|
|
|
// Copyright (c) 2011-present, Facebook, Inc. All rights reserved.
|
|
|
|
// This source code is licensed under both the GPLv2 (found in the
|
|
|
|
// COPYING file in the root directory) and Apache 2.0 License
|
|
|
|
// (found in the LICENSE.Apache file in the root directory).
|
|
|
|
//
|
|
|
|
// 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_test_util.h"
|
|
|
|
|
|
|
|
#include "cache/cache_reservation_manager.h"
|
|
|
|
#include "db/forward_iterator.h"
|
|
|
|
#include "env/mock_env.h"
|
Eliminate unnecessary (slow) block cache Ref()ing in MultiGet (#9899)
Summary:
When MultiGet() determines that multiple query keys can be
served by examining the same data block in block cache (one Lookup()),
each PinnableSlice referring to data in that data block needs to hold
on to the block in cache so that they can be released at arbitrary
times by the API user. Historically this is accomplished with extra
calls to Ref() on the Handle from Lookup(), with each PinnableSlice
cleanup calling Release() on the Handle, but this creates extra
contention on the block cache for the extra Ref()s and Release()es,
especially because they hit the same cache shard repeatedly.
In the case of merge operands (possibly more cases?), the problem was
compounded by doing an extra Ref()+eventual Release() for each merge
operand for a key reusing a block (which could be the same key!), rather
than one Ref() per key. (Note: the non-shared case with `biter` was
already one per key.)
This change optimizes MultiGet not to rely on these extra, contentious
Ref()+Release() calls by instead, in the shared block case, wrapping
the cache Release() cleanup in a refcounted object referenced by the
PinnableSlices, such that after the last wrapped reference is released,
the cache entry is Release()ed. Relaxed atomic refcounts should be
much faster than mutex-guarded Ref() and Release(), and much less prone
to a performance cliff when MultiGet() does a lot of block sharing.
Note that I did not use std::shared_ptr, because that would require an
extra indirection object (shared_ptr itself new/delete) in order to
associate a ref increment/decrement with a Cleanable cleanup entry. (If
I assumed it was the size of two pointers, I could do some hackery to
make it work without the extra indirection, but that's too fragile.)
Some details:
* Fixed (removed) extra block cache tracing entries in cases of cache
entry reuse in MultiGet, but it's likely that in some other cases traces
are missing (XXX comment inserted)
* Moved existing implementations for cleanable.h from iterator.cc to
new cleanable.cc
* Improved API comments on Cleanable
* Added a public SharedCleanablePtr class to cleanable.h in case others
could benefit from the same pattern (potentially many Cleanables and/or
smart pointers referencing a shared Cleanable)
* Add a typedef for MultiGetContext::Mask
* Some variable renaming for clarity
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9899
Test Plan:
Added unit tests for SharedCleanablePtr.
Greatly enhanced ability of existing tests to detect cache use-after-free.
* Release PinnableSlices from MultiGet as they are read rather than in
bulk (in db_test_util wrapper).
* In ASAN build, default to using a trivially small LRUCache for block_cache
so that entries are immediately erased when unreferenced. (Updated two
tests that depend on caching.) New ASAN testsuite running time seems
OK to me.
If I introduce a bug into my implementation where we skip the shared
cleanups on block reuse, ASAN detects the bug in
`db_basic_test *MultiGet*`. If I remove either of the above testing
enhancements, the bug is not detected.
Consider for follow-up work: manipulate or randomize ordering of
PinnableSlice use and release from MultiGet db_test_util wrapper. But in
typical cases, natural ordering gives pretty good functional coverage.
Performance test:
In the extreme (but possible) case of MultiGetting the same or adjacent keys
in a batch, throughput can improve by an order of magnitude.
`./db_bench -benchmarks=multireadrandom -db=/dev/shm/testdb -readonly -num=5 -duration=10 -threads=20 -multiread_batched -batch_size=200`
Before ops/sec, num=5: 1,384,394
Before ops/sec, num=500: 6,423,720
After ops/sec, num=500: 10,658,794
After ops/sec, num=5: 16,027,257
Also note that previously, with high parallelism, having query keys
concentrated in a single block was worse than spreading them out a bit. Now
concentrated in a single block is faster than spread out, which is hopefully
consistent with natural expectation.
Random query performance: with num=1000000, over 999 x 10s runs running before & after simultaneously (each -threads=12):
Before: multireadrandom [AVG 999 runs] : 1088699 (± 7344) ops/sec; 120.4 (± 0.8 ) MB/sec
After: multireadrandom [AVG 999 runs] : 1090402 (± 7230) ops/sec; 120.6 (± 0.8 ) MB/sec
Possibly better, possibly in the noise.
Reviewed By: anand1976
Differential Revision: D35907003
Pulled By: pdillinger
fbshipit-source-id: bbd244d703649a8ca12d476f2d03853ed9d1a17e
3 years ago
|
|
|
#include "port/lang.h"
|
|
|
|
#include "rocksdb/cache.h"
|
|
|
|
#include "rocksdb/convenience.h"
|
|
|
|
#include "rocksdb/env_encryption.h"
|
Experimental support for SST unique IDs (#8990)
Summary:
* New public header unique_id.h and function GetUniqueIdFromTableProperties
which computes a universally unique identifier based on table properties
of table files from recent RocksDB versions.
* Generation of DB session IDs is refactored so that they are
guaranteed unique in the lifetime of a process running RocksDB.
(SemiStructuredUniqueIdGen, new test included.) Along with file numbers,
this enables SST unique IDs to be guaranteed unique among SSTs generated
in a single process, and "better than random" between processes.
See https://github.com/pdillinger/unique_id
* In addition to public API producing 'external' unique IDs, there is a function
for producing 'internal' unique IDs, with functions for converting between the
two. In short, the external ID is "safe" for things people might do with it, and
the internal ID enables more "power user" features for the future. Specifically,
the external ID goes through a hashing layer so that any subset of bits in the
external ID can be used as a hash of the full ID, while also preserving
uniqueness guarantees in the first 128 bits (bijective both on first 128 bits
and on full 192 bits).
Intended follow-up:
* Use the internal unique IDs in cache keys. (Avoid conflicts with https://github.com/facebook/rocksdb/issues/8912) (The file offset can be XORed into
the third 64-bit value of the unique ID.)
* Publish the external unique IDs in FileStorageInfo (https://github.com/facebook/rocksdb/issues/8968)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8990
Test Plan:
Unit tests added, and checking of unique ids in stress test.
NOTE in stress test we do not generate nearly enough files to thoroughly
stress uniqueness, but the test trims off pieces of the ID to check for
uniqueness so that we can infer (with some assumptions) stronger
properties in the aggregate.
Reviewed By: zhichao-cao, mrambacher
Differential Revision: D31582865
Pulled By: pdillinger
fbshipit-source-id: 1f620c4c86af9abe2a8d177b9ccf2ad2b9f48243
3 years ago
|
|
|
#include "rocksdb/unique_id.h"
|
|
|
|
#include "rocksdb/utilities/object_registry.h"
|
|
|
|
#include "table/format.h"
|
|
|
|
#include "util/random.h"
|
|
|
|
|
|
|
|
namespace ROCKSDB_NAMESPACE {
|
|
|
|
|
|
|
|
namespace {
|
|
|
|
int64_t MaybeCurrentTime(Env* env) {
|
|
|
|
int64_t time = 1337346000; // arbitrary fallback default
|
|
|
|
env->GetCurrentTime(&time).PermitUncheckedError();
|
|
|
|
return time;
|
|
|
|
}
|
|
|
|
} // anonymous namespace
|
|
|
|
|
|
|
|
// Special Env used to delay background operations
|
|
|
|
|
Fix+clean up handling of mock sleeps (#7101)
Summary:
We have a number of tests hanging on MacOS and windows due to
mishandling of code for mock sleeps. In addition, the code was in
terrible shape because the same variable (addon_time_) would sometimes
refer to microseconds and sometimes to seconds. One test even assumed it
was nanoseconds but was written to pass anyway.
This has been cleaned up so that DB tests generally use a SpecialEnv
function to mock sleep, for either some number of microseconds or seconds
depending on the function called. But to call one of these, the test must first
call SetMockSleep (precondition enforced with assertion), which also turns
sleeps in RocksDB into mock sleeps. To also removes accounting for actual
clock time, call SetTimeElapseOnlySleepOnReopen, which implies
SetMockSleep (on DB re-open). This latter setting only works by applying
on DB re-open, otherwise havoc can ensue if Env goes back in time with
DB open.
More specifics:
Removed some unused test classes, and updated comments on the general
problem.
Fixed DBSSTTest.GetTotalSstFilesSize using a sync point callback instead
of mock time. For this we have the only modification to production code,
inserting a sync point callback in flush_job.cc, which is not a change to
production behavior.
Removed unnecessary resetting of mock times to 0 in many tests. RocksDB
deals in relative time. Any behaviors relying on absolute date/time are likely
a bug. (The above test DBSSTTest.GetTotalSstFilesSize was the only one
clearly injecting a specific absolute time for actual testing convenience.) Just
in case I misunderstood some test, I put this note in each replacement:
// NOTE: Presumed unnecessary and removed: resetting mock time in env
Strengthened some tests like MergeTestTime, MergeCompactionTimeTest, and
FilterCompactionTimeTest in db_test.cc
stats_history_test and blob_db_test are each their own beast, rather deeply
dependent on MockTimeEnv. Each gets its own variant of a work-around for
TimedWait in a mock time environment. (Reduces redundancy and
inconsistency in stats_history_test.)
Intended follow-up:
Remove TimedWait from the public API of InstrumentedCondVar, and only
make that accessible through Env by passing in an InstrumentedCondVar and
a deadline. Then the Env implementations mocking time can fix this problem
without using sync points. (Test infrastructure using sync points interferes
with individual tests' control over sync points.)
With that change, we can simplify/consolidate the scattered work-arounds.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7101
Test Plan: make check on Linux and MacOS
Reviewed By: zhichao-cao
Differential Revision: D23032815
Pulled By: pdillinger
fbshipit-source-id: 7f33967ada8b83011fb54e8279365c008bd6610b
4 years ago
|
|
|
SpecialEnv::SpecialEnv(Env* base, bool time_elapse_only_sleep)
|
|
|
|
: EnvWrapper(base),
|
|
|
|
maybe_starting_time_(MaybeCurrentTime(base)),
|
|
|
|
rnd_(301),
|
|
|
|
sleep_counter_(this),
|
Fix+clean up handling of mock sleeps (#7101)
Summary:
We have a number of tests hanging on MacOS and windows due to
mishandling of code for mock sleeps. In addition, the code was in
terrible shape because the same variable (addon_time_) would sometimes
refer to microseconds and sometimes to seconds. One test even assumed it
was nanoseconds but was written to pass anyway.
This has been cleaned up so that DB tests generally use a SpecialEnv
function to mock sleep, for either some number of microseconds or seconds
depending on the function called. But to call one of these, the test must first
call SetMockSleep (precondition enforced with assertion), which also turns
sleeps in RocksDB into mock sleeps. To also removes accounting for actual
clock time, call SetTimeElapseOnlySleepOnReopen, which implies
SetMockSleep (on DB re-open). This latter setting only works by applying
on DB re-open, otherwise havoc can ensue if Env goes back in time with
DB open.
More specifics:
Removed some unused test classes, and updated comments on the general
problem.
Fixed DBSSTTest.GetTotalSstFilesSize using a sync point callback instead
of mock time. For this we have the only modification to production code,
inserting a sync point callback in flush_job.cc, which is not a change to
production behavior.
Removed unnecessary resetting of mock times to 0 in many tests. RocksDB
deals in relative time. Any behaviors relying on absolute date/time are likely
a bug. (The above test DBSSTTest.GetTotalSstFilesSize was the only one
clearly injecting a specific absolute time for actual testing convenience.) Just
in case I misunderstood some test, I put this note in each replacement:
// NOTE: Presumed unnecessary and removed: resetting mock time in env
Strengthened some tests like MergeTestTime, MergeCompactionTimeTest, and
FilterCompactionTimeTest in db_test.cc
stats_history_test and blob_db_test are each their own beast, rather deeply
dependent on MockTimeEnv. Each gets its own variant of a work-around for
TimedWait in a mock time environment. (Reduces redundancy and
inconsistency in stats_history_test.)
Intended follow-up:
Remove TimedWait from the public API of InstrumentedCondVar, and only
make that accessible through Env by passing in an InstrumentedCondVar and
a deadline. Then the Env implementations mocking time can fix this problem
without using sync points. (Test infrastructure using sync points interferes
with individual tests' control over sync points.)
With that change, we can simplify/consolidate the scattered work-arounds.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7101
Test Plan: make check on Linux and MacOS
Reviewed By: zhichao-cao
Differential Revision: D23032815
Pulled By: pdillinger
fbshipit-source-id: 7f33967ada8b83011fb54e8279365c008bd6610b
4 years ago
|
|
|
time_elapse_only_sleep_(time_elapse_only_sleep),
|
|
|
|
no_slowdown_(time_elapse_only_sleep) {
|
|
|
|
delay_sstable_sync_.store(false, std::memory_order_release);
|
|
|
|
drop_writes_.store(false, std::memory_order_release);
|
|
|
|
no_space_.store(false, std::memory_order_release);
|
|
|
|
non_writable_.store(false, std::memory_order_release);
|
|
|
|
count_random_reads_ = false;
|
|
|
|
count_sequential_reads_ = false;
|
|
|
|
manifest_sync_error_.store(false, std::memory_order_release);
|
|
|
|
manifest_write_error_.store(false, std::memory_order_release);
|
|
|
|
log_write_error_.store(false, std::memory_order_release);
|
Handle rename() failure in non-local FS (#8192)
Summary:
In a distributed environment, a file `rename()` operation can succeed on server (remote)
side, but the client can somehow return non-ok status to RocksDB. Possible reasons include
network partition, connection issue, etc. This happens in `rocksdb::SetCurrentFile()`, which
can be called in `LogAndApply() -> ProcessManifestWrites()` if RocksDB tries to switch to a
new MANIFEST. We currently always delete the new MANIFEST if an error occurs.
This is problematic in distributed world. If the server-side successfully updates the CURRENT
file via renaming, then a subsequent `DB::Open()` will try to look for the new MANIFEST and fail.
As a fix, we can track the execution result of IO operations on the new MANIFEST.
- If IO operations on the new MANIFEST fail, then we know the CURRENT must point to the original
MANIFEST. Therefore, it is safe to remove the new MANIFEST.
- If IO operations on the new MANIFEST all succeed, but somehow we end up in the clean up
code block, then we do not know whether CURRENT points to the new or old MANIFEST. (For local
POSIX-compliant FS, it should still point to old MANIFEST, but it does not matter if we keep the
new MANIFEST.) Therefore, we keep the new MANIFEST.
- Any future `LogAndApply()` will switch to a new MANIFEST and update CURRENT.
- If process reopens the db immediately after the failure, then the CURRENT file can point
to either the new MANIFEST or the old one, both of which exist. Therefore, recovery can
succeed and ignore the other.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8192
Test Plan: make check
Reviewed By: zhichao-cao
Differential Revision: D27804648
Pulled By: riversand963
fbshipit-source-id: 9c16f2a5ce41bc6aadf085e48449b19ede8423e4
4 years ago
|
|
|
no_file_overwrite_.store(false, std::memory_order_release);
|
|
|
|
random_file_open_counter_.store(0, std::memory_order_relaxed);
|
|
|
|
delete_count_.store(0, std::memory_order_relaxed);
|
|
|
|
num_open_wal_file_.store(0);
|
|
|
|
log_write_slowdown_ = 0;
|
|
|
|
bytes_written_ = 0;
|
|
|
|
sync_counter_ = 0;
|
|
|
|
non_writeable_rate_ = 0;
|
|
|
|
new_writable_count_ = 0;
|
|
|
|
non_writable_count_ = 0;
|
|
|
|
table_write_callback_ = nullptr;
|
|
|
|
}
|
|
|
|
DBTestBase::DBTestBase(const std::string path, bool env_do_fsync)
|
|
|
|
: mem_env_(nullptr), encrypted_env_(nullptr), option_config_(kDefault) {
|
|
|
|
Env* base_env = Env::Default();
|
|
|
|
ConfigOptions config_options;
|
|
|
|
EXPECT_OK(test::CreateEnvFromSystem(config_options, &base_env, &env_guard_));
|
|
|
|
EXPECT_NE(nullptr, base_env);
|
|
|
|
if (getenv("MEM_ENV")) {
|
|
|
|
mem_env_ = MockEnv::Create(base_env, base_env->GetSystemClock());
|
|
|
|
}
|
|
|
|
if (getenv("ENCRYPTED_ENV")) {
|
|
|
|
std::shared_ptr<EncryptionProvider> provider;
|
|
|
|
std::string provider_id = getenv("ENCRYPTED_ENV");
|
|
|
|
if (provider_id.find("=") == std::string::npos &&
|
|
|
|
!EndsWith(provider_id, "://test")) {
|
|
|
|
provider_id = provider_id + "://test";
|
|
|
|
}
|
|
|
|
EXPECT_OK(EncryptionProvider::CreateFromString(ConfigOptions(), provider_id,
|
|
|
|
&provider));
|
|
|
|
encrypted_env_ = NewEncryptedEnv(mem_env_ ? mem_env_ : base_env, provider);
|
|
|
|
}
|
|
|
|
env_ = new SpecialEnv(encrypted_env_ ? encrypted_env_
|
|
|
|
: (mem_env_ ? mem_env_ : base_env));
|
|
|
|
env_->SetBackgroundThreads(1, Env::LOW);
|
|
|
|
env_->SetBackgroundThreads(1, Env::HIGH);
|
|
|
|
env_->skip_fsync_ = !env_do_fsync;
|
|
|
|
dbname_ = test::PerThreadDBPath(env_, path);
|
|
|
|
alternative_wal_dir_ = dbname_ + "/wal";
|
|
|
|
alternative_db_log_dir_ = dbname_ + "/db_log_dir";
|
|
|
|
auto options = CurrentOptions();
|
|
|
|
options.env = env_;
|
|
|
|
auto delete_options = options;
|
|
|
|
delete_options.wal_dir = alternative_wal_dir_;
|
|
|
|
EXPECT_OK(DestroyDB(dbname_, delete_options));
|
|
|
|
// Destroy it for not alternative WAL dir is used.
|
|
|
|
EXPECT_OK(DestroyDB(dbname_, options));
|
|
|
|
db_ = nullptr;
|
|
|
|
Reopen(options);
|
|
|
|
Random::GetTLSInstance()->Reset(0xdeadbeef);
|
|
|
|
}
|
|
|
|
|
|
|
|
DBTestBase::~DBTestBase() {
|
|
|
|
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->DisableProcessing();
|
|
|
|
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->LoadDependency({});
|
|
|
|
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearAllCallBacks();
|
|
|
|
Close();
|
|
|
|
Options options;
|
|
|
|
options.db_paths.emplace_back(dbname_, 0);
|
|
|
|
options.db_paths.emplace_back(dbname_ + "_2", 0);
|
|
|
|
options.db_paths.emplace_back(dbname_ + "_3", 0);
|
|
|
|
options.db_paths.emplace_back(dbname_ + "_4", 0);
|
|
|
|
options.env = env_;
|
|
|
|
|
|
|
|
if (getenv("KEEP_DB")) {
|
|
|
|
printf("DB is still at %s\n", dbname_.c_str());
|
|
|
|
} else {
|
|
|
|
EXPECT_OK(DestroyDB(dbname_, options));
|
|
|
|
}
|
|
|
|
delete env_;
|
|
|
|
}
|
|
|
|
|
|
|
|
bool DBTestBase::ShouldSkipOptions(int option_config, int skip_mask) {
|
|
|
|
|
|
|
|
if ((skip_mask & kSkipUniversalCompaction) &&
|
|
|
|
(option_config == kUniversalCompaction ||
|
|
|
|
option_config == kUniversalCompactionMultiLevel ||
|
|
|
|
option_config == kUniversalSubcompactions)) {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
if ((skip_mask & kSkipMergePut) && option_config == kMergePut) {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
if ((skip_mask & kSkipNoSeekToLast) &&
|
|
|
|
(option_config == kHashLinkList || option_config == kHashSkipList)) {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
if ((skip_mask & kSkipPlainTable) &&
|
|
|
|
(option_config == kPlainTableAllBytesPrefix ||
|
|
|
|
option_config == kPlainTableFirstBytePrefix ||
|
|
|
|
option_config == kPlainTableCappedPrefix ||
|
|
|
|
option_config == kPlainTableCappedPrefixNonMmap)) {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
if ((skip_mask & kSkipHashIndex) &&
|
|
|
|
(option_config == kBlockBasedTableWithPrefixHashIndex ||
|
|
|
|
option_config == kBlockBasedTableWithWholeKeyHashIndex)) {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
if ((skip_mask & kSkipFIFOCompaction) && option_config == kFIFOCompaction) {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
if ((skip_mask & kSkipMmapReads) && option_config == kWalDirAndMmapReads) {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Switch to a fresh database with the next option configuration to
|
|
|
|
// test. Return false if there are no more configurations to test.
|
|
|
|
bool DBTestBase::ChangeOptions(int skip_mask) {
|
|
|
|
for (option_config_++; option_config_ < kEnd; option_config_++) {
|
|
|
|
if (ShouldSkipOptions(option_config_, skip_mask)) {
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (option_config_ >= kEnd) {
|
|
|
|
Destroy(last_options_);
|
|
|
|
return false;
|
|
|
|
} else {
|
|
|
|
auto options = CurrentOptions();
|
|
|
|
options.create_if_missing = true;
|
|
|
|
DestroyAndReopen(options);
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Switch between different compaction styles.
|
|
|
|
bool DBTestBase::ChangeCompactOptions() {
|
|
|
|
if (option_config_ == kDefault) {
|
|
|
|
option_config_ = kUniversalCompaction;
|
|
|
|
Destroy(last_options_);
|
|
|
|
auto options = CurrentOptions();
|
|
|
|
options.create_if_missing = true;
|
|
|
|
Reopen(options);
|
|
|
|
return true;
|
|
|
|
} else if (option_config_ == kUniversalCompaction) {
|
|
|
|
option_config_ = kUniversalCompactionMultiLevel;
|
|
|
|
Destroy(last_options_);
|
|
|
|
auto options = CurrentOptions();
|
|
|
|
options.create_if_missing = true;
|
|
|
|
Reopen(options);
|
|
|
|
return true;
|
|
|
|
} else if (option_config_ == kUniversalCompactionMultiLevel) {
|
|
|
|
option_config_ = kLevelSubcompactions;
|
|
|
|
Destroy(last_options_);
|
|
|
|
auto options = CurrentOptions();
|
|
|
|
assert(options.max_subcompactions > 1);
|
|
|
|
Reopen(options);
|
|
|
|
return true;
|
|
|
|
} else if (option_config_ == kLevelSubcompactions) {
|
|
|
|
option_config_ = kUniversalSubcompactions;
|
|
|
|
Destroy(last_options_);
|
|
|
|
auto options = CurrentOptions();
|
|
|
|
assert(options.max_subcompactions > 1);
|
|
|
|
Reopen(options);
|
|
|
|
return true;
|
|
|
|
} else {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Switch between different WAL settings
|
|
|
|
bool DBTestBase::ChangeWalOptions() {
|
|
|
|
if (option_config_ == kDefault) {
|
|
|
|
option_config_ = kDBLogDir;
|
|
|
|
Destroy(last_options_);
|
|
|
|
auto options = CurrentOptions();
|
|
|
|
Destroy(options);
|
|
|
|
options.create_if_missing = true;
|
|
|
|
Reopen(options);
|
|
|
|
return true;
|
|
|
|
} else if (option_config_ == kDBLogDir) {
|
|
|
|
option_config_ = kWalDirAndMmapReads;
|
|
|
|
Destroy(last_options_);
|
|
|
|
auto options = CurrentOptions();
|
|
|
|
Destroy(options);
|
|
|
|
options.create_if_missing = true;
|
|
|
|
Reopen(options);
|
|
|
|
return true;
|
|
|
|
} else if (option_config_ == kWalDirAndMmapReads) {
|
|
|
|
option_config_ = kRecycleLogFiles;
|
|
|
|
Destroy(last_options_);
|
|
|
|
auto options = CurrentOptions();
|
|
|
|
Destroy(options);
|
|
|
|
Reopen(options);
|
|
|
|
return true;
|
|
|
|
} else {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Switch between different filter policy
|
|
|
|
// Jump from kDefault to kFilter to kFullFilter
|
|
|
|
bool DBTestBase::ChangeFilterOptions() {
|
|
|
|
if (option_config_ == kDefault) {
|
|
|
|
option_config_ = kFilter;
|
|
|
|
} else if (option_config_ == kFilter) {
|
|
|
|
option_config_ = kFullFilterWithNewTableReaderForCompactions;
|
|
|
|
} else if (option_config_ == kFullFilterWithNewTableReaderForCompactions) {
|
|
|
|
option_config_ = kPartitionedFilterWithNewTableReaderForCompactions;
|
|
|
|
} else {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
Destroy(last_options_);
|
|
|
|
|
|
|
|
auto options = CurrentOptions();
|
|
|
|
options.create_if_missing = true;
|
|
|
|
TryReopen(options);
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Switch between different DB options for file ingestion tests.
|
|
|
|
bool DBTestBase::ChangeOptionsForFileIngestionTest() {
|
|
|
|
if (option_config_ == kDefault) {
|
|
|
|
option_config_ = kUniversalCompaction;
|
|
|
|
Destroy(last_options_);
|
|
|
|
auto options = CurrentOptions();
|
|
|
|
options.create_if_missing = true;
|
|
|
|
TryReopen(options);
|
|
|
|
return true;
|
|
|
|
} else if (option_config_ == kUniversalCompaction) {
|
|
|
|
option_config_ = kUniversalCompactionMultiLevel;
|
|
|
|
Destroy(last_options_);
|
|
|
|
auto options = CurrentOptions();
|
|
|
|
options.create_if_missing = true;
|
|
|
|
TryReopen(options);
|
|
|
|
return true;
|
|
|
|
} else if (option_config_ == kUniversalCompactionMultiLevel) {
|
|
|
|
option_config_ = kLevelSubcompactions;
|
|
|
|
Destroy(last_options_);
|
|
|
|
auto options = CurrentOptions();
|
|
|
|
assert(options.max_subcompactions > 1);
|
|
|
|
TryReopen(options);
|
|
|
|
return true;
|
|
|
|
} else if (option_config_ == kLevelSubcompactions) {
|
|
|
|
option_config_ = kUniversalSubcompactions;
|
|
|
|
Destroy(last_options_);
|
|
|
|
auto options = CurrentOptions();
|
|
|
|
assert(options.max_subcompactions > 1);
|
|
|
|
TryReopen(options);
|
|
|
|
return true;
|
|
|
|
} else if (option_config_ == kUniversalSubcompactions) {
|
|
|
|
option_config_ = kDirectIO;
|
|
|
|
Destroy(last_options_);
|
|
|
|
auto options = CurrentOptions();
|
|
|
|
TryReopen(options);
|
|
|
|
return true;
|
|
|
|
} else {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Return the current option configuration.
|
|
|
|
Options DBTestBase::CurrentOptions(
|
|
|
|
const anon::OptionsOverride& options_override) const {
|
|
|
|
return GetOptions(option_config_, GetDefaultOptions(), options_override);
|
|
|
|
}
|
|
|
|
|
|
|
|
Options DBTestBase::CurrentOptions(
|
|
|
|
const Options& default_options,
|
|
|
|
const anon::OptionsOverride& options_override) const {
|
|
|
|
return GetOptions(option_config_, default_options, options_override);
|
|
|
|
}
|
|
|
|
|
Fix many tests to run with MEM_ENV and ENCRYPTED_ENV; Introduce a MemoryFileSystem class (#7566)
Summary:
This PR does a few things:
1. The MockFileSystem class was split out from the MockEnv. This change would theoretically allow a MockFileSystem to be used by other Environments as well (if we created a means of constructing one). The MockFileSystem implements a FileSystem in its entirety and does not rely on any Wrapper implementation.
2. Make the RocksDB test suite work when MOCK_ENV=1 and ENCRYPTED_ENV=1 are set. To accomplish this, a few things were needed:
- The tests that tried to use the "wrong" environment (Env::Default() instead of env_) were updated
- The MockFileSystem was changed to support the features it was missing or mishandled (such as recursively deleting files in a directory or supporting renaming of a directory).
3. Updated the test framework to have a ROCKSDB_GTEST_SKIP macro. This can be used to flag tests that are skipped. Currently, this defaults to doing nothing (marks the test as SUCCESS) but will mark the tests as SKIPPED when RocksDB is upgraded to a version of gtest that supports this (gtest-1.10).
I have run a full "make check" with MEM_ENV, ENCRYPTED_ENV, both, and neither under both MacOS and RedHat. A few tests were disabled/skipped for the MEM/ENCRYPTED cases. The error_handler_fs_test fails/hangs for MEM_ENV (presumably a timing problem) and I will introduce another PR/issue to track that problem. (I will also push a change to disable those tests soon). There is one more test in DBTest2 that also fails which I need to investigate or skip before this PR is merged.
Theoretically, this PR should also allow the test suite to run against an Env loaded from the registry, though I do not have one to try it with currently.
Finally, once this is accepted, it would be nice if there was a CircleCI job to run these tests on a checkin so this effort does not become stale. I do not know how to do that, so if someone could write that job, it would be appreciated :)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7566
Reviewed By: zhichao-cao
Differential Revision: D24408980
Pulled By: jay-zhuang
fbshipit-source-id: 911b1554a4d0da06fd51feca0c090a4abdcb4a5f
4 years ago
|
|
|
Options DBTestBase::GetDefaultOptions() const {
|
|
|
|
Options options;
|
|
|
|
options.write_buffer_size = 4090 * 4096;
|
|
|
|
options.target_file_size_base = 2 * 1024 * 1024;
|
|
|
|
options.max_bytes_for_level_base = 10 * 1024 * 1024;
|
|
|
|
options.max_open_files = 5000;
|
|
|
|
options.wal_recovery_mode = WALRecoveryMode::kTolerateCorruptedTailRecords;
|
|
|
|
options.compaction_pri = CompactionPri::kByCompensatedSize;
|
Fix many tests to run with MEM_ENV and ENCRYPTED_ENV; Introduce a MemoryFileSystem class (#7566)
Summary:
This PR does a few things:
1. The MockFileSystem class was split out from the MockEnv. This change would theoretically allow a MockFileSystem to be used by other Environments as well (if we created a means of constructing one). The MockFileSystem implements a FileSystem in its entirety and does not rely on any Wrapper implementation.
2. Make the RocksDB test suite work when MOCK_ENV=1 and ENCRYPTED_ENV=1 are set. To accomplish this, a few things were needed:
- The tests that tried to use the "wrong" environment (Env::Default() instead of env_) were updated
- The MockFileSystem was changed to support the features it was missing or mishandled (such as recursively deleting files in a directory or supporting renaming of a directory).
3. Updated the test framework to have a ROCKSDB_GTEST_SKIP macro. This can be used to flag tests that are skipped. Currently, this defaults to doing nothing (marks the test as SUCCESS) but will mark the tests as SKIPPED when RocksDB is upgraded to a version of gtest that supports this (gtest-1.10).
I have run a full "make check" with MEM_ENV, ENCRYPTED_ENV, both, and neither under both MacOS and RedHat. A few tests were disabled/skipped for the MEM/ENCRYPTED cases. The error_handler_fs_test fails/hangs for MEM_ENV (presumably a timing problem) and I will introduce another PR/issue to track that problem. (I will also push a change to disable those tests soon). There is one more test in DBTest2 that also fails which I need to investigate or skip before this PR is merged.
Theoretically, this PR should also allow the test suite to run against an Env loaded from the registry, though I do not have one to try it with currently.
Finally, once this is accepted, it would be nice if there was a CircleCI job to run these tests on a checkin so this effort does not become stale. I do not know how to do that, so if someone could write that job, it would be appreciated :)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7566
Reviewed By: zhichao-cao
Differential Revision: D24408980
Pulled By: jay-zhuang
fbshipit-source-id: 911b1554a4d0da06fd51feca0c090a4abdcb4a5f
4 years ago
|
|
|
options.env = env_;
|
|
|
|
if (!env_->skip_fsync_) {
|
|
|
|
options.track_and_verify_wals_in_manifest = true;
|
|
|
|
}
|
|
|
|
return options;
|
|
|
|
}
|
|
|
|
|
|
|
|
Options DBTestBase::GetOptions(
|
|
|
|
int option_config, const Options& default_options,
|
|
|
|
const anon::OptionsOverride& options_override) const {
|
|
|
|
// this redundant copy is to minimize code change w/o having lint error.
|
|
|
|
Options options = default_options;
|
|
|
|
BlockBasedTableOptions table_options;
|
|
|
|
bool set_block_based_table_factory = true;
|
|
|
|
#if !defined(OS_MACOSX) && !defined(OS_WIN) && !defined(OS_SOLARIS) && \
|
|
|
|
!defined(OS_AIX)
|
|
|
|
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearCallBack(
|
|
|
|
"NewRandomAccessFile:O_DIRECT");
|
|
|
|
ROCKSDB_NAMESPACE::SyncPoint::GetInstance()->ClearCallBack(
|
|
|
|
"NewWritableFile:O_DIRECT");
|
|
|
|
#endif
|
Eliminate unnecessary (slow) block cache Ref()ing in MultiGet (#9899)
Summary:
When MultiGet() determines that multiple query keys can be
served by examining the same data block in block cache (one Lookup()),
each PinnableSlice referring to data in that data block needs to hold
on to the block in cache so that they can be released at arbitrary
times by the API user. Historically this is accomplished with extra
calls to Ref() on the Handle from Lookup(), with each PinnableSlice
cleanup calling Release() on the Handle, but this creates extra
contention on the block cache for the extra Ref()s and Release()es,
especially because they hit the same cache shard repeatedly.
In the case of merge operands (possibly more cases?), the problem was
compounded by doing an extra Ref()+eventual Release() for each merge
operand for a key reusing a block (which could be the same key!), rather
than one Ref() per key. (Note: the non-shared case with `biter` was
already one per key.)
This change optimizes MultiGet not to rely on these extra, contentious
Ref()+Release() calls by instead, in the shared block case, wrapping
the cache Release() cleanup in a refcounted object referenced by the
PinnableSlices, such that after the last wrapped reference is released,
the cache entry is Release()ed. Relaxed atomic refcounts should be
much faster than mutex-guarded Ref() and Release(), and much less prone
to a performance cliff when MultiGet() does a lot of block sharing.
Note that I did not use std::shared_ptr, because that would require an
extra indirection object (shared_ptr itself new/delete) in order to
associate a ref increment/decrement with a Cleanable cleanup entry. (If
I assumed it was the size of two pointers, I could do some hackery to
make it work without the extra indirection, but that's too fragile.)
Some details:
* Fixed (removed) extra block cache tracing entries in cases of cache
entry reuse in MultiGet, but it's likely that in some other cases traces
are missing (XXX comment inserted)
* Moved existing implementations for cleanable.h from iterator.cc to
new cleanable.cc
* Improved API comments on Cleanable
* Added a public SharedCleanablePtr class to cleanable.h in case others
could benefit from the same pattern (potentially many Cleanables and/or
smart pointers referencing a shared Cleanable)
* Add a typedef for MultiGetContext::Mask
* Some variable renaming for clarity
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9899
Test Plan:
Added unit tests for SharedCleanablePtr.
Greatly enhanced ability of existing tests to detect cache use-after-free.
* Release PinnableSlices from MultiGet as they are read rather than in
bulk (in db_test_util wrapper).
* In ASAN build, default to using a trivially small LRUCache for block_cache
so that entries are immediately erased when unreferenced. (Updated two
tests that depend on caching.) New ASAN testsuite running time seems
OK to me.
If I introduce a bug into my implementation where we skip the shared
cleanups on block reuse, ASAN detects the bug in
`db_basic_test *MultiGet*`. If I remove either of the above testing
enhancements, the bug is not detected.
Consider for follow-up work: manipulate or randomize ordering of
PinnableSlice use and release from MultiGet db_test_util wrapper. But in
typical cases, natural ordering gives pretty good functional coverage.
Performance test:
In the extreme (but possible) case of MultiGetting the same or adjacent keys
in a batch, throughput can improve by an order of magnitude.
`./db_bench -benchmarks=multireadrandom -db=/dev/shm/testdb -readonly -num=5 -duration=10 -threads=20 -multiread_batched -batch_size=200`
Before ops/sec, num=5: 1,384,394
Before ops/sec, num=500: 6,423,720
After ops/sec, num=500: 10,658,794
After ops/sec, num=5: 16,027,257
Also note that previously, with high parallelism, having query keys
concentrated in a single block was worse than spreading them out a bit. Now
concentrated in a single block is faster than spread out, which is hopefully
consistent with natural expectation.
Random query performance: with num=1000000, over 999 x 10s runs running before & after simultaneously (each -threads=12):
Before: multireadrandom [AVG 999 runs] : 1088699 (± 7344) ops/sec; 120.4 (± 0.8 ) MB/sec
After: multireadrandom [AVG 999 runs] : 1090402 (± 7230) ops/sec; 120.6 (± 0.8 ) MB/sec
Possibly better, possibly in the noise.
Reviewed By: anand1976
Differential Revision: D35907003
Pulled By: pdillinger
fbshipit-source-id: bbd244d703649a8ca12d476f2d03853ed9d1a17e
3 years ago
|
|
|
// kMustFreeHeapAllocations -> indicates ASAN build
|
|
|
|
if (kMustFreeHeapAllocations && !options_override.full_block_cache) {
|
|
|
|
// Detecting block cache use-after-free is normally difficult in unit
|
|
|
|
// tests, because as a cache, it tends to keep unreferenced entries in
|
|
|
|
// memory, and we normally want unit tests to take advantage of block
|
|
|
|
// cache for speed. However, we also want a strong chance of detecting
|
|
|
|
// block cache use-after-free in unit tests in ASAN builds, so for ASAN
|
|
|
|
// builds we use a trivially small block cache to which entries can be
|
|
|
|
// added but are immediately freed on no more references.
|
|
|
|
table_options.block_cache = NewLRUCache(/* too small */ 1);
|
|
|
|
}
|
|
|
|
|
|
|
|
bool can_allow_mmap = IsMemoryMappedAccessSupported();
|
|
|
|
switch (option_config) {
|
|
|
|
case kHashSkipList:
|
|
|
|
options.prefix_extractor.reset(NewFixedPrefixTransform(1));
|
|
|
|
options.memtable_factory.reset(NewHashSkipListRepFactory(16));
|
|
|
|
options.allow_concurrent_memtable_write = false;
|
|
|
|
options.unordered_write = false;
|
|
|
|
break;
|
|
|
|
case kPlainTableFirstBytePrefix:
|
Fix many tests to run with MEM_ENV and ENCRYPTED_ENV; Introduce a MemoryFileSystem class (#7566)
Summary:
This PR does a few things:
1. The MockFileSystem class was split out from the MockEnv. This change would theoretically allow a MockFileSystem to be used by other Environments as well (if we created a means of constructing one). The MockFileSystem implements a FileSystem in its entirety and does not rely on any Wrapper implementation.
2. Make the RocksDB test suite work when MOCK_ENV=1 and ENCRYPTED_ENV=1 are set. To accomplish this, a few things were needed:
- The tests that tried to use the "wrong" environment (Env::Default() instead of env_) were updated
- The MockFileSystem was changed to support the features it was missing or mishandled (such as recursively deleting files in a directory or supporting renaming of a directory).
3. Updated the test framework to have a ROCKSDB_GTEST_SKIP macro. This can be used to flag tests that are skipped. Currently, this defaults to doing nothing (marks the test as SUCCESS) but will mark the tests as SKIPPED when RocksDB is upgraded to a version of gtest that supports this (gtest-1.10).
I have run a full "make check" with MEM_ENV, ENCRYPTED_ENV, both, and neither under both MacOS and RedHat. A few tests were disabled/skipped for the MEM/ENCRYPTED cases. The error_handler_fs_test fails/hangs for MEM_ENV (presumably a timing problem) and I will introduce another PR/issue to track that problem. (I will also push a change to disable those tests soon). There is one more test in DBTest2 that also fails which I need to investigate or skip before this PR is merged.
Theoretically, this PR should also allow the test suite to run against an Env loaded from the registry, though I do not have one to try it with currently.
Finally, once this is accepted, it would be nice if there was a CircleCI job to run these tests on a checkin so this effort does not become stale. I do not know how to do that, so if someone could write that job, it would be appreciated :)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7566
Reviewed By: zhichao-cao
Differential Revision: D24408980
Pulled By: jay-zhuang
fbshipit-source-id: 911b1554a4d0da06fd51feca0c090a4abdcb4a5f
4 years ago
|
|
|
options.table_factory.reset(NewPlainTableFactory());
|
|
|
|
options.prefix_extractor.reset(NewFixedPrefixTransform(1));
|
|
|
|
options.allow_mmap_reads = can_allow_mmap;
|
|
|
|
options.max_sequential_skip_in_iterations = 999999;
|
|
|
|
set_block_based_table_factory = false;
|
|
|
|
break;
|
|
|
|
case kPlainTableCappedPrefix:
|
Fix many tests to run with MEM_ENV and ENCRYPTED_ENV; Introduce a MemoryFileSystem class (#7566)
Summary:
This PR does a few things:
1. The MockFileSystem class was split out from the MockEnv. This change would theoretically allow a MockFileSystem to be used by other Environments as well (if we created a means of constructing one). The MockFileSystem implements a FileSystem in its entirety and does not rely on any Wrapper implementation.
2. Make the RocksDB test suite work when MOCK_ENV=1 and ENCRYPTED_ENV=1 are set. To accomplish this, a few things were needed:
- The tests that tried to use the "wrong" environment (Env::Default() instead of env_) were updated
- The MockFileSystem was changed to support the features it was missing or mishandled (such as recursively deleting files in a directory or supporting renaming of a directory).
3. Updated the test framework to have a ROCKSDB_GTEST_SKIP macro. This can be used to flag tests that are skipped. Currently, this defaults to doing nothing (marks the test as SUCCESS) but will mark the tests as SKIPPED when RocksDB is upgraded to a version of gtest that supports this (gtest-1.10).
I have run a full "make check" with MEM_ENV, ENCRYPTED_ENV, both, and neither under both MacOS and RedHat. A few tests were disabled/skipped for the MEM/ENCRYPTED cases. The error_handler_fs_test fails/hangs for MEM_ENV (presumably a timing problem) and I will introduce another PR/issue to track that problem. (I will also push a change to disable those tests soon). There is one more test in DBTest2 that also fails which I need to investigate or skip before this PR is merged.
Theoretically, this PR should also allow the test suite to run against an Env loaded from the registry, though I do not have one to try it with currently.
Finally, once this is accepted, it would be nice if there was a CircleCI job to run these tests on a checkin so this effort does not become stale. I do not know how to do that, so if someone could write that job, it would be appreciated :)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7566
Reviewed By: zhichao-cao
Differential Revision: D24408980
Pulled By: jay-zhuang
fbshipit-source-id: 911b1554a4d0da06fd51feca0c090a4abdcb4a5f
4 years ago
|
|
|
options.table_factory.reset(NewPlainTableFactory());
|
|
|
|
options.prefix_extractor.reset(NewCappedPrefixTransform(8));
|
|
|
|
options.allow_mmap_reads = can_allow_mmap;
|
|
|
|
options.max_sequential_skip_in_iterations = 999999;
|
|
|
|
set_block_based_table_factory = false;
|
|
|
|
break;
|
|
|
|
case kPlainTableCappedPrefixNonMmap:
|
Fix many tests to run with MEM_ENV and ENCRYPTED_ENV; Introduce a MemoryFileSystem class (#7566)
Summary:
This PR does a few things:
1. The MockFileSystem class was split out from the MockEnv. This change would theoretically allow a MockFileSystem to be used by other Environments as well (if we created a means of constructing one). The MockFileSystem implements a FileSystem in its entirety and does not rely on any Wrapper implementation.
2. Make the RocksDB test suite work when MOCK_ENV=1 and ENCRYPTED_ENV=1 are set. To accomplish this, a few things were needed:
- The tests that tried to use the "wrong" environment (Env::Default() instead of env_) were updated
- The MockFileSystem was changed to support the features it was missing or mishandled (such as recursively deleting files in a directory or supporting renaming of a directory).
3. Updated the test framework to have a ROCKSDB_GTEST_SKIP macro. This can be used to flag tests that are skipped. Currently, this defaults to doing nothing (marks the test as SUCCESS) but will mark the tests as SKIPPED when RocksDB is upgraded to a version of gtest that supports this (gtest-1.10).
I have run a full "make check" with MEM_ENV, ENCRYPTED_ENV, both, and neither under both MacOS and RedHat. A few tests were disabled/skipped for the MEM/ENCRYPTED cases. The error_handler_fs_test fails/hangs for MEM_ENV (presumably a timing problem) and I will introduce another PR/issue to track that problem. (I will also push a change to disable those tests soon). There is one more test in DBTest2 that also fails which I need to investigate or skip before this PR is merged.
Theoretically, this PR should also allow the test suite to run against an Env loaded from the registry, though I do not have one to try it with currently.
Finally, once this is accepted, it would be nice if there was a CircleCI job to run these tests on a checkin so this effort does not become stale. I do not know how to do that, so if someone could write that job, it would be appreciated :)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7566
Reviewed By: zhichao-cao
Differential Revision: D24408980
Pulled By: jay-zhuang
fbshipit-source-id: 911b1554a4d0da06fd51feca0c090a4abdcb4a5f
4 years ago
|
|
|
options.table_factory.reset(NewPlainTableFactory());
|
|
|
|
options.prefix_extractor.reset(NewCappedPrefixTransform(8));
|
|
|
|
options.allow_mmap_reads = false;
|
|
|
|
options.max_sequential_skip_in_iterations = 999999;
|
|
|
|
set_block_based_table_factory = false;
|
|
|
|
break;
|
|
|
|
case kPlainTableAllBytesPrefix:
|
Fix many tests to run with MEM_ENV and ENCRYPTED_ENV; Introduce a MemoryFileSystem class (#7566)
Summary:
This PR does a few things:
1. The MockFileSystem class was split out from the MockEnv. This change would theoretically allow a MockFileSystem to be used by other Environments as well (if we created a means of constructing one). The MockFileSystem implements a FileSystem in its entirety and does not rely on any Wrapper implementation.
2. Make the RocksDB test suite work when MOCK_ENV=1 and ENCRYPTED_ENV=1 are set. To accomplish this, a few things were needed:
- The tests that tried to use the "wrong" environment (Env::Default() instead of env_) were updated
- The MockFileSystem was changed to support the features it was missing or mishandled (such as recursively deleting files in a directory or supporting renaming of a directory).
3. Updated the test framework to have a ROCKSDB_GTEST_SKIP macro. This can be used to flag tests that are skipped. Currently, this defaults to doing nothing (marks the test as SUCCESS) but will mark the tests as SKIPPED when RocksDB is upgraded to a version of gtest that supports this (gtest-1.10).
I have run a full "make check" with MEM_ENV, ENCRYPTED_ENV, both, and neither under both MacOS and RedHat. A few tests were disabled/skipped for the MEM/ENCRYPTED cases. The error_handler_fs_test fails/hangs for MEM_ENV (presumably a timing problem) and I will introduce another PR/issue to track that problem. (I will also push a change to disable those tests soon). There is one more test in DBTest2 that also fails which I need to investigate or skip before this PR is merged.
Theoretically, this PR should also allow the test suite to run against an Env loaded from the registry, though I do not have one to try it with currently.
Finally, once this is accepted, it would be nice if there was a CircleCI job to run these tests on a checkin so this effort does not become stale. I do not know how to do that, so if someone could write that job, it would be appreciated :)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7566
Reviewed By: zhichao-cao
Differential Revision: D24408980
Pulled By: jay-zhuang
fbshipit-source-id: 911b1554a4d0da06fd51feca0c090a4abdcb4a5f
4 years ago
|
|
|
options.table_factory.reset(NewPlainTableFactory());
|
|
|
|
options.prefix_extractor.reset(NewNoopTransform());
|
|
|
|
options.allow_mmap_reads = can_allow_mmap;
|
|
|
|
options.max_sequential_skip_in_iterations = 999999;
|
|
|
|
set_block_based_table_factory = false;
|
|
|
|
break;
|
|
|
|
case kVectorRep:
|
|
|
|
options.memtable_factory.reset(new VectorRepFactory(100));
|
|
|
|
options.allow_concurrent_memtable_write = false;
|
|
|
|
options.unordered_write = false;
|
|
|
|
break;
|
|
|
|
case kHashLinkList:
|
|
|
|
options.prefix_extractor.reset(NewFixedPrefixTransform(1));
|
|
|
|
options.memtable_factory.reset(
|
|
|
|
NewHashLinkListRepFactory(4, 0, 3, true, 4));
|
|
|
|
options.allow_concurrent_memtable_write = false;
|
|
|
|
options.unordered_write = false;
|
|
|
|
break;
|
|
|
|
case kDirectIO: {
|
|
|
|
options.use_direct_reads = true;
|
|
|
|
options.use_direct_io_for_flush_and_compaction = true;
|
|
|
|
options.compaction_readahead_size = 2 * 1024 * 1024;
|
|
|
|
SetupSyncPointsToMockDirectIO();
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
case kMergePut:
|
|
|
|
options.merge_operator = MergeOperators::CreatePutOperator();
|
|
|
|
break;
|
|
|
|
case kFilter:
|
|
|
|
table_options.filter_policy.reset(NewBloomFilterPolicy(10, true));
|
|
|
|
break;
|
|
|
|
case kFullFilterWithNewTableReaderForCompactions:
|
|
|
|
table_options.filter_policy.reset(NewBloomFilterPolicy(10, false));
|
|
|
|
options.compaction_readahead_size = 10 * 1024 * 1024;
|
|
|
|
break;
|
|
|
|
case kPartitionedFilterWithNewTableReaderForCompactions:
|
|
|
|
table_options.filter_policy.reset(NewBloomFilterPolicy(10, false));
|
|
|
|
table_options.partition_filters = true;
|
|
|
|
table_options.index_type =
|
|
|
|
BlockBasedTableOptions::IndexType::kTwoLevelIndexSearch;
|
|
|
|
options.compaction_readahead_size = 10 * 1024 * 1024;
|
|
|
|
break;
|
|
|
|
case kUncompressed:
|
|
|
|
options.compression = kNoCompression;
|
|
|
|
break;
|
|
|
|
case kNumLevel_3:
|
|
|
|
options.num_levels = 3;
|
|
|
|
break;
|
|
|
|
case kDBLogDir:
|
|
|
|
options.db_log_dir = alternative_db_log_dir_;
|
|
|
|
break;
|
|
|
|
case kWalDirAndMmapReads:
|
|
|
|
options.wal_dir = alternative_wal_dir_;
|
|
|
|
// mmap reads should be orthogonal to WalDir setting, so we piggyback to
|
|
|
|
// this option config to test mmap reads as well
|
|
|
|
options.allow_mmap_reads = can_allow_mmap;
|
|
|
|
break;
|
|
|
|
case kManifestFileSize:
|
|
|
|
options.max_manifest_file_size = 50; // 50 bytes
|
|
|
|
break;
|
|
|
|
case kPerfOptions:
|
|
|
|
options.delayed_write_rate = 8 * 1024 * 1024;
|
|
|
|
options.report_bg_io_stats = true;
|
|
|
|
// TODO(3.13) -- test more options
|
|
|
|
break;
|
|
|
|
case kUniversalCompaction:
|
|
|
|
options.compaction_style = kCompactionStyleUniversal;
|
|
|
|
options.num_levels = 1;
|
|
|
|
break;
|
|
|
|
case kUniversalCompactionMultiLevel:
|
|
|
|
options.compaction_style = kCompactionStyleUniversal;
|
|
|
|
options.num_levels = 8;
|
|
|
|
break;
|
|
|
|
case kInfiniteMaxOpenFiles:
|
|
|
|
options.max_open_files = -1;
|
|
|
|
break;
|
|
|
|
case kCRC32cChecksum: {
|
|
|
|
// Old default was CRC32c, but XXH3 (new default) is faster on common
|
|
|
|
// hardware
|
|
|
|
table_options.checksum = kCRC32c;
|
|
|
|
// Thrown in here for basic coverage:
|
|
|
|
options.DisableExtraChecks();
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
case kFIFOCompaction: {
|
|
|
|
options.compaction_style = kCompactionStyleFIFO;
|
|
|
|
options.max_open_files = -1;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
case kBlockBasedTableWithPrefixHashIndex: {
|
|
|
|
table_options.index_type = BlockBasedTableOptions::kHashSearch;
|
|
|
|
options.prefix_extractor.reset(NewFixedPrefixTransform(1));
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
case kBlockBasedTableWithWholeKeyHashIndex: {
|
|
|
|
table_options.index_type = BlockBasedTableOptions::kHashSearch;
|
|
|
|
options.prefix_extractor.reset(NewNoopTransform());
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
case kBlockBasedTableWithPartitionedIndex: {
|
|
|
|
table_options.format_version = 3;
|
|
|
|
table_options.index_type = BlockBasedTableOptions::kTwoLevelIndexSearch;
|
|
|
|
options.prefix_extractor.reset(NewNoopTransform());
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
case kBlockBasedTableWithPartitionedIndexFormat4: {
|
|
|
|
table_options.format_version = 4;
|
|
|
|
// Format 4 changes the binary index format. Since partitioned index is a
|
|
|
|
// super-set of simple indexes, we are also using kTwoLevelIndexSearch to
|
|
|
|
// test this format.
|
|
|
|
table_options.index_type = BlockBasedTableOptions::kTwoLevelIndexSearch;
|
|
|
|
// The top-level index in partition filters are also affected by format 4.
|
|
|
|
table_options.filter_policy.reset(NewBloomFilterPolicy(10, false));
|
|
|
|
table_options.partition_filters = true;
|
|
|
|
table_options.index_block_restart_interval = 8;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
case kBlockBasedTableWithIndexRestartInterval: {
|
|
|
|
table_options.index_block_restart_interval = 8;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
case kBlockBasedTableWithLatestFormat: {
|
|
|
|
// In case different from default
|
|
|
|
table_options.format_version = kLatestFormatVersion;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
case kOptimizeFiltersForHits: {
|
|
|
|
options.optimize_filters_for_hits = true;
|
|
|
|
set_block_based_table_factory = true;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
case kRowCache: {
|
|
|
|
options.row_cache = NewLRUCache(1024 * 1024);
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
case kRecycleLogFiles: {
|
|
|
|
options.recycle_log_file_num = 2;
|
|
|
|
break;
|
|
|
|
}
|
Parallelize L0-L1 Compaction: Restructure Compaction Job
Summary:
As of now compactions involving files from Level 0 and Level 1 are single
threaded because the files in L0, although sorted, are not range partitioned like
the other levels. This means that during L0-L1 compaction each file from L1
needs to be merged with potentially all the files from L0.
This attempt to parallelize the L0-L1 compaction assigns a thread and a
corresponding iterator to each L1 file that then considers only the key range
found in that L1 file and only the L0 files that have those keys (and only the
specific portion of those L0 files in which those keys are found). In this way
the overlap is minimized and potentially eliminated between different iterators
focusing on the same files.
The first step is to restructure the compaction logic to break L0-L1 compactions
into multiple, smaller, sequential compactions. Eventually each of these smaller
jobs will be run simultaneously. Areas to pay extra attention to are
# Correct aggregation of compaction job statistics across multiple threads
# Proper opening/closing of output files (make sure each thread's is unique)
# Keys that span multiple L1 files
# Skewed distributions of keys within L0 files
Test Plan: Make and run db_test (newer version has separate compaction tests) and compaction_job_stats_test
Reviewers: igor, noetzli, anthony, sdong, yhchiang
Reviewed By: yhchiang
Subscribers: MarkCallaghan, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D42699
10 years ago
|
|
|
case kLevelSubcompactions: {
|
|
|
|
options.max_subcompactions = 4;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
case kUniversalSubcompactions: {
|
|
|
|
options.compaction_style = kCompactionStyleUniversal;
|
|
|
|
options.num_levels = 8;
|
|
|
|
options.max_subcompactions = 4;
|
Parallelize L0-L1 Compaction: Restructure Compaction Job
Summary:
As of now compactions involving files from Level 0 and Level 1 are single
threaded because the files in L0, although sorted, are not range partitioned like
the other levels. This means that during L0-L1 compaction each file from L1
needs to be merged with potentially all the files from L0.
This attempt to parallelize the L0-L1 compaction assigns a thread and a
corresponding iterator to each L1 file that then considers only the key range
found in that L1 file and only the L0 files that have those keys (and only the
specific portion of those L0 files in which those keys are found). In this way
the overlap is minimized and potentially eliminated between different iterators
focusing on the same files.
The first step is to restructure the compaction logic to break L0-L1 compactions
into multiple, smaller, sequential compactions. Eventually each of these smaller
jobs will be run simultaneously. Areas to pay extra attention to are
# Correct aggregation of compaction job statistics across multiple threads
# Proper opening/closing of output files (make sure each thread's is unique)
# Keys that span multiple L1 files
# Skewed distributions of keys within L0 files
Test Plan: Make and run db_test (newer version has separate compaction tests) and compaction_job_stats_test
Reviewers: igor, noetzli, anthony, sdong, yhchiang
Reviewed By: yhchiang
Subscribers: MarkCallaghan, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D42699
10 years ago
|
|
|
break;
|
|
|
|
}
|
|
|
|
case kConcurrentSkipList: {
|
|
|
|
options.allow_concurrent_memtable_write = true;
|
|
|
|
options.enable_write_thread_adaptive_yield = true;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
case kPipelinedWrite: {
|
|
|
|
options.enable_pipelined_write = true;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
case kConcurrentWALWrites: {
|
|
|
|
// This options optimize 2PC commit path
|
|
|
|
options.two_write_queues = true;
|
|
|
|
options.manual_wal_flush = true;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
case kUnorderedWrite: {
|
|
|
|
options.allow_concurrent_memtable_write = false;
|
|
|
|
options.unordered_write = false;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
default:
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (options_override.filter_policy) {
|
|
|
|
table_options.filter_policy = options_override.filter_policy;
|
|
|
|
table_options.partition_filters = options_override.partition_filters;
|
|
|
|
table_options.metadata_block_size = options_override.metadata_block_size;
|
|
|
|
}
|
|
|
|
if (set_block_based_table_factory) {
|
|
|
|
options.table_factory.reset(NewBlockBasedTableFactory(table_options));
|
|
|
|
}
|
|
|
|
options.env = env_;
|
|
|
|
options.create_if_missing = true;
|
|
|
|
options.fail_if_options_file_error = true;
|
|
|
|
return options;
|
|
|
|
}
|
|
|
|
|
|
|
|
void DBTestBase::CreateColumnFamilies(const std::vector<std::string>& cfs,
|
|
|
|
const Options& options) {
|
|
|
|
ColumnFamilyOptions cf_opts(options);
|
|
|
|
size_t cfi = handles_.size();
|
|
|
|
handles_.resize(cfi + cfs.size());
|
|
|
|
for (auto cf : cfs) {
|
|
|
|
Status s = db_->CreateColumnFamily(cf_opts, cf, &handles_[cfi++]);
|
|
|
|
ASSERT_OK(s);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void DBTestBase::CreateAndReopenWithCF(const std::vector<std::string>& cfs,
|
|
|
|
const Options& options) {
|
|
|
|
CreateColumnFamilies(cfs, options);
|
|
|
|
std::vector<std::string> cfs_plus_default = cfs;
|
|
|
|
cfs_plus_default.insert(cfs_plus_default.begin(), kDefaultColumnFamilyName);
|
|
|
|
ReopenWithColumnFamilies(cfs_plus_default, options);
|
|
|
|
}
|
|
|
|
|
|
|
|
void DBTestBase::ReopenWithColumnFamilies(const std::vector<std::string>& cfs,
|
|
|
|
const std::vector<Options>& options) {
|
|
|
|
ASSERT_OK(TryReopenWithColumnFamilies(cfs, options));
|
|
|
|
}
|
|
|
|
|
|
|
|
void DBTestBase::ReopenWithColumnFamilies(const std::vector<std::string>& cfs,
|
|
|
|
const Options& options) {
|
|
|
|
ASSERT_OK(TryReopenWithColumnFamilies(cfs, options));
|
|
|
|
}
|
|
|
|
|
Fix+clean up handling of mock sleeps (#7101)
Summary:
We have a number of tests hanging on MacOS and windows due to
mishandling of code for mock sleeps. In addition, the code was in
terrible shape because the same variable (addon_time_) would sometimes
refer to microseconds and sometimes to seconds. One test even assumed it
was nanoseconds but was written to pass anyway.
This has been cleaned up so that DB tests generally use a SpecialEnv
function to mock sleep, for either some number of microseconds or seconds
depending on the function called. But to call one of these, the test must first
call SetMockSleep (precondition enforced with assertion), which also turns
sleeps in RocksDB into mock sleeps. To also removes accounting for actual
clock time, call SetTimeElapseOnlySleepOnReopen, which implies
SetMockSleep (on DB re-open). This latter setting only works by applying
on DB re-open, otherwise havoc can ensue if Env goes back in time with
DB open.
More specifics:
Removed some unused test classes, and updated comments on the general
problem.
Fixed DBSSTTest.GetTotalSstFilesSize using a sync point callback instead
of mock time. For this we have the only modification to production code,
inserting a sync point callback in flush_job.cc, which is not a change to
production behavior.
Removed unnecessary resetting of mock times to 0 in many tests. RocksDB
deals in relative time. Any behaviors relying on absolute date/time are likely
a bug. (The above test DBSSTTest.GetTotalSstFilesSize was the only one
clearly injecting a specific absolute time for actual testing convenience.) Just
in case I misunderstood some test, I put this note in each replacement:
// NOTE: Presumed unnecessary and removed: resetting mock time in env
Strengthened some tests like MergeTestTime, MergeCompactionTimeTest, and
FilterCompactionTimeTest in db_test.cc
stats_history_test and blob_db_test are each their own beast, rather deeply
dependent on MockTimeEnv. Each gets its own variant of a work-around for
TimedWait in a mock time environment. (Reduces redundancy and
inconsistency in stats_history_test.)
Intended follow-up:
Remove TimedWait from the public API of InstrumentedCondVar, and only
make that accessible through Env by passing in an InstrumentedCondVar and
a deadline. Then the Env implementations mocking time can fix this problem
without using sync points. (Test infrastructure using sync points interferes
with individual tests' control over sync points.)
With that change, we can simplify/consolidate the scattered work-arounds.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7101
Test Plan: make check on Linux and MacOS
Reviewed By: zhichao-cao
Differential Revision: D23032815
Pulled By: pdillinger
fbshipit-source-id: 7f33967ada8b83011fb54e8279365c008bd6610b
4 years ago
|
|
|
void DBTestBase::SetTimeElapseOnlySleepOnReopen(DBOptions* options) {
|
|
|
|
time_elapse_only_sleep_on_reopen_ = true;
|
|
|
|
|
|
|
|
// Need to disable stats dumping and persisting which also use
|
|
|
|
// RepeatableThread, which uses InstrumentedCondVar::TimedWaitInternal.
|
|
|
|
// With time_elapse_only_sleep_, this can hang on some platforms (MacOS)
|
|
|
|
// because (a) on some platforms, pthread_cond_timedwait does not appear
|
|
|
|
// to release the lock for other threads to operate if the deadline time
|
|
|
|
// is already passed, and (b) TimedWait calls are currently a bad abstraction
|
|
|
|
// because the deadline parameter is usually computed from Env time,
|
|
|
|
// but is interpreted in real clock time.
|
|
|
|
options->stats_dump_period_sec = 0;
|
|
|
|
options->stats_persist_period_sec = 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
void DBTestBase::MaybeInstallTimeElapseOnlySleep(const DBOptions& options) {
|
|
|
|
if (time_elapse_only_sleep_on_reopen_) {
|
|
|
|
assert(options.env == env_ ||
|
|
|
|
static_cast_with_check<CompositeEnvWrapper>(options.env)
|
|
|
|
->env_target() == env_);
|
|
|
|
assert(options.stats_dump_period_sec == 0);
|
|
|
|
assert(options.stats_persist_period_sec == 0);
|
|
|
|
// We cannot set these before destroying the last DB because they might
|
|
|
|
// cause a deadlock or similar without the appropriate options set in
|
|
|
|
// the DB.
|
|
|
|
env_->time_elapse_only_sleep_ = true;
|
|
|
|
env_->no_slowdown_ = true;
|
|
|
|
} else {
|
|
|
|
// Going back in same test run is not yet supported, so no
|
|
|
|
// reset in this case.
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
Status DBTestBase::TryReopenWithColumnFamilies(
|
|
|
|
const std::vector<std::string>& cfs, const std::vector<Options>& options) {
|
|
|
|
Close();
|
|
|
|
EXPECT_EQ(cfs.size(), options.size());
|
|
|
|
std::vector<ColumnFamilyDescriptor> column_families;
|
|
|
|
for (size_t i = 0; i < cfs.size(); ++i) {
|
|
|
|
column_families.push_back(ColumnFamilyDescriptor(cfs[i], options[i]));
|
|
|
|
}
|
|
|
|
DBOptions db_opts = DBOptions(options[0]);
|
|
|
|
last_options_ = options[0];
|
Fix+clean up handling of mock sleeps (#7101)
Summary:
We have a number of tests hanging on MacOS and windows due to
mishandling of code for mock sleeps. In addition, the code was in
terrible shape because the same variable (addon_time_) would sometimes
refer to microseconds and sometimes to seconds. One test even assumed it
was nanoseconds but was written to pass anyway.
This has been cleaned up so that DB tests generally use a SpecialEnv
function to mock sleep, for either some number of microseconds or seconds
depending on the function called. But to call one of these, the test must first
call SetMockSleep (precondition enforced with assertion), which also turns
sleeps in RocksDB into mock sleeps. To also removes accounting for actual
clock time, call SetTimeElapseOnlySleepOnReopen, which implies
SetMockSleep (on DB re-open). This latter setting only works by applying
on DB re-open, otherwise havoc can ensue if Env goes back in time with
DB open.
More specifics:
Removed some unused test classes, and updated comments on the general
problem.
Fixed DBSSTTest.GetTotalSstFilesSize using a sync point callback instead
of mock time. For this we have the only modification to production code,
inserting a sync point callback in flush_job.cc, which is not a change to
production behavior.
Removed unnecessary resetting of mock times to 0 in many tests. RocksDB
deals in relative time. Any behaviors relying on absolute date/time are likely
a bug. (The above test DBSSTTest.GetTotalSstFilesSize was the only one
clearly injecting a specific absolute time for actual testing convenience.) Just
in case I misunderstood some test, I put this note in each replacement:
// NOTE: Presumed unnecessary and removed: resetting mock time in env
Strengthened some tests like MergeTestTime, MergeCompactionTimeTest, and
FilterCompactionTimeTest in db_test.cc
stats_history_test and blob_db_test are each their own beast, rather deeply
dependent on MockTimeEnv. Each gets its own variant of a work-around for
TimedWait in a mock time environment. (Reduces redundancy and
inconsistency in stats_history_test.)
Intended follow-up:
Remove TimedWait from the public API of InstrumentedCondVar, and only
make that accessible through Env by passing in an InstrumentedCondVar and
a deadline. Then the Env implementations mocking time can fix this problem
without using sync points. (Test infrastructure using sync points interferes
with individual tests' control over sync points.)
With that change, we can simplify/consolidate the scattered work-arounds.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7101
Test Plan: make check on Linux and MacOS
Reviewed By: zhichao-cao
Differential Revision: D23032815
Pulled By: pdillinger
fbshipit-source-id: 7f33967ada8b83011fb54e8279365c008bd6610b
4 years ago
|
|
|
MaybeInstallTimeElapseOnlySleep(db_opts);
|
|
|
|
return DB::Open(db_opts, dbname_, column_families, &handles_, &db_);
|
|
|
|
}
|
|
|
|
|
|
|
|
Status DBTestBase::TryReopenWithColumnFamilies(
|
|
|
|
const std::vector<std::string>& cfs, const Options& options) {
|
|
|
|
Close();
|
|
|
|
std::vector<Options> v_opts(cfs.size(), options);
|
|
|
|
return TryReopenWithColumnFamilies(cfs, v_opts);
|
|
|
|
}
|
|
|
|
|
|
|
|
void DBTestBase::Reopen(const Options& options) {
|
|
|
|
ASSERT_OK(TryReopen(options));
|
|
|
|
}
|
|
|
|
|
|
|
|
void DBTestBase::Close() {
|
|
|
|
for (auto h : handles_) {
|
|
|
|
EXPECT_OK(db_->DestroyColumnFamilyHandle(h));
|
|
|
|
}
|
|
|
|
handles_.clear();
|
|
|
|
delete db_;
|
|
|
|
db_ = nullptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
void DBTestBase::DestroyAndReopen(const Options& options) {
|
|
|
|
// Destroy using last options
|
|
|
|
Destroy(last_options_);
|
|
|
|
Reopen(options);
|
|
|
|
}
|
|
|
|
|
|
|
|
void DBTestBase::Destroy(const Options& options, bool delete_cf_paths) {
|
|
|
|
std::vector<ColumnFamilyDescriptor> column_families;
|
|
|
|
if (delete_cf_paths) {
|
|
|
|
for (size_t i = 0; i < handles_.size(); ++i) {
|
|
|
|
ColumnFamilyDescriptor cfdescriptor;
|
|
|
|
handles_[i]->GetDescriptor(&cfdescriptor).PermitUncheckedError();
|
|
|
|
column_families.push_back(cfdescriptor);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
Close();
|
|
|
|
ASSERT_OK(DestroyDB(dbname_, options, column_families));
|
|
|
|
}
|
|
|
|
|
|
|
|
Status DBTestBase::ReadOnlyReopen(const Options& options) {
|
Fix+clean up handling of mock sleeps (#7101)
Summary:
We have a number of tests hanging on MacOS and windows due to
mishandling of code for mock sleeps. In addition, the code was in
terrible shape because the same variable (addon_time_) would sometimes
refer to microseconds and sometimes to seconds. One test even assumed it
was nanoseconds but was written to pass anyway.
This has been cleaned up so that DB tests generally use a SpecialEnv
function to mock sleep, for either some number of microseconds or seconds
depending on the function called. But to call one of these, the test must first
call SetMockSleep (precondition enforced with assertion), which also turns
sleeps in RocksDB into mock sleeps. To also removes accounting for actual
clock time, call SetTimeElapseOnlySleepOnReopen, which implies
SetMockSleep (on DB re-open). This latter setting only works by applying
on DB re-open, otherwise havoc can ensue if Env goes back in time with
DB open.
More specifics:
Removed some unused test classes, and updated comments on the general
problem.
Fixed DBSSTTest.GetTotalSstFilesSize using a sync point callback instead
of mock time. For this we have the only modification to production code,
inserting a sync point callback in flush_job.cc, which is not a change to
production behavior.
Removed unnecessary resetting of mock times to 0 in many tests. RocksDB
deals in relative time. Any behaviors relying on absolute date/time are likely
a bug. (The above test DBSSTTest.GetTotalSstFilesSize was the only one
clearly injecting a specific absolute time for actual testing convenience.) Just
in case I misunderstood some test, I put this note in each replacement:
// NOTE: Presumed unnecessary and removed: resetting mock time in env
Strengthened some tests like MergeTestTime, MergeCompactionTimeTest, and
FilterCompactionTimeTest in db_test.cc
stats_history_test and blob_db_test are each their own beast, rather deeply
dependent on MockTimeEnv. Each gets its own variant of a work-around for
TimedWait in a mock time environment. (Reduces redundancy and
inconsistency in stats_history_test.)
Intended follow-up:
Remove TimedWait from the public API of InstrumentedCondVar, and only
make that accessible through Env by passing in an InstrumentedCondVar and
a deadline. Then the Env implementations mocking time can fix this problem
without using sync points. (Test infrastructure using sync points interferes
with individual tests' control over sync points.)
With that change, we can simplify/consolidate the scattered work-arounds.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7101
Test Plan: make check on Linux and MacOS
Reviewed By: zhichao-cao
Differential Revision: D23032815
Pulled By: pdillinger
fbshipit-source-id: 7f33967ada8b83011fb54e8279365c008bd6610b
4 years ago
|
|
|
MaybeInstallTimeElapseOnlySleep(options);
|
|
|
|
return DB::OpenForReadOnly(options, dbname_, &db_);
|
|
|
|
}
|
|
|
|
|
|
|
|
Status DBTestBase::TryReopen(const Options& options) {
|
|
|
|
Close();
|
|
|
|
last_options_.table_factory.reset();
|
|
|
|
// Note: operator= is an unsafe approach here since it destructs
|
|
|
|
// std::shared_ptr in the same order of their creation, in contrast to
|
|
|
|
// destructors which destructs them in the opposite order of creation. One
|
Fix many tests to run with MEM_ENV and ENCRYPTED_ENV; Introduce a MemoryFileSystem class (#7566)
Summary:
This PR does a few things:
1. The MockFileSystem class was split out from the MockEnv. This change would theoretically allow a MockFileSystem to be used by other Environments as well (if we created a means of constructing one). The MockFileSystem implements a FileSystem in its entirety and does not rely on any Wrapper implementation.
2. Make the RocksDB test suite work when MOCK_ENV=1 and ENCRYPTED_ENV=1 are set. To accomplish this, a few things were needed:
- The tests that tried to use the "wrong" environment (Env::Default() instead of env_) were updated
- The MockFileSystem was changed to support the features it was missing or mishandled (such as recursively deleting files in a directory or supporting renaming of a directory).
3. Updated the test framework to have a ROCKSDB_GTEST_SKIP macro. This can be used to flag tests that are skipped. Currently, this defaults to doing nothing (marks the test as SUCCESS) but will mark the tests as SKIPPED when RocksDB is upgraded to a version of gtest that supports this (gtest-1.10).
I have run a full "make check" with MEM_ENV, ENCRYPTED_ENV, both, and neither under both MacOS and RedHat. A few tests were disabled/skipped for the MEM/ENCRYPTED cases. The error_handler_fs_test fails/hangs for MEM_ENV (presumably a timing problem) and I will introduce another PR/issue to track that problem. (I will also push a change to disable those tests soon). There is one more test in DBTest2 that also fails which I need to investigate or skip before this PR is merged.
Theoretically, this PR should also allow the test suite to run against an Env loaded from the registry, though I do not have one to try it with currently.
Finally, once this is accepted, it would be nice if there was a CircleCI job to run these tests on a checkin so this effort does not become stale. I do not know how to do that, so if someone could write that job, it would be appreciated :)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7566
Reviewed By: zhichao-cao
Differential Revision: D24408980
Pulled By: jay-zhuang
fbshipit-source-id: 911b1554a4d0da06fd51feca0c090a4abdcb4a5f
4 years ago
|
|
|
// particular problem is that the cache destructor might invoke callback
|
|
|
|
// functions that use Option members such as statistics. To work around this
|
Fix many tests to run with MEM_ENV and ENCRYPTED_ENV; Introduce a MemoryFileSystem class (#7566)
Summary:
This PR does a few things:
1. The MockFileSystem class was split out from the MockEnv. This change would theoretically allow a MockFileSystem to be used by other Environments as well (if we created a means of constructing one). The MockFileSystem implements a FileSystem in its entirety and does not rely on any Wrapper implementation.
2. Make the RocksDB test suite work when MOCK_ENV=1 and ENCRYPTED_ENV=1 are set. To accomplish this, a few things were needed:
- The tests that tried to use the "wrong" environment (Env::Default() instead of env_) were updated
- The MockFileSystem was changed to support the features it was missing or mishandled (such as recursively deleting files in a directory or supporting renaming of a directory).
3. Updated the test framework to have a ROCKSDB_GTEST_SKIP macro. This can be used to flag tests that are skipped. Currently, this defaults to doing nothing (marks the test as SUCCESS) but will mark the tests as SKIPPED when RocksDB is upgraded to a version of gtest that supports this (gtest-1.10).
I have run a full "make check" with MEM_ENV, ENCRYPTED_ENV, both, and neither under both MacOS and RedHat. A few tests were disabled/skipped for the MEM/ENCRYPTED cases. The error_handler_fs_test fails/hangs for MEM_ENV (presumably a timing problem) and I will introduce another PR/issue to track that problem. (I will also push a change to disable those tests soon). There is one more test in DBTest2 that also fails which I need to investigate or skip before this PR is merged.
Theoretically, this PR should also allow the test suite to run against an Env loaded from the registry, though I do not have one to try it with currently.
Finally, once this is accepted, it would be nice if there was a CircleCI job to run these tests on a checkin so this effort does not become stale. I do not know how to do that, so if someone could write that job, it would be appreciated :)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7566
Reviewed By: zhichao-cao
Differential Revision: D24408980
Pulled By: jay-zhuang
fbshipit-source-id: 911b1554a4d0da06fd51feca0c090a4abdcb4a5f
4 years ago
|
|
|
// problem, we manually call destructor of table_factory which eventually
|
|
|
|
// clears the block cache.
|
|
|
|
last_options_ = options;
|
Fix+clean up handling of mock sleeps (#7101)
Summary:
We have a number of tests hanging on MacOS and windows due to
mishandling of code for mock sleeps. In addition, the code was in
terrible shape because the same variable (addon_time_) would sometimes
refer to microseconds and sometimes to seconds. One test even assumed it
was nanoseconds but was written to pass anyway.
This has been cleaned up so that DB tests generally use a SpecialEnv
function to mock sleep, for either some number of microseconds or seconds
depending on the function called. But to call one of these, the test must first
call SetMockSleep (precondition enforced with assertion), which also turns
sleeps in RocksDB into mock sleeps. To also removes accounting for actual
clock time, call SetTimeElapseOnlySleepOnReopen, which implies
SetMockSleep (on DB re-open). This latter setting only works by applying
on DB re-open, otherwise havoc can ensue if Env goes back in time with
DB open.
More specifics:
Removed some unused test classes, and updated comments on the general
problem.
Fixed DBSSTTest.GetTotalSstFilesSize using a sync point callback instead
of mock time. For this we have the only modification to production code,
inserting a sync point callback in flush_job.cc, which is not a change to
production behavior.
Removed unnecessary resetting of mock times to 0 in many tests. RocksDB
deals in relative time. Any behaviors relying on absolute date/time are likely
a bug. (The above test DBSSTTest.GetTotalSstFilesSize was the only one
clearly injecting a specific absolute time for actual testing convenience.) Just
in case I misunderstood some test, I put this note in each replacement:
// NOTE: Presumed unnecessary and removed: resetting mock time in env
Strengthened some tests like MergeTestTime, MergeCompactionTimeTest, and
FilterCompactionTimeTest in db_test.cc
stats_history_test and blob_db_test are each their own beast, rather deeply
dependent on MockTimeEnv. Each gets its own variant of a work-around for
TimedWait in a mock time environment. (Reduces redundancy and
inconsistency in stats_history_test.)
Intended follow-up:
Remove TimedWait from the public API of InstrumentedCondVar, and only
make that accessible through Env by passing in an InstrumentedCondVar and
a deadline. Then the Env implementations mocking time can fix this problem
without using sync points. (Test infrastructure using sync points interferes
with individual tests' control over sync points.)
With that change, we can simplify/consolidate the scattered work-arounds.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/7101
Test Plan: make check on Linux and MacOS
Reviewed By: zhichao-cao
Differential Revision: D23032815
Pulled By: pdillinger
fbshipit-source-id: 7f33967ada8b83011fb54e8279365c008bd6610b
4 years ago
|
|
|
MaybeInstallTimeElapseOnlySleep(options);
|
|
|
|
return DB::Open(options, dbname_, &db_);
|
|
|
|
}
|
|
|
|
|
|
|
|
bool DBTestBase::IsDirectIOSupported() {
|
|
|
|
return test::IsDirectIOSupported(env_, dbname_);
|
|
|
|
}
|
|
|
|
|
|
|
|
bool DBTestBase::IsMemoryMappedAccessSupported() const {
|
|
|
|
return (!encrypted_env_);
|
|
|
|
}
|
|
|
|
|
|
|
|
Status DBTestBase::Flush(int cf) {
|
|
|
|
if (cf == 0) {
|
|
|
|
return db_->Flush(FlushOptions());
|
|
|
|
} else {
|
|
|
|
return db_->Flush(FlushOptions(), handles_[cf]);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
Status DBTestBase::Flush(const std::vector<int>& cf_ids) {
|
|
|
|
std::vector<ColumnFamilyHandle*> cfhs;
|
|
|
|
std::for_each(cf_ids.begin(), cf_ids.end(),
|
|
|
|
[&cfhs, this](int id) { cfhs.emplace_back(handles_[id]); });
|
|
|
|
return db_->Flush(FlushOptions(), cfhs);
|
|
|
|
}
|
|
|
|
|
|
|
|
Status DBTestBase::Put(const Slice& k, const Slice& v, WriteOptions wo) {
|
|
|
|
if (kMergePut == option_config_) {
|
|
|
|
return db_->Merge(wo, k, v);
|
|
|
|
} else {
|
|
|
|
return db_->Put(wo, k, v);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
Status DBTestBase::Put(int cf, const Slice& k, const Slice& v,
|
|
|
|
WriteOptions wo) {
|
|
|
|
if (kMergePut == option_config_) {
|
|
|
|
return db_->Merge(wo, handles_[cf], k, v);
|
|
|
|
} else {
|
|
|
|
return db_->Put(wo, handles_[cf], k, v);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
Status DBTestBase::Merge(const Slice& k, const Slice& v, WriteOptions wo) {
|
|
|
|
return db_->Merge(wo, k, v);
|
|
|
|
}
|
|
|
|
|
|
|
|
Status DBTestBase::Merge(int cf, const Slice& k, const Slice& v,
|
|
|
|
WriteOptions wo) {
|
|
|
|
return db_->Merge(wo, handles_[cf], k, v);
|
|
|
|
}
|
|
|
|
|
|
|
|
Status DBTestBase::Delete(const std::string& k) {
|
|
|
|
return db_->Delete(WriteOptions(), k);
|
|
|
|
}
|
|
|
|
|
|
|
|
Status DBTestBase::Delete(int cf, const std::string& k) {
|
|
|
|
return db_->Delete(WriteOptions(), handles_[cf], k);
|
|
|
|
}
|
|
|
|
|
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 DBTestBase::SingleDelete(const std::string& k) {
|
|
|
|
return db_->SingleDelete(WriteOptions(), k);
|
|
|
|
}
|
|
|
|
|
|
|
|
Status DBTestBase::SingleDelete(int cf, const std::string& k) {
|
|
|
|
return db_->SingleDelete(WriteOptions(), handles_[cf], k);
|
|
|
|
}
|
|
|
|
|
|
|
|
std::string DBTestBase::Get(const std::string& k, const Snapshot* snapshot) {
|
|
|
|
ReadOptions options;
|
|
|
|
options.verify_checksums = true;
|
|
|
|
options.snapshot = snapshot;
|
|
|
|
std::string result;
|
|
|
|
Status s = db_->Get(options, k, &result);
|
|
|
|
if (s.IsNotFound()) {
|
|
|
|
result = "NOT_FOUND";
|
|
|
|
} else if (!s.ok()) {
|
|
|
|
result = s.ToString();
|
|
|
|
}
|
|
|
|
return result;
|
|
|
|
}
|
|
|
|
|
|
|
|
std::string DBTestBase::Get(int cf, const std::string& k,
|
|
|
|
const Snapshot* snapshot) {
|
|
|
|
ReadOptions options;
|
|
|
|
options.verify_checksums = true;
|
|
|
|
options.snapshot = snapshot;
|
|
|
|
std::string result;
|
|
|
|
Status s = db_->Get(options, handles_[cf], k, &result);
|
|
|
|
if (s.IsNotFound()) {
|
|
|
|
result = "NOT_FOUND";
|
|
|
|
} else if (!s.ok()) {
|
|
|
|
result = s.ToString();
|
|
|
|
}
|
|
|
|
return result;
|
|
|
|
}
|
|
|
|
|
|
|
|
std::vector<std::string> DBTestBase::MultiGet(std::vector<int> cfs,
|
|
|
|
const std::vector<std::string>& k,
|
|
|
|
const Snapshot* snapshot,
|
Multi file concurrency in MultiGet using coroutines and async IO (#9968)
Summary:
This PR implements a coroutine version of batched MultiGet in order to concurrently read from multiple SST files in a level using async IO, thus reducing the latency of the MultiGet. The API from the user perspective is still synchronous and single threaded, with the RocksDB part of the processing happening in the context of the caller's thread. In Version::MultiGet, the decision is made whether to call synchronous or coroutine code.
A good way to review this PR is to review the first 4 commits in order - de773b3, 70c2f70, 10b50e1, and 377a597 - before reviewing the rest.
TODO:
1. Figure out how to build it in CircleCI (requires some dependencies to be installed)
2. Do some stress testing with coroutines enabled
No regression in synchronous MultiGet between this branch and main -
```
./db_bench -use_existing_db=true --db=/data/mysql/rocksdb/prefix_scan -benchmarks="readseq,multireadrandom" -key_size=32 -value_size=512 -num=5000000 -batch_size=64 -multiread_batched=true -use_direct_reads=false -duration=60 -ops_between_duration_checks=1 -readonly=true -adaptive_readahead=true -threads=16 -cache_size=10485760000 -async_io=false -multiread_stride=40000 -statistics
```
Branch - ```multireadrandom : 4.025 micros/op 3975111 ops/sec 60.001 seconds 238509056 operations; 2062.3 MB/s (14767808 of 14767808 found)```
Main - ```multireadrandom : 3.987 micros/op 4013216 ops/sec 60.001 seconds 240795392 operations; 2082.1 MB/s (15231040 of 15231040 found)```
More benchmarks in various scenarios are given below. The measurements were taken with ```async_io=false``` (no coroutines) and ```async_io=true``` (use coroutines). For an IO bound workload (with every key requiring an IO), the coroutines version shows a clear benefit, being ~2.6X faster. For CPU bound workloads, the coroutines version has ~6-15% higher CPU utilization, depending on how many keys overlap an SST file.
1. Single thread IO bound workload on remote storage with sparse MultiGet batch keys (~1 key overlap/file) -
No coroutines - ```multireadrandom : 831.774 micros/op 1202 ops/sec 60.001 seconds 72136 operations; 0.6 MB/s (72136 of 72136 found)```
Using coroutines - ```multireadrandom : 318.742 micros/op 3137 ops/sec 60.003 seconds 188248 operations; 1.6 MB/s (188248 of 188248 found)```
2. Single thread CPU bound workload (all data cached) with ~1 key overlap/file -
No coroutines - ```multireadrandom : 4.127 micros/op 242322 ops/sec 60.000 seconds 14539384 operations; 125.7 MB/s (14539384 of 14539384 found)```
Using coroutines - ```multireadrandom : 4.741 micros/op 210935 ops/sec 60.000 seconds 12656176 operations; 109.4 MB/s (12656176 of 12656176 found)```
3. Single thread CPU bound workload with ~2 key overlap/file -
No coroutines - ```multireadrandom : 3.717 micros/op 269000 ops/sec 60.000 seconds 16140024 operations; 139.6 MB/s (16140024 of 16140024 found)```
Using coroutines - ```multireadrandom : 4.146 micros/op 241204 ops/sec 60.000 seconds 14472296 operations; 125.1 MB/s (14472296 of 14472296 found)```
4. CPU bound multi-threaded (16 threads) with ~4 key overlap/file -
No coroutines - ```multireadrandom : 4.534 micros/op 3528792 ops/sec 60.000 seconds 211728728 operations; 1830.7 MB/s (12737024 of 12737024 found) ```
Using coroutines - ```multireadrandom : 4.872 micros/op 3283812 ops/sec 60.000 seconds 197030096 operations; 1703.6 MB/s (12548032 of 12548032 found) ```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9968
Reviewed By: akankshamahajan15
Differential Revision: D36348563
Pulled By: anand1976
fbshipit-source-id: c0ce85a505fd26ebfbb09786cbd7f25202038696
3 years ago
|
|
|
const bool batched,
|
|
|
|
const bool async) {
|
|
|
|
ReadOptions options;
|
|
|
|
options.verify_checksums = true;
|
|
|
|
options.snapshot = snapshot;
|
Multi file concurrency in MultiGet using coroutines and async IO (#9968)
Summary:
This PR implements a coroutine version of batched MultiGet in order to concurrently read from multiple SST files in a level using async IO, thus reducing the latency of the MultiGet. The API from the user perspective is still synchronous and single threaded, with the RocksDB part of the processing happening in the context of the caller's thread. In Version::MultiGet, the decision is made whether to call synchronous or coroutine code.
A good way to review this PR is to review the first 4 commits in order - de773b3, 70c2f70, 10b50e1, and 377a597 - before reviewing the rest.
TODO:
1. Figure out how to build it in CircleCI (requires some dependencies to be installed)
2. Do some stress testing with coroutines enabled
No regression in synchronous MultiGet between this branch and main -
```
./db_bench -use_existing_db=true --db=/data/mysql/rocksdb/prefix_scan -benchmarks="readseq,multireadrandom" -key_size=32 -value_size=512 -num=5000000 -batch_size=64 -multiread_batched=true -use_direct_reads=false -duration=60 -ops_between_duration_checks=1 -readonly=true -adaptive_readahead=true -threads=16 -cache_size=10485760000 -async_io=false -multiread_stride=40000 -statistics
```
Branch - ```multireadrandom : 4.025 micros/op 3975111 ops/sec 60.001 seconds 238509056 operations; 2062.3 MB/s (14767808 of 14767808 found)```
Main - ```multireadrandom : 3.987 micros/op 4013216 ops/sec 60.001 seconds 240795392 operations; 2082.1 MB/s (15231040 of 15231040 found)```
More benchmarks in various scenarios are given below. The measurements were taken with ```async_io=false``` (no coroutines) and ```async_io=true``` (use coroutines). For an IO bound workload (with every key requiring an IO), the coroutines version shows a clear benefit, being ~2.6X faster. For CPU bound workloads, the coroutines version has ~6-15% higher CPU utilization, depending on how many keys overlap an SST file.
1. Single thread IO bound workload on remote storage with sparse MultiGet batch keys (~1 key overlap/file) -
No coroutines - ```multireadrandom : 831.774 micros/op 1202 ops/sec 60.001 seconds 72136 operations; 0.6 MB/s (72136 of 72136 found)```
Using coroutines - ```multireadrandom : 318.742 micros/op 3137 ops/sec 60.003 seconds 188248 operations; 1.6 MB/s (188248 of 188248 found)```
2. Single thread CPU bound workload (all data cached) with ~1 key overlap/file -
No coroutines - ```multireadrandom : 4.127 micros/op 242322 ops/sec 60.000 seconds 14539384 operations; 125.7 MB/s (14539384 of 14539384 found)```
Using coroutines - ```multireadrandom : 4.741 micros/op 210935 ops/sec 60.000 seconds 12656176 operations; 109.4 MB/s (12656176 of 12656176 found)```
3. Single thread CPU bound workload with ~2 key overlap/file -
No coroutines - ```multireadrandom : 3.717 micros/op 269000 ops/sec 60.000 seconds 16140024 operations; 139.6 MB/s (16140024 of 16140024 found)```
Using coroutines - ```multireadrandom : 4.146 micros/op 241204 ops/sec 60.000 seconds 14472296 operations; 125.1 MB/s (14472296 of 14472296 found)```
4. CPU bound multi-threaded (16 threads) with ~4 key overlap/file -
No coroutines - ```multireadrandom : 4.534 micros/op 3528792 ops/sec 60.000 seconds 211728728 operations; 1830.7 MB/s (12737024 of 12737024 found) ```
Using coroutines - ```multireadrandom : 4.872 micros/op 3283812 ops/sec 60.000 seconds 197030096 operations; 1703.6 MB/s (12548032 of 12548032 found) ```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9968
Reviewed By: akankshamahajan15
Differential Revision: D36348563
Pulled By: anand1976
fbshipit-source-id: c0ce85a505fd26ebfbb09786cbd7f25202038696
3 years ago
|
|
|
options.async_io = async;
|
|
|
|
std::vector<ColumnFamilyHandle*> handles;
|
|
|
|
std::vector<Slice> keys;
|
|
|
|
std::vector<std::string> result;
|
|
|
|
|
|
|
|
for (unsigned int i = 0; i < cfs.size(); ++i) {
|
|
|
|
handles.push_back(handles_[cfs[i]]);
|
|
|
|
keys.push_back(k[i]);
|
|
|
|
}
|
|
|
|
std::vector<Status> s;
|
|
|
|
if (!batched) {
|
|
|
|
s = db_->MultiGet(options, handles, keys, &result);
|
Eliminate unnecessary (slow) block cache Ref()ing in MultiGet (#9899)
Summary:
When MultiGet() determines that multiple query keys can be
served by examining the same data block in block cache (one Lookup()),
each PinnableSlice referring to data in that data block needs to hold
on to the block in cache so that they can be released at arbitrary
times by the API user. Historically this is accomplished with extra
calls to Ref() on the Handle from Lookup(), with each PinnableSlice
cleanup calling Release() on the Handle, but this creates extra
contention on the block cache for the extra Ref()s and Release()es,
especially because they hit the same cache shard repeatedly.
In the case of merge operands (possibly more cases?), the problem was
compounded by doing an extra Ref()+eventual Release() for each merge
operand for a key reusing a block (which could be the same key!), rather
than one Ref() per key. (Note: the non-shared case with `biter` was
already one per key.)
This change optimizes MultiGet not to rely on these extra, contentious
Ref()+Release() calls by instead, in the shared block case, wrapping
the cache Release() cleanup in a refcounted object referenced by the
PinnableSlices, such that after the last wrapped reference is released,
the cache entry is Release()ed. Relaxed atomic refcounts should be
much faster than mutex-guarded Ref() and Release(), and much less prone
to a performance cliff when MultiGet() does a lot of block sharing.
Note that I did not use std::shared_ptr, because that would require an
extra indirection object (shared_ptr itself new/delete) in order to
associate a ref increment/decrement with a Cleanable cleanup entry. (If
I assumed it was the size of two pointers, I could do some hackery to
make it work without the extra indirection, but that's too fragile.)
Some details:
* Fixed (removed) extra block cache tracing entries in cases of cache
entry reuse in MultiGet, but it's likely that in some other cases traces
are missing (XXX comment inserted)
* Moved existing implementations for cleanable.h from iterator.cc to
new cleanable.cc
* Improved API comments on Cleanable
* Added a public SharedCleanablePtr class to cleanable.h in case others
could benefit from the same pattern (potentially many Cleanables and/or
smart pointers referencing a shared Cleanable)
* Add a typedef for MultiGetContext::Mask
* Some variable renaming for clarity
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9899
Test Plan:
Added unit tests for SharedCleanablePtr.
Greatly enhanced ability of existing tests to detect cache use-after-free.
* Release PinnableSlices from MultiGet as they are read rather than in
bulk (in db_test_util wrapper).
* In ASAN build, default to using a trivially small LRUCache for block_cache
so that entries are immediately erased when unreferenced. (Updated two
tests that depend on caching.) New ASAN testsuite running time seems
OK to me.
If I introduce a bug into my implementation where we skip the shared
cleanups on block reuse, ASAN detects the bug in
`db_basic_test *MultiGet*`. If I remove either of the above testing
enhancements, the bug is not detected.
Consider for follow-up work: manipulate or randomize ordering of
PinnableSlice use and release from MultiGet db_test_util wrapper. But in
typical cases, natural ordering gives pretty good functional coverage.
Performance test:
In the extreme (but possible) case of MultiGetting the same or adjacent keys
in a batch, throughput can improve by an order of magnitude.
`./db_bench -benchmarks=multireadrandom -db=/dev/shm/testdb -readonly -num=5 -duration=10 -threads=20 -multiread_batched -batch_size=200`
Before ops/sec, num=5: 1,384,394
Before ops/sec, num=500: 6,423,720
After ops/sec, num=500: 10,658,794
After ops/sec, num=5: 16,027,257
Also note that previously, with high parallelism, having query keys
concentrated in a single block was worse than spreading them out a bit. Now
concentrated in a single block is faster than spread out, which is hopefully
consistent with natural expectation.
Random query performance: with num=1000000, over 999 x 10s runs running before & after simultaneously (each -threads=12):
Before: multireadrandom [AVG 999 runs] : 1088699 (± 7344) ops/sec; 120.4 (± 0.8 ) MB/sec
After: multireadrandom [AVG 999 runs] : 1090402 (± 7230) ops/sec; 120.6 (± 0.8 ) MB/sec
Possibly better, possibly in the noise.
Reviewed By: anand1976
Differential Revision: D35907003
Pulled By: pdillinger
fbshipit-source-id: bbd244d703649a8ca12d476f2d03853ed9d1a17e
3 years ago
|
|
|
for (size_t i = 0; i < s.size(); ++i) {
|
|
|
|
if (s[i].IsNotFound()) {
|
|
|
|
result[i] = "NOT_FOUND";
|
|
|
|
} else if (!s[i].ok()) {
|
|
|
|
result[i] = s[i].ToString();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
std::vector<PinnableSlice> pin_values(cfs.size());
|
|
|
|
result.resize(cfs.size());
|
|
|
|
s.resize(cfs.size());
|
|
|
|
db_->MultiGet(options, cfs.size(), handles.data(), keys.data(),
|
|
|
|
pin_values.data(), s.data());
|
Eliminate unnecessary (slow) block cache Ref()ing in MultiGet (#9899)
Summary:
When MultiGet() determines that multiple query keys can be
served by examining the same data block in block cache (one Lookup()),
each PinnableSlice referring to data in that data block needs to hold
on to the block in cache so that they can be released at arbitrary
times by the API user. Historically this is accomplished with extra
calls to Ref() on the Handle from Lookup(), with each PinnableSlice
cleanup calling Release() on the Handle, but this creates extra
contention on the block cache for the extra Ref()s and Release()es,
especially because they hit the same cache shard repeatedly.
In the case of merge operands (possibly more cases?), the problem was
compounded by doing an extra Ref()+eventual Release() for each merge
operand for a key reusing a block (which could be the same key!), rather
than one Ref() per key. (Note: the non-shared case with `biter` was
already one per key.)
This change optimizes MultiGet not to rely on these extra, contentious
Ref()+Release() calls by instead, in the shared block case, wrapping
the cache Release() cleanup in a refcounted object referenced by the
PinnableSlices, such that after the last wrapped reference is released,
the cache entry is Release()ed. Relaxed atomic refcounts should be
much faster than mutex-guarded Ref() and Release(), and much less prone
to a performance cliff when MultiGet() does a lot of block sharing.
Note that I did not use std::shared_ptr, because that would require an
extra indirection object (shared_ptr itself new/delete) in order to
associate a ref increment/decrement with a Cleanable cleanup entry. (If
I assumed it was the size of two pointers, I could do some hackery to
make it work without the extra indirection, but that's too fragile.)
Some details:
* Fixed (removed) extra block cache tracing entries in cases of cache
entry reuse in MultiGet, but it's likely that in some other cases traces
are missing (XXX comment inserted)
* Moved existing implementations for cleanable.h from iterator.cc to
new cleanable.cc
* Improved API comments on Cleanable
* Added a public SharedCleanablePtr class to cleanable.h in case others
could benefit from the same pattern (potentially many Cleanables and/or
smart pointers referencing a shared Cleanable)
* Add a typedef for MultiGetContext::Mask
* Some variable renaming for clarity
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9899
Test Plan:
Added unit tests for SharedCleanablePtr.
Greatly enhanced ability of existing tests to detect cache use-after-free.
* Release PinnableSlices from MultiGet as they are read rather than in
bulk (in db_test_util wrapper).
* In ASAN build, default to using a trivially small LRUCache for block_cache
so that entries are immediately erased when unreferenced. (Updated two
tests that depend on caching.) New ASAN testsuite running time seems
OK to me.
If I introduce a bug into my implementation where we skip the shared
cleanups on block reuse, ASAN detects the bug in
`db_basic_test *MultiGet*`. If I remove either of the above testing
enhancements, the bug is not detected.
Consider for follow-up work: manipulate or randomize ordering of
PinnableSlice use and release from MultiGet db_test_util wrapper. But in
typical cases, natural ordering gives pretty good functional coverage.
Performance test:
In the extreme (but possible) case of MultiGetting the same or adjacent keys
in a batch, throughput can improve by an order of magnitude.
`./db_bench -benchmarks=multireadrandom -db=/dev/shm/testdb -readonly -num=5 -duration=10 -threads=20 -multiread_batched -batch_size=200`
Before ops/sec, num=5: 1,384,394
Before ops/sec, num=500: 6,423,720
After ops/sec, num=500: 10,658,794
After ops/sec, num=5: 16,027,257
Also note that previously, with high parallelism, having query keys
concentrated in a single block was worse than spreading them out a bit. Now
concentrated in a single block is faster than spread out, which is hopefully
consistent with natural expectation.
Random query performance: with num=1000000, over 999 x 10s runs running before & after simultaneously (each -threads=12):
Before: multireadrandom [AVG 999 runs] : 1088699 (± 7344) ops/sec; 120.4 (± 0.8 ) MB/sec
After: multireadrandom [AVG 999 runs] : 1090402 (± 7230) ops/sec; 120.6 (± 0.8 ) MB/sec
Possibly better, possibly in the noise.
Reviewed By: anand1976
Differential Revision: D35907003
Pulled By: pdillinger
fbshipit-source-id: bbd244d703649a8ca12d476f2d03853ed9d1a17e
3 years ago
|
|
|
for (size_t i = 0; i < s.size(); ++i) {
|
|
|
|
if (s[i].IsNotFound()) {
|
|
|
|
result[i] = "NOT_FOUND";
|
|
|
|
} else if (!s[i].ok()) {
|
|
|
|
result[i] = s[i].ToString();
|
|
|
|
} else {
|
|
|
|
result[i].assign(pin_values[i].data(), pin_values[i].size());
|
Eliminate unnecessary (slow) block cache Ref()ing in MultiGet (#9899)
Summary:
When MultiGet() determines that multiple query keys can be
served by examining the same data block in block cache (one Lookup()),
each PinnableSlice referring to data in that data block needs to hold
on to the block in cache so that they can be released at arbitrary
times by the API user. Historically this is accomplished with extra
calls to Ref() on the Handle from Lookup(), with each PinnableSlice
cleanup calling Release() on the Handle, but this creates extra
contention on the block cache for the extra Ref()s and Release()es,
especially because they hit the same cache shard repeatedly.
In the case of merge operands (possibly more cases?), the problem was
compounded by doing an extra Ref()+eventual Release() for each merge
operand for a key reusing a block (which could be the same key!), rather
than one Ref() per key. (Note: the non-shared case with `biter` was
already one per key.)
This change optimizes MultiGet not to rely on these extra, contentious
Ref()+Release() calls by instead, in the shared block case, wrapping
the cache Release() cleanup in a refcounted object referenced by the
PinnableSlices, such that after the last wrapped reference is released,
the cache entry is Release()ed. Relaxed atomic refcounts should be
much faster than mutex-guarded Ref() and Release(), and much less prone
to a performance cliff when MultiGet() does a lot of block sharing.
Note that I did not use std::shared_ptr, because that would require an
extra indirection object (shared_ptr itself new/delete) in order to
associate a ref increment/decrement with a Cleanable cleanup entry. (If
I assumed it was the size of two pointers, I could do some hackery to
make it work without the extra indirection, but that's too fragile.)
Some details:
* Fixed (removed) extra block cache tracing entries in cases of cache
entry reuse in MultiGet, but it's likely that in some other cases traces
are missing (XXX comment inserted)
* Moved existing implementations for cleanable.h from iterator.cc to
new cleanable.cc
* Improved API comments on Cleanable
* Added a public SharedCleanablePtr class to cleanable.h in case others
could benefit from the same pattern (potentially many Cleanables and/or
smart pointers referencing a shared Cleanable)
* Add a typedef for MultiGetContext::Mask
* Some variable renaming for clarity
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9899
Test Plan:
Added unit tests for SharedCleanablePtr.
Greatly enhanced ability of existing tests to detect cache use-after-free.
* Release PinnableSlices from MultiGet as they are read rather than in
bulk (in db_test_util wrapper).
* In ASAN build, default to using a trivially small LRUCache for block_cache
so that entries are immediately erased when unreferenced. (Updated two
tests that depend on caching.) New ASAN testsuite running time seems
OK to me.
If I introduce a bug into my implementation where we skip the shared
cleanups on block reuse, ASAN detects the bug in
`db_basic_test *MultiGet*`. If I remove either of the above testing
enhancements, the bug is not detected.
Consider for follow-up work: manipulate or randomize ordering of
PinnableSlice use and release from MultiGet db_test_util wrapper. But in
typical cases, natural ordering gives pretty good functional coverage.
Performance test:
In the extreme (but possible) case of MultiGetting the same or adjacent keys
in a batch, throughput can improve by an order of magnitude.
`./db_bench -benchmarks=multireadrandom -db=/dev/shm/testdb -readonly -num=5 -duration=10 -threads=20 -multiread_batched -batch_size=200`
Before ops/sec, num=5: 1,384,394
Before ops/sec, num=500: 6,423,720
After ops/sec, num=500: 10,658,794
After ops/sec, num=5: 16,027,257
Also note that previously, with high parallelism, having query keys
concentrated in a single block was worse than spreading them out a bit. Now
concentrated in a single block is faster than spread out, which is hopefully
consistent with natural expectation.
Random query performance: with num=1000000, over 999 x 10s runs running before & after simultaneously (each -threads=12):
Before: multireadrandom [AVG 999 runs] : 1088699 (± 7344) ops/sec; 120.4 (± 0.8 ) MB/sec
After: multireadrandom [AVG 999 runs] : 1090402 (± 7230) ops/sec; 120.6 (± 0.8 ) MB/sec
Possibly better, possibly in the noise.
Reviewed By: anand1976
Differential Revision: D35907003
Pulled By: pdillinger
fbshipit-source-id: bbd244d703649a8ca12d476f2d03853ed9d1a17e
3 years ago
|
|
|
// Increase likelihood of detecting potential use-after-free bugs with
|
|
|
|
// PinnableSlices tracking the same resource
|
|
|
|
pin_values[i].Reset();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return result;
|
|
|
|
}
|
|
|
|
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
6 years ago
|
|
|
std::vector<std::string> DBTestBase::MultiGet(const std::vector<std::string>& k,
|
Multi file concurrency in MultiGet using coroutines and async IO (#9968)
Summary:
This PR implements a coroutine version of batched MultiGet in order to concurrently read from multiple SST files in a level using async IO, thus reducing the latency of the MultiGet. The API from the user perspective is still synchronous and single threaded, with the RocksDB part of the processing happening in the context of the caller's thread. In Version::MultiGet, the decision is made whether to call synchronous or coroutine code.
A good way to review this PR is to review the first 4 commits in order - de773b3, 70c2f70, 10b50e1, and 377a597 - before reviewing the rest.
TODO:
1. Figure out how to build it in CircleCI (requires some dependencies to be installed)
2. Do some stress testing with coroutines enabled
No regression in synchronous MultiGet between this branch and main -
```
./db_bench -use_existing_db=true --db=/data/mysql/rocksdb/prefix_scan -benchmarks="readseq,multireadrandom" -key_size=32 -value_size=512 -num=5000000 -batch_size=64 -multiread_batched=true -use_direct_reads=false -duration=60 -ops_between_duration_checks=1 -readonly=true -adaptive_readahead=true -threads=16 -cache_size=10485760000 -async_io=false -multiread_stride=40000 -statistics
```
Branch - ```multireadrandom : 4.025 micros/op 3975111 ops/sec 60.001 seconds 238509056 operations; 2062.3 MB/s (14767808 of 14767808 found)```
Main - ```multireadrandom : 3.987 micros/op 4013216 ops/sec 60.001 seconds 240795392 operations; 2082.1 MB/s (15231040 of 15231040 found)```
More benchmarks in various scenarios are given below. The measurements were taken with ```async_io=false``` (no coroutines) and ```async_io=true``` (use coroutines). For an IO bound workload (with every key requiring an IO), the coroutines version shows a clear benefit, being ~2.6X faster. For CPU bound workloads, the coroutines version has ~6-15% higher CPU utilization, depending on how many keys overlap an SST file.
1. Single thread IO bound workload on remote storage with sparse MultiGet batch keys (~1 key overlap/file) -
No coroutines - ```multireadrandom : 831.774 micros/op 1202 ops/sec 60.001 seconds 72136 operations; 0.6 MB/s (72136 of 72136 found)```
Using coroutines - ```multireadrandom : 318.742 micros/op 3137 ops/sec 60.003 seconds 188248 operations; 1.6 MB/s (188248 of 188248 found)```
2. Single thread CPU bound workload (all data cached) with ~1 key overlap/file -
No coroutines - ```multireadrandom : 4.127 micros/op 242322 ops/sec 60.000 seconds 14539384 operations; 125.7 MB/s (14539384 of 14539384 found)```
Using coroutines - ```multireadrandom : 4.741 micros/op 210935 ops/sec 60.000 seconds 12656176 operations; 109.4 MB/s (12656176 of 12656176 found)```
3. Single thread CPU bound workload with ~2 key overlap/file -
No coroutines - ```multireadrandom : 3.717 micros/op 269000 ops/sec 60.000 seconds 16140024 operations; 139.6 MB/s (16140024 of 16140024 found)```
Using coroutines - ```multireadrandom : 4.146 micros/op 241204 ops/sec 60.000 seconds 14472296 operations; 125.1 MB/s (14472296 of 14472296 found)```
4. CPU bound multi-threaded (16 threads) with ~4 key overlap/file -
No coroutines - ```multireadrandom : 4.534 micros/op 3528792 ops/sec 60.000 seconds 211728728 operations; 1830.7 MB/s (12737024 of 12737024 found) ```
Using coroutines - ```multireadrandom : 4.872 micros/op 3283812 ops/sec 60.000 seconds 197030096 operations; 1703.6 MB/s (12548032 of 12548032 found) ```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9968
Reviewed By: akankshamahajan15
Differential Revision: D36348563
Pulled By: anand1976
fbshipit-source-id: c0ce85a505fd26ebfbb09786cbd7f25202038696
3 years ago
|
|
|
const Snapshot* snapshot,
|
|
|
|
const bool async) {
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
6 years ago
|
|
|
ReadOptions options;
|
|
|
|
options.verify_checksums = true;
|
|
|
|
options.snapshot = snapshot;
|
Multi file concurrency in MultiGet using coroutines and async IO (#9968)
Summary:
This PR implements a coroutine version of batched MultiGet in order to concurrently read from multiple SST files in a level using async IO, thus reducing the latency of the MultiGet. The API from the user perspective is still synchronous and single threaded, with the RocksDB part of the processing happening in the context of the caller's thread. In Version::MultiGet, the decision is made whether to call synchronous or coroutine code.
A good way to review this PR is to review the first 4 commits in order - de773b3, 70c2f70, 10b50e1, and 377a597 - before reviewing the rest.
TODO:
1. Figure out how to build it in CircleCI (requires some dependencies to be installed)
2. Do some stress testing with coroutines enabled
No regression in synchronous MultiGet between this branch and main -
```
./db_bench -use_existing_db=true --db=/data/mysql/rocksdb/prefix_scan -benchmarks="readseq,multireadrandom" -key_size=32 -value_size=512 -num=5000000 -batch_size=64 -multiread_batched=true -use_direct_reads=false -duration=60 -ops_between_duration_checks=1 -readonly=true -adaptive_readahead=true -threads=16 -cache_size=10485760000 -async_io=false -multiread_stride=40000 -statistics
```
Branch - ```multireadrandom : 4.025 micros/op 3975111 ops/sec 60.001 seconds 238509056 operations; 2062.3 MB/s (14767808 of 14767808 found)```
Main - ```multireadrandom : 3.987 micros/op 4013216 ops/sec 60.001 seconds 240795392 operations; 2082.1 MB/s (15231040 of 15231040 found)```
More benchmarks in various scenarios are given below. The measurements were taken with ```async_io=false``` (no coroutines) and ```async_io=true``` (use coroutines). For an IO bound workload (with every key requiring an IO), the coroutines version shows a clear benefit, being ~2.6X faster. For CPU bound workloads, the coroutines version has ~6-15% higher CPU utilization, depending on how many keys overlap an SST file.
1. Single thread IO bound workload on remote storage with sparse MultiGet batch keys (~1 key overlap/file) -
No coroutines - ```multireadrandom : 831.774 micros/op 1202 ops/sec 60.001 seconds 72136 operations; 0.6 MB/s (72136 of 72136 found)```
Using coroutines - ```multireadrandom : 318.742 micros/op 3137 ops/sec 60.003 seconds 188248 operations; 1.6 MB/s (188248 of 188248 found)```
2. Single thread CPU bound workload (all data cached) with ~1 key overlap/file -
No coroutines - ```multireadrandom : 4.127 micros/op 242322 ops/sec 60.000 seconds 14539384 operations; 125.7 MB/s (14539384 of 14539384 found)```
Using coroutines - ```multireadrandom : 4.741 micros/op 210935 ops/sec 60.000 seconds 12656176 operations; 109.4 MB/s (12656176 of 12656176 found)```
3. Single thread CPU bound workload with ~2 key overlap/file -
No coroutines - ```multireadrandom : 3.717 micros/op 269000 ops/sec 60.000 seconds 16140024 operations; 139.6 MB/s (16140024 of 16140024 found)```
Using coroutines - ```multireadrandom : 4.146 micros/op 241204 ops/sec 60.000 seconds 14472296 operations; 125.1 MB/s (14472296 of 14472296 found)```
4. CPU bound multi-threaded (16 threads) with ~4 key overlap/file -
No coroutines - ```multireadrandom : 4.534 micros/op 3528792 ops/sec 60.000 seconds 211728728 operations; 1830.7 MB/s (12737024 of 12737024 found) ```
Using coroutines - ```multireadrandom : 4.872 micros/op 3283812 ops/sec 60.000 seconds 197030096 operations; 1703.6 MB/s (12548032 of 12548032 found) ```
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9968
Reviewed By: akankshamahajan15
Differential Revision: D36348563
Pulled By: anand1976
fbshipit-source-id: c0ce85a505fd26ebfbb09786cbd7f25202038696
3 years ago
|
|
|
options.async_io = async;
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
6 years ago
|
|
|
std::vector<Slice> keys;
|
Eliminate unnecessary (slow) block cache Ref()ing in MultiGet (#9899)
Summary:
When MultiGet() determines that multiple query keys can be
served by examining the same data block in block cache (one Lookup()),
each PinnableSlice referring to data in that data block needs to hold
on to the block in cache so that they can be released at arbitrary
times by the API user. Historically this is accomplished with extra
calls to Ref() on the Handle from Lookup(), with each PinnableSlice
cleanup calling Release() on the Handle, but this creates extra
contention on the block cache for the extra Ref()s and Release()es,
especially because they hit the same cache shard repeatedly.
In the case of merge operands (possibly more cases?), the problem was
compounded by doing an extra Ref()+eventual Release() for each merge
operand for a key reusing a block (which could be the same key!), rather
than one Ref() per key. (Note: the non-shared case with `biter` was
already one per key.)
This change optimizes MultiGet not to rely on these extra, contentious
Ref()+Release() calls by instead, in the shared block case, wrapping
the cache Release() cleanup in a refcounted object referenced by the
PinnableSlices, such that after the last wrapped reference is released,
the cache entry is Release()ed. Relaxed atomic refcounts should be
much faster than mutex-guarded Ref() and Release(), and much less prone
to a performance cliff when MultiGet() does a lot of block sharing.
Note that I did not use std::shared_ptr, because that would require an
extra indirection object (shared_ptr itself new/delete) in order to
associate a ref increment/decrement with a Cleanable cleanup entry. (If
I assumed it was the size of two pointers, I could do some hackery to
make it work without the extra indirection, but that's too fragile.)
Some details:
* Fixed (removed) extra block cache tracing entries in cases of cache
entry reuse in MultiGet, but it's likely that in some other cases traces
are missing (XXX comment inserted)
* Moved existing implementations for cleanable.h from iterator.cc to
new cleanable.cc
* Improved API comments on Cleanable
* Added a public SharedCleanablePtr class to cleanable.h in case others
could benefit from the same pattern (potentially many Cleanables and/or
smart pointers referencing a shared Cleanable)
* Add a typedef for MultiGetContext::Mask
* Some variable renaming for clarity
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9899
Test Plan:
Added unit tests for SharedCleanablePtr.
Greatly enhanced ability of existing tests to detect cache use-after-free.
* Release PinnableSlices from MultiGet as they are read rather than in
bulk (in db_test_util wrapper).
* In ASAN build, default to using a trivially small LRUCache for block_cache
so that entries are immediately erased when unreferenced. (Updated two
tests that depend on caching.) New ASAN testsuite running time seems
OK to me.
If I introduce a bug into my implementation where we skip the shared
cleanups on block reuse, ASAN detects the bug in
`db_basic_test *MultiGet*`. If I remove either of the above testing
enhancements, the bug is not detected.
Consider for follow-up work: manipulate or randomize ordering of
PinnableSlice use and release from MultiGet db_test_util wrapper. But in
typical cases, natural ordering gives pretty good functional coverage.
Performance test:
In the extreme (but possible) case of MultiGetting the same or adjacent keys
in a batch, throughput can improve by an order of magnitude.
`./db_bench -benchmarks=multireadrandom -db=/dev/shm/testdb -readonly -num=5 -duration=10 -threads=20 -multiread_batched -batch_size=200`
Before ops/sec, num=5: 1,384,394
Before ops/sec, num=500: 6,423,720
After ops/sec, num=500: 10,658,794
After ops/sec, num=5: 16,027,257
Also note that previously, with high parallelism, having query keys
concentrated in a single block was worse than spreading them out a bit. Now
concentrated in a single block is faster than spread out, which is hopefully
consistent with natural expectation.
Random query performance: with num=1000000, over 999 x 10s runs running before & after simultaneously (each -threads=12):
Before: multireadrandom [AVG 999 runs] : 1088699 (± 7344) ops/sec; 120.4 (± 0.8 ) MB/sec
After: multireadrandom [AVG 999 runs] : 1090402 (± 7230) ops/sec; 120.6 (± 0.8 ) MB/sec
Possibly better, possibly in the noise.
Reviewed By: anand1976
Differential Revision: D35907003
Pulled By: pdillinger
fbshipit-source-id: bbd244d703649a8ca12d476f2d03853ed9d1a17e
3 years ago
|
|
|
std::vector<std::string> result(k.size());
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
6 years ago
|
|
|
std::vector<Status> statuses(k.size());
|
|
|
|
std::vector<PinnableSlice> pin_values(k.size());
|
|
|
|
|
Eliminate unnecessary (slow) block cache Ref()ing in MultiGet (#9899)
Summary:
When MultiGet() determines that multiple query keys can be
served by examining the same data block in block cache (one Lookup()),
each PinnableSlice referring to data in that data block needs to hold
on to the block in cache so that they can be released at arbitrary
times by the API user. Historically this is accomplished with extra
calls to Ref() on the Handle from Lookup(), with each PinnableSlice
cleanup calling Release() on the Handle, but this creates extra
contention on the block cache for the extra Ref()s and Release()es,
especially because they hit the same cache shard repeatedly.
In the case of merge operands (possibly more cases?), the problem was
compounded by doing an extra Ref()+eventual Release() for each merge
operand for a key reusing a block (which could be the same key!), rather
than one Ref() per key. (Note: the non-shared case with `biter` was
already one per key.)
This change optimizes MultiGet not to rely on these extra, contentious
Ref()+Release() calls by instead, in the shared block case, wrapping
the cache Release() cleanup in a refcounted object referenced by the
PinnableSlices, such that after the last wrapped reference is released,
the cache entry is Release()ed. Relaxed atomic refcounts should be
much faster than mutex-guarded Ref() and Release(), and much less prone
to a performance cliff when MultiGet() does a lot of block sharing.
Note that I did not use std::shared_ptr, because that would require an
extra indirection object (shared_ptr itself new/delete) in order to
associate a ref increment/decrement with a Cleanable cleanup entry. (If
I assumed it was the size of two pointers, I could do some hackery to
make it work without the extra indirection, but that's too fragile.)
Some details:
* Fixed (removed) extra block cache tracing entries in cases of cache
entry reuse in MultiGet, but it's likely that in some other cases traces
are missing (XXX comment inserted)
* Moved existing implementations for cleanable.h from iterator.cc to
new cleanable.cc
* Improved API comments on Cleanable
* Added a public SharedCleanablePtr class to cleanable.h in case others
could benefit from the same pattern (potentially many Cleanables and/or
smart pointers referencing a shared Cleanable)
* Add a typedef for MultiGetContext::Mask
* Some variable renaming for clarity
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9899
Test Plan:
Added unit tests for SharedCleanablePtr.
Greatly enhanced ability of existing tests to detect cache use-after-free.
* Release PinnableSlices from MultiGet as they are read rather than in
bulk (in db_test_util wrapper).
* In ASAN build, default to using a trivially small LRUCache for block_cache
so that entries are immediately erased when unreferenced. (Updated two
tests that depend on caching.) New ASAN testsuite running time seems
OK to me.
If I introduce a bug into my implementation where we skip the shared
cleanups on block reuse, ASAN detects the bug in
`db_basic_test *MultiGet*`. If I remove either of the above testing
enhancements, the bug is not detected.
Consider for follow-up work: manipulate or randomize ordering of
PinnableSlice use and release from MultiGet db_test_util wrapper. But in
typical cases, natural ordering gives pretty good functional coverage.
Performance test:
In the extreme (but possible) case of MultiGetting the same or adjacent keys
in a batch, throughput can improve by an order of magnitude.
`./db_bench -benchmarks=multireadrandom -db=/dev/shm/testdb -readonly -num=5 -duration=10 -threads=20 -multiread_batched -batch_size=200`
Before ops/sec, num=5: 1,384,394
Before ops/sec, num=500: 6,423,720
After ops/sec, num=500: 10,658,794
After ops/sec, num=5: 16,027,257
Also note that previously, with high parallelism, having query keys
concentrated in a single block was worse than spreading them out a bit. Now
concentrated in a single block is faster than spread out, which is hopefully
consistent with natural expectation.
Random query performance: with num=1000000, over 999 x 10s runs running before & after simultaneously (each -threads=12):
Before: multireadrandom [AVG 999 runs] : 1088699 (± 7344) ops/sec; 120.4 (± 0.8 ) MB/sec
After: multireadrandom [AVG 999 runs] : 1090402 (± 7230) ops/sec; 120.6 (± 0.8 ) MB/sec
Possibly better, possibly in the noise.
Reviewed By: anand1976
Differential Revision: D35907003
Pulled By: pdillinger
fbshipit-source-id: bbd244d703649a8ca12d476f2d03853ed9d1a17e
3 years ago
|
|
|
for (size_t i = 0; i < k.size(); ++i) {
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
6 years ago
|
|
|
keys.push_back(k[i]);
|
|
|
|
}
|
|
|
|
db_->MultiGet(options, dbfull()->DefaultColumnFamily(), keys.size(),
|
|
|
|
keys.data(), pin_values.data(), statuses.data());
|
Eliminate unnecessary (slow) block cache Ref()ing in MultiGet (#9899)
Summary:
When MultiGet() determines that multiple query keys can be
served by examining the same data block in block cache (one Lookup()),
each PinnableSlice referring to data in that data block needs to hold
on to the block in cache so that they can be released at arbitrary
times by the API user. Historically this is accomplished with extra
calls to Ref() on the Handle from Lookup(), with each PinnableSlice
cleanup calling Release() on the Handle, but this creates extra
contention on the block cache for the extra Ref()s and Release()es,
especially because they hit the same cache shard repeatedly.
In the case of merge operands (possibly more cases?), the problem was
compounded by doing an extra Ref()+eventual Release() for each merge
operand for a key reusing a block (which could be the same key!), rather
than one Ref() per key. (Note: the non-shared case with `biter` was
already one per key.)
This change optimizes MultiGet not to rely on these extra, contentious
Ref()+Release() calls by instead, in the shared block case, wrapping
the cache Release() cleanup in a refcounted object referenced by the
PinnableSlices, such that after the last wrapped reference is released,
the cache entry is Release()ed. Relaxed atomic refcounts should be
much faster than mutex-guarded Ref() and Release(), and much less prone
to a performance cliff when MultiGet() does a lot of block sharing.
Note that I did not use std::shared_ptr, because that would require an
extra indirection object (shared_ptr itself new/delete) in order to
associate a ref increment/decrement with a Cleanable cleanup entry. (If
I assumed it was the size of two pointers, I could do some hackery to
make it work without the extra indirection, but that's too fragile.)
Some details:
* Fixed (removed) extra block cache tracing entries in cases of cache
entry reuse in MultiGet, but it's likely that in some other cases traces
are missing (XXX comment inserted)
* Moved existing implementations for cleanable.h from iterator.cc to
new cleanable.cc
* Improved API comments on Cleanable
* Added a public SharedCleanablePtr class to cleanable.h in case others
could benefit from the same pattern (potentially many Cleanables and/or
smart pointers referencing a shared Cleanable)
* Add a typedef for MultiGetContext::Mask
* Some variable renaming for clarity
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9899
Test Plan:
Added unit tests for SharedCleanablePtr.
Greatly enhanced ability of existing tests to detect cache use-after-free.
* Release PinnableSlices from MultiGet as they are read rather than in
bulk (in db_test_util wrapper).
* In ASAN build, default to using a trivially small LRUCache for block_cache
so that entries are immediately erased when unreferenced. (Updated two
tests that depend on caching.) New ASAN testsuite running time seems
OK to me.
If I introduce a bug into my implementation where we skip the shared
cleanups on block reuse, ASAN detects the bug in
`db_basic_test *MultiGet*`. If I remove either of the above testing
enhancements, the bug is not detected.
Consider for follow-up work: manipulate or randomize ordering of
PinnableSlice use and release from MultiGet db_test_util wrapper. But in
typical cases, natural ordering gives pretty good functional coverage.
Performance test:
In the extreme (but possible) case of MultiGetting the same or adjacent keys
in a batch, throughput can improve by an order of magnitude.
`./db_bench -benchmarks=multireadrandom -db=/dev/shm/testdb -readonly -num=5 -duration=10 -threads=20 -multiread_batched -batch_size=200`
Before ops/sec, num=5: 1,384,394
Before ops/sec, num=500: 6,423,720
After ops/sec, num=500: 10,658,794
After ops/sec, num=5: 16,027,257
Also note that previously, with high parallelism, having query keys
concentrated in a single block was worse than spreading them out a bit. Now
concentrated in a single block is faster than spread out, which is hopefully
consistent with natural expectation.
Random query performance: with num=1000000, over 999 x 10s runs running before & after simultaneously (each -threads=12):
Before: multireadrandom [AVG 999 runs] : 1088699 (± 7344) ops/sec; 120.4 (± 0.8 ) MB/sec
After: multireadrandom [AVG 999 runs] : 1090402 (± 7230) ops/sec; 120.6 (± 0.8 ) MB/sec
Possibly better, possibly in the noise.
Reviewed By: anand1976
Differential Revision: D35907003
Pulled By: pdillinger
fbshipit-source-id: bbd244d703649a8ca12d476f2d03853ed9d1a17e
3 years ago
|
|
|
for (size_t i = 0; i < statuses.size(); ++i) {
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
6 years ago
|
|
|
if (statuses[i].IsNotFound()) {
|
|
|
|
result[i] = "NOT_FOUND";
|
Eliminate unnecessary (slow) block cache Ref()ing in MultiGet (#9899)
Summary:
When MultiGet() determines that multiple query keys can be
served by examining the same data block in block cache (one Lookup()),
each PinnableSlice referring to data in that data block needs to hold
on to the block in cache so that they can be released at arbitrary
times by the API user. Historically this is accomplished with extra
calls to Ref() on the Handle from Lookup(), with each PinnableSlice
cleanup calling Release() on the Handle, but this creates extra
contention on the block cache for the extra Ref()s and Release()es,
especially because they hit the same cache shard repeatedly.
In the case of merge operands (possibly more cases?), the problem was
compounded by doing an extra Ref()+eventual Release() for each merge
operand for a key reusing a block (which could be the same key!), rather
than one Ref() per key. (Note: the non-shared case with `biter` was
already one per key.)
This change optimizes MultiGet not to rely on these extra, contentious
Ref()+Release() calls by instead, in the shared block case, wrapping
the cache Release() cleanup in a refcounted object referenced by the
PinnableSlices, such that after the last wrapped reference is released,
the cache entry is Release()ed. Relaxed atomic refcounts should be
much faster than mutex-guarded Ref() and Release(), and much less prone
to a performance cliff when MultiGet() does a lot of block sharing.
Note that I did not use std::shared_ptr, because that would require an
extra indirection object (shared_ptr itself new/delete) in order to
associate a ref increment/decrement with a Cleanable cleanup entry. (If
I assumed it was the size of two pointers, I could do some hackery to
make it work without the extra indirection, but that's too fragile.)
Some details:
* Fixed (removed) extra block cache tracing entries in cases of cache
entry reuse in MultiGet, but it's likely that in some other cases traces
are missing (XXX comment inserted)
* Moved existing implementations for cleanable.h from iterator.cc to
new cleanable.cc
* Improved API comments on Cleanable
* Added a public SharedCleanablePtr class to cleanable.h in case others
could benefit from the same pattern (potentially many Cleanables and/or
smart pointers referencing a shared Cleanable)
* Add a typedef for MultiGetContext::Mask
* Some variable renaming for clarity
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9899
Test Plan:
Added unit tests for SharedCleanablePtr.
Greatly enhanced ability of existing tests to detect cache use-after-free.
* Release PinnableSlices from MultiGet as they are read rather than in
bulk (in db_test_util wrapper).
* In ASAN build, default to using a trivially small LRUCache for block_cache
so that entries are immediately erased when unreferenced. (Updated two
tests that depend on caching.) New ASAN testsuite running time seems
OK to me.
If I introduce a bug into my implementation where we skip the shared
cleanups on block reuse, ASAN detects the bug in
`db_basic_test *MultiGet*`. If I remove either of the above testing
enhancements, the bug is not detected.
Consider for follow-up work: manipulate or randomize ordering of
PinnableSlice use and release from MultiGet db_test_util wrapper. But in
typical cases, natural ordering gives pretty good functional coverage.
Performance test:
In the extreme (but possible) case of MultiGetting the same or adjacent keys
in a batch, throughput can improve by an order of magnitude.
`./db_bench -benchmarks=multireadrandom -db=/dev/shm/testdb -readonly -num=5 -duration=10 -threads=20 -multiread_batched -batch_size=200`
Before ops/sec, num=5: 1,384,394
Before ops/sec, num=500: 6,423,720
After ops/sec, num=500: 10,658,794
After ops/sec, num=5: 16,027,257
Also note that previously, with high parallelism, having query keys
concentrated in a single block was worse than spreading them out a bit. Now
concentrated in a single block is faster than spread out, which is hopefully
consistent with natural expectation.
Random query performance: with num=1000000, over 999 x 10s runs running before & after simultaneously (each -threads=12):
Before: multireadrandom [AVG 999 runs] : 1088699 (± 7344) ops/sec; 120.4 (± 0.8 ) MB/sec
After: multireadrandom [AVG 999 runs] : 1090402 (± 7230) ops/sec; 120.6 (± 0.8 ) MB/sec
Possibly better, possibly in the noise.
Reviewed By: anand1976
Differential Revision: D35907003
Pulled By: pdillinger
fbshipit-source-id: bbd244d703649a8ca12d476f2d03853ed9d1a17e
3 years ago
|
|
|
} else if (!statuses[i].ok()) {
|
|
|
|
result[i] = statuses[i].ToString();
|
|
|
|
} else {
|
|
|
|
result[i].assign(pin_values[i].data(), pin_values[i].size());
|
|
|
|
// Increase likelihood of detecting potential use-after-free bugs with
|
|
|
|
// PinnableSlices tracking the same resource
|
|
|
|
pin_values[i].Reset();
|
Introduce a new MultiGet batching implementation (#5011)
Summary:
This PR introduces a new MultiGet() API, with the underlying implementation grouping keys based on SST file and batching lookups in a file. The reason for the new API is twofold - the definition allows callers to allocate storage for status and values on stack instead of std::vector, as well as return values as PinnableSlices in order to avoid copying, and it keeps the original MultiGet() implementation intact while we experiment with batching.
Batching is useful when there is some spatial locality to the keys being queries, as well as larger batch sizes. The main benefits are due to -
1. Fewer function calls, especially to BlockBasedTableReader::MultiGet() and FullFilterBlockReader::KeysMayMatch()
2. Bloom filter cachelines can be prefetched, hiding the cache miss latency
The next step is to optimize the binary searches in the level_storage_info, index blocks and data blocks, since we could reduce the number of key comparisons if the keys are relatively close to each other. The batching optimizations also need to be extended to other formats, such as PlainTable and filter formats. This also needs to be added to db_stress.
Benchmark results from db_bench for various batch size/locality of reference combinations are given below. Locality was simulated by offsetting the keys in a batch by a stride length. Each SST file is about 8.6MB uncompressed and key/value size is 16/100 uncompressed. To focus on the cpu benefit of batching, the runs were single threaded and bound to the same cpu to eliminate interference from other system events. The results show a 10-25% improvement in micros/op from smaller to larger batch sizes (4 - 32).
Batch Sizes
1 | 2 | 4 | 8 | 16 | 32
Random pattern (Stride length 0)
4.158 | 4.109 | 4.026 | 4.05 | 4.1 | 4.074 - Get
4.438 | 4.302 | 4.165 | 4.122 | 4.096 | 4.075 - MultiGet (no batching)
4.461 | 4.256 | 4.277 | 4.11 | 4.182 | 4.14 - MultiGet (w/ batching)
Good locality (Stride length 16)
4.048 | 3.659 | 3.248 | 2.99 | 2.84 | 2.753
4.429 | 3.728 | 3.406 | 3.053 | 2.911 | 2.781
4.452 | 3.45 | 2.833 | 2.451 | 2.233 | 2.135
Good locality (Stride length 256)
4.066 | 3.786 | 3.581 | 3.447 | 3.415 | 3.232
4.406 | 4.005 | 3.644 | 3.49 | 3.381 | 3.268
4.393 | 3.649 | 3.186 | 2.882 | 2.676 | 2.62
Medium locality (Stride length 4096)
4.012 | 3.922 | 3.768 | 3.61 | 3.582 | 3.555
4.364 | 4.057 | 3.791 | 3.65 | 3.57 | 3.465
4.479 | 3.758 | 3.316 | 3.077 | 2.959 | 2.891
dbbench command used (on a DB with 4 levels, 12 million keys)-
TEST_TMPDIR=/dev/shm numactl -C 10 ./db_bench.tmp -use_existing_db=true -benchmarks="readseq,multireadrandom" -write_buffer_size=4194304 -target_file_size_base=4194304 -max_bytes_for_level_base=16777216 -num=12000000 -reads=12000000 -duration=90 -threads=1 -compression_type=none -cache_size=4194304000 -batch_size=32 -disable_auto_compactions=true -bloom_bits=10 -cache_index_and_filter_blocks=true -pin_l0_filter_and_index_blocks_in_cache=true -multiread_batched=true -multiread_stride=4
Pull Request resolved: https://github.com/facebook/rocksdb/pull/5011
Differential Revision: D14348703
Pulled By: anand1976
fbshipit-source-id: 774406dab3776d979c809522a67bedac6c17f84b
6 years ago
|
|
|
}
|
|
|
|
}
|
|
|
|
return result;
|
|
|
|
}
|
|
|
|
|
|
|
|
Status DBTestBase::Get(const std::string& k, PinnableSlice* v) {
|
|
|
|
ReadOptions options;
|
|
|
|
options.verify_checksums = true;
|
|
|
|
Status s = dbfull()->Get(options, dbfull()->DefaultColumnFamily(), k, v);
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
uint64_t DBTestBase::GetNumSnapshots() {
|
|
|
|
uint64_t int_num;
|
|
|
|
EXPECT_TRUE(dbfull()->GetIntProperty("rocksdb.num-snapshots", &int_num));
|
|
|
|
return int_num;
|
|
|
|
}
|
|
|
|
|
|
|
|
uint64_t DBTestBase::GetTimeOldestSnapshots() {
|
|
|
|
uint64_t int_num;
|
|
|
|
EXPECT_TRUE(
|
|
|
|
dbfull()->GetIntProperty("rocksdb.oldest-snapshot-time", &int_num));
|
|
|
|
return int_num;
|
|
|
|
}
|
|
|
|
|
|
|
|
uint64_t DBTestBase::GetSequenceOldestSnapshots() {
|
|
|
|
uint64_t int_num;
|
|
|
|
EXPECT_TRUE(
|
|
|
|
dbfull()->GetIntProperty("rocksdb.oldest-snapshot-sequence", &int_num));
|
|
|
|
return int_num;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Return a string that contains all key,value pairs in order,
|
|
|
|
// formatted like "(k1->v1)(k2->v2)".
|
|
|
|
std::string DBTestBase::Contents(int cf) {
|
|
|
|
std::vector<std::string> forward;
|
|
|
|
std::string result;
|
|
|
|
Iterator* iter = (cf == 0) ? db_->NewIterator(ReadOptions())
|
|
|
|
: db_->NewIterator(ReadOptions(), handles_[cf]);
|
|
|
|
for (iter->SeekToFirst(); iter->Valid(); iter->Next()) {
|
|
|
|
std::string s = IterStatus(iter);
|
|
|
|
result.push_back('(');
|
|
|
|
result.append(s);
|
|
|
|
result.push_back(')');
|
|
|
|
forward.push_back(s);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Check reverse iteration results are the reverse of forward results
|
|
|
|
unsigned int matched = 0;
|
|
|
|
for (iter->SeekToLast(); iter->Valid(); iter->Prev()) {
|
|
|
|
EXPECT_LT(matched, forward.size());
|
|
|
|
EXPECT_EQ(IterStatus(iter), forward[forward.size() - matched - 1]);
|
|
|
|
matched++;
|
|
|
|
}
|
|
|
|
EXPECT_EQ(matched, forward.size());
|
|
|
|
|
|
|
|
delete iter;
|
|
|
|
return result;
|
|
|
|
}
|
|
|
|
|
|
|
|
void DBTestBase::CheckAllEntriesWithFifoReopen(
|
|
|
|
const std::string& expected_value, const Slice& user_key, int cf,
|
|
|
|
const std::vector<std::string>& cfs, const Options& options) {
|
|
|
|
ASSERT_EQ(AllEntriesFor(user_key, cf), expected_value);
|
|
|
|
|
|
|
|
std::vector<std::string> cfs_plus_default = cfs;
|
|
|
|
cfs_plus_default.insert(cfs_plus_default.begin(), kDefaultColumnFamilyName);
|
|
|
|
|
|
|
|
Options fifo_options(options);
|
|
|
|
fifo_options.compaction_style = kCompactionStyleFIFO;
|
|
|
|
fifo_options.max_open_files = -1;
|
|
|
|
fifo_options.disable_auto_compactions = true;
|
|
|
|
ASSERT_OK(TryReopenWithColumnFamilies(cfs_plus_default, fifo_options));
|
|
|
|
ASSERT_EQ(AllEntriesFor(user_key, cf), expected_value);
|
|
|
|
|
|
|
|
ASSERT_OK(TryReopenWithColumnFamilies(cfs_plus_default, options));
|
|
|
|
ASSERT_EQ(AllEntriesFor(user_key, cf), expected_value);
|
|
|
|
}
|
|
|
|
|
|
|
|
std::string DBTestBase::AllEntriesFor(const Slice& user_key, int cf) {
|
|
|
|
Arena arena;
|
|
|
|
auto options = CurrentOptions();
|
|
|
|
InternalKeyComparator icmp(options.comparator);
|
|
|
|
ReadOptions read_options;
|
|
|
|
ScopedArenaIterator iter;
|
|
|
|
if (cf == 0) {
|
Skip swaths of range tombstone covered keys in merging iterator (2022 edition) (#10449)
Summary:
Delete range logic is moved from `DBIter` to `MergingIterator`, and `MergingIterator` will seek to the end of a range deletion if possible instead of scanning through each key and check with `RangeDelAggregator`.
With the invariant that a key in level L (consider memtable as the first level, each immutable and L0 as a separate level) has a larger sequence number than all keys in any level >L, a range tombstone `[start, end)` from level L covers all keys in its range in any level >L. This property motivates optimizations in iterator:
- in `Seek(target)`, if level L has a range tombstone `[start, end)` that covers `target.UserKey`, then for all levels > L, we can do Seek() on `end` instead of `target` to skip some range tombstone covered keys.
- in `Next()/Prev()`, if the current key is covered by a range tombstone `[start, end)` from level L, we can do `Seek` to `end` for all levels > L.
This PR implements the above optimizations in `MergingIterator`. As all range tombstone covered keys are now skipped in `MergingIterator`, the range tombstone logic is removed from `DBIter`. The idea in this PR is similar to https://github.com/facebook/rocksdb/issues/7317, but this PR leaves `InternalIterator` interface mostly unchanged. **Credit**: the cascading seek optimization and the sentinel key (discussed below) are inspired by [Pebble](https://github.com/cockroachdb/pebble/blob/master/merging_iter.go) and suggested by ajkr in https://github.com/facebook/rocksdb/issues/7317. The two optimizations are mostly implemented in `SeekImpl()/SeekForPrevImpl()` and `IsNextDeleted()/IsPrevDeleted()` in `merging_iterator.cc`. See comments for each method for more detail.
One notable change is that the minHeap/maxHeap used by `MergingIterator` now contains range tombstone end keys besides point key iterators. This helps to reduce the number of key comparisons. For example, for a range tombstone `[start, end)`, a `start` and an `end` `HeapItem` are inserted into the heap. When a `HeapItem` for range tombstone start key is popped from the minHeap, we know this range tombstone becomes "active" in the sense that, before the range tombstone's end key is popped from the minHeap, all the keys popped from this heap is covered by the range tombstone's internal key range `[start, end)`.
Another major change, *delete range sentinel key*, is made to `LevelIterator`. Before this PR, when all point keys in an SST file are iterated through in `MergingIterator`, a level iterator would advance to the next SST file in its level. In the case when an SST file has a range tombstone that covers keys beyond the SST file's last point key, advancing to the next SST file would lose this range tombstone. Consequently, `MergingIterator` could return keys that should have been deleted by some range tombstone. We prevent this by pretending that file boundaries in each SST file are sentinel keys. A `LevelIterator` now only advance the file iterator once the sentinel key is processed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10449
Test Plan:
- Added many unit tests in db_range_del_test
- Stress test: `./db_stress --readpercent=5 --prefixpercent=19 --writepercent=20 -delpercent=10 --iterpercent=44 --delrangepercent=2`
- Additional iterator stress test is added to verify against iterators against expected state: https://github.com/facebook/rocksdb/issues/10538. This is based on ajkr's previous attempt https://github.com/facebook/rocksdb/pull/5506#issuecomment-506021913.
```
python3 ./tools/db_crashtest.py blackbox --simple --write_buffer_size=524288 --target_file_size_base=524288 --max_bytes_for_level_base=2097152 --compression_type=none --max_background_compactions=8 --value_size_mult=33 --max_key=5000000 --interval=10 --duration=7200 --delrangepercent=3 --delpercent=9 --iterpercent=25 --writepercent=60 --readpercent=3 --prefixpercent=0 --num_iterations=1000 --range_deletion_width=100 --verify_iterator_with_expected_state_one_in=1
```
- Performance benchmark: I used a similar setup as in the blog [post](http://rocksdb.org/blog/2018/11/21/delete-range.html) that introduced DeleteRange, "a database with 5 million data keys, and 10000 range tombstones (ignoring those dropped during compaction) that were written in regular intervals after 4.5 million data keys were written". As expected, the performance with this PR depends on the range tombstone width.
```
# Setup:
TEST_TMPDIR=/dev/shm ./db_bench_main --benchmarks=fillrandom --writes=4500000 --num=5000000
TEST_TMPDIR=/dev/shm ./db_bench_main --benchmarks=overwrite --writes=500000 --num=5000000 --use_existing_db=true --writes_per_range_tombstone=50
# Scan entire DB
TEST_TMPDIR=/dev/shm ./db_bench_main --benchmarks=readseq[-X5] --use_existing_db=true --num=5000000 --disable_auto_compactions=true
# Short range scan (10 Next())
TEST_TMPDIR=/dev/shm/width-100/ ./db_bench_main --benchmarks=seekrandom[-X5] --use_existing_db=true --num=500000 --reads=100000 --seek_nexts=10 --disable_auto_compactions=true
# Long range scan(1000 Next())
TEST_TMPDIR=/dev/shm/width-100/ ./db_bench_main --benchmarks=seekrandom[-X5] --use_existing_db=true --num=500000 --reads=2500 --seek_nexts=1000 --disable_auto_compactions=true
```
Avg over of 10 runs (some slower tests had fews runs):
For the first column (tombstone), 0 means no range tombstone, 100-10000 means width of the 10k range tombstones, and 1 means there is a single range tombstone in the entire DB (width is 1000). The 1 tombstone case is to test regression when there's very few range tombstones in the DB, as no range tombstone is likely to take a different code path than with range tombstones.
- Scan entire DB
| tombstone width | Pre-PR ops/sec | Post-PR ops/sec | ±% |
| ------------- | ------------- | ------------- | ------------- |
| 0 range tombstone |2525600 (± 43564) |2486917 (± 33698) |-1.53% |
| 100 |1853835 (± 24736) |2073884 (± 32176) |+11.87% |
| 1000 |422415 (± 7466) |1115801 (± 22781) |+164.15% |
| 10000 |22384 (± 227) |227919 (± 6647) |+918.22% |
| 1 range tombstone |2176540 (± 39050) |2434954 (± 24563) |+11.87% |
- Short range scan
| tombstone width | Pre-PR ops/sec | Post-PR ops/sec | ±% |
| ------------- | ------------- | ------------- | ------------- |
| 0 range tombstone |35398 (± 533) |35338 (± 569) |-0.17% |
| 100 |28276 (± 664) |31684 (± 331) |+12.05% |
| 1000 |7637 (± 77) |25422 (± 277) |+232.88% |
| 10000 |1367 |28667 |+1997.07% |
| 1 range tombstone |32618 (± 581) |32748 (± 506) |+0.4% |
- Long range scan
| tombstone width | Pre-PR ops/sec | Post-PR ops/sec | ±% |
| ------------- | ------------- | ------------- | ------------- |
| 0 range tombstone |2262 (± 33) |2353 (± 20) |+4.02% |
| 100 |1696 (± 26) |1926 (± 18) |+13.56% |
| 1000 |410 (± 6) |1255 (± 29) |+206.1% |
| 10000 |25 |414 |+1556.0% |
| 1 range tombstone |1957 (± 30) |2185 (± 44) |+11.65% |
- Microbench does not show significant regression: https://gist.github.com/cbi42/59f280f85a59b678e7e5d8561e693b61
Reviewed By: ajkr
Differential Revision: D38450331
Pulled By: cbi42
fbshipit-source-id: b5ef12e8d8c289ed2e163ccdf277f5039b511fca
2 years ago
|
|
|
iter.set(dbfull()->NewInternalIterator(read_options, &arena,
|
|
|
|
kMaxSequenceNumber));
|
|
|
|
} else {
|
Skip swaths of range tombstone covered keys in merging iterator (2022 edition) (#10449)
Summary:
Delete range logic is moved from `DBIter` to `MergingIterator`, and `MergingIterator` will seek to the end of a range deletion if possible instead of scanning through each key and check with `RangeDelAggregator`.
With the invariant that a key in level L (consider memtable as the first level, each immutable and L0 as a separate level) has a larger sequence number than all keys in any level >L, a range tombstone `[start, end)` from level L covers all keys in its range in any level >L. This property motivates optimizations in iterator:
- in `Seek(target)`, if level L has a range tombstone `[start, end)` that covers `target.UserKey`, then for all levels > L, we can do Seek() on `end` instead of `target` to skip some range tombstone covered keys.
- in `Next()/Prev()`, if the current key is covered by a range tombstone `[start, end)` from level L, we can do `Seek` to `end` for all levels > L.
This PR implements the above optimizations in `MergingIterator`. As all range tombstone covered keys are now skipped in `MergingIterator`, the range tombstone logic is removed from `DBIter`. The idea in this PR is similar to https://github.com/facebook/rocksdb/issues/7317, but this PR leaves `InternalIterator` interface mostly unchanged. **Credit**: the cascading seek optimization and the sentinel key (discussed below) are inspired by [Pebble](https://github.com/cockroachdb/pebble/blob/master/merging_iter.go) and suggested by ajkr in https://github.com/facebook/rocksdb/issues/7317. The two optimizations are mostly implemented in `SeekImpl()/SeekForPrevImpl()` and `IsNextDeleted()/IsPrevDeleted()` in `merging_iterator.cc`. See comments for each method for more detail.
One notable change is that the minHeap/maxHeap used by `MergingIterator` now contains range tombstone end keys besides point key iterators. This helps to reduce the number of key comparisons. For example, for a range tombstone `[start, end)`, a `start` and an `end` `HeapItem` are inserted into the heap. When a `HeapItem` for range tombstone start key is popped from the minHeap, we know this range tombstone becomes "active" in the sense that, before the range tombstone's end key is popped from the minHeap, all the keys popped from this heap is covered by the range tombstone's internal key range `[start, end)`.
Another major change, *delete range sentinel key*, is made to `LevelIterator`. Before this PR, when all point keys in an SST file are iterated through in `MergingIterator`, a level iterator would advance to the next SST file in its level. In the case when an SST file has a range tombstone that covers keys beyond the SST file's last point key, advancing to the next SST file would lose this range tombstone. Consequently, `MergingIterator` could return keys that should have been deleted by some range tombstone. We prevent this by pretending that file boundaries in each SST file are sentinel keys. A `LevelIterator` now only advance the file iterator once the sentinel key is processed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10449
Test Plan:
- Added many unit tests in db_range_del_test
- Stress test: `./db_stress --readpercent=5 --prefixpercent=19 --writepercent=20 -delpercent=10 --iterpercent=44 --delrangepercent=2`
- Additional iterator stress test is added to verify against iterators against expected state: https://github.com/facebook/rocksdb/issues/10538. This is based on ajkr's previous attempt https://github.com/facebook/rocksdb/pull/5506#issuecomment-506021913.
```
python3 ./tools/db_crashtest.py blackbox --simple --write_buffer_size=524288 --target_file_size_base=524288 --max_bytes_for_level_base=2097152 --compression_type=none --max_background_compactions=8 --value_size_mult=33 --max_key=5000000 --interval=10 --duration=7200 --delrangepercent=3 --delpercent=9 --iterpercent=25 --writepercent=60 --readpercent=3 --prefixpercent=0 --num_iterations=1000 --range_deletion_width=100 --verify_iterator_with_expected_state_one_in=1
```
- Performance benchmark: I used a similar setup as in the blog [post](http://rocksdb.org/blog/2018/11/21/delete-range.html) that introduced DeleteRange, "a database with 5 million data keys, and 10000 range tombstones (ignoring those dropped during compaction) that were written in regular intervals after 4.5 million data keys were written". As expected, the performance with this PR depends on the range tombstone width.
```
# Setup:
TEST_TMPDIR=/dev/shm ./db_bench_main --benchmarks=fillrandom --writes=4500000 --num=5000000
TEST_TMPDIR=/dev/shm ./db_bench_main --benchmarks=overwrite --writes=500000 --num=5000000 --use_existing_db=true --writes_per_range_tombstone=50
# Scan entire DB
TEST_TMPDIR=/dev/shm ./db_bench_main --benchmarks=readseq[-X5] --use_existing_db=true --num=5000000 --disable_auto_compactions=true
# Short range scan (10 Next())
TEST_TMPDIR=/dev/shm/width-100/ ./db_bench_main --benchmarks=seekrandom[-X5] --use_existing_db=true --num=500000 --reads=100000 --seek_nexts=10 --disable_auto_compactions=true
# Long range scan(1000 Next())
TEST_TMPDIR=/dev/shm/width-100/ ./db_bench_main --benchmarks=seekrandom[-X5] --use_existing_db=true --num=500000 --reads=2500 --seek_nexts=1000 --disable_auto_compactions=true
```
Avg over of 10 runs (some slower tests had fews runs):
For the first column (tombstone), 0 means no range tombstone, 100-10000 means width of the 10k range tombstones, and 1 means there is a single range tombstone in the entire DB (width is 1000). The 1 tombstone case is to test regression when there's very few range tombstones in the DB, as no range tombstone is likely to take a different code path than with range tombstones.
- Scan entire DB
| tombstone width | Pre-PR ops/sec | Post-PR ops/sec | ±% |
| ------------- | ------------- | ------------- | ------------- |
| 0 range tombstone |2525600 (± 43564) |2486917 (± 33698) |-1.53% |
| 100 |1853835 (± 24736) |2073884 (± 32176) |+11.87% |
| 1000 |422415 (± 7466) |1115801 (± 22781) |+164.15% |
| 10000 |22384 (± 227) |227919 (± 6647) |+918.22% |
| 1 range tombstone |2176540 (± 39050) |2434954 (± 24563) |+11.87% |
- Short range scan
| tombstone width | Pre-PR ops/sec | Post-PR ops/sec | ±% |
| ------------- | ------------- | ------------- | ------------- |
| 0 range tombstone |35398 (± 533) |35338 (± 569) |-0.17% |
| 100 |28276 (± 664) |31684 (± 331) |+12.05% |
| 1000 |7637 (± 77) |25422 (± 277) |+232.88% |
| 10000 |1367 |28667 |+1997.07% |
| 1 range tombstone |32618 (± 581) |32748 (± 506) |+0.4% |
- Long range scan
| tombstone width | Pre-PR ops/sec | Post-PR ops/sec | ±% |
| ------------- | ------------- | ------------- | ------------- |
| 0 range tombstone |2262 (± 33) |2353 (± 20) |+4.02% |
| 100 |1696 (± 26) |1926 (± 18) |+13.56% |
| 1000 |410 (± 6) |1255 (± 29) |+206.1% |
| 10000 |25 |414 |+1556.0% |
| 1 range tombstone |1957 (± 30) |2185 (± 44) |+11.65% |
- Microbench does not show significant regression: https://gist.github.com/cbi42/59f280f85a59b678e7e5d8561e693b61
Reviewed By: ajkr
Differential Revision: D38450331
Pulled By: cbi42
fbshipit-source-id: b5ef12e8d8c289ed2e163ccdf277f5039b511fca
2 years ago
|
|
|
iter.set(dbfull()->NewInternalIterator(read_options, &arena,
|
|
|
|
kMaxSequenceNumber, handles_[cf]));
|
|
|
|
}
|
|
|
|
InternalKey target(user_key, kMaxSequenceNumber, kTypeValue);
|
|
|
|
iter->Seek(target.Encode());
|
|
|
|
std::string result;
|
|
|
|
if (!iter->status().ok()) {
|
|
|
|
result = iter->status().ToString();
|
|
|
|
} else {
|
|
|
|
result = "[ ";
|
|
|
|
bool first = true;
|
|
|
|
while (iter->Valid()) {
|
|
|
|
ParsedInternalKey ikey(Slice(), 0, kTypeValue);
|
|
|
|
if (ParseInternalKey(iter->key(), &ikey, true /* log_err_key */) !=
|
|
|
|
Status::OK()) {
|
|
|
|
result += "CORRUPTED";
|
|
|
|
} else {
|
|
|
|
if (!last_options_.comparator->Equal(ikey.user_key, user_key)) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
if (!first) {
|
|
|
|
result += ", ";
|
|
|
|
}
|
|
|
|
first = false;
|
|
|
|
switch (ikey.type) {
|
|
|
|
case kTypeValue:
|
|
|
|
result += iter->value().ToString();
|
|
|
|
break;
|
|
|
|
case kTypeMerge:
|
|
|
|
// keep it the same as kTypeValue for testing kMergePut
|
|
|
|
result += iter->value().ToString();
|
|
|
|
break;
|
|
|
|
case kTypeDeletion:
|
|
|
|
result += "DEL";
|
|
|
|
break;
|
Support for SingleDelete()
Summary:
This patch fixes #7460559. It introduces SingleDelete as a new database
operation. This operation can be used to delete keys that were never
overwritten (no put following another put of the same key). If an overwritten
key is single deleted the behavior is undefined. Single deletion of a
non-existent key has no effect but multiple consecutive single deletions are
not allowed (see limitations).
In contrast to the conventional Delete() operation, the deletion entry is
removed along with the value when the two are lined up in a compaction. Note:
The semantics are similar to @igor's prototype that allowed to have this
behavior on the granularity of a column family (
https://reviews.facebook.net/D42093 ). This new patch, however, is more
aggressive when it comes to removing tombstones: It removes the SingleDelete
together with the value whenever there is no snapshot between them while the
older patch only did this when the sequence number of the deletion was older
than the earliest snapshot.
Most of the complex additions are in the Compaction Iterator, all other changes
should be relatively straightforward. The patch also includes basic support for
single deletions in db_stress and db_bench.
Limitations:
- Not compatible with cuckoo hash tables
- Single deletions cannot be used in combination with merges and normal
deletions on the same key (other keys are not affected by this)
- Consecutive single deletions are currently not allowed (and older version of
this patch supported this so it could be resurrected if needed)
Test Plan: make all check
Reviewers: yhchiang, sdong, rven, anthony, yoshinorim, igor
Reviewed By: igor
Subscribers: maykov, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D43179
9 years ago
|
|
|
case kTypeSingleDeletion:
|
|
|
|
result += "SDEL";
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
assert(false);
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
iter->Next();
|
|
|
|
}
|
|
|
|
if (!first) {
|
|
|
|
result += " ";
|
|
|
|
}
|
|
|
|
result += "]";
|
|
|
|
}
|
|
|
|
return result;
|
|
|
|
}
|
|
|
|
|
|
|
|
int DBTestBase::NumSortedRuns(int cf) {
|
|
|
|
ColumnFamilyMetaData cf_meta;
|
|
|
|
if (cf == 0) {
|
|
|
|
db_->GetColumnFamilyMetaData(&cf_meta);
|
|
|
|
} else {
|
|
|
|
db_->GetColumnFamilyMetaData(handles_[cf], &cf_meta);
|
|
|
|
}
|
|
|
|
int num_sr = static_cast<int>(cf_meta.levels[0].files.size());
|
|
|
|
for (size_t i = 1U; i < cf_meta.levels.size(); i++) {
|
|
|
|
if (cf_meta.levels[i].files.size() > 0) {
|
|
|
|
num_sr++;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return num_sr;
|
|
|
|
}
|
|
|
|
|
|
|
|
uint64_t DBTestBase::TotalSize(int cf) {
|
|
|
|
ColumnFamilyMetaData cf_meta;
|
|
|
|
if (cf == 0) {
|
|
|
|
db_->GetColumnFamilyMetaData(&cf_meta);
|
|
|
|
} else {
|
|
|
|
db_->GetColumnFamilyMetaData(handles_[cf], &cf_meta);
|
|
|
|
}
|
|
|
|
return cf_meta.size;
|
|
|
|
}
|
|
|
|
|
|
|
|
uint64_t DBTestBase::SizeAtLevel(int level) {
|
|
|
|
std::vector<LiveFileMetaData> metadata;
|
|
|
|
db_->GetLiveFilesMetaData(&metadata);
|
|
|
|
uint64_t sum = 0;
|
|
|
|
for (const auto& m : metadata) {
|
|
|
|
if (m.level == level) {
|
|
|
|
sum += m.size;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return sum;
|
|
|
|
}
|
|
|
|
|
|
|
|
size_t DBTestBase::TotalLiveFiles(int cf) {
|
|
|
|
ColumnFamilyMetaData cf_meta;
|
|
|
|
if (cf == 0) {
|
|
|
|
db_->GetColumnFamilyMetaData(&cf_meta);
|
|
|
|
} else {
|
|
|
|
db_->GetColumnFamilyMetaData(handles_[cf], &cf_meta);
|
|
|
|
}
|
|
|
|
size_t num_files = 0;
|
|
|
|
for (auto& level : cf_meta.levels) {
|
|
|
|
num_files += level.files.size();
|
|
|
|
}
|
|
|
|
return num_files;
|
|
|
|
}
|
|
|
|
|
|
|
|
size_t DBTestBase::CountLiveFiles() {
|
|
|
|
std::vector<LiveFileMetaData> metadata;
|
|
|
|
db_->GetLiveFilesMetaData(&metadata);
|
|
|
|
return metadata.size();
|
|
|
|
}
|
|
|
|
|
|
|
|
int DBTestBase::NumTableFilesAtLevel(int level, int cf) {
|
|
|
|
std::string property;
|
|
|
|
if (cf == 0) {
|
|
|
|
// default cfd
|
|
|
|
EXPECT_TRUE(db_->GetProperty(
|
|
|
|
"rocksdb.num-files-at-level" + std::to_string(level), &property));
|
|
|
|
} else {
|
|
|
|
EXPECT_TRUE(db_->GetProperty(
|
|
|
|
handles_[cf], "rocksdb.num-files-at-level" + std::to_string(level),
|
|
|
|
&property));
|
|
|
|
}
|
|
|
|
return atoi(property.c_str());
|
|
|
|
}
|
|
|
|
|
|
|
|
double DBTestBase::CompressionRatioAtLevel(int level, int cf) {
|
|
|
|
std::string property;
|
|
|
|
if (cf == 0) {
|
|
|
|
// default cfd
|
|
|
|
EXPECT_TRUE(db_->GetProperty(
|
|
|
|
"rocksdb.compression-ratio-at-level" + std::to_string(level),
|
|
|
|
&property));
|
|
|
|
} else {
|
|
|
|
EXPECT_TRUE(db_->GetProperty(
|
|
|
|
handles_[cf],
|
|
|
|
"rocksdb.compression-ratio-at-level" + std::to_string(level),
|
|
|
|
&property));
|
|
|
|
}
|
|
|
|
return std::stod(property);
|
|
|
|
}
|
|
|
|
|
|
|
|
int DBTestBase::TotalTableFiles(int cf, int levels) {
|
|
|
|
if (levels == -1) {
|
|
|
|
levels = (cf == 0) ? db_->NumberLevels() : db_->NumberLevels(handles_[1]);
|
|
|
|
}
|
|
|
|
int result = 0;
|
|
|
|
for (int level = 0; level < levels; level++) {
|
|
|
|
result += NumTableFilesAtLevel(level, cf);
|
|
|
|
}
|
|
|
|
return result;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Return spread of files per level
|
|
|
|
std::string DBTestBase::FilesPerLevel(int cf) {
|
|
|
|
int num_levels =
|
|
|
|
(cf == 0) ? db_->NumberLevels() : db_->NumberLevels(handles_[1]);
|
|
|
|
std::string result;
|
|
|
|
size_t last_non_zero_offset = 0;
|
|
|
|
for (int level = 0; level < num_levels; level++) {
|
|
|
|
int f = NumTableFilesAtLevel(level, cf);
|
|
|
|
char buf[100];
|
|
|
|
snprintf(buf, sizeof(buf), "%s%d", (level ? "," : ""), f);
|
|
|
|
result += buf;
|
|
|
|
if (f > 0) {
|
|
|
|
last_non_zero_offset = result.size();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
result.resize(last_non_zero_offset);
|
|
|
|
return result;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
std::vector<uint64_t> DBTestBase::GetBlobFileNumbers() {
|
|
|
|
VersionSet* const versions = dbfull()->GetVersionSet();
|
|
|
|
assert(versions);
|
|
|
|
|
|
|
|
ColumnFamilyData* const cfd = versions->GetColumnFamilySet()->GetDefault();
|
|
|
|
assert(cfd);
|
|
|
|
|
|
|
|
Version* const current = cfd->current();
|
|
|
|
assert(current);
|
|
|
|
|
|
|
|
const VersionStorageInfo* const storage_info = current->storage_info();
|
|
|
|
assert(storage_info);
|
|
|
|
|
|
|
|
const auto& blob_files = storage_info->GetBlobFiles();
|
|
|
|
|
|
|
|
std::vector<uint64_t> result;
|
|
|
|
result.reserve(blob_files.size());
|
|
|
|
|
|
|
|
for (const auto& blob_file : blob_files) {
|
Use a sorted vector instead of a map to store blob file metadata (#9526)
Summary:
The patch replaces `std::map` with a sorted `std::vector` for
`VersionStorageInfo::blob_files_` and preallocates the space
for the `vector` before saving the `BlobFileMetaData` into the
new `VersionStorageInfo` in `VersionBuilder::Rep::SaveBlobFilesTo`.
These changes reduce the time the DB mutex is held while
saving new `Version`s, and using a sorted `vector` also makes
lookups faster thanks to better memory locality.
In addition, the patch introduces helper methods
`VersionStorageInfo::GetBlobFileMetaData` and
`VersionStorageInfo::GetBlobFileMetaDataLB` that can be used by
clients to perform lookups in the `vector`, and does some general
cleanup in the parts of code where blob file metadata are used.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9526
Test Plan:
Ran `make check` and the crash test script for a while.
Performance was tested using a load-optimized benchmark (`fillseq` with vector memtable, no WAL) and small file sizes so that a significant number of files are produced:
```
numactl --interleave=all ./db_bench --benchmarks=fillseq --allow_concurrent_memtable_write=false --level0_file_num_compaction_trigger=4 --level0_slowdown_writes_trigger=20 --level0_stop_writes_trigger=30 --max_background_jobs=8 --max_write_buffer_number=8 --db=/data/ltamasi-dbbench --wal_dir=/data/ltamasi-dbbench --num=800000000 --num_levels=8 --key_size=20 --value_size=400 --block_size=8192 --cache_size=51539607552 --cache_numshardbits=6 --compression_max_dict_bytes=0 --compression_ratio=0.5 --compression_type=lz4 --bytes_per_sync=8388608 --cache_index_and_filter_blocks=1 --cache_high_pri_pool_ratio=0.5 --benchmark_write_rate_limit=0 --write_buffer_size=16777216 --target_file_size_base=16777216 --max_bytes_for_level_base=67108864 --verify_checksum=1 --delete_obsolete_files_period_micros=62914560 --max_bytes_for_level_multiplier=8 --statistics=0 --stats_per_interval=1 --stats_interval_seconds=20 --histogram=1 --memtablerep=skip_list --bloom_bits=10 --open_files=-1 --subcompactions=1 --compaction_style=0 --min_level_to_compress=3 --level_compaction_dynamic_level_bytes=true --pin_l0_filter_and_index_blocks_in_cache=1 --soft_pending_compaction_bytes_limit=167503724544 --hard_pending_compaction_bytes_limit=335007449088 --min_level_to_compress=0 --use_existing_db=0 --sync=0 --threads=1 --memtablerep=vector --allow_concurrent_memtable_write=false --disable_wal=1 --enable_blob_files=1 --blob_file_size=16777216 --min_blob_size=0 --blob_compression_type=lz4 --enable_blob_garbage_collection=1 --seed=<some value>
```
Final statistics before the patch:
```
Cumulative writes: 0 writes, 700M keys, 0 commit groups, 0.0 writes per commit group, ingest: 284.62 GB, 121.27 MB/s
Interval writes: 0 writes, 334K keys, 0 commit groups, 0.0 writes per commit group, ingest: 139.28 MB, 72.46 MB/s
```
With the patch:
```
Cumulative writes: 0 writes, 760M keys, 0 commit groups, 0.0 writes per commit group, ingest: 308.66 GB, 131.52 MB/s
Interval writes: 0 writes, 445K keys, 0 commit groups, 0.0 writes per commit group, ingest: 185.35 MB, 93.15 MB/s
```
Total time to complete the benchmark is 2611 seconds with the patch, down from 2986 secs.
Reviewed By: riversand963
Differential Revision: D34082728
Pulled By: ltamasi
fbshipit-source-id: fc598abf676dce436734d06bb9d2d99a26a004fc
3 years ago
|
|
|
assert(blob_file);
|
|
|
|
result.emplace_back(blob_file->GetBlobFileNumber());
|
|
|
|
}
|
|
|
|
|
|
|
|
return result;
|
|
|
|
}
|
|
|
|
|
|
|
|
size_t DBTestBase::CountFiles() {
|
|
|
|
size_t count = 0;
|
|
|
|
std::vector<std::string> files;
|
|
|
|
if (env_->GetChildren(dbname_, &files).ok()) {
|
|
|
|
count += files.size();
|
|
|
|
}
|
|
|
|
|
|
|
|
if (dbname_ != last_options_.wal_dir) {
|
|
|
|
if (env_->GetChildren(last_options_.wal_dir, &files).ok()) {
|
|
|
|
count += files.size();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return count;
|
|
|
|
};
|
|
|
|
|
|
|
|
Status DBTestBase::CountFiles(size_t* count) {
|
|
|
|
std::vector<std::string> files;
|
|
|
|
Status s = env_->GetChildren(dbname_, &files);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
size_t files_count = files.size();
|
|
|
|
|
|
|
|
if (dbname_ != last_options_.wal_dir) {
|
|
|
|
s = env_->GetChildren(last_options_.wal_dir, &files);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
*count = files_count + files.size();
|
|
|
|
}
|
|
|
|
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
Status DBTestBase::Size(const Slice& start, const Slice& limit, int cf,
|
|
|
|
uint64_t* size) {
|
|
|
|
Range r(start, limit);
|
|
|
|
if (cf == 0) {
|
|
|
|
return db_->GetApproximateSizes(&r, 1, size);
|
|
|
|
} else {
|
|
|
|
return db_->GetApproximateSizes(handles_[1], &r, 1, size);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void DBTestBase::Compact(int cf, const Slice& start, const Slice& limit,
|
|
|
|
uint32_t target_path_id) {
|
|
|
|
CompactRangeOptions compact_options;
|
|
|
|
compact_options.target_path_id = target_path_id;
|
|
|
|
ASSERT_OK(db_->CompactRange(compact_options, handles_[cf], &start, &limit));
|
|
|
|
}
|
|
|
|
|
|
|
|
void DBTestBase::Compact(int cf, const Slice& start, const Slice& limit) {
|
|
|
|
ASSERT_OK(
|
|
|
|
db_->CompactRange(CompactRangeOptions(), handles_[cf], &start, &limit));
|
|
|
|
}
|
|
|
|
|
|
|
|
void DBTestBase::Compact(const Slice& start, const Slice& limit) {
|
|
|
|
ASSERT_OK(db_->CompactRange(CompactRangeOptions(), &start, &limit));
|
|
|
|
}
|
|
|
|
|
|
|
|
// Do n memtable compactions, each of which produces an sstable
|
|
|
|
// covering the range [small,large].
|
|
|
|
void DBTestBase::MakeTables(int n, const std::string& small,
|
|
|
|
const std::string& large, int cf) {
|
|
|
|
for (int i = 0; i < n; i++) {
|
|
|
|
ASSERT_OK(Put(cf, small, "begin"));
|
|
|
|
ASSERT_OK(Put(cf, large, "end"));
|
|
|
|
ASSERT_OK(Flush(cf));
|
|
|
|
MoveFilesToLevel(n - i - 1, cf);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Prevent pushing of new sstables into deeper levels by adding
|
|
|
|
// tables that cover a specified range to all levels.
|
|
|
|
void DBTestBase::FillLevels(const std::string& smallest,
|
|
|
|
const std::string& largest, int cf) {
|
|
|
|
MakeTables(db_->NumberLevels(handles_[cf]), smallest, largest, cf);
|
|
|
|
}
|
|
|
|
|
|
|
|
void DBTestBase::MoveFilesToLevel(int level, int cf) {
|
|
|
|
for (int l = 0; l < level; ++l) {
|
|
|
|
if (cf > 0) {
|
|
|
|
EXPECT_OK(dbfull()->TEST_CompactRange(l, nullptr, nullptr, handles_[cf]));
|
|
|
|
} else {
|
|
|
|
EXPECT_OK(dbfull()->TEST_CompactRange(l, nullptr, nullptr));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void DBTestBase::DumpFileCounts(const char* label) {
|
|
|
|
fprintf(stderr, "---\n%s:\n", label);
|
|
|
|
fprintf(stderr, "maxoverlap: %" PRIu64 "\n",
|
|
|
|
dbfull()->TEST_MaxNextLevelOverlappingBytes());
|
|
|
|
for (int level = 0; level < db_->NumberLevels(); level++) {
|
|
|
|
int num = NumTableFilesAtLevel(level);
|
|
|
|
if (num > 0) {
|
|
|
|
fprintf(stderr, " level %3d : %d files\n", level, num);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
std::string DBTestBase::DumpSSTableList() {
|
|
|
|
std::string property;
|
|
|
|
db_->GetProperty("rocksdb.sstables", &property);
|
|
|
|
return property;
|
|
|
|
}
|
|
|
|
|
|
|
|
void DBTestBase::GetSstFiles(Env* env, std::string path,
|
|
|
|
std::vector<std::string>* files) {
|
|
|
|
EXPECT_OK(env->GetChildren(path, files));
|
|
|
|
|
|
|
|
files->erase(std::remove_if(files->begin(), files->end(),
|
|
|
|
[](std::string name) {
|
|
|
|
uint64_t number;
|
|
|
|
FileType type;
|
|
|
|
return !(ParseFileName(name, &number, &type) &&
|
|
|
|
type == kTableFile);
|
|
|
|
}),
|
|
|
|
files->end());
|
|
|
|
}
|
|
|
|
|
|
|
|
int DBTestBase::GetSstFileCount(std::string path) {
|
|
|
|
std::vector<std::string> files;
|
|
|
|
DBTestBase::GetSstFiles(env_, path, &files);
|
|
|
|
return static_cast<int>(files.size());
|
|
|
|
}
|
|
|
|
|
|
|
|
// this will generate non-overlapping files since it keeps increasing key_idx
|
|
|
|
void DBTestBase::GenerateNewFile(int cf, Random* rnd, int* key_idx,
|
|
|
|
bool nowait) {
|
|
|
|
for (int i = 0; i < KNumKeysByGenerateNewFile; i++) {
|
|
|
|
ASSERT_OK(Put(cf, Key(*key_idx), rnd->RandomString((i == 99) ? 1 : 990)));
|
|
|
|
(*key_idx)++;
|
|
|
|
}
|
|
|
|
if (!nowait) {
|
|
|
|
ASSERT_OK(dbfull()->TEST_WaitForFlushMemTable());
|
|
|
|
ASSERT_OK(dbfull()->TEST_WaitForCompact());
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// this will generate non-overlapping files since it keeps increasing key_idx
|
|
|
|
void DBTestBase::GenerateNewFile(Random* rnd, int* key_idx, bool nowait) {
|
|
|
|
for (int i = 0; i < KNumKeysByGenerateNewFile; i++) {
|
|
|
|
ASSERT_OK(Put(Key(*key_idx), rnd->RandomString((i == 99) ? 1 : 990)));
|
|
|
|
(*key_idx)++;
|
|
|
|
}
|
|
|
|
if (!nowait) {
|
|
|
|
ASSERT_OK(dbfull()->TEST_WaitForFlushMemTable());
|
|
|
|
ASSERT_OK(dbfull()->TEST_WaitForCompact());
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
const int DBTestBase::kNumKeysByGenerateNewRandomFile = 51;
|
|
|
|
|
|
|
|
void DBTestBase::GenerateNewRandomFile(Random* rnd, bool nowait) {
|
|
|
|
for (int i = 0; i < kNumKeysByGenerateNewRandomFile; i++) {
|
|
|
|
ASSERT_OK(Put("key" + rnd->RandomString(7), rnd->RandomString(2000)));
|
|
|
|
}
|
|
|
|
ASSERT_OK(Put("key" + rnd->RandomString(7), rnd->RandomString(200)));
|
|
|
|
if (!nowait) {
|
|
|
|
ASSERT_OK(dbfull()->TEST_WaitForFlushMemTable());
|
|
|
|
ASSERT_OK(dbfull()->TEST_WaitForCompact());
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
std::string DBTestBase::IterStatus(Iterator* iter) {
|
|
|
|
std::string result;
|
|
|
|
if (iter->Valid()) {
|
|
|
|
result = iter->key().ToString() + "->" + iter->value().ToString();
|
|
|
|
} else {
|
|
|
|
result = "(invalid)";
|
|
|
|
}
|
|
|
|
return result;
|
|
|
|
}
|
|
|
|
|
|
|
|
Options DBTestBase::OptionsForLogIterTest() {
|
|
|
|
Options options = CurrentOptions();
|
|
|
|
options.create_if_missing = true;
|
|
|
|
options.WAL_ttl_seconds = 1000;
|
|
|
|
return options;
|
|
|
|
}
|
|
|
|
|
|
|
|
std::string DBTestBase::DummyString(size_t len, char c) {
|
|
|
|
return std::string(len, c);
|
|
|
|
}
|
|
|
|
|
|
|
|
void DBTestBase::VerifyIterLast(std::string expected_key, int cf) {
|
|
|
|
Iterator* iter;
|
|
|
|
ReadOptions ro;
|
|
|
|
if (cf == 0) {
|
|
|
|
iter = db_->NewIterator(ro);
|
|
|
|
} else {
|
|
|
|
iter = db_->NewIterator(ro, handles_[cf]);
|
|
|
|
}
|
|
|
|
iter->SeekToLast();
|
|
|
|
ASSERT_EQ(IterStatus(iter), expected_key);
|
|
|
|
delete iter;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Used to test InplaceUpdate
|
|
|
|
|
|
|
|
// If previous value is nullptr or delta is > than previous value,
|
|
|
|
// sets newValue with delta
|
|
|
|
// If previous value is not empty,
|
|
|
|
// updates previous value with 'b' string of previous value size - 1.
|
|
|
|
UpdateStatus DBTestBase::updateInPlaceSmallerSize(char* prevValue,
|
|
|
|
uint32_t* prevSize,
|
|
|
|
Slice delta,
|
|
|
|
std::string* newValue) {
|
|
|
|
if (prevValue == nullptr) {
|
|
|
|
*newValue = std::string(delta.size(), 'c');
|
|
|
|
return UpdateStatus::UPDATED;
|
|
|
|
} else {
|
|
|
|
*prevSize = *prevSize - 1;
|
|
|
|
std::string str_b = std::string(*prevSize, 'b');
|
|
|
|
memcpy(prevValue, str_b.c_str(), str_b.size());
|
|
|
|
return UpdateStatus::UPDATED_INPLACE;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
UpdateStatus DBTestBase::updateInPlaceSmallerVarintSize(char* prevValue,
|
|
|
|
uint32_t* prevSize,
|
|
|
|
Slice delta,
|
|
|
|
std::string* newValue) {
|
|
|
|
if (prevValue == nullptr) {
|
|
|
|
*newValue = std::string(delta.size(), 'c');
|
|
|
|
return UpdateStatus::UPDATED;
|
|
|
|
} else {
|
|
|
|
*prevSize = 1;
|
|
|
|
std::string str_b = std::string(*prevSize, 'b');
|
|
|
|
memcpy(prevValue, str_b.c_str(), str_b.size());
|
|
|
|
return UpdateStatus::UPDATED_INPLACE;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
UpdateStatus DBTestBase::updateInPlaceLargerSize(char* /*prevValue*/,
|
|
|
|
uint32_t* /*prevSize*/,
|
|
|
|
Slice delta,
|
|
|
|
std::string* newValue) {
|
|
|
|
*newValue = std::string(delta.size(), 'c');
|
|
|
|
return UpdateStatus::UPDATED;
|
|
|
|
}
|
|
|
|
|
|
|
|
UpdateStatus DBTestBase::updateInPlaceNoAction(char* /*prevValue*/,
|
|
|
|
uint32_t* /*prevSize*/,
|
|
|
|
Slice /*delta*/,
|
|
|
|
std::string* /*newValue*/) {
|
|
|
|
return UpdateStatus::UPDATE_FAILED;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Utility method to test InplaceUpdate
|
|
|
|
void DBTestBase::validateNumberOfEntries(int numValues, int cf) {
|
|
|
|
Arena arena;
|
|
|
|
auto options = CurrentOptions();
|
|
|
|
InternalKeyComparator icmp(options.comparator);
|
|
|
|
ReadOptions read_options;
|
|
|
|
ScopedArenaIterator iter;
|
|
|
|
if (cf != 0) {
|
Skip swaths of range tombstone covered keys in merging iterator (2022 edition) (#10449)
Summary:
Delete range logic is moved from `DBIter` to `MergingIterator`, and `MergingIterator` will seek to the end of a range deletion if possible instead of scanning through each key and check with `RangeDelAggregator`.
With the invariant that a key in level L (consider memtable as the first level, each immutable and L0 as a separate level) has a larger sequence number than all keys in any level >L, a range tombstone `[start, end)` from level L covers all keys in its range in any level >L. This property motivates optimizations in iterator:
- in `Seek(target)`, if level L has a range tombstone `[start, end)` that covers `target.UserKey`, then for all levels > L, we can do Seek() on `end` instead of `target` to skip some range tombstone covered keys.
- in `Next()/Prev()`, if the current key is covered by a range tombstone `[start, end)` from level L, we can do `Seek` to `end` for all levels > L.
This PR implements the above optimizations in `MergingIterator`. As all range tombstone covered keys are now skipped in `MergingIterator`, the range tombstone logic is removed from `DBIter`. The idea in this PR is similar to https://github.com/facebook/rocksdb/issues/7317, but this PR leaves `InternalIterator` interface mostly unchanged. **Credit**: the cascading seek optimization and the sentinel key (discussed below) are inspired by [Pebble](https://github.com/cockroachdb/pebble/blob/master/merging_iter.go) and suggested by ajkr in https://github.com/facebook/rocksdb/issues/7317. The two optimizations are mostly implemented in `SeekImpl()/SeekForPrevImpl()` and `IsNextDeleted()/IsPrevDeleted()` in `merging_iterator.cc`. See comments for each method for more detail.
One notable change is that the minHeap/maxHeap used by `MergingIterator` now contains range tombstone end keys besides point key iterators. This helps to reduce the number of key comparisons. For example, for a range tombstone `[start, end)`, a `start` and an `end` `HeapItem` are inserted into the heap. When a `HeapItem` for range tombstone start key is popped from the minHeap, we know this range tombstone becomes "active" in the sense that, before the range tombstone's end key is popped from the minHeap, all the keys popped from this heap is covered by the range tombstone's internal key range `[start, end)`.
Another major change, *delete range sentinel key*, is made to `LevelIterator`. Before this PR, when all point keys in an SST file are iterated through in `MergingIterator`, a level iterator would advance to the next SST file in its level. In the case when an SST file has a range tombstone that covers keys beyond the SST file's last point key, advancing to the next SST file would lose this range tombstone. Consequently, `MergingIterator` could return keys that should have been deleted by some range tombstone. We prevent this by pretending that file boundaries in each SST file are sentinel keys. A `LevelIterator` now only advance the file iterator once the sentinel key is processed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10449
Test Plan:
- Added many unit tests in db_range_del_test
- Stress test: `./db_stress --readpercent=5 --prefixpercent=19 --writepercent=20 -delpercent=10 --iterpercent=44 --delrangepercent=2`
- Additional iterator stress test is added to verify against iterators against expected state: https://github.com/facebook/rocksdb/issues/10538. This is based on ajkr's previous attempt https://github.com/facebook/rocksdb/pull/5506#issuecomment-506021913.
```
python3 ./tools/db_crashtest.py blackbox --simple --write_buffer_size=524288 --target_file_size_base=524288 --max_bytes_for_level_base=2097152 --compression_type=none --max_background_compactions=8 --value_size_mult=33 --max_key=5000000 --interval=10 --duration=7200 --delrangepercent=3 --delpercent=9 --iterpercent=25 --writepercent=60 --readpercent=3 --prefixpercent=0 --num_iterations=1000 --range_deletion_width=100 --verify_iterator_with_expected_state_one_in=1
```
- Performance benchmark: I used a similar setup as in the blog [post](http://rocksdb.org/blog/2018/11/21/delete-range.html) that introduced DeleteRange, "a database with 5 million data keys, and 10000 range tombstones (ignoring those dropped during compaction) that were written in regular intervals after 4.5 million data keys were written". As expected, the performance with this PR depends on the range tombstone width.
```
# Setup:
TEST_TMPDIR=/dev/shm ./db_bench_main --benchmarks=fillrandom --writes=4500000 --num=5000000
TEST_TMPDIR=/dev/shm ./db_bench_main --benchmarks=overwrite --writes=500000 --num=5000000 --use_existing_db=true --writes_per_range_tombstone=50
# Scan entire DB
TEST_TMPDIR=/dev/shm ./db_bench_main --benchmarks=readseq[-X5] --use_existing_db=true --num=5000000 --disable_auto_compactions=true
# Short range scan (10 Next())
TEST_TMPDIR=/dev/shm/width-100/ ./db_bench_main --benchmarks=seekrandom[-X5] --use_existing_db=true --num=500000 --reads=100000 --seek_nexts=10 --disable_auto_compactions=true
# Long range scan(1000 Next())
TEST_TMPDIR=/dev/shm/width-100/ ./db_bench_main --benchmarks=seekrandom[-X5] --use_existing_db=true --num=500000 --reads=2500 --seek_nexts=1000 --disable_auto_compactions=true
```
Avg over of 10 runs (some slower tests had fews runs):
For the first column (tombstone), 0 means no range tombstone, 100-10000 means width of the 10k range tombstones, and 1 means there is a single range tombstone in the entire DB (width is 1000). The 1 tombstone case is to test regression when there's very few range tombstones in the DB, as no range tombstone is likely to take a different code path than with range tombstones.
- Scan entire DB
| tombstone width | Pre-PR ops/sec | Post-PR ops/sec | ±% |
| ------------- | ------------- | ------------- | ------------- |
| 0 range tombstone |2525600 (± 43564) |2486917 (± 33698) |-1.53% |
| 100 |1853835 (± 24736) |2073884 (± 32176) |+11.87% |
| 1000 |422415 (± 7466) |1115801 (± 22781) |+164.15% |
| 10000 |22384 (± 227) |227919 (± 6647) |+918.22% |
| 1 range tombstone |2176540 (± 39050) |2434954 (± 24563) |+11.87% |
- Short range scan
| tombstone width | Pre-PR ops/sec | Post-PR ops/sec | ±% |
| ------------- | ------------- | ------------- | ------------- |
| 0 range tombstone |35398 (± 533) |35338 (± 569) |-0.17% |
| 100 |28276 (± 664) |31684 (± 331) |+12.05% |
| 1000 |7637 (± 77) |25422 (± 277) |+232.88% |
| 10000 |1367 |28667 |+1997.07% |
| 1 range tombstone |32618 (± 581) |32748 (± 506) |+0.4% |
- Long range scan
| tombstone width | Pre-PR ops/sec | Post-PR ops/sec | ±% |
| ------------- | ------------- | ------------- | ------------- |
| 0 range tombstone |2262 (± 33) |2353 (± 20) |+4.02% |
| 100 |1696 (± 26) |1926 (± 18) |+13.56% |
| 1000 |410 (± 6) |1255 (± 29) |+206.1% |
| 10000 |25 |414 |+1556.0% |
| 1 range tombstone |1957 (± 30) |2185 (± 44) |+11.65% |
- Microbench does not show significant regression: https://gist.github.com/cbi42/59f280f85a59b678e7e5d8561e693b61
Reviewed By: ajkr
Differential Revision: D38450331
Pulled By: cbi42
fbshipit-source-id: b5ef12e8d8c289ed2e163ccdf277f5039b511fca
2 years ago
|
|
|
iter.set(dbfull()->NewInternalIterator(read_options, &arena,
|
|
|
|
kMaxSequenceNumber, handles_[cf]));
|
|
|
|
} else {
|
Skip swaths of range tombstone covered keys in merging iterator (2022 edition) (#10449)
Summary:
Delete range logic is moved from `DBIter` to `MergingIterator`, and `MergingIterator` will seek to the end of a range deletion if possible instead of scanning through each key and check with `RangeDelAggregator`.
With the invariant that a key in level L (consider memtable as the first level, each immutable and L0 as a separate level) has a larger sequence number than all keys in any level >L, a range tombstone `[start, end)` from level L covers all keys in its range in any level >L. This property motivates optimizations in iterator:
- in `Seek(target)`, if level L has a range tombstone `[start, end)` that covers `target.UserKey`, then for all levels > L, we can do Seek() on `end` instead of `target` to skip some range tombstone covered keys.
- in `Next()/Prev()`, if the current key is covered by a range tombstone `[start, end)` from level L, we can do `Seek` to `end` for all levels > L.
This PR implements the above optimizations in `MergingIterator`. As all range tombstone covered keys are now skipped in `MergingIterator`, the range tombstone logic is removed from `DBIter`. The idea in this PR is similar to https://github.com/facebook/rocksdb/issues/7317, but this PR leaves `InternalIterator` interface mostly unchanged. **Credit**: the cascading seek optimization and the sentinel key (discussed below) are inspired by [Pebble](https://github.com/cockroachdb/pebble/blob/master/merging_iter.go) and suggested by ajkr in https://github.com/facebook/rocksdb/issues/7317. The two optimizations are mostly implemented in `SeekImpl()/SeekForPrevImpl()` and `IsNextDeleted()/IsPrevDeleted()` in `merging_iterator.cc`. See comments for each method for more detail.
One notable change is that the minHeap/maxHeap used by `MergingIterator` now contains range tombstone end keys besides point key iterators. This helps to reduce the number of key comparisons. For example, for a range tombstone `[start, end)`, a `start` and an `end` `HeapItem` are inserted into the heap. When a `HeapItem` for range tombstone start key is popped from the minHeap, we know this range tombstone becomes "active" in the sense that, before the range tombstone's end key is popped from the minHeap, all the keys popped from this heap is covered by the range tombstone's internal key range `[start, end)`.
Another major change, *delete range sentinel key*, is made to `LevelIterator`. Before this PR, when all point keys in an SST file are iterated through in `MergingIterator`, a level iterator would advance to the next SST file in its level. In the case when an SST file has a range tombstone that covers keys beyond the SST file's last point key, advancing to the next SST file would lose this range tombstone. Consequently, `MergingIterator` could return keys that should have been deleted by some range tombstone. We prevent this by pretending that file boundaries in each SST file are sentinel keys. A `LevelIterator` now only advance the file iterator once the sentinel key is processed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10449
Test Plan:
- Added many unit tests in db_range_del_test
- Stress test: `./db_stress --readpercent=5 --prefixpercent=19 --writepercent=20 -delpercent=10 --iterpercent=44 --delrangepercent=2`
- Additional iterator stress test is added to verify against iterators against expected state: https://github.com/facebook/rocksdb/issues/10538. This is based on ajkr's previous attempt https://github.com/facebook/rocksdb/pull/5506#issuecomment-506021913.
```
python3 ./tools/db_crashtest.py blackbox --simple --write_buffer_size=524288 --target_file_size_base=524288 --max_bytes_for_level_base=2097152 --compression_type=none --max_background_compactions=8 --value_size_mult=33 --max_key=5000000 --interval=10 --duration=7200 --delrangepercent=3 --delpercent=9 --iterpercent=25 --writepercent=60 --readpercent=3 --prefixpercent=0 --num_iterations=1000 --range_deletion_width=100 --verify_iterator_with_expected_state_one_in=1
```
- Performance benchmark: I used a similar setup as in the blog [post](http://rocksdb.org/blog/2018/11/21/delete-range.html) that introduced DeleteRange, "a database with 5 million data keys, and 10000 range tombstones (ignoring those dropped during compaction) that were written in regular intervals after 4.5 million data keys were written". As expected, the performance with this PR depends on the range tombstone width.
```
# Setup:
TEST_TMPDIR=/dev/shm ./db_bench_main --benchmarks=fillrandom --writes=4500000 --num=5000000
TEST_TMPDIR=/dev/shm ./db_bench_main --benchmarks=overwrite --writes=500000 --num=5000000 --use_existing_db=true --writes_per_range_tombstone=50
# Scan entire DB
TEST_TMPDIR=/dev/shm ./db_bench_main --benchmarks=readseq[-X5] --use_existing_db=true --num=5000000 --disable_auto_compactions=true
# Short range scan (10 Next())
TEST_TMPDIR=/dev/shm/width-100/ ./db_bench_main --benchmarks=seekrandom[-X5] --use_existing_db=true --num=500000 --reads=100000 --seek_nexts=10 --disable_auto_compactions=true
# Long range scan(1000 Next())
TEST_TMPDIR=/dev/shm/width-100/ ./db_bench_main --benchmarks=seekrandom[-X5] --use_existing_db=true --num=500000 --reads=2500 --seek_nexts=1000 --disable_auto_compactions=true
```
Avg over of 10 runs (some slower tests had fews runs):
For the first column (tombstone), 0 means no range tombstone, 100-10000 means width of the 10k range tombstones, and 1 means there is a single range tombstone in the entire DB (width is 1000). The 1 tombstone case is to test regression when there's very few range tombstones in the DB, as no range tombstone is likely to take a different code path than with range tombstones.
- Scan entire DB
| tombstone width | Pre-PR ops/sec | Post-PR ops/sec | ±% |
| ------------- | ------------- | ------------- | ------------- |
| 0 range tombstone |2525600 (± 43564) |2486917 (± 33698) |-1.53% |
| 100 |1853835 (± 24736) |2073884 (± 32176) |+11.87% |
| 1000 |422415 (± 7466) |1115801 (± 22781) |+164.15% |
| 10000 |22384 (± 227) |227919 (± 6647) |+918.22% |
| 1 range tombstone |2176540 (± 39050) |2434954 (± 24563) |+11.87% |
- Short range scan
| tombstone width | Pre-PR ops/sec | Post-PR ops/sec | ±% |
| ------------- | ------------- | ------------- | ------------- |
| 0 range tombstone |35398 (± 533) |35338 (± 569) |-0.17% |
| 100 |28276 (± 664) |31684 (± 331) |+12.05% |
| 1000 |7637 (± 77) |25422 (± 277) |+232.88% |
| 10000 |1367 |28667 |+1997.07% |
| 1 range tombstone |32618 (± 581) |32748 (± 506) |+0.4% |
- Long range scan
| tombstone width | Pre-PR ops/sec | Post-PR ops/sec | ±% |
| ------------- | ------------- | ------------- | ------------- |
| 0 range tombstone |2262 (± 33) |2353 (± 20) |+4.02% |
| 100 |1696 (± 26) |1926 (± 18) |+13.56% |
| 1000 |410 (± 6) |1255 (± 29) |+206.1% |
| 10000 |25 |414 |+1556.0% |
| 1 range tombstone |1957 (± 30) |2185 (± 44) |+11.65% |
- Microbench does not show significant regression: https://gist.github.com/cbi42/59f280f85a59b678e7e5d8561e693b61
Reviewed By: ajkr
Differential Revision: D38450331
Pulled By: cbi42
fbshipit-source-id: b5ef12e8d8c289ed2e163ccdf277f5039b511fca
2 years ago
|
|
|
iter.set(dbfull()->NewInternalIterator(read_options, &arena,
|
|
|
|
kMaxSequenceNumber));
|
|
|
|
}
|
|
|
|
iter->SeekToFirst();
|
|
|
|
ASSERT_OK(iter->status());
|
|
|
|
int seq = numValues;
|
|
|
|
while (iter->Valid()) {
|
|
|
|
ParsedInternalKey ikey;
|
|
|
|
ikey.clear();
|
|
|
|
ASSERT_OK(ParseInternalKey(iter->key(), &ikey, true /* log_err_key */));
|
|
|
|
|
|
|
|
// checks sequence number for updates
|
|
|
|
ASSERT_EQ(ikey.sequence, (unsigned)seq--);
|
|
|
|
iter->Next();
|
|
|
|
}
|
|
|
|
ASSERT_EQ(0, seq);
|
|
|
|
}
|
|
|
|
|
|
|
|
void DBTestBase::CopyFile(const std::string& source,
|
|
|
|
const std::string& destination, uint64_t size) {
|
|
|
|
const EnvOptions soptions;
|
|
|
|
std::unique_ptr<SequentialFile> srcfile;
|
|
|
|
ASSERT_OK(env_->NewSequentialFile(source, &srcfile, soptions));
|
|
|
|
std::unique_ptr<WritableFile> destfile;
|
|
|
|
ASSERT_OK(env_->NewWritableFile(destination, &destfile, soptions));
|
|
|
|
|
|
|
|
if (size == 0) {
|
|
|
|
// default argument means copy everything
|
|
|
|
ASSERT_OK(env_->GetFileSize(source, &size));
|
|
|
|
}
|
|
|
|
|
|
|
|
char buffer[4096];
|
|
|
|
Slice slice;
|
|
|
|
while (size > 0) {
|
|
|
|
uint64_t one = std::min(uint64_t(sizeof(buffer)), size);
|
|
|
|
ASSERT_OK(srcfile->Read(one, &slice, buffer));
|
|
|
|
ASSERT_OK(destfile->Append(slice));
|
|
|
|
size -= slice.size();
|
|
|
|
}
|
|
|
|
ASSERT_OK(destfile->Close());
|
|
|
|
}
|
|
|
|
|
|
|
|
Status DBTestBase::GetAllDataFiles(
|
|
|
|
const FileType file_type, std::unordered_map<std::string, uint64_t>* files,
|
|
|
|
uint64_t* total_size /* = nullptr */) {
|
|
|
|
if (total_size) {
|
|
|
|
*total_size = 0;
|
|
|
|
}
|
|
|
|
std::vector<std::string> children;
|
|
|
|
Status s = env_->GetChildren(dbname_, &children);
|
|
|
|
if (s.ok()) {
|
|
|
|
for (auto& file_name : children) {
|
|
|
|
uint64_t number;
|
|
|
|
FileType type;
|
|
|
|
if (ParseFileName(file_name, &number, &type) && type == file_type) {
|
|
|
|
std::string file_path = dbname_ + "/" + file_name;
|
|
|
|
uint64_t file_size = 0;
|
|
|
|
s = env_->GetFileSize(file_path, &file_size);
|
|
|
|
if (!s.ok()) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
(*files)[file_path] = file_size;
|
|
|
|
if (total_size) {
|
|
|
|
*total_size += file_size;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
std::vector<std::uint64_t> DBTestBase::ListTableFiles(Env* env,
|
|
|
|
const std::string& path) {
|
|
|
|
std::vector<std::string> files;
|
|
|
|
std::vector<uint64_t> file_numbers;
|
|
|
|
EXPECT_OK(env->GetChildren(path, &files));
|
|
|
|
uint64_t number;
|
|
|
|
FileType type;
|
|
|
|
for (size_t i = 0; i < files.size(); ++i) {
|
|
|
|
if (ParseFileName(files[i], &number, &type)) {
|
|
|
|
if (type == kTableFile) {
|
|
|
|
file_numbers.push_back(number);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return file_numbers;
|
|
|
|
}
|
|
|
|
|
|
|
|
void DBTestBase::VerifyDBFromMap(std::map<std::string, std::string> true_data,
|
|
|
|
size_t* total_reads_res, bool tailing_iter,
|
|
|
|
std::map<std::string, Status> status) {
|
|
|
|
size_t total_reads = 0;
|
|
|
|
|
Introduce FullMergeV2 (eliminate memcpy from merge operators)
Summary:
This diff update the code to pin the merge operator operands while the merge operation is done, so that we can eliminate the memcpy cost, to do that we need a new public API for FullMerge that replace the std::deque<std::string> with std::vector<Slice>
This diff is stacked on top of D56493 and D56511
In this diff we
- Update FullMergeV2 arguments to be encapsulated in MergeOperationInput and MergeOperationOutput which will make it easier to add new arguments in the future
- Replace std::deque<std::string> with std::vector<Slice> to pass operands
- Replace MergeContext std::deque with std::vector (based on a simple benchmark I ran https://gist.github.com/IslamAbdelRahman/78fc86c9ab9f52b1df791e58943fb187)
- Allow FullMergeV2 output to be an existing operand
```
[Everything in Memtable | 10K operands | 10 KB each | 1 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --merge_keys=10000 --num=10000 --disable_auto_compactions --value_size=10240 --write_buffer_size=1000000000
[FullMergeV2]
readseq : 0.607 micros/op 1648235 ops/sec; 16121.2 MB/s
readseq : 0.478 micros/op 2091546 ops/sec; 20457.2 MB/s
readseq : 0.252 micros/op 3972081 ops/sec; 38850.5 MB/s
readseq : 0.237 micros/op 4218328 ops/sec; 41259.0 MB/s
readseq : 0.247 micros/op 4043927 ops/sec; 39553.2 MB/s
[master]
readseq : 3.935 micros/op 254140 ops/sec; 2485.7 MB/s
readseq : 3.722 micros/op 268657 ops/sec; 2627.7 MB/s
readseq : 3.149 micros/op 317605 ops/sec; 3106.5 MB/s
readseq : 3.125 micros/op 320024 ops/sec; 3130.1 MB/s
readseq : 4.075 micros/op 245374 ops/sec; 2400.0 MB/s
```
```
[Everything in Memtable | 10K operands | 10 KB each | 10 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --merge_keys=1000 --num=10000 --disable_auto_compactions --value_size=10240 --write_buffer_size=1000000000
[FullMergeV2]
readseq : 3.472 micros/op 288018 ops/sec; 2817.1 MB/s
readseq : 2.304 micros/op 434027 ops/sec; 4245.2 MB/s
readseq : 1.163 micros/op 859845 ops/sec; 8410.0 MB/s
readseq : 1.192 micros/op 838926 ops/sec; 8205.4 MB/s
readseq : 1.250 micros/op 800000 ops/sec; 7824.7 MB/s
[master]
readseq : 24.025 micros/op 41623 ops/sec; 407.1 MB/s
readseq : 18.489 micros/op 54086 ops/sec; 529.0 MB/s
readseq : 18.693 micros/op 53495 ops/sec; 523.2 MB/s
readseq : 23.621 micros/op 42335 ops/sec; 414.1 MB/s
readseq : 18.775 micros/op 53262 ops/sec; 521.0 MB/s
```
```
[Everything in Block cache | 10K operands | 10 KB each | 1 operand per key]
[FullMergeV2]
$ DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --num=100000 --db="/dev/shm/merge-random-10K-10KB" --cache_size=1000000000 --use_existing_db --disable_auto_compactions
readseq : 14.741 micros/op 67837 ops/sec; 663.5 MB/s
readseq : 1.029 micros/op 971446 ops/sec; 9501.6 MB/s
readseq : 0.974 micros/op 1026229 ops/sec; 10037.4 MB/s
readseq : 0.965 micros/op 1036080 ops/sec; 10133.8 MB/s
readseq : 0.943 micros/op 1060657 ops/sec; 10374.2 MB/s
[master]
readseq : 16.735 micros/op 59755 ops/sec; 584.5 MB/s
readseq : 3.029 micros/op 330151 ops/sec; 3229.2 MB/s
readseq : 3.136 micros/op 318883 ops/sec; 3119.0 MB/s
readseq : 3.065 micros/op 326245 ops/sec; 3191.0 MB/s
readseq : 3.014 micros/op 331813 ops/sec; 3245.4 MB/s
```
```
[Everything in Block cache | 10K operands | 10 KB each | 10 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --num=100000 --db="/dev/shm/merge-random-10-operands-10K-10KB" --cache_size=1000000000 --use_existing_db --disable_auto_compactions
[FullMergeV2]
readseq : 24.325 micros/op 41109 ops/sec; 402.1 MB/s
readseq : 1.470 micros/op 680272 ops/sec; 6653.7 MB/s
readseq : 1.231 micros/op 812347 ops/sec; 7945.5 MB/s
readseq : 1.091 micros/op 916590 ops/sec; 8965.1 MB/s
readseq : 1.109 micros/op 901713 ops/sec; 8819.6 MB/s
[master]
readseq : 27.257 micros/op 36687 ops/sec; 358.8 MB/s
readseq : 4.443 micros/op 225073 ops/sec; 2201.4 MB/s
readseq : 5.830 micros/op 171526 ops/sec; 1677.7 MB/s
readseq : 4.173 micros/op 239635 ops/sec; 2343.8 MB/s
readseq : 4.150 micros/op 240963 ops/sec; 2356.8 MB/s
```
Test Plan: COMPILE_WITH_ASAN=1 make check -j64
Reviewers: yhchiang, andrewkr, sdong
Reviewed By: sdong
Subscribers: lovro, andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D57075
9 years ago
|
|
|
for (auto& kv : true_data) {
|
|
|
|
Status s = status[kv.first];
|
|
|
|
if (s.ok()) {
|
|
|
|
ASSERT_EQ(Get(kv.first), kv.second);
|
|
|
|
} else {
|
|
|
|
std::string value;
|
|
|
|
ASSERT_EQ(s, db_->Get(ReadOptions(), kv.first, &value));
|
|
|
|
}
|
|
|
|
total_reads++;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Normal Iterator
|
|
|
|
{
|
|
|
|
int iter_cnt = 0;
|
|
|
|
ReadOptions ro;
|
|
|
|
ro.total_order_seek = true;
|
|
|
|
Iterator* iter = db_->NewIterator(ro);
|
|
|
|
// Verify Iterator::Next()
|
|
|
|
iter_cnt = 0;
|
|
|
|
auto data_iter = true_data.begin();
|
|
|
|
Status s;
|
|
|
|
for (iter->SeekToFirst(); iter->Valid(); iter->Next(), data_iter++) {
|
|
|
|
ASSERT_EQ(iter->key().ToString(), data_iter->first);
|
|
|
|
Status current_status = status[data_iter->first];
|
|
|
|
if (!current_status.ok()) {
|
|
|
|
s = current_status;
|
|
|
|
}
|
|
|
|
ASSERT_EQ(iter->status(), s);
|
|
|
|
if (current_status.ok()) {
|
|
|
|
ASSERT_EQ(iter->value().ToString(), data_iter->second);
|
|
|
|
}
|
|
|
|
iter_cnt++;
|
|
|
|
total_reads++;
|
|
|
|
}
|
|
|
|
ASSERT_EQ(data_iter, true_data.end())
|
|
|
|
<< iter_cnt << " / " << true_data.size();
|
|
|
|
delete iter;
|
|
|
|
|
|
|
|
// Verify Iterator::Prev()
|
|
|
|
// Use a new iterator to make sure its status is clean.
|
|
|
|
iter = db_->NewIterator(ro);
|
|
|
|
iter_cnt = 0;
|
|
|
|
s = Status::OK();
|
|
|
|
auto data_rev = true_data.rbegin();
|
|
|
|
for (iter->SeekToLast(); iter->Valid(); iter->Prev(), data_rev++) {
|
|
|
|
ASSERT_EQ(iter->key().ToString(), data_rev->first);
|
|
|
|
Status current_status = status[data_rev->first];
|
|
|
|
if (!current_status.ok()) {
|
|
|
|
s = current_status;
|
|
|
|
}
|
|
|
|
ASSERT_EQ(iter->status(), s);
|
|
|
|
if (current_status.ok()) {
|
|
|
|
ASSERT_EQ(iter->value().ToString(), data_rev->second);
|
|
|
|
}
|
|
|
|
iter_cnt++;
|
|
|
|
total_reads++;
|
|
|
|
}
|
|
|
|
ASSERT_EQ(data_rev, true_data.rend())
|
|
|
|
<< iter_cnt << " / " << true_data.size();
|
|
|
|
|
|
|
|
// Verify Iterator::Seek()
|
|
|
|
for (auto kv : true_data) {
|
|
|
|
iter->Seek(kv.first);
|
|
|
|
ASSERT_EQ(kv.first, iter->key().ToString());
|
|
|
|
ASSERT_EQ(kv.second, iter->value().ToString());
|
|
|
|
total_reads++;
|
|
|
|
}
|
|
|
|
delete iter;
|
Introduce FullMergeV2 (eliminate memcpy from merge operators)
Summary:
This diff update the code to pin the merge operator operands while the merge operation is done, so that we can eliminate the memcpy cost, to do that we need a new public API for FullMerge that replace the std::deque<std::string> with std::vector<Slice>
This diff is stacked on top of D56493 and D56511
In this diff we
- Update FullMergeV2 arguments to be encapsulated in MergeOperationInput and MergeOperationOutput which will make it easier to add new arguments in the future
- Replace std::deque<std::string> with std::vector<Slice> to pass operands
- Replace MergeContext std::deque with std::vector (based on a simple benchmark I ran https://gist.github.com/IslamAbdelRahman/78fc86c9ab9f52b1df791e58943fb187)
- Allow FullMergeV2 output to be an existing operand
```
[Everything in Memtable | 10K operands | 10 KB each | 1 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --merge_keys=10000 --num=10000 --disable_auto_compactions --value_size=10240 --write_buffer_size=1000000000
[FullMergeV2]
readseq : 0.607 micros/op 1648235 ops/sec; 16121.2 MB/s
readseq : 0.478 micros/op 2091546 ops/sec; 20457.2 MB/s
readseq : 0.252 micros/op 3972081 ops/sec; 38850.5 MB/s
readseq : 0.237 micros/op 4218328 ops/sec; 41259.0 MB/s
readseq : 0.247 micros/op 4043927 ops/sec; 39553.2 MB/s
[master]
readseq : 3.935 micros/op 254140 ops/sec; 2485.7 MB/s
readseq : 3.722 micros/op 268657 ops/sec; 2627.7 MB/s
readseq : 3.149 micros/op 317605 ops/sec; 3106.5 MB/s
readseq : 3.125 micros/op 320024 ops/sec; 3130.1 MB/s
readseq : 4.075 micros/op 245374 ops/sec; 2400.0 MB/s
```
```
[Everything in Memtable | 10K operands | 10 KB each | 10 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --merge_keys=1000 --num=10000 --disable_auto_compactions --value_size=10240 --write_buffer_size=1000000000
[FullMergeV2]
readseq : 3.472 micros/op 288018 ops/sec; 2817.1 MB/s
readseq : 2.304 micros/op 434027 ops/sec; 4245.2 MB/s
readseq : 1.163 micros/op 859845 ops/sec; 8410.0 MB/s
readseq : 1.192 micros/op 838926 ops/sec; 8205.4 MB/s
readseq : 1.250 micros/op 800000 ops/sec; 7824.7 MB/s
[master]
readseq : 24.025 micros/op 41623 ops/sec; 407.1 MB/s
readseq : 18.489 micros/op 54086 ops/sec; 529.0 MB/s
readseq : 18.693 micros/op 53495 ops/sec; 523.2 MB/s
readseq : 23.621 micros/op 42335 ops/sec; 414.1 MB/s
readseq : 18.775 micros/op 53262 ops/sec; 521.0 MB/s
```
```
[Everything in Block cache | 10K operands | 10 KB each | 1 operand per key]
[FullMergeV2]
$ DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --num=100000 --db="/dev/shm/merge-random-10K-10KB" --cache_size=1000000000 --use_existing_db --disable_auto_compactions
readseq : 14.741 micros/op 67837 ops/sec; 663.5 MB/s
readseq : 1.029 micros/op 971446 ops/sec; 9501.6 MB/s
readseq : 0.974 micros/op 1026229 ops/sec; 10037.4 MB/s
readseq : 0.965 micros/op 1036080 ops/sec; 10133.8 MB/s
readseq : 0.943 micros/op 1060657 ops/sec; 10374.2 MB/s
[master]
readseq : 16.735 micros/op 59755 ops/sec; 584.5 MB/s
readseq : 3.029 micros/op 330151 ops/sec; 3229.2 MB/s
readseq : 3.136 micros/op 318883 ops/sec; 3119.0 MB/s
readseq : 3.065 micros/op 326245 ops/sec; 3191.0 MB/s
readseq : 3.014 micros/op 331813 ops/sec; 3245.4 MB/s
```
```
[Everything in Block cache | 10K operands | 10 KB each | 