= mutable_cf_options->write_buffer_size *
mutable_cf_options->max_write_buffer_number;
Log(InfoLogLevel::INFO_LEVEL, immutable_db_options_.info_log,
"Dropped column family with id %u\n", cfd->GetID());
if (!options_persist_status.ok()) {
if (immutable_db_options_.fail_if_options_file_error) {
s = Status::IOError(
"ColumnFamily has been dropped, but unable to persist "
"options in DropColumnFamily()",
options_persist_status.ToString().c_str());
}
Warn(immutable_db_options_.info_log,
"Unable to persist options in DropColumnFamily() -- %s",
options_persist_status.ToString().c_str());
}
} else {
Log(InfoLogLevel::ERROR_LEVEL, immutable_db_options_.info_log,
"Dropping column family with id %u FAILED -- %s\n", cfd->GetID(),
s.ToString().c_str());
}
return s;
}
bool DBImpl::KeyMayExist(const ReadOptions& read_options,
ColumnFamilyHandle* column_family, const Slice& key,
std::string* value, bool* value_found) {
if (value_found != nullptr) {
// falsify later if key-may-exist but can't fetch value
*value_found = true;
}
ReadOptions roptions = read_options;
roptions.read_tier = kBlockCacheTier; // read from block cache only
auto s = GetImpl(roptions, column_family, key, value, value_found);
// If block_cache is enabled and the index block of the table didn't
// not present in block_cache, the return value will be Status::Incomplete.
// In this case, key may still exist in the table.
return s.ok() || s.IsIncomplete();
}
Iterator* DBImpl::NewIterator(const ReadOptions& read_options,
ColumnFamilyHandle* column_family) {
if (read_options.read_tier == kPersistedTier) {
return NewErrorIterator(Status::NotSupported(
"ReadTier::kPersistedData is not yet supported in iterators."));
}
auto cfh = reinterpret_cast<ColumnFamilyHandleImpl*>(column_family);
auto cfd = cfh->cfd();
XFUNC_TEST("", "managed_new", managed_new1, xf_manage_new,
reinterpret_cast<DBImpl*>(this),
const_cast<ReadOptions*>(&read_options), is_snapshot_supported_);
if (read_options.managed) {
#ifdef ROCKSDB_LITE
// not supported in lite version
return NewErrorIterator(Status::InvalidArgument(
"Managed Iterators not supported in RocksDBLite."));
#else
if ((read_options.tailing) || (read_options.snapshot != nullptr) ||
(is_snapshot_supported_)) {
return new ManagedIterator(this, read_options, cfd);
}
// Managed iter not supported
return NewErrorIterator(Status::InvalidArgument(
"Managed Iterators not supported without snapshots."));
#endif
} else if (read_options.tailing) {
#ifdef ROCKSDB_LITE
// not supported in lite version
return nullptr;
#else
SuperVersion* sv = cfd->GetReferencedSuperVersion(&mutex_);
auto iter = new ForwardIterator(this, read_options, cfd, sv);
return NewDBIterator(
env_, *cfd->ioptions(), cfd->user_comparator(), iter,
kMaxSequenceNumber,
sv->mutable_cf_options.max_sequential_skip_in_iterations,
sv->version_number, read_options.iterate_upper_bound,
read_options.prefix_same_as_start, read_options.pin_data);
#endif
} else {
SequenceNumber latest_snapshot = versions_->LastSequence();
SuperVersion* sv = cfd->GetReferencedSuperVersion(&mutex_);
auto snapshot =
read_options.snapshot != nullptr
? reinterpret_cast<const SnapshotImpl*>(
read_options.snapshot)->number_
: latest_snapshot;
// Try to generate a DB iterator tree in continuous memory area to be
// cache friendly. Here is an example of result:
// +-------------------------------+
// | |
// | ArenaWrappedDBIter |
// | + |
// | +---> Inner Iterator ------------+
// | | | |
// | | +-- -- -- -- -- -- -- --+ |
// | +--- | Arena | |
// | | | |
// | Allocated Memory: | |
// | | +-------------------+ |
// | | | DBIter | <---+
// | | + |
// | | | +-> iter_ ------------+
// | | | | |
// | | +-------------------+ |
// | | | MergingIterator | <---+
// | | + |
// | | | +->child iter1 ------------+
// | | | | | |
// | | +->child iter2 ----------+ |
// | | | | | | |
// | | | +->child iter3 --------+ | |
// | | | | | |
// | | +-------------------+ | | |
// | | | Iterator1 | <--------+
// | | +-------------------+ | |
// | | | Iterator2 | <------+
// | | +-------------------+ |
// | | | Iterator3 | <----+
// | | +-------------------+
// | | |
// +-------+-----------------------+
//
// ArenaWrappedDBIter inlines an arena area where all the iterators in
// the iterator tree are allocated in the order of being accessed when
// querying.
// Laying out the iterators in the order of being accessed makes it more
// likely that any iterator pointer is close to the iterator it points to so
// that they are likely to be in the same cache line and/or page.
ArenaWrappedDBIter* db_iter = NewArenaWrappedDbIterator(
env_, *cfd->ioptions(), cfd->user_comparator(), snapshot,
sv->mutable_cf_options.max_sequential_skip_in_iterations,
sv->version_number, read_options.iterate_upper_bound,
read_options.prefix_same_as_start, read_options.pin_data,
read_options.total_order_seek);
InternalIterator* internal_iter =
NewInternalIterator(read_options, cfd, sv, db_iter->GetArena());
db_iter->SetIterUnderDBIter(internal_iter);
return db_iter;
}
// To stop compiler from complaining
return nullptr;
}
Status DBImpl::NewIterators(
const ReadOptions& read_options,
const std::vector<ColumnFamilyHandle*>& column_families,
std::vector<Iterator*>* iterators) {
if (read_options.read_tier == kPersistedTier) {
return Status::NotSupported(
"ReadTier::kPersistedData is not yet supported in iterators.");
}
iterators->clear();
iterators->reserve(column_families.size());
XFUNC_TEST("", "managed_new", managed_new1, xf_manage_new,
reinterpret_cast<DBImpl*>(this),
const_cast<ReadOptions*>(&read_options), is_snapshot_supported_);
if (read_options.managed) {
#ifdef ROCKSDB_LITE
return Status::InvalidArgument(
"Managed interator not supported in RocksDB lite");
#else
if ((!read_options.tailing) && (read_options.snapshot == nullptr) &&
(!is_snapshot_supported_)) {
return Status::InvalidArgument(
"Managed interator not supported without snapshots");
}
for (auto cfh : column_families) {
auto cfd = reinterpret_cast<ColumnFamilyHandleImpl*>(cfh)->cfd();
auto iter = new ManagedIterator(this, read_options, cfd);
iterators->push_back(iter);
}
#endif
} else if (read_options.tailing) {
#ifdef ROCKSDB_LITE
return Status::InvalidArgument(
"Tailing interator not supported in RocksDB lite");
#else
for (auto cfh : column_families) {
auto cfd = reinterpret_cast<ColumnFamilyHandleImpl*>(cfh)->cfd();
SuperVersion* sv = cfd->GetReferencedSuperVersion(&mutex_);
auto iter = new ForwardIterator(this, read_options, cfd, sv);
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
iterators->push_back(NewDBIterator(
env_, *cfd->ioptions(), cfd->user_comparator(), iter,
kMaxSequenceNumber,
sv->mutable_cf_options.max_sequential_skip_in_iterations,
sv->version_number, nullptr, false, read_options.pin_data));
}
#endif
} else {
SequenceNumber latest_snapshot = versions_->LastSequence();
for (size_t i = 0; i < column_families.size(); ++i) {
auto* cfd = reinterpret_cast<ColumnFamilyHandleImpl*>(
column_families[i])->cfd();
SuperVersion* sv = cfd->GetReferencedSuperVersion(&mutex_);
auto snapshot =
read_options.snapshot != nullptr
? reinterpret_cast<const SnapshotImpl*>(
read_options.snapshot)->number_
: latest_snapshot;
ArenaWrappedDBIter* db_iter = NewArenaWrappedDbIterator(
env_, *cfd->ioptions(), cfd->user_comparator(), snapshot,
sv->mutable_cf_options.max_sequential_skip_in_iterations,
sv->version_number, nullptr, false, read_options.pin_data);
InternalIterator* internal_iter =
NewInternalIterator(read_options, cfd, sv, db_iter->GetArena());
db_iter->SetIterUnderDBIter(internal_iter);
iterators->push_back(db_iter);
}
}
return Status::OK();
}
const Snapshot* DBImpl::GetSnapshot() { return GetSnapshotImpl(false); }
#ifndef ROCKSDB_LITE
const Snapshot* DBImpl::GetSnapshotForWriteConflictBoundary() {
return GetSnapshotImpl(true);
}
#endif // ROCKSDB_LITE
const Snapshot* DBImpl::GetSnapshotImpl(bool is_write_conflict_boundary) {
int64_t unix_time = 0;
env_->GetCurrentTime(&unix_time); // Ignore error
SnapshotImpl* s = new SnapshotImpl;
InstrumentedMutexLock l(&mutex_);
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
10 years ago
// returns null if the underlying memtable does not support snapshot.
if (!is_snapshot_supported_) {
delete s;
return nullptr;
}
return snapshots_.New(s, versions_->LastSequence(), unix_time,
is_write_conflict_boundary);
}
void DBImpl::ReleaseSnapshot(const Snapshot* s) {
const SnapshotImpl* casted_s = reinterpret_cast<const SnapshotImpl*>(s);
{
InstrumentedMutexLock l(&mutex_);
snapshots_.Delete(casted_s);
}
delete casted_s;
}
// Convenience methods
Status DBImpl::Put(const WriteOptions& o, ColumnFamilyHandle* column_family,
const Slice& key, const Slice& val) {
return DB::Put(o, column_family, key, val);
}
Status DBImpl::Merge(const WriteOptions& o, ColumnFamilyHandle* column_family,
const Slice& key, const Slice& val) {
auto cfh = reinterpret_cast<ColumnFamilyHandleImpl*>(column_family);
if (!cfh->cfd()->ioptions()->merge_operator) {
return Status::NotSupported("Provide a merge_operator when opening DB");
} else {
return DB::Merge(o, column_family, key, val);
}
}
Status DBImpl::Delete(const WriteOptions& write_options,
ColumnFamilyHandle* column_family, const Slice& key) {
return DB::Delete(write_options, column_family, 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
Status DBImpl::SingleDelete(const WriteOptions& write_options,
ColumnFamilyHandle* column_family,
const Slice& key) {
return DB::SingleDelete(write_options, column_family, key);
}
Status DBImpl::Write(const WriteOptions& write_options, WriteBatch* my_batch) {
return WriteImpl(write_options, my_batch, nullptr, nullptr);
}
#ifndef ROCKSDB_LITE
Status DBImpl::WriteWithCallback(const WriteOptions& write_options,
WriteBatch* my_batch,
WriteCallback* callback) {
return WriteImpl(write_options, my_batch, callback, nullptr);
}
#endif // ROCKSDB_LITE
Status DBImpl::WriteImpl(const WriteOptions& write_options,
WriteBatch* my_batch, WriteCallback* callback,
uint64_t* log_used, uint64_t log_ref,
bool disable_memtable) {
if (my_batch == nullptr) {
return Status::Corruption("Batch is nullptr!");
}
if (write_options.timeout_hint_us != 0) {
return Status::InvalidArgument("timeout_hint_us is deprecated");
}
Status status;
bool xfunc_attempted_write = false;
XFUNC_TEST("transaction", "transaction_xftest_write_impl",
xf_transaction_write1, xf_transaction_write, write_options,
immutable_db_options_, my_batch, callback, this, &status,
&xfunc_attempted_write);
if (xfunc_attempted_write) {
// Test already did the write
return status;
}
PERF_TIMER_GUARD(write_pre_and_post_process_time);
WriteThread::Writer w;
w.batch = my_batch;
w.sync = write_options.sync;
w.disableWAL = write_options.disableWAL;
w.disable_memtable = disable_memtable;
w.in_batch_group = false;
w.callback = callback;
w.log_ref = log_ref;
if (!write_options.disableWAL) {
RecordTick(stats_, WRITE_WITH_WAL);
}
StopWatch write_sw(env_, immutable_db_options_.statistics.get(), DB_WRITE);
write_thread_.JoinBatchGroup(&w);
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 (w.state == WriteThread::STATE_PARALLEL_FOLLOWER) {
// we are a non-leader in a parallel group
PERF_TIMER_GUARD(write_memtable_time);
if (log_used != nullptr) {
*log_used = w.log_used;
}
if (w.ShouldWriteToMemtable()) {
ColumnFamilyMemTablesImpl column_family_memtables(
versions_->GetColumnFamilySet());
WriteBatchInternal::SetSequence(w.batch, w.sequence);
w.status = WriteBatchInternal::InsertInto(
&w, &column_family_memtables, &flush_scheduler_,
write_options.ignore_missing_column_families, 0 /*log_number*/, this,
true /*concurrent_memtable_writes*/);
}
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 (write_thread_.CompleteParallelWorker(&w)) {
// we're responsible for early exit
auto last_sequence = w.parallel_group->last_sequence;
support for concurrent adds to memtable Summary: This diff adds support for concurrent adds to the skiplist memtable implementations. Memory allocation is made thread-safe by the addition of a spinlock, with small per-core buffers to avoid contention. Concurrent memtable writes are made via an additional method and don't impose a performance overhead on the non-concurrent case, so parallelism can be selected on a per-batch basis. Write thread synchronization is an increasing bottleneck for higher levels of concurrency, so this diff adds --enable_write_thread_adaptive_yield (default off). This feature causes threads joining a write batch group to spin for a short time (default 100 usec) using sched_yield, rather than going to sleep on a mutex. If the timing of the yield calls indicates that another thread has actually run during the yield then spinning is avoided. This option improves performance for concurrent situations even without parallel adds, although it has the potential to increase CPU usage (and the heuristic adaptation is not yet mature). Parallel writes are not currently compatible with inplace updates, update callbacks, or delete filtering. Enable it with --allow_concurrent_memtable_write (and --enable_write_thread_adaptive_yield). Parallel memtable writes are performance neutral when there is no actual parallelism, and in my experiments (SSD server-class Linux and varying contention and key sizes for fillrandom) they are always a performance win when there is more than one thread. Statistics are updated earlier in the write path, dropping the number of DB mutex acquisitions from 2 to 1 for almost all cases. This diff was motivated and inspired by Yahoo's cLSM work. It is more conservative than cLSM: RocksDB's write batch group leader role is preserved (along with all of the existing flush and write throttling logic) and concurrent writers are blocked until all memtable insertions have completed and the sequence number has been advanced, to preserve linearizability. My test config is "db_bench -benchmarks=fillrandom -threads=$T -batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T -level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999 -disable_auto_compactions --max_write_buffer_number=8 -max_background_flushes=8 --disable_wal --write_buffer_size=160000000 --block_size=16384 --allow_concurrent_memtable_write" on a two-socket Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1 thread I get ~440Kops/sec. Peak performance for 1 socket (numactl -N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance across both sockets happens at 30 threads, and is ~900Kops/sec, although with fewer threads there is less performance loss when the system has background work. Test Plan: 1. concurrent stress tests for InlineSkipList and DynamicBloom 2. make clean; make check 3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench 4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench 5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench 6. make clean; OPT=-DROCKSDB_LITE make check 7. verify no perf regressions when disabled Reviewers: igor, sdong Reviewed By: sdong Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba Differential Revision: https://reviews.facebook.net/D50589
9 years ago
SetTickerCount(stats_, SEQUENCE_NUMBER, last_sequence);
versions_->SetLastSequence(last_sequence);
write_thread_.EarlyExitParallelGroup(&w);
}
assert(w.state == WriteThread::STATE_COMPLETED);
// STATE_COMPLETED conditional below handles exit
status = w.FinalStatus();
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 (w.state == WriteThread::STATE_COMPLETED) {
if (log_used != nullptr) {
*log_used = w.log_used;
}
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
// write is complete and leader has updated sequence
RecordTick(stats_, WRITE_DONE_BY_OTHER);
return w.FinalStatus();
}
// else we are the leader of the write batch group
support for concurrent adds to memtable Summary: This diff adds support for concurrent adds to the skiplist memtable implementations. Memory allocation is made thread-safe by the addition of a spinlock, with small per-core buffers to avoid contention. Concurrent memtable writes are made via an additional method and don't impose a performance overhead on the non-concurrent case, so parallelism can be selected on a per-batch basis. Write thread synchronization is an increasing bottleneck for higher levels of concurrency, so this diff adds --enable_write_thread_adaptive_yield (default off). This feature causes threads joining a write batch group to spin for a short time (default 100 usec) using sched_yield, rather than going to sleep on a mutex. If the timing of the yield calls indicates that another thread has actually run during the yield then spinning is avoided. This option improves performance for concurrent situations even without parallel adds, although it has the potential to increase CPU usage (and the heuristic adaptation is not yet mature). Parallel writes are not currently compatible with inplace updates, update callbacks, or delete filtering. Enable it with --allow_concurrent_memtable_write (and --enable_write_thread_adaptive_yield). Parallel memtable writes are performance neutral when there is no actual parallelism, and in my experiments (SSD server-class Linux and varying contention and key sizes for fillrandom) they are always a performance win when there is more than one thread. Statistics are updated earlier in the write path, dropping the number of DB mutex acquisitions from 2 to 1 for almost all cases. This diff was motivated and inspired by Yahoo's cLSM work. It is more conservative than cLSM: RocksDB's write batch group leader role is preserved (along with all of the existing flush and write throttling logic) and concurrent writers are blocked until all memtable insertions have completed and the sequence number has been advanced, to preserve linearizability. My test config is "db_bench -benchmarks=fillrandom -threads=$T -batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T -level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999 -disable_auto_compactions --max_write_buffer_number=8 -max_background_flushes=8 --disable_wal --write_buffer_size=160000000 --block_size=16384 --allow_concurrent_memtable_write" on a two-socket Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1 thread I get ~440Kops/sec. Peak performance for 1 socket (numactl -N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance across both sockets happens at 30 threads, and is ~900Kops/sec, although with fewer threads there is less performance loss when the system has background work. Test Plan: 1. concurrent stress tests for InlineSkipList and DynamicBloom 2. make clean; make check 3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench 4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench 5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench 6. make clean; OPT=-DROCKSDB_LITE make check 7. verify no perf regressions when disabled Reviewers: igor, sdong Reviewed By: sdong Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba Differential Revision: https://reviews.facebook.net/D50589
9 years ago
assert(w.state == WriteThread::STATE_GROUP_LEADER);
WriteContext context;
mutex_.Lock();
if (!write_options.disableWAL) {
default_cf_internal_stats_->AddDBStats(InternalStats::WRITE_WITH_WAL, 1);
}
RecordTick(stats_, WRITE_DONE_BY_SELF);
default_cf_internal_stats_->AddDBStats(InternalStats::WRITE_DONE_BY_SELF, 1);
// Once reaches this point, the current writer "w" will try to do its write
// job. It may also pick up some of the remaining writers in the "writers_"
// when it finds suitable, and finish them in the same write batch.
// This is how a write job could be done by the other writer.
assert(!single_column_family_mode_ ||
versions_->GetColumnFamilySet()->NumberOfColumnFamilies() == 1);
uint64_t max_total_wal_size = (immutable_db_options_.max_total_wal_size == 0)
? 4 * max_total_in_memory_state_
: immutable_db_options_.max_total_wal_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
9 years ago
if (UNLIKELY(!single_column_family_mode_ &&
alive_log_files_.begin()->getting_flushed == false &&
total_log_size_ > max_total_wal_size)) {
uint64_t flush_column_family_if_log_file = alive_log_files_.begin()->number;
alive_log_files_.begin()->getting_flushed = true;
Log(InfoLogLevel::INFO_LEVEL, immutable_db_options_.info_log,
"Flushing all column families with data in WAL number %" PRIu64
". Total log size is %" PRIu64 " while max_total_wal_size is %" PRIu64,
flush_column_family_if_log_file, total_log_size_, max_total_wal_size);
// no need to refcount because drop is happening in write thread, so can't
// happen while we're in the write thread
for (auto cfd : *versions_->GetColumnFamilySet()) {
if (cfd->IsDropped()) {
continue;
}
if (cfd->GetLogNumber() <= flush_column_family_if_log_file) {
status = SwitchMemtable(cfd, &context);
if (!status.ok()) {
break;
}
cfd->imm()->FlushRequested();
Rewritten system for scheduling background work Summary: When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue. The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction. Here are the performance results: Command: ./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000 --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333 Before the patch: fillrandom : 26.950 micros/op 37105 ops/sec; 4.1 MB/s After the patch: fillrandom : 17.404 micros/op 57456 ops/sec; 6.4 MB/s Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got: fillrandom : 7.590 micros/op 131758 ops/sec; 14.6 MB/s Test Plan: make check two stress tests: Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 max_background_flushes=0, to verify that this case also works correctly ./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: ljin, rven, yhchiang, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30123
10 years ago
SchedulePendingFlush(cfd);
}
}
MaybeScheduleFlushOrCompaction();
} else if (UNLIKELY(write_buffer_manager_->ShouldFlush())) {
// Before a new memtable is added in SwitchMemtable(),
// write_buffer_manager_->ShouldFlush() will keep returning true. If another
// thread is writing to another DB with the same write buffer, they may also
// be flushed. We may end up with flushing much more DBs than needed. It's
// suboptimal but still correct.
Log(InfoLogLevel::INFO_LEVEL, immutable_db_options_.info_log,
"Flushing column family with largest mem table size. Write buffer is "
"using %" PRIu64 " bytes out of a total of %" PRIu64 ".",
write_buffer_manager_->memory_usage(),
write_buffer_manager_->buffer_size());
// no need to refcount because drop is happening in write thread, so can't
// happen while we're in the write thread
ColumnFamilyData* largest_cfd = nullptr;
size_t largest_cfd_size = 0;
for (auto cfd : *versions_->GetColumnFamilySet()) {
if (cfd->IsDropped()) {
continue;
}
if (!cfd->mem()->IsEmpty()) {
// We only consider active mem table, hoping immutable memtable is
// already in the process of flushing.
size_t cfd_size = cfd->mem()->ApproximateMemoryUsage();
if (largest_cfd == nullptr || cfd_size > largest_cfd_size) {
largest_cfd = cfd;
largest_cfd_size = cfd_size;
}
}
}
if (largest_cfd != nullptr) {
status = SwitchMemtable(largest_cfd, &context);
if (status.ok()) {
largest_cfd->imm()->FlushRequested();
SchedulePendingFlush(largest_cfd);
MaybeScheduleFlushOrCompaction();
}
}
}
if (UNLIKELY(status.ok() && !bg_error_.ok())) {
status = bg_error_;
}
if (UNLIKELY(status.ok() && !flush_scheduler_.Empty())) {
status = ScheduleFlushes(&context);
}
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 (UNLIKELY(status.ok() && (write_controller_.IsStopped() ||
write_controller_.NeedsDelay()))) {
PERF_TIMER_STOP(write_pre_and_post_process_time);
PERF_TIMER_GUARD(write_delay_time);
// We don't know size of curent batch so that we always use the size
// for previous one. It might create a fairness issue that expiration
// might happen for smaller writes but larger writes can go through.
// Can optimize it if it is an issue.
status = DelayWrite(last_batch_group_size_);
PERF_TIMER_START(write_pre_and_post_process_time);
}
uint64_t last_sequence = versions_->LastSequence();
WriteThread::Writer* last_writer = &w;
autovector<WriteThread::Writer*> write_group;
bool need_log_sync = !write_options.disableWAL && write_options.sync;
bool need_log_dir_sync = need_log_sync && !log_dir_synced_;
if (status.ok()) {
if (need_log_sync) {
while (logs_.front().getting_synced) {
log_sync_cv_.Wait();
}
for (auto& log : logs_) {
assert(!log.getting_synced);
log.getting_synced = true;
}
}
// Add to log and apply to memtable. We can release the lock
// during this phase since &w is currently responsible for logging
// and protects against concurrent loggers and concurrent writes
// into memtables
}
mutex_.Unlock();
// At this point the mutex is unlocked
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 exit_completed_early = false;
last_batch_group_size_ =
write_thread_.EnterAsBatchGroupLeader(&w, &last_writer, &write_group);
support for concurrent adds to memtable Summary: This diff adds support for concurrent adds to the skiplist memtable implementations. Memory allocation is made thread-safe by the addition of a spinlock, with small per-core buffers to avoid contention. Concurrent memtable writes are made via an additional method and don't impose a performance overhead on the non-concurrent case, so parallelism can be selected on a per-batch basis. Write thread synchronization is an increasing bottleneck for higher levels of concurrency, so this diff adds --enable_write_thread_adaptive_yield (default off). This feature causes threads joining a write batch group to spin for a short time (default 100 usec) using sched_yield, rather than going to sleep on a mutex. If the timing of the yield calls indicates that another thread has actually run during the yield then spinning is avoided. This option improves performance for concurrent situations even without parallel adds, although it has the potential to increase CPU usage (and the heuristic adaptation is not yet mature). Parallel writes are not currently compatible with inplace updates, update callbacks, or delete filtering. Enable it with --allow_concurrent_memtable_write (and --enable_write_thread_adaptive_yield). Parallel memtable writes are performance neutral when there is no actual parallelism, and in my experiments (SSD server-class Linux and varying contention and key sizes for fillrandom) they are always a performance win when there is more than one thread. Statistics are updated earlier in the write path, dropping the number of DB mutex acquisitions from 2 to 1 for almost all cases. This diff was motivated and inspired by Yahoo's cLSM work. It is more conservative than cLSM: RocksDB's write batch group leader role is preserved (along with all of the existing flush and write throttling logic) and concurrent writers are blocked until all memtable insertions have completed and the sequence number has been advanced, to preserve linearizability. My test config is "db_bench -benchmarks=fillrandom -threads=$T -batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T -level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999 -disable_auto_compactions --max_write_buffer_number=8 -max_background_flushes=8 --disable_wal --write_buffer_size=160000000 --block_size=16384 --allow_concurrent_memtable_write" on a two-socket Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1 thread I get ~440Kops/sec. Peak performance for 1 socket (numactl -N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance across both sockets happens at 30 threads, and is ~900Kops/sec, although with fewer threads there is less performance loss when the system has background work. Test Plan: 1. concurrent stress tests for InlineSkipList and DynamicBloom 2. make clean; make check 3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench 4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench 5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench 6. make clean; OPT=-DROCKSDB_LITE make check 7. verify no perf regressions when disabled Reviewers: igor, sdong Reviewed By: sdong Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba Differential Revision: https://reviews.facebook.net/D50589
9 years ago
if (status.ok()) {
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
// Rules for when we can update the memtable concurrently
// 1. supported by memtable
// 2. Puts are not okay if inplace_update_support
// 3. Deletes or SingleDeletes are not okay if filtering deletes
// (controlled by both batch and memtable setting)
// 4. Merges are not okay
//
// Rules 1..3 are enforced by checking the options
// during startup (CheckConcurrentWritesSupported), so if
// options.allow_concurrent_memtable_write is true then they can be
// assumed to be true. Rule 4 is checked for each batch. We could
// relax rules 2 and 3 if we could prevent write batches from referring
// more than once to a particular key.
bool parallel = immutable_db_options_.allow_concurrent_memtable_write &&
write_group.size() > 1;
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
int total_count = 0;
uint64_t total_byte_size = 0;
for (auto writer : write_group) {
if (writer->CheckCallback(this)) {
if (writer->ShouldWriteToMemtable()) {
total_count += WriteBatchInternal::Count(writer->batch);
parallel = parallel && !writer->batch->HasMerge();
}
if (writer->ShouldWriteToWAL()) {
total_byte_size = WriteBatchInternal::AppendedByteSize(
total_byte_size, WriteBatchInternal::ByteSize(writer->batch));
}
}
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
}
const SequenceNumber current_sequence = last_sequence + 1;
last_sequence += total_count;
// Record statistics
RecordTick(stats_, NUMBER_KEYS_WRITTEN, total_count);
RecordTick(stats_, BYTES_WRITTEN, total_byte_size);
MeasureTime(stats_, BYTES_PER_WRITE, total_byte_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
9 years ago
PERF_TIMER_STOP(write_pre_and_post_process_time);
support for concurrent adds to memtable Summary: This diff adds support for concurrent adds to the skiplist memtable implementations. Memory allocation is made thread-safe by the addition of a spinlock, with small per-core buffers to avoid contention. Concurrent memtable writes are made via an additional method and don't impose a performance overhead on the non-concurrent case, so parallelism can be selected on a per-batch basis. Write thread synchronization is an increasing bottleneck for higher levels of concurrency, so this diff adds --enable_write_thread_adaptive_yield (default off). This feature causes threads joining a write batch group to spin for a short time (default 100 usec) using sched_yield, rather than going to sleep on a mutex. If the timing of the yield calls indicates that another thread has actually run during the yield then spinning is avoided. This option improves performance for concurrent situations even without parallel adds, although it has the potential to increase CPU usage (and the heuristic adaptation is not yet mature). Parallel writes are not currently compatible with inplace updates, update callbacks, or delete filtering. Enable it with --allow_concurrent_memtable_write (and --enable_write_thread_adaptive_yield). Parallel memtable writes are performance neutral when there is no actual parallelism, and in my experiments (SSD server-class Linux and varying contention and key sizes for fillrandom) they are always a performance win when there is more than one thread. Statistics are updated earlier in the write path, dropping the number of DB mutex acquisitions from 2 to 1 for almost all cases. This diff was motivated and inspired by Yahoo's cLSM work. It is more conservative than cLSM: RocksDB's write batch group leader role is preserved (along with all of the existing flush and write throttling logic) and concurrent writers are blocked until all memtable insertions have completed and the sequence number has been advanced, to preserve linearizability. My test config is "db_bench -benchmarks=fillrandom -threads=$T -batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T -level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999 -disable_auto_compactions --max_write_buffer_number=8 -max_background_flushes=8 --disable_wal --write_buffer_size=160000000 --block_size=16384 --allow_concurrent_memtable_write" on a two-socket Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1 thread I get ~440Kops/sec. Peak performance for 1 socket (numactl -N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance across both sockets happens at 30 threads, and is ~900Kops/sec, although with fewer threads there is less performance loss when the system has background work. Test Plan: 1. concurrent stress tests for InlineSkipList and DynamicBloom 2. make clean; make check 3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench 4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench 5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench 6. make clean; OPT=-DROCKSDB_LITE make check 7. verify no perf regressions when disabled Reviewers: igor, sdong Reviewed By: sdong Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba Differential Revision: https://reviews.facebook.net/D50589
9 years ago
if (write_options.disableWAL) {
has_unpersisted_data_ = true;
support for concurrent adds to memtable Summary: This diff adds support for concurrent adds to the skiplist memtable implementations. Memory allocation is made thread-safe by the addition of a spinlock, with small per-core buffers to avoid contention. Concurrent memtable writes are made via an additional method and don't impose a performance overhead on the non-concurrent case, so parallelism can be selected on a per-batch basis. Write thread synchronization is an increasing bottleneck for higher levels of concurrency, so this diff adds --enable_write_thread_adaptive_yield (default off). This feature causes threads joining a write batch group to spin for a short time (default 100 usec) using sched_yield, rather than going to sleep on a mutex. If the timing of the yield calls indicates that another thread has actually run during the yield then spinning is avoided. This option improves performance for concurrent situations even without parallel adds, although it has the potential to increase CPU usage (and the heuristic adaptation is not yet mature). Parallel writes are not currently compatible with inplace updates, update callbacks, or delete filtering. Enable it with --allow_concurrent_memtable_write (and --enable_write_thread_adaptive_yield). Parallel memtable writes are performance neutral when there is no actual parallelism, and in my experiments (SSD server-class Linux and varying contention and key sizes for fillrandom) they are always a performance win when there is more than one thread. Statistics are updated earlier in the write path, dropping the number of DB mutex acquisitions from 2 to 1 for almost all cases. This diff was motivated and inspired by Yahoo's cLSM work. It is more conservative than cLSM: RocksDB's write batch group leader role is preserved (along with all of the existing flush and write throttling logic) and concurrent writers are blocked until all memtable insertions have completed and the sequence number has been advanced, to preserve linearizability. My test config is "db_bench -benchmarks=fillrandom -threads=$T -batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T -level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999 -disable_auto_compactions --max_write_buffer_number=8 -max_background_flushes=8 --disable_wal --write_buffer_size=160000000 --block_size=16384 --allow_concurrent_memtable_write" on a two-socket Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1 thread I get ~440Kops/sec. Peak performance for 1 socket (numactl -N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance across both sockets happens at 30 threads, and is ~900Kops/sec, although with fewer threads there is less performance loss when the system has background work. Test Plan: 1. concurrent stress tests for InlineSkipList and DynamicBloom 2. make clean; make check 3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench 4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench 5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench 6. make clean; OPT=-DROCKSDB_LITE make check 7. verify no perf regressions when disabled Reviewers: igor, sdong Reviewed By: sdong Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba Differential Revision: https://reviews.facebook.net/D50589
9 years ago
}
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
uint64_t log_size = 0;
if (!write_options.disableWAL) {
PERF_TIMER_GUARD(write_wal_time);
support for concurrent adds to memtable Summary: This diff adds support for concurrent adds to the skiplist memtable implementations. Memory allocation is made thread-safe by the addition of a spinlock, with small per-core buffers to avoid contention. Concurrent memtable writes are made via an additional method and don't impose a performance overhead on the non-concurrent case, so parallelism can be selected on a per-batch basis. Write thread synchronization is an increasing bottleneck for higher levels of concurrency, so this diff adds --enable_write_thread_adaptive_yield (default off). This feature causes threads joining a write batch group to spin for a short time (default 100 usec) using sched_yield, rather than going to sleep on a mutex. If the timing of the yield calls indicates that another thread has actually run during the yield then spinning is avoided. This option improves performance for concurrent situations even without parallel adds, although it has the potential to increase CPU usage (and the heuristic adaptation is not yet mature). Parallel writes are not currently compatible with inplace updates, update callbacks, or delete filtering. Enable it with --allow_concurrent_memtable_write (and --enable_write_thread_adaptive_yield). Parallel memtable writes are performance neutral when there is no actual parallelism, and in my experiments (SSD server-class Linux and varying contention and key sizes for fillrandom) they are always a performance win when there is more than one thread. Statistics are updated earlier in the write path, dropping the number of DB mutex acquisitions from 2 to 1 for almost all cases. This diff was motivated and inspired by Yahoo's cLSM work. It is more conservative than cLSM: RocksDB's write batch group leader role is preserved (along with all of the existing flush and write throttling logic) and concurrent writers are blocked until all memtable insertions have completed and the sequence number has been advanced, to preserve linearizability. My test config is "db_bench -benchmarks=fillrandom -threads=$T -batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T -level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999 -disable_auto_compactions --max_write_buffer_number=8 -max_background_flushes=8 --disable_wal --write_buffer_size=160000000 --block_size=16384 --allow_concurrent_memtable_write" on a two-socket Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1 thread I get ~440Kops/sec. Peak performance for 1 socket (numactl -N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance across both sockets happens at 30 threads, and is ~900Kops/sec, although with fewer threads there is less performance loss when the system has background work. Test Plan: 1. concurrent stress tests for InlineSkipList and DynamicBloom 2. make clean; make check 3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench 4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench 5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench 6. make clean; OPT=-DROCKSDB_LITE make check 7. verify no perf regressions when disabled Reviewers: igor, sdong Reviewed By: sdong Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba Differential Revision: https://reviews.facebook.net/D50589
9 years ago
WriteBatch* merged_batch = nullptr;
if (write_group.size() == 1 && write_group[0]->ShouldWriteToWAL() &&
write_group[0]->batch->GetWalTerminationPoint().is_cleared()) {
// we simply write the first WriteBatch to WAL if the group only
// contains one batch, that batch should be written to the WAL,
// and the batch is not wanting to be truncated
merged_batch = write_group[0]->batch;
write_group[0]->log_used = logfile_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
} else {
// WAL needs all of the batches flattened into a single batch.
// We could avoid copying here with an iov-like AddRecord
// interface
merged_batch = &tmp_batch_;
for (auto writer : write_group) {
if (writer->ShouldWriteToWAL()) {
WriteBatchInternal::Append(merged_batch, writer->batch,
/*WAL_only*/ true);
}
writer->log_used = logfile_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 (log_used != nullptr) {
*log_used = logfile_number_;
}
WriteBatchInternal::SetSequence(merged_batch, current_sequence);
support for concurrent adds to memtable Summary: This diff adds support for concurrent adds to the skiplist memtable implementations. Memory allocation is made thread-safe by the addition of a spinlock, with small per-core buffers to avoid contention. Concurrent memtable writes are made via an additional method and don't impose a performance overhead on the non-concurrent case, so parallelism can be selected on a per-batch basis. Write thread synchronization is an increasing bottleneck for higher levels of concurrency, so this diff adds --enable_write_thread_adaptive_yield (default off). This feature causes threads joining a write batch group to spin for a short time (default 100 usec) using sched_yield, rather than going to sleep on a mutex. If the timing of the yield calls indicates that another thread has actually run during the yield then spinning is avoided. This option improves performance for concurrent situations even without parallel adds, although it has the potential to increase CPU usage (and the heuristic adaptation is not yet mature). Parallel writes are not currently compatible with inplace updates, update callbacks, or delete filtering. Enable it with --allow_concurrent_memtable_write (and --enable_write_thread_adaptive_yield). Parallel memtable writes are performance neutral when there is no actual parallelism, and in my experiments (SSD server-class Linux and varying contention and key sizes for fillrandom) they are always a performance win when there is more than one thread. Statistics are updated earlier in the write path, dropping the number of DB mutex acquisitions from 2 to 1 for almost all cases. This diff was motivated and inspired by Yahoo's cLSM work. It is more conservative than cLSM: RocksDB's write batch group leader role is preserved (along with all of the existing flush and write throttling logic) and concurrent writers are blocked until all memtable insertions have completed and the sequence number has been advanced, to preserve linearizability. My test config is "db_bench -benchmarks=fillrandom -threads=$T -batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T -level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999 -disable_auto_compactions --max_write_buffer_number=8 -max_background_flushes=8 --disable_wal --write_buffer_size=160000000 --block_size=16384 --allow_concurrent_memtable_write" on a two-socket Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1 thread I get ~440Kops/sec. Peak performance for 1 socket (numactl -N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance across both sockets happens at 30 threads, and is ~900Kops/sec, although with fewer threads there is less performance loss when the system has background work. Test Plan: 1. concurrent stress tests for InlineSkipList and DynamicBloom 2. make clean; make check 3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench 4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench 5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench 6. make clean; OPT=-DROCKSDB_LITE make check 7. verify no perf regressions when disabled Reviewers: igor, sdong Reviewed By: sdong Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba Differential Revision: https://reviews.facebook.net/D50589
9 years ago
Slice log_entry = WriteBatchInternal::Contents(merged_batch);
status = logs_.back().writer->AddRecord(log_entry);
total_log_size_ += log_entry.size();
alive_log_files_.back().AddSize(log_entry.size());
log_empty_ = false;
log_size = log_entry.size();
RecordTick(stats_, WAL_FILE_BYTES, log_size);
if (status.ok() && need_log_sync) {
RecordTick(stats_, WAL_FILE_SYNCED);
StopWatch sw(env_, stats_, WAL_FILE_SYNC_MICROS);
// It's safe to access logs_ with unlocked mutex_ here because:
// - we've set getting_synced=true for all logs,
// so other threads won't pop from logs_ while we're here,
// - only writer thread can push to logs_, and we're in
// writer thread, so no one will push to logs_,
// - as long as other threads don't modify it, it's safe to read
// from std::deque from multiple threads concurrently.
for (auto& log : logs_) {
status = log.writer->file()->Sync(immutable_db_options_.use_fsync);
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 (!status.ok()) {
break;
}
}
if (status.ok() && need_log_dir_sync) {
// We only sync WAL directory the first time WAL syncing is
// requested, so that in case users never turn on WAL sync,
// we can avoid the disk I/O in the write code path.
status = directories_.GetWalDir()->Fsync();
}
}
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 (merged_batch == &tmp_batch_) {
tmp_batch_.Clear();
}
}
if (status.ok()) {
PERF_TIMER_GUARD(write_memtable_time);
support for concurrent adds to memtable Summary: This diff adds support for concurrent adds to the skiplist memtable implementations. Memory allocation is made thread-safe by the addition of a spinlock, with small per-core buffers to avoid contention. Concurrent memtable writes are made via an additional method and don't impose a performance overhead on the non-concurrent case, so parallelism can be selected on a per-batch basis. Write thread synchronization is an increasing bottleneck for higher levels of concurrency, so this diff adds --enable_write_thread_adaptive_yield (default off). This feature causes threads joining a write batch group to spin for a short time (default 100 usec) using sched_yield, rather than going to sleep on a mutex. If the timing of the yield calls indicates that another thread has actually run during the yield then spinning is avoided. This option improves performance for concurrent situations even without parallel adds, although it has the potential to increase CPU usage (and the heuristic adaptation is not yet mature). Parallel writes are not currently compatible with inplace updates, update callbacks, or delete filtering. Enable it with --allow_concurrent_memtable_write (and --enable_write_thread_adaptive_yield). Parallel memtable writes are performance neutral when there is no actual parallelism, and in my experiments (SSD server-class Linux and varying contention and key sizes for fillrandom) they are always a performance win when there is more than one thread. Statistics are updated earlier in the write path, dropping the number of DB mutex acquisitions from 2 to 1 for almost all cases. This diff was motivated and inspired by Yahoo's cLSM work. It is more conservative than cLSM: RocksDB's write batch group leader role is preserved (along with all of the existing flush and write throttling logic) and concurrent writers are blocked until all memtable insertions have completed and the sequence number has been advanced, to preserve linearizability. My test config is "db_bench -benchmarks=fillrandom -threads=$T -batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T -level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999 -disable_auto_compactions --max_write_buffer_number=8 -max_background_flushes=8 --disable_wal --write_buffer_size=160000000 --block_size=16384 --allow_concurrent_memtable_write" on a two-socket Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1 thread I get ~440Kops/sec. Peak performance for 1 socket (numactl -N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance across both sockets happens at 30 threads, and is ~900Kops/sec, although with fewer threads there is less performance loss when the system has background work. Test Plan: 1. concurrent stress tests for InlineSkipList and DynamicBloom 2. make clean; make check 3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench 4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench 5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench 6. make clean; OPT=-DROCKSDB_LITE make check 7. verify no perf regressions when disabled Reviewers: igor, sdong Reviewed By: sdong Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba Differential Revision: https://reviews.facebook.net/D50589
9 years ago
{
// Update stats while we are an exclusive group leader, so we know
// that nobody else can be writing to these particular stats.
// We're optimistic, updating the stats before we successfully
// commit. That lets us release our leader status early in
// some cases.
auto stats = default_cf_internal_stats_;
stats->AddDBStats(InternalStats::BYTES_WRITTEN, total_byte_size);
stats->AddDBStats(InternalStats::NUMBER_KEYS_WRITTEN, total_count);
if (!write_options.disableWAL) {
if (write_options.sync) {
stats->AddDBStats(InternalStats::WAL_FILE_SYNCED, 1);
}
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
stats->AddDBStats(InternalStats::WAL_FILE_BYTES, log_size);
}
uint64_t for_other = write_group.size() - 1;
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 (for_other > 0) {
stats->AddDBStats(InternalStats::WRITE_DONE_BY_OTHER, for_other);
if (!write_options.disableWAL) {
stats->AddDBStats(InternalStats::WRITE_WITH_WAL, for_other);
}
}
}
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 (!parallel) {
status = WriteBatchInternal::InsertInto(
write_group, current_sequence, column_family_memtables_.get(),
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
&flush_scheduler_, write_options.ignore_missing_column_families,
0 /*log_number*/, this);
if (status.ok()) {
// There were no write failures. Set leader's status
// in case the write callback returned a non-ok status.
status = w.FinalStatus();
}
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
} else {
WriteThread::ParallelGroup pg;
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
pg.leader = &w;
pg.last_writer = last_writer;
pg.last_sequence = last_sequence;
support for concurrent adds to memtable Summary: This diff adds support for concurrent adds to the skiplist memtable implementations. Memory allocation is made thread-safe by the addition of a spinlock, with small per-core buffers to avoid contention. Concurrent memtable writes are made via an additional method and don't impose a performance overhead on the non-concurrent case, so parallelism can be selected on a per-batch basis. Write thread synchronization is an increasing bottleneck for higher levels of concurrency, so this diff adds --enable_write_thread_adaptive_yield (default off). This feature causes threads joining a write batch group to spin for a short time (default 100 usec) using sched_yield, rather than going to sleep on a mutex. If the timing of the yield calls indicates that another thread has actually run during the yield then spinning is avoided. This option improves performance for concurrent situations even without parallel adds, although it has the potential to increase CPU usage (and the heuristic adaptation is not yet mature). Parallel writes are not currently compatible with inplace updates, update callbacks, or delete filtering. Enable it with --allow_concurrent_memtable_write (and --enable_write_thread_adaptive_yield). Parallel memtable writes are performance neutral when there is no actual parallelism, and in my experiments (SSD server-class Linux and varying contention and key sizes for fillrandom) they are always a performance win when there is more than one thread. Statistics are updated earlier in the write path, dropping the number of DB mutex acquisitions from 2 to 1 for almost all cases. This diff was motivated and inspired by Yahoo's cLSM work. It is more conservative than cLSM: RocksDB's write batch group leader role is preserved (along with all of the existing flush and write throttling logic) and concurrent writers are blocked until all memtable insertions have completed and the sequence number has been advanced, to preserve linearizability. My test config is "db_bench -benchmarks=fillrandom -threads=$T -batch_size=1 -memtablerep=skip_list -value_size=100 --num=1000000/$T -level0_slowdown_writes_trigger=9999 -level0_stop_writes_trigger=9999 -disable_auto_compactions --max_write_buffer_number=8 -max_background_flushes=8 --disable_wal --write_buffer_size=160000000 --block_size=16384 --allow_concurrent_memtable_write" on a two-socket Xeon E5-2660 @ 2.2Ghz with lots of memory and an SSD hard drive. With 1 thread I get ~440Kops/sec. Peak performance for 1 socket (numactl -N1) is slightly more than 1Mops/sec, at 16 threads. Peak performance across both sockets happens at 30 threads, and is ~900Kops/sec, although with fewer threads there is less performance loss when the system has background work. Test Plan: 1. concurrent stress tests for InlineSkipList and DynamicBloom 2. make clean; make check 3. make clean; DISABLE_JEMALLOC=1 make valgrind_check; valgrind db_bench 4. make clean; COMPILE_WITH_TSAN=1 make all check; db_bench 5. make clean; COMPILE_WITH_ASAN=1 make all check; db_bench 6. make clean; OPT=-DROCKSDB_LITE make check 7. verify no perf regressions when disabled Reviewers: igor, sdong Reviewed By: sdong Subscribers: MarkCallaghan, IslamAbdelRahman, anthony, yhchiang, rven, sdong, guyg8, kradhakrishnan, dhruba Differential Revision: https://reviews.facebook.net/D50589
9 years ago
pg.early_exit_allowed = !need_log_sync;
pg.running.store(static_cast<uint32_t>(write_group.size()),
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
9 years ago
write_thread_.LaunchParallelFollowers(&pg, current_sequence);
if (w.ShouldWriteToMemtable()) {
// do leader write
ColumnFamilyMemTablesImpl column_family_memtables(
versions_->GetColumnFamilySet());
assert(w.sequence == current_sequence);
WriteBatchInternal::SetSequence(w.batch, w.sequence);
w.status = WriteBatchInternal::InsertInto(
&w, &column_family_memtables, &flush_scheduler_,
write_options.ignore_missing_column_families, 0 /*log_number*/,
this, true /*concurrent_memtable_writes*/);
}
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
// CompleteParallelWorker returns true if this thread should
// handle exit, false means somebody else did
exit_completed_early = !write_thread_.CompleteParallelWorker(&w);
status = w.FinalStatus();
}
if (!exit_completed_early && w.status.ok()) {
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
SetTickerCount(stats_, SEQUENCE_NUMBER, last_sequence);
versions_->SetLastSequence(last_sequence);
if (!need_log_sync) {
write_thread_.ExitAsBatchGroupLeader(&w, last_writer, w.status);
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
exit_completed_early = true;
}
make internal stats independent of statistics Summary: also make it aware of column family output from db_bench ``` ** Compaction Stats [default] ** Level Files Size(MB) Score Read(GB) Rn(GB) Rnp1(GB) Write(GB) Wnew(GB) RW-Amp W-Amp Rd(MB/s) Wr(MB/s) Rn(cnt) Rnp1(cnt) Wnp1(cnt) Wnew(cnt) Comp(sec) Comp(cnt) Avg(sec) Stall(sec) Stall(cnt) Avg(ms) ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ L0 14 956 0.9 0.0 0.0 0.0 2.7 2.7 0.0 0.0 0.0 111.6 0 0 0 0 24 40 0.612 75.20 492387 0.15 L1 21 2001 2.0 5.7 2.0 3.7 5.3 1.6 5.4 2.6 71.2 65.7 31 43 55 12 82 2 41.242 43.72 41183 1.06 L2 217 18974 1.9 16.5 2.0 14.4 15.1 0.7 15.6 7.4 70.1 64.3 17 182 185 3 241 16 15.052 0.00 0 0.00 L3 1641 188245 1.8 9.1 1.1 8.0 8.5 0.5 15.4 7.4 61.3 57.2 9 75 76 1 152 9 16.887 0.00 0 0.00 L4 4447 449025 0.4 13.4 4.8 8.6 9.1 0.5 4.7 1.9 77.8 52.7 38 79 100 21 176 38 4.639 0.00 0 0.00 Sum 6340 659201 0.0 44.7 10.0 34.7 40.6 6.0 32.0 15.2 67.7 61.6 95 379 416 37 676 105 6.439 118.91 533570 0.22 Int 0 0 0.0 1.2 0.4 0.8 1.3 0.5 5.2 2.7 59.1 65.6 3 7 9 2 20 10 2.003 0.00 0 0.00 Stalls(secs): 75.197 level0_slowdown, 0.000 level0_numfiles, 0.000 memtable_compaction, 43.717 leveln_slowdown Stalls(count): 492387 level0_slowdown, 0 level0_numfiles, 0 memtable_compaction, 41183 leveln_slowdown ** DB Stats ** Uptime(secs): 202.1 total, 13.5 interval Cumulative writes: 6291456 writes, 6291456 batches, 1.0 writes per batch, 4.90 ingest GB Cumulative WAL: 6291456 writes, 6291456 syncs, 1.00 writes per sync, 4.90 GB written Interval writes: 1048576 writes, 1048576 batches, 1.0 writes per batch, 836.0 ingest MB Interval WAL: 1048576 writes, 1048576 syncs, 1.00 writes per sync, 0.82 MB written Test Plan: ran it Reviewers: sdong, yhchiang, igor Reviewed By: igor Subscribers: leveldb Differential Revision: https://reviews.facebook.net/D19917
10 years ago
}
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
// A non-OK status here indicates that the state implied by the
// WAL has diverged from the in-memory state. This could be
// because of a corrupt write_batch (very bad), or because the
// client specified an invalid column family and didn't specify
// ignore_missing_column_families.
//
// Is setting bg_error_ enough here? This will at least stop
// compaction and fail any further writes.
if (!status.ok() && bg_error_.ok() && !w.CallbackFailed()) {
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
bg_error_ = status;
}
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
}
}
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
PERF_TIMER_START(write_pre_and_post_process_time);
if (immutable_db_options_.paranoid_checks && !status.ok() &&
!w.CallbackFailed() && !status.IsBusy()) {
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
mutex_.Lock();
if (bg_error_.ok()) {
bg_error_ = status; // stop compaction & fail any further writes
}
mutex_.Unlock();
}
if (need_log_sync) {
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
mutex_.Lock();
MarkLogsSynced(logfile_number_, need_log_dir_sync, status);
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
mutex_.Unlock();
}
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 (!exit_completed_early) {
write_thread_.ExitAsBatchGroupLeader(&w, last_writer, w.status);
}
Rewritten system for scheduling background work Summary: When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue. The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction. Here are the performance results: Command: ./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000 --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333 Before the patch: fillrandom : 26.950 micros/op 37105 ops/sec; 4.1 MB/s After the patch: fillrandom : 17.404 micros/op 57456 ops/sec; 6.4 MB/s Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got: fillrandom : 7.590 micros/op 131758 ops/sec; 14.6 MB/s Test Plan: make check two stress tests: Big number of compactions and flushes: ./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 max_background_flushes=0, to verify that this case also works correctly ./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000 Reviewers: ljin, rven, yhchiang, sdong Reviewed By: sdong Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D30123
10 years ago
return status;
}
// REQUIRES: mutex_ is held
// REQUIRES: this thread is currently at the front of the writer queue
Status DBImpl::DelayWrite(uint64_t num_bytes) {
uint64_t time_delayed = 0;
bool delayed = false;
{
StopWatch sw(env_, stats_, WRITE_STALL, &time_delayed);
auto delay = write_controller_.GetDelay(env_, num_bytes);
if (delay > 0) {
mutex_.Unlock();
delayed = true;
TEST_SYNC_POINT("DBImpl::DelayWrite:Sleep");
// hopefully we don't have to sleep more than 2 billion microseconds
env_->SleepForMicroseconds(static_cast<int>(delay));
mutex_.Lock();
}
while (bg_error_.ok() && write_controller_.IsStopped()) {
delayed = true;
TEST_SYNC_POINT("DBImpl::DelayWrite:Wait");
bg_cv_.Wait();
}
}
if (delayed) {
default_cf_internal_stats_->AddDBStats(InternalStats::WRITE_STALL_MICROS,
time_delayed);
RecordTick(stats_, STALL_MICROS, time_delayed);
}
return bg_error_;
}
Status DBImpl::ScheduleFlushes(WriteContext* context) {
ColumnFamilyData* cfd;
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
while ((cfd = flush_scheduler_.TakeNextColumnFamily()) != nullptr) {
auto status = SwitchMemtable(cfd, context);
if (cfd->Unref()) {
delete cfd;
}
if (!status.ok()) {
return status;
}
}
return Status::OK();
}
#ifndef ROCKSDB_LITE
void DBImpl::NotifyOnMemTableSealed(ColumnFamilyData* cfd,
const MemTableInfo& mem_table_info) {
if (immutable_db_options_.listeners.size() == 0U) {
return;
}
if (shutting_down_.load(std::memory_order_acquire)) {
return;
}
for (auto listener : immutable_db_options_.listeners) {
listener->OnMemTableSealed(mem_table_info);
}
}
#endif // ROCKSDB_LITE
// REQUIRES: mutex_ is held
// REQUIRES: this thread is currently at the front of the writer queue
Status DBImpl::SwitchMemtable(ColumnFamilyData* cfd, WriteContext* context) {
mutex_.AssertHeld();
unique_ptr<WritableFile> lfile;
log::Writer* new_log = nullptr;
MemTable* new_mem = nullptr;
// Attempt to switch to a new memtable and trigger flush of old.
// Do this without holding the dbmutex lock.
assert(versions_->prev_log_number() == 0);
bool creating_new_log = !log_empty_;
uint64_t recycle_log_number = 0;
if (creating_new_log && immutable_db_options_.recycle_log_file_num &&
!log_recycle_files.empty()) {
recycle_log_number = log_recycle_files.front();
log_recycle_files.pop_front();
}
uint64_t new_log_number =
creating_new_log ? versions_->NewFileNumber() : logfile_number_;
SuperVersion* new_superversion = nullptr;
const MutableCFOptions mutable_cf_options = *cfd->GetLatestMutableCFOptions();
// Set current_memtble_info for memtable sealed callback
#ifndef ROCKSDB_LITE
MemTableInfo memtable_info;
memtable_info.cf_name = cfd->GetName();
memtable_info.first_seqno = cfd->mem()->GetFirstSequenceNumber();
memtable_info.earliest_seqno = cfd->mem()->GetEarliestSequenceNumber();
memtable_info.num_entries = cfd->mem()->num_entries();
memtable_info.num_deletes = cfd->mem()->num_deletes();
#endif // ROCKSDB_LITE
// Log this later after lock release. It may be outdated, e.g., if background
// flush happens before logging, but that should be ok.
int num_imm_unflushed = cfd->imm()->NumNotFlushed();
DBOptions db_options =
BuildDBOptions(immutable_db_options_, mutable_db_options_);
mutex_.Unlock();
Status s;
{
if (creating_new_log) {
EnvOptions opt_env_opt =
env_->OptimizeForLogWrite(env_options_, db_options);
if (recycle_log_number) {
Log(InfoLogLevel::INFO_LEVEL, immutable_db_options_.info_log,
"reusing log %" PRIu64 " from recycle list\n", recycle_log_number);
s = env_->ReuseWritableFile(
LogFileName(immutable_db_options_.wal_dir, new_log_number),
LogFileName(immutable_db_options_.wal_dir, recycle_log_number),
&lfile, opt_env_opt);
} else {
s = NewWritableFile(
env_, LogFileName(immutable_db_options_.wal_dir, new_log_number),
&lfile, opt_env_opt);
}
if (s.ok()) {
// Our final size should be less than write_buffer_size
// (compression, etc) but err on the side of caution.
lfile->SetPreallocationBlockSize(
GetWalPreallocateBlockSize(mutable_cf_options.write_buffer_size));
unique_ptr<WritableFileWriter> file_writer(
new WritableFileWriter(std::move(lfile), opt_env_opt));
new_log =
new log::Writer(std::move(file_writer), new_log_number,
immutable_db_options_.recycle_log_file_num > 0);
}
}
if (s.ok()) {
SequenceNumber seq = versions_->LastSequence();
new_mem = cfd->ConstructNewMemtable(mutable_cf_options, seq);
new_superversion = new SuperVersion();
}
#ifndef ROCKSDB_LITE
// PLEASE NOTE: We assume that there are no failable operations
// after lock is acquired below since we are already notifying
// client about mem table becoming immutable.
NotifyOnMemTableSealed(cfd, memtable_info);
#endif //ROCKSDB_LITE
}
Log(InfoLogLevel::INFO_LEVEL, immutable_db_options_.info_log,
"[%s] New memtable created with log file: #%" PRIu64
". Immutable memtables: %d.\n",
cfd->GetName().c_str(), new_log_number, num_imm_unflushed);
mutex_.Lock();
if (!s.ok()) {
// how do we fail if we're not creating new log?
assert(creating_new_log);
assert(!new_mem);
assert(!new_log);
return s;
}
if (creating_new_log) {
logfile_number_ = new_log_number;
assert(new_log != nullptr);
log_empty_ = true;
log_dir_synced_ = false;
logs_.emplace_back(logfile_number_, new_log);
alive_log_files_.push_back(LogFileNumberSize(logfile_number_));
for (auto loop_cfd : *versions_->GetColumnFamilySet()) {
// all this is just optimization to delete logs that
// are no longer needed -- if CF is empty, that means it
// doesn't need that particular log to stay alive, so we just
// advance the log number. no need to persist this in the manifest
if (loop_cfd->mem()->GetFirstSequenceNumber() == 0 &&
Support saving history in memtable_list Summary: For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit. After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much. This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit). However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions. Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here. Reviewers: sdong, rven, igor Reviewed By: igor Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D37443
9 years ago
loop_cfd->imm()->NumNotFlushed() == 0) {
loop_cfd->SetLogNumber(logfile_number_);
}
}
}
cfd->mem()->SetNextLogNumber(logfile_number_);
Support saving history in memtable_list Summary: For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit. After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much. This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit). However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions. Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here. Reviewers: sdong, rven, igor Reviewed By: igor Subscribers: dhruba, leveldb Differential Revision: https://reviews.facebook.net/D37443
9 years ago
cfd->imm()->Add(cfd->mem(), &context->memtables_to_free_);
new_mem->Ref();
cfd->SetMemtable(new_mem);
context->superversions_to_free_.push_back(InstallSuperVersionAndScheduleWork(
cfd, new_superversion, mutable_cf_options));
return s;
}
#ifndef ROCKSDB_LITE
Status DBImpl::GetPropertiesOfAllTables(ColumnFamilyHandle* column_family,
TablePropertiesCollection* props) {
auto cfh = reinterpret_cast<ColumnFamilyHandleImpl*>(column_family);
auto cfd = cfh->cfd();
// Increment the ref count
mutex_.Lock();
auto version = cfd->current();
version->Ref();
mutex_.Unlock();
auto s = version->GetPropertiesOfAllTables(props);
// Decrement the ref count
mutex_.Lock();
version->Unref();
mutex_.Unlock();
return s;
}
Status DBImpl::GetPropertiesOfTablesInRange(ColumnFamilyHandle* column_family,
const Range* range, std::size_t n,
TablePropertiesCollection* props) {
auto cfh = reinterpret_cast<ColumnFamilyHandleImpl*>(column_family);
auto cfd = cfh->cfd();
// Increment the ref count
mutex_.Lock();
auto version = cfd->current();
version->Ref();
mutex_.Unlock();
auto s = version->GetPropertiesOfTablesInRange(range, n, props);
// Decrement the ref count
mutex_.Lock();
version->Unref();
mutex_.Unlock();
return s;
}
#endif // ROCKSDB_LITE
[RocksDB] BackupableDB Summary: In this diff I present you BackupableDB v1. You can easily use it to backup your DB and it will do incremental snapshots for you. Let's first describe how you would use BackupableDB. It's inheriting StackableDB interface so you can easily construct it with your DB object -- it will add a method RollTheSnapshot() to the DB object. When you call RollTheSnapshot(), current snapshot of the DB will be stored in the backup dir. To restore, you can just call RestoreDBFromBackup() on a BackupableDB (which is a static method) and it will restore all files from the backup dir. In the next version, it will even support automatic backuping every X minutes. There are multiple things you can configure: 1. backup_env and db_env can be different, which is awesome because then you can easily backup to HDFS or wherever you feel like. 2. sync - if true, it *guarantees* backup consistency on machine reboot 3. number of snapshots to keep - this will keep last N snapshots around if you want, for some reason, be able to restore from an earlier snapshot. All the backuping is done in incremental fashion - if we already have 00010.sst, we will not copy it again. *IMPORTANT* -- This is based on assumption that 00010.sst never changes - two files named 00010.sst from the same DB will always be exactly the same. Is this true? I always copy manifest, current and log files. 4. You can decide if you want to flush the memtables before you backup, or you're fine with backing up the log files -- either way, you get a complete and consistent view of the database at a time of backup. 5. More things you can find in BackupableDBOptions Here is the directory structure I use: backup_dir/CURRENT_SNAPSHOT - just 4 bytes holding the latest snapshot 0, 1, 2, ... - files containing serialized version of each snapshot - containing a list of files files/*.sst - sst files shared between snapshots - if one snapshot references 00010.sst and another one needs to backup it from the DB, it will just reference the same file files/ 0/, 1/, 2/, ... - snapshot directories containing private snapshot files - current, manifest and log files All the files are ref counted and deleted immediatelly when they get out of scope. Some other stuff in this diff: 1. Added GetEnv() method to the DB. Discussed with @haobo and we agreed that it seems right thing to do. 2. Fixed StackableDB interface. The way it was set up before, I was not able to implement BackupableDB. Test Plan: I have a unittest, but please don't look at this yet. I just hacked it up to help me with debugging. I will write a lot of good tests and update the diff. Also, `make asan_check` Reviewers: dhruba, haobo, emayanke Reviewed By: dhruba CC: leveldb, haobo Differential Revision: https://reviews.facebook.net/D14295
11 years ago
const std::string& DBImpl::GetName() const {
return dbname_;
}
Env* DBImpl::GetEnv() const {
return env_;
}
Options DBImpl::GetOptions(ColumnFamilyHandle* column_family) const {
InstrumentedMutexLock l(&mutex_);
auto cfh = reinterpret_cast<ColumnFamilyHandleImpl*>(column_family);
return Options(BuildDBOptions(immutable_db_options_, mutable_db_options_),
cfh->cfd()->GetLatestCFOptions());
}
DBOptions DBImpl::GetDBOptions() const {
InstrumentedMutexLock l(&mutex_);
return BuildDBOptions(immutable_db_options_, mutable_db_options_);
}
bool DBImpl::GetProperty(ColumnFamilyHandle* column_family,
const Slice& property, std::string* value) {
const DBPropertyInfo* property_info = GetPropertyInfo(property);
value->clear();
auto cfd = reinterpret_cast<ColumnFamilyHandleImpl*>(column_family)->cfd();
if (property_info == nullptr) {
return false;
} else if (property_info->handle_int) {
uint64_t int_value;
bool ret_value =
GetIntPropertyInternal(cfd, *property_info, false, &int_value);
if (ret_value) {
*value = ToString(int_value);
}
return ret_value;
} else if (property_info->handle_string) {
InstrumentedMutexLock l(&mutex_);
return cfd->internal_stats()->GetStringProperty(*property_info, property,
value);
}
// Shouldn't reach here since exactly one of handle_string and handle_int
// should be non-nullptr.
assert(false);
return false;
}
bool DBImpl::GetIntProperty(ColumnFamilyHandle* column_family,
const Slice& property, uint64_t* value) {
const DBPropertyInfo* property_info = GetPropertyInfo(property);
if (property_info == nullptr || property_info->handle_int == nullptr) {
return false;
}
auto cfd = reinterpret_cast<ColumnFamilyHandleImpl*>(column_family)->cfd();
return GetIntPropertyInternal(cfd, *property_info, false, value);
}
bool DBImpl::GetIntPropertyInternal(ColumnFamilyData* cfd,
const DBPropertyInfo& property_info,
bool is_locked, uint64_t* value) {
assert(property_info.handle_int != nullptr);
if (!property_info.need_out_of_mutex) {
if (is_locked) {
mutex_.AssertHeld();
return cfd->internal_stats()->GetIntProperty(property_info, value, this);
} else {
InstrumentedMutexLock l(&mutex_);
return cfd->internal_stats()->GetIntProperty(property_info, value, this);
}
} else {
SuperVersion* sv = nullptr;
if (!is_locked) {
sv = GetAndRefSuperVersion(cfd);
} else {
sv = cfd->GetSuperVersion();
}
bool ret = cfd->internal_stats()->GetIntPropertyOutOfMutex(
property_info, sv->current, value);
if (!is_locked) {
ReturnAndCleanupSuperVersion(cfd, sv);
}
return ret;
}
}
bool DBImpl::GetAggregatedIntProperty(const Slice& property,
uint64_t* aggregated_value) {
const DBPropertyInfo* property_info = GetPropertyInfo(property);
if (property_info == nullptr || property_info->handle_int == nullptr) {
return false;
}
uint64_t sum = 0;
{
// Needs mutex to protect the list of column families.
InstrumentedMutexLock l(&mutex_);
uint64_t value;
for (auto* cfd : *versions_->GetColumnFamilySet()) {
if (GetIntPropertyInternal(cfd, *property_info, true, &value)) {
sum += value;
} else {
return false;
}
}
}
*aggregated_value = sum;
return true;
}
SuperVersion* DBImpl::GetAndRefSuperVersion(ColumnFamilyData* cfd) {
// TODO(ljin): consider using GetReferencedSuperVersion() directly
return cfd->GetThreadLocalSuperVersion(&mutex_);
}
// REQUIRED: this function should only be called on the write thread or if the
// mutex is held.
SuperVersion* DBImpl::GetAndRefSuperVersion(uint32_t column_family_id) {
auto column_family_set = versions_->GetColumnFamilySet();
auto cfd = column_family_set->GetColumnFamily(column_family_id);
if (!cfd) {
return nullptr;
}
return GetAndRefSuperVersion(cfd);
}
// REQUIRED: mutex is NOT held
SuperVersion* DBImpl::GetAndRefSuperVersionUnlocked(uint32_t column_family_id) {
ColumnFamilyData* cfd;
{
InstrumentedMutexLock l(&mutex_);
auto column_family_set = versions_->GetColumnFamilySet();
cfd = column_family_set->GetColumnFamily(column_family_id);
}
if (!cfd) {
return nullptr;
}
return GetAndRefSuperVersion(cfd);
}
void DBImpl::ReturnAndCleanupSuperVersion(ColumnFamilyData* cfd,
SuperVersion* sv) {
bool unref_sv = !cfd->ReturnThreadLocalSuperVersion(sv);
if (unref_sv) {
// Release SuperVersion
if (sv->Unref()) {
{
InstrumentedMutexLock l(&mutex_);
sv->Cleanup();
}
delete sv;
RecordTick(stats_, NUMBER_SUPERVERSION_CLEANUPS);
}
RecordTick(stats_, NUMBER_SUPERVERSION_RELEASES);
}
}
// REQUIRED: this function should only be called on the write thread.
void DBImpl::ReturnAndCleanupSuperVersion(uint32_t column_family_id,
SuperVersion* sv) {
auto column_family_set = versions_->GetColumnFamilySet();
auto cfd = column_family_set->GetColumnFamily(column_family_id);
// If SuperVersion is held, and we successfully fetched a cfd using
// GetAndRefSuperVersion(), it must still exist.
assert(cfd != nullptr);
ReturnAndCleanupSuperVersion(cfd, sv);
}
// REQUIRED: Mutex should NOT be held.
void DBImpl::ReturnAndCleanupSuperVersionUnlocked(uint32_t column_family_id,
SuperVersion* sv) {
ColumnFamilyData* cfd;
{
InstrumentedMutexLock l(&mutex_);
auto column_family_set = versions_->GetColumnFamilySet();
cfd = column_family_set->GetColumnFamily(column_family_id);
}
// If SuperVersion is held, and we successfully fetched a cfd using
// GetAndRefSuperVersion(), it must still exist.
assert(cfd != nullptr);
ReturnAndCleanupSuperVersion(cfd, sv);
}
// REQUIRED: this function should only be called on the write thread or if the
// mutex is held.
ColumnFamilyHandle* DBImpl::GetColumnFamilyHandle(uint32_t column_family_id) {
ColumnFamilyMemTables* cf_memtables = column_family_memtables_.get();
if (!cf_memtables->Seek(column_family_id)) {
return nullptr;
}
return cf_memtables->GetColumnFamilyHandle();
}
// REQUIRED: mutex is NOT held.
ColumnFamilyHandle* DBImpl::GetColumnFamilyHandleUnlocked(
uint32_t column_family_id) {
ColumnFamilyMemTables* cf_memtables = column_family_memtables_.get();
InstrumentedMutexLock l(&mutex_);
if (!cf_memtables->Seek(column_family_id)) {
return nullptr;
}
return cf_memtables->GetColumnFamilyHandle();
}
void DBImpl::GetApproximateSizes(ColumnFamilyHandle* column_family,
const Range* range, int n, uint64_t* sizes,
bool include_memtable) {
Version* v;
auto cfh = reinterpret_cast<ColumnFamilyHandleImpl*>(column_family);
auto cfd = cfh->cfd();
SuperVersion* sv = GetAndRefSuperVersion(cfd);
v = sv->current;
for (int i = 0; i < n; i++) {
// Convert user_key into a corresponding internal key.
InternalKey k1(range[i].start, kMaxSequenceNumber, kValueTypeForSeek);
InternalKey k2(range[i].limit, kMaxSequenceNumber, kValueTypeForSeek);
sizes[i] = versions_->ApproximateSize(v, k1.Encode(), k2.Encode());
if (include_memtable) {
sizes[i] += sv->mem->ApproximateSize(k1.Encode(), k2.Encode());
sizes[i] += sv->imm->ApproximateSize(k1.Encode(), k2.Encode());
}
}
ReturnAndCleanupSuperVersion(cfd, sv);
}
std::list<uint64_t>::iterator
DBImpl::CaptureCurrentFileNumberInPendingOutputs() {
// We need to remember the iterator of our insert, because after the
// background job is done, we need to remove that element from
// pending_outputs_.
pending_outputs_.push_back(versions_->current_next_file_number());
auto pending_outputs_inserted_elem = pending_outputs_.end();
--pending_outputs_inserted_elem;
return pending_outputs_inserted_elem;
}
void DBImpl::ReleaseFileNumberFromPendingOutputs(
std::list<uint64_t>::iterator v) {
pending_outputs_.erase(v);
}
#ifndef ROCKSDB_LITE
Status DBImpl::GetUpdatesSince(
SequenceNumber seq, unique_ptr<TransactionLogIterator>* iter,
const TransactionLogIterator::ReadOptions& read_options) {
RecordTick(stats_, GET_UPDATES_SINCE_CALLS);
if (seq > versions_->LastSequence()) {
return Status::NotFound("Requested sequence not yet written in the db");
}
return wal_manager_.GetUpdatesSince(seq, iter, read_options, versions_.get());
}
Status DBImpl::DeleteFile(std::string name) {
uint64_t number;
FileType type;
WalFileType log_type;
if (!ParseFileName(name, &number, &type, &log_type) ||
(type != kTableFile && type != kLogFile)) {
Log(InfoLogLevel::ERROR_LEVEL, immutable_db_options_.info_log,
"DeleteFile %s failed.\n", name.c_str());
return Status::InvalidArgument("Invalid file name");
}
Status status;
if (type == kLogFile) {
// Only allow deleting archived log files
if (log_type != kArchivedLogFile) {
Log(InfoLogLevel::ERROR_LEVEL, immutable_db_options_.info_log,
"DeleteFile %s failed - not archived log.\n", name.c_str());
return Status::NotSupported("Delete only supported for archived logs");
}
status =
env_->DeleteFile(immutable_db_options_.wal_dir + "/" + name.c_str());
if (!status.ok()) {
Log(InfoLogLevel::ERROR_LEVEL, immutable_db_options_.info_log,
"DeleteFile %s failed -- %s.\n", name.c_str(),
status.ToString().c_str());
}
return status;
}
int level;
FileMetaData* metadata;
ColumnFamilyData* cfd;
VersionEdit edit;
JobContext job_context(next_job_id_.fetch_add(1), true);
{
InstrumentedMutexLock l(&mutex_);
status = versions_->GetMetadataForFile(number, &level, &metadata, &cfd);
if (!status.ok()) {
Log(InfoLogLevel::WARN_LEVEL, immutable_db_options_.info_log,
"DeleteFile %s failed. File not found\n", name.c_str());
job_context.Clean();
return Status::InvalidArgument("File not found");
}
assert(level < cfd->NumberLevels());
// If the file is being compacted no need to delete.
if (metadata->being_compacted) {
Log(InfoLogLevel::INFO_LEVEL, immutable_db_options_.info_log,
"DeleteFile %s Skipped. File about to be compacted\n", name.c_str());
job_context.Clean();
return Status::OK();
}
// Only the files in the last level can be deleted externally.
// This is to make sure that any deletion tombstones are not
// lost. Check that the level passed is the last level.
auto* vstoreage = cfd->current()->storage_info();
for (int i = level + 1; i < cfd->NumberLevels(); i++) {
if (vstoreage->NumLevelFiles(i) != 0) {
Log(InfoLogLevel::WARN_LEVEL, immutable_db_options_.info_log,
"DeleteFile %s FAILED. File not in last level\n", name.c_str());
job_context.Clean();
return Status::InvalidArgument("File not in last level");
}
}
// if level == 0, it has to be the oldest file
if (level == 0 &&
vstoreage->LevelFiles(0).back()->fd.GetNumber() != number) {
Log(InfoLogLevel::WARN_LEVEL, immutable_db_options_.info_log,
"DeleteFile %s failed ---"
" target file in level 0 must be the oldest.",
name.c_str());
job_context.Clean();
return Status::InvalidArgument("File in level 0, but not oldest");
}
edit.SetColumnFamily(cfd->GetID());
edit.DeleteFile(level, number);
status = versions_->LogAndApply(cfd, *cfd->GetLatestMutableCFOptions(),
&edit, &mutex_, directories_.GetDbDir());
if (status.ok()) {
InstallSuperVersionAndScheduleWorkWrapper(
cfd, &job_context, *cfd->GetLatestMutableCFOptions());
}
FindObsoleteFiles(&job_context, false);
} // lock released here
LogFlush(immutable_db_options_.info_log);
// remove files outside the db-lock
if (job_context.HaveSomethingToDelete()) {
// Call PurgeObsoleteFiles() without holding mutex.
PurgeObsoleteFiles(job_context);
}
job_context.Clean();
return status;
}
Status DBImpl::DeleteFilesInRange(ColumnFamilyHandle* column_family,
const Slice* begin, const Slice* end) {
Status status;
auto cfh = reinterpret_cast<ColumnFamilyHandleImpl*>(column_family);
ColumnFamilyData* cfd = cfh->cfd();
VersionEdit edit;
std::vector<FileMetaData*> deleted_files;
JobContext job_context(next_job_id_.fetch_add(1), true);
{
InstrumentedMutexLock l(&mutex_);
Version* input_version = cfd->current();
auto* vstorage = input_version->storage_info();
for (int i = 1; i < cfd->NumberLevels(); i++) {
if (vstorage->LevelFiles(i).empty() ||
!vstorage->OverlapInLevel(i, begin, end)) {
continue;
}
std::vector<FileMetaData*> level_files;
InternalKey begin_storage, end_storage, *begin_key, *end_key;
if (begin == nullptr) {
begin_key = nullptr;
} else {
begin_storage.SetMaxPossibleForUserKey(*begin);
begin_key = &begin_storage;
}
if (end == nullptr) {
end_key = nullptr;
} else {
end_storage.SetMinPossibleForUserKey(*end);
end_key = &end_storage;
}
vstorage->GetOverlappingInputs(i, begin_key, end_key, &level_files, -1,
nullptr, false);
FileMetaData* level_file;
for (uint32_t j = 0; j < level_files.size(); j++) {
level_file = level_files[j];
if (((begin == nullptr) ||
(cfd->internal_comparator().user_comparator()->Compare(
level_file->smallest.user_key(), *begin) >= 0)) &&
((end == nullptr) ||
(cfd->internal_comparator().user_comparator()->Compare(
level_file->largest.user_key(), *end) <= 0))) {
if (level_file->being_compacted) {
continue;
}
edit.SetColumnFamily(cfd->GetID());
edit.DeleteFile(i, level_file->fd.GetNumber());
deleted_files.push_back(level_file);
level_file->being_compacted = true;
}
}
}
if (edit.GetDeletedFiles().empty()) {
job_context.Clean();
return Status::OK();
}
input_version->Ref();
status = versions_->LogAndApply(cfd, *cfd->GetLatestMutableCFOptions(),
&edit, &mutex_, directories_.GetDbDir());
if (status.ok()) {
InstallSuperVersionAndScheduleWorkWrapper(
cfd, &job_context, *cfd->GetLatestMutableCFOptions());
}
for (auto* deleted_file : deleted_files) {
deleted_file->being_compacted = false;
}
input_version->Unref();
FindObsoleteFiles(&job_context, false);
} // lock released here
LogFlush(immutable_db_options_.info_log);
// remove files outside the db-lock
if (job_context.HaveSomethingToDelete()) {
// Call PurgeObsoleteFiles() without holding mutex.
PurgeObsoleteFiles(job_context);
}
job_context.Clean();
return status;
}
void DBImpl::GetLiveFilesMetaData(std::vector<LiveFileMetaData>* metadata) {
InstrumentedMutexLock l(&mutex_);
versions_->GetLiveFilesMetaData(metadata);
}
void DBImpl::GetColumnFamilyMetaData(
ColumnFamilyHandle* column_family,
ColumnFamilyMetaData* cf_meta) {
assert(column_family);
auto* cfd = reinterpret_cast<ColumnFamilyHandleImpl*>(column_family)->cfd();
auto* sv = GetAndRefSuperVersion(cfd);
sv->current->GetColumnFamilyMetaData(cf_meta);
ReturnAndCleanupSuperVersion(cfd, sv);
}
#endif // ROCKSDB_LITE
Status DBImpl::CheckConsistency() {
mutex_.AssertHeld();
std::vector<LiveFileMetaData> metadata;
versions_->GetLiveFilesMetaData(&metadata);
std::string corruption_messages;
for (const auto& md : metadata) {
// md.name has a leading "/".
std::string file_path = md.db_path + md.name;
uint64_t fsize = 0;
Status s = env_->GetFileSize(file_path, &fsize);
if (!s.ok() &&
env_->GetFileSize(Rocks2LevelTableFileName(file_path), &fsize).ok()) {
s = Status::OK();
}
if (!s.ok()) {
corruption_messages +=
"Can't access " + md.name + ": " + s.ToString() + "\n";
} else if (fsize != md.size) {
corruption_messages += "Sst file size mismatch: " + file_path +
". Size recorded in manifest " +
ToString(md.size) + ", actual size " +
ToString(fsize) + "\n";
}
}
if (corruption_messages.size() == 0) {
return Status::OK();
} else {
return Status::Corruption(corruption_messages);
}
}
Status DBImpl::GetDbIdentity(std::string& identity) const {
std::string idfilename = IdentityFileName(dbname_);
const EnvOptions soptions;
unique_ptr<SequentialFileReader> id_file_reader;
Status s;
{
unique_ptr<SequentialFile> idfile;
s = env_->NewSequentialFile(idfilename, &idfile, soptions);
if (!s.ok()) {
return s;
}
id_file_reader.reset(new SequentialFileReader(std::move(idfile)));
}
uint64_t file_size;
s = env_->GetFileSize(idfilename, &file_size);
if (!s.ok()) {
return s;
}
char* buffer = reinterpret_cast<char*>(alloca(file_size));
Slice id;
s = id_file_reader->Read(static_cast<size_t>(file_size), &id, buffer);
if (!s.ok()) {
return s;
}
identity.assign(id.ToString());
// If last character is '\n' remove it from identity
if (identity.size() > 0 && identity.back() == '\n') {
identity.pop_back();
}
return s;
}
// Default implementations of convenience methods that subclasses of DB
// can call if they wish
Status DB::Put(const WriteOptions& opt, ColumnFamilyHandle* column_family,
const Slice& key, const Slice& value) {
// Pre-allocate size of write batch conservatively.
// 8 bytes are taken by header, 4 bytes for count, 1 byte for type,
// and we allocate 11 extra bytes for key length, as well as value length.
WriteBatch batch(key.size() + value.size() + 24);
batch.Put(column_family, key, value);
return Write(opt, &batch);
}
Status DB::Delete(const WriteOptions& opt, ColumnFamilyHandle* column_family,
const Slice& key) {
WriteBatch batch;
batch.Delete(column_family, key);
return Write(opt, &batch);
}
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
Status DB::SingleDelete(const WriteOptions& opt,
ColumnFamilyHandle* column_family, const Slice& key) {
WriteBatch batch;
batch.SingleDelete(column_family, key);
return Write(opt, &batch);
}
Status DB::Merge(const WriteOptions& opt, ColumnFamilyHandle* column_family,
const Slice& key, const Slice& value) {
WriteBatch batch;
batch.Merge(column_family, key, value);
return Write(opt, &batch);
}
// Default implementation -- returns not supported status
Status DB::CreateColumnFamily(const ColumnFamilyOptions& cf_options,
const std::string& column_family_name,
ColumnFamilyHandle** handle) {
return Status::NotSupported("");
}
Status DB::DropColumnFamily(ColumnFamilyHandle* column_family) {
return Status::NotSupported("");
}
Status DB::DestroyColumnFamilyHandle(ColumnFamilyHandle* column_family) {
delete column_family;
return Status::OK();
}
DB::~DB() { }
Status DB::Open(const Options& options, const std::string& dbname, DB** dbptr) {
DBOptions db_options(options);
ColumnFamilyOptions cf_options(options);
std::vector<ColumnFamilyDescriptor> column_families;
column_families.push_back(
ColumnFamilyDescriptor(kDefaultColumnFamilyName, cf_options));
std::vector<ColumnFamilyHandle*> handles;
Status s = DB::Open(db_options, dbname, column_families, &handles, dbptr);
if (s.ok()) {
assert(handles.size() == 1);
// i can delete the handle since DBImpl is always holding a reference to
// default column family
delete handles[0];
}
return s;
}
Status DB::Open(const DBOptions& db_options, const std::string& dbname,
const std::vector<ColumnFamilyDescriptor>& column_families,
std::vector<ColumnFamilyHandle*>* handles, DB** dbptr) {
Status s = SanitizeOptionsByTable(db_options, column_families);
if (!s.ok()) {
return s;
}
s = ValidateOptions(db_options, column_families);
if (!s.ok()) {
return s;
}
*dbptr = nullptr;
handles->clear();
size_t max_write_buffer_size = 0;
for (auto cf : column_families) {
max_write_buffer_size =
std::max(max_write_buffer_size, cf.options.write_buffer_size);
}
DBImpl* impl = new DBImpl(db_options, dbname);
s = impl->env_->CreateDirIfMissing(impl->immutable_db_options_.wal_dir);
if (s.ok()) {
for (auto db_path : impl->immutable_db_options_.db_paths) {
s = impl->env_->CreateDirIfMissing(db_path.path);
if (!s.ok()) {
break;
}
}
}
if (!s.ok()) {
delete impl;
return s;
}
s = impl->CreateArchivalDirectory();
if (!s.ok()) {
delete impl;
return s;
}
impl->mutex_.Lock();
// Handles create_if_missing, error_if_exists
s = impl->Recover(column_families);
if (s.ok()) {
uint64_t new_log_number = impl->versions_->NewFileNumber();
unique_ptr<WritableFile> lfile;
EnvOptions soptions(db_options);
EnvOptions opt_env_options =
impl->immutable_db_options_.env->OptimizeForLogWrite(
soptions, BuildDBOptions(impl->immutable_db_options_,
impl->mutable_db_options_));
s = NewWritableFile(
impl->immutable_db_options_.env,
LogFileName(impl->immutable_db_options_.wal_dir, new_log_number),
&lfile, opt_env_options);
if (s.ok()) {
lfile->SetPreallocationBlockSize(
impl->GetWalPreallocateBlockSize(max_write_buffer_size));
impl->logfile_number_ = new_log_number;
unique_ptr<WritableFileWriter> file_writer(
new WritableFileWriter(std::move(lfile), opt_env_options));
impl->logs_.emplace_back(
new_log_number,
new log::Writer(
std::move(file_writer), new_log_number,
impl->immutable_db_options_.recycle_log_file_num > 0));
// set column family handles
for (auto cf : column_families) {
auto cfd =
impl->versions_->GetColumnFamilySet()->GetColumnFamily(cf.name);
if (cfd != nullptr) {
handles->push_back(
new ColumnFamilyHandleImpl(cfd, impl, &impl->mutex_));
impl->NewThreadStatusCfInfo(cfd);
} else {
if (db_options.create_missing_column_families) {
// missing column family, create it
ColumnFamilyHandle* handle;
impl->mutex_.Unlock();
s = impl->CreateColumnFamily(cf.options, cf.name, &handle);
impl->mutex_.Lock();
if (s.ok()) {
handles->push_back(handle);
} else {
break;
}
} else {
s = Status::InvalidArgument("Column family not found: ", cf.name);
break;
}
}
}
}
if (s.ok()) {
for (auto cfd : *impl->versions_->GetColumnFamilySet()) {
delete impl->InstallSuperVersionAndScheduleWork(
cfd, nullptr, *cfd->GetLatestMutableCFOptions());
}
impl->alive_log_files_.push_back(
DBImpl::LogFileNumberSize(impl->logfile_number_));
impl->DeleteObsoleteFiles();
s = impl->directories_.GetDbDir()->Fsync();
}
}
if (s.ok()) {
for (auto cfd : *impl->versions_->GetColumnFamilySet()) {
if (cfd->ioptions()->compaction_style == kCompactionStyleFIFO) {
auto* vstorage = cfd->current()->storage_info();
for (int i = 1; i < vstorage->num_levels(); ++i) {
int num_files = vstorage->NumLevelFiles(i);
if (num_files > 0) {
s = Status::InvalidArgument(
"Not all files are at level 0. Cannot "
"open with FIFO compaction style.");
break;
}
}
}
if (!cfd->mem()->IsSnapshotSupported()) {
impl->is_snapshot_supported_ = false;
}
if (cfd->ioptions()->merge_operator != nullptr &&
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
10 years ago
!cfd->mem()->IsMergeOperatorSupported()) {
s = Status::InvalidArgument(
"The memtable of column family %s does not support merge operator "
"its options.merge_operator is non-null", cfd->GetName().c_str());
}
if (!s.ok()) {
break;
}
}
}
TEST_SYNC_POINT("DBImpl::Open:Opened");
Status persist_options_status;
if (s.ok()) {
// Persist RocksDB Options before scheduling the compaction.
// The WriteOptionsFile() will release and lock the mutex internally.
persist_options_status = impl->WriteOptionsFile();
*dbptr = impl;
impl->opened_successfully_ = true;
impl->MaybeScheduleFlushOrCompaction();
}
impl->mutex_.Unlock();
auto sfm = static_cast<SstFileManagerImpl*>(
impl->immutable_db_options_.sst_file_manager.get());
if (s.ok() && sfm) {
// Notify SstFileManager about all sst files that already exist in
// db_paths[0] when the DB is opened.
auto& db_path = impl->immutable_db_options_.db_paths[0];
std::vector<std::string> existing_files;
impl->immutable_db_options_.env->GetChildren(db_path.path, &existing_files);
for (auto& file_name : existing_files) {
uint64_t file_number;
FileType file_type;
std::string file_path = db_path.path + "/" + file_name;
if (ParseFileName(file_name, &file_number, &file_type) &&
file_type == kTableFile) {
sfm->OnAddFile(file_path);
}
}
}
if (s.ok()) {
Log(InfoLogLevel::INFO_LEVEL, impl->immutable_db_options_.info_log,
"DB pointer %p", impl);
LogFlush(impl->immutable_db_options_.info_log);
if (!persist_options_status.ok()) {
if (db_options.fail_if_options_file_error) {
s = Status::IOError(
"DB::Open() failed --- Unable to persist Options file",
persist_options_status.ToString());
}
Warn(impl->immutable_db_options_.info_log,
"Unable to persist options in DB::Open() -- %s",
persist_options_status.ToString().c_str());
}
}
if (!s.ok()) {
for (auto* h : *handles) {
delete h;
}
handles->clear();
delete impl;
*dbptr = nullptr;
}
return s;
}
Status DB::ListColumnFamilies(const DBOptions& db_options,
const std::string& name,
std::vector<std::string>* column_families) {
return VersionSet::ListColumnFamilies(column_families, name, db_options.env);
}
Snapshot::~Snapshot() {
}
Status DestroyDB(const std::string& dbname, const Options& options) {
const ImmutableDBOptions soptions(SanitizeOptions(dbname, options));
Env* env = soptions.env;
std::vector<std::string> filenames;
// Ignore error in case directory does not exist
env->GetChildren(dbname, &filenames);
FileLock* lock;
const std::string lockname = LockFileName(dbname);
Status result = env->LockFile(lockname, &lock);
if (result.ok()) {
uint64_t number;
FileType type;
InfoLogPrefix info_log_prefix(!soptions.db_log_dir.empty(), dbname);
for (size_t i = 0; i < filenames.size(); i++) {
if (ParseFileName(filenames[i], &number, info_log_prefix.prefix, &type) &&
type != kDBLockFile) { // Lock file will be deleted at end
Status del;
std::string path_to_delete = dbname + "/" + filenames[i];
if (type == kMetaDatabase) {
del = DestroyDB(path_to_delete, options);
} else if (type == kTableFile) {
del = DeleteSSTFile(&soptions, path_to_delete, 0);
} else {
del = env->DeleteFile(path_to_delete);
}
if (result.ok() && !del.ok()) {
result = del;
}
}
}
for (size_t path_id = 0; path_id < options.db_paths.size(); path_id++) {
const auto& db_path = options.db_paths[path_id];
env->GetChildren(db_path.path, &filenames);
for (size_t i = 0; i < filenames.size(); i++) {
if (ParseFileName(filenames[i], &number, &type) &&
type == kTableFile) { // Lock file will be deleted at end
std::string table_path = db_path.path + "/" + filenames[i];
Status del = DeleteSSTFile(&soptions, table_path,
static_cast<uint32_t>(path_id));
if (result.ok() && !del.ok()) {
result = del;
}
}
}
}
std::vector<std::string> walDirFiles;
std::string archivedir = ArchivalDirectory(dbname);
if (dbname != soptions.wal_dir) {
env->GetChildren(soptions.wal_dir, &walDirFiles);
archivedir = ArchivalDirectory(soptions.wal_dir);
}
// Delete log files in the WAL dir
for (const auto& file : walDirFiles) {
if (ParseFileName(file, &number, &type) && type == kLogFile) {
Status del = env->DeleteFile(soptions.wal_dir + "/" + file);
if (result.ok() && !del.ok()) {
result = del;
}
}
}
std::vector<std::string> archiveFiles;
env->GetChildren(archivedir, &archiveFiles);
// Delete archival files.
for (size_t i = 0; i < archiveFiles.size(); ++i) {
if (ParseFileName(archiveFiles[i], &number, &type) &&
type == kLogFile) {
Status del = env->DeleteFile(archivedir + "/" + archiveFiles[i]);
if (result.ok() && !del.ok()) {
result = del;
}
}
}
// ignore case where no archival directory is present.
env->DeleteDir(archivedir);
env->UnlockFile(lock); // Ignore error since state is already gone
env->DeleteFile(lockname);
env->DeleteDir(dbname); // Ignore error in case dir contains other files
env->DeleteDir(soptions.wal_dir);
}
return result;
}
Status DBImpl::WriteOptionsFile() {
#ifndef ROCKSDB_LITE
mutex_.AssertHeld();
std::vector<std::string> cf_names;
std::vector<ColumnFamilyOptions> cf_opts;
// This part requires mutex to protect the column family options
for (auto cfd : *versions_->GetColumnFamilySet()) {
if (cfd->IsDropped()) {
continue;
}
cf_names.push_back(cfd->GetName());
cf_opts.push_back(cfd->GetLatestCFOptions());
}
// Unlock during expensive operations. New writes cannot get here
// because the single write thread ensures all new writes get queued.
DBOptions db_options =
BuildDBOptions(immutable_db_options_, mutable_db_options_);
mutex_.Unlock();
std::string file_name =
TempOptionsFileName(GetName(), versions_->NewFileNumber());
Status s =
PersistRocksDBOptions(db_options, cf_names, cf_opts, file_name, GetEnv());
if (s.ok()) {
s = RenameTempFileToOptionsFile(file_name);
}
mutex_.Lock();
return s;
#else
return Status::OK();
#endif // !ROCKSDB_LITE
}
#ifndef ROCKSDB_LITE
namespace {
void DeleteOptionsFilesHelper(const std::map<uint64_t, std::string>& filenames,
const size_t num_files_to_keep,
const std::shared_ptr<Logger>& info_log,
Env* env) {
if (filenames.size() <= num_files_to_keep) {
return;
}
for (auto iter = std::next(filenames.begin(), num_files_to_keep);
iter != filenames.end(); ++iter) {
if (!env->DeleteFile(iter->second).ok()) {
Warn(info_log, "Unable to delete options file %s", iter->second.c_str());
}
}
}
} // namespace
#endif // !ROCKSDB_LITE
Status DBImpl::DeleteObsoleteOptionsFiles() {
#ifndef ROCKSDB_LITE
std::vector<std::string> filenames;
// use ordered map to store keep the filenames sorted from the newest
// to the oldest.
std::map<uint64_t, std::string> options_filenames;
Status s;
s = GetEnv()->GetChildren(GetName(), &filenames);
if (!s.ok()) {
return s;
}
for (auto& filename : filenames) {
uint64_t file_number;
FileType type;
if (ParseFileName(filename, &file_number, &type) && type == kOptionsFile) {
options_filenames.insert(
{std::numeric_limits<uint64_t>::max() - file_number,
GetName() + "/" + filename});
}
}
// Keeps the latest 2 Options file
const size_t kNumOptionsFilesKept = 2;
DeleteOptionsFilesHelper(options_filenames, kNumOptionsFilesKept,
immutable_db_options_.info_log, GetEnv());
return Status::OK();
#else
return Status::OK();
#endif // !ROCKSDB_LITE
}
Status DBImpl::RenameTempFileToOptionsFile(const std::string& file_name) {
#ifndef ROCKSDB_LITE
Status s;
versions_->options_file_number_ = versions_->NewFileNumber();
std::string options_file_name =
OptionsFileName(GetName(), versions_->options_file_number_);
// Retry if the file name happen to conflict with an existing one.
s = GetEnv()->RenameFile(file_name, options_file_name);
DeleteObsoleteOptionsFiles();
return s;
#else
return Status::OK();
#endif // !ROCKSDB_LITE
}
#if ROCKSDB_USING_THREAD_STATUS
void DBImpl::NewThreadStatusCfInfo(
ColumnFamilyData* cfd) const {
if (immutable_db_options_.enable_thread_tracking) {
ThreadStatusUtil::NewColumnFamilyInfo(this, cfd, cfd->GetName(),
cfd->ioptions()->env);
}
}
void DBImpl::EraseThreadStatusCfInfo(
ColumnFamilyData* cfd) const {
if (immutable_db_options_.enable_thread_tracking) {
ThreadStatusUtil::EraseColumnFamilyInfo(cfd);
}
}
void DBImpl::EraseThreadStatusDbInfo() const {
if (immutable_db_options_.enable_thread_tracking) {
ThreadStatusUtil::EraseDatabaseInfo(this);
}
}
#else
void DBImpl::NewThreadStatusCfInfo(
ColumnFamilyData* cfd) const {
}
void DBImpl::EraseThreadStatusCfInfo(
ColumnFamilyData* cfd) const {
}
void DBImpl::EraseThreadStatusDbInfo() const {
}
#endif // ROCKSDB_USING_THREAD_STATUS
//
// A global method that can dump out the build version
void DumpRocksDBBuildVersion(Logger * log) {
#if !defined(IOS_CROSS_COMPILE)
// if we compile with Xcode, we don't run build_detect_vesion, so we don't
// generate util/build_version.cc
Header(log, "RocksDB version: %d.%d.%d\n", ROCKSDB_MAJOR, ROCKSDB_MINOR,
ROCKSDB_PATCH);
Header(log, "Git sha %s", rocksdb_build_git_sha);
Header(log, "Compile date %s", rocksdb_build_compile_date);
#endif
}
#ifndef ROCKSDB_LITE
SequenceNumber DBImpl::GetEarliestMemTableSequenceNumber(SuperVersion* sv,
bool include_history) {
// Find the earliest sequence number that we know we can rely on reading
// from the memtable without needing to check sst files.
SequenceNumber earliest_seq =
sv->imm->GetEarliestSequenceNumber(include_history);
if (earliest_seq == kMaxSequenceNumber) {
earliest_seq = sv->mem->GetEarliestSequenceNumber();
}
assert(sv->mem->GetEarliestSequenceNumber() >= earliest_seq);
return earliest_seq;
}
#endif // ROCKSDB_LITE
#ifndef ROCKSDB_LITE
Status DBImpl::GetLatestSequenceForKey(SuperVersion* sv, const Slice& key,
bool cache_only, SequenceNumber* seq,
bool* found_record_for_key) {
Status s;
MergeContext merge_context;
SequenceNumber current_seq = versions_->LastSequence();
LookupKey lkey(key, current_seq);
*seq = kMaxSequenceNumber;
*found_record_for_key = false;
// Check if there is a record for this key in the latest memtable
sv->mem->Get(lkey, nullptr, &s, &merge_context, seq);
if (!(s.ok() || s.IsNotFound() || s.IsMergeInProgress())) {
// unexpected error reading memtable.
Log(InfoLogLevel::ERROR_LEVEL, immutable_db_options_.info_log,
"Unexpected status returned from MemTable::Get: %s\n",
s.ToString().c_str());
return s;
}
if (*seq != kMaxSequenceNumber) {
// Found a sequence number, no need to check immutable memtables
*found_record_for_key = true;
return Status::OK();
}
// Check if there is a record for this key in the immutable memtables
sv->imm->Get(lkey, nullptr, &s, &merge_context, seq);
if (!(s.ok() || s.IsNotFound() || s.IsMergeInProgress())) {
// unexpected error reading memtable.
Log(InfoLogLevel::ERROR_LEVEL, immutable_db_options_.info_log,
"Unexpected status returned from MemTableList::Get: %s\n",
s.ToString().c_str());
return s;
}
if (*seq != kMaxSequenceNumber) {
// Found a sequence number, no need to check memtable history
*found_record_for_key = true;
return Status::OK();
}
// Check if there is a record for this key in the immutable memtables
sv->imm->GetFromHistory(lkey, nullptr, &s, &merge_context, seq);
if (!(s.ok() || s.IsNotFound() || s.IsMergeInProgress())) {
// unexpected error reading memtable.
Log(InfoLogLevel::ERROR_LEVEL, immutable_db_options_.info_log,
"Unexpected status returned from MemTableList::GetFromHistory: %s\n",
s.ToString().c_str());
return s;
}
if (*seq != kMaxSequenceNumber) {
// Found a sequence number, no need to check SST files
*found_record_for_key = true;
return Status::OK();
}
// TODO(agiardullo): possible optimization: consider checking cached
// SST files if cache_only=true?
if (!cache_only) {
// Check tables
ReadOptions read_options;
sv->current->Get(read_options, lkey, nullptr, &s, &merge_context,
nullptr /* value_found */, found_record_for_key, seq);
if (!(s.ok() || s.IsNotFound() || s.IsMergeInProgress())) {
// unexpected error reading SST files
Log(InfoLogLevel::ERROR_LEVEL, immutable_db_options_.info_log,
"Unexpected status returned from Version::Get: %s\n",
s.ToString().c_str());
return s;
}
}
return Status::OK();
}
Status DBImpl::IngestExternalFile(
ColumnFamilyHandle* column_family,
const std::vector<std::string>& external_files,
const IngestExternalFileOptions& ingestion_options) {
Status status;
auto cfh = reinterpret_cast<ColumnFamilyHandleImpl*>(column_family);
auto cfd = cfh->cfd();
ExternalSstFileIngestionJob ingestion_job(env_, versions_.get(), cfd,
immutable_db_options_, env_options_,
&snapshots_, ingestion_options);
// Make sure that bg cleanup wont delete the files that we are ingesting
std::list<uint64_t>::iterator pending_output_elem;
{
InstrumentedMutexLock l(&mutex_);
pending_output_elem = CaptureCurrentFileNumberInPendingOutputs();
}
status = ingestion_job.Prepare(external_files);
if (!status.ok()) {
return status;
}
{
// Lock db mutex
InstrumentedMutexLock l(&mutex_);
TEST_SYNC_POINT("DBImpl::AddFile:MutexLock");
num_running_ingest_file_++;
// Stop writes to the DB
WriteThread::Writer w;
write_thread_.EnterUnbatched(&w, &mutex_);
// Figure out if we need to flush the memtable first
bool need_flush = false;
status = ingestion_job.NeedsFlush(&need_flush);
if (status.ok() && need_flush) {
mutex_.Unlock();
status = FlushMemTable(cfd, FlushOptions(), true /* writes_stopped */);
mutex_.Lock();
}
// Run the ingestion job
if (status.ok()) {
status = ingestion_job.Run();
}
// Install job edit [Mutex will be unlocked here]
auto mutable_cf_options = cfd->GetLatestMutableCFOptions();
if (status.ok()) {
status =
versions_->LogAndApply(cfd, *mutable_cf_options, ingestion_job.edit(),
&mutex_, directories_.GetDbDir());
}
if (status.ok()) {
delete InstallSuperVersionAndScheduleWork(cfd, nullptr,
*mutable_cf_options);
}
// Resume writes to the DB
write_thread_.ExitUnbatched(&w);
// Update stats
if (status.ok()) {
ingestion_job.UpdateStats();
}
ReleaseFileNumberFromPendingOutputs(pending_output_elem);
num_running_ingest_file_--;
if (num_running_ingest_file_ == 0) {
bg_cv_.SignalAll();
}
TEST_SYNC_POINT("DBImpl::AddFile:MutexUnlock");
}
// mutex_ is unlocked here
// Cleanup
ingestion_job.Cleanup(status);
return status;
}
void DBImpl::WaitForIngestFile() {
mutex_.AssertHeld();
while (num_running_ingest_file_ > 0) {
bg_cv_.Wait();
}
}
#endif // ROCKSDB_LITE
} // namespace rocksdb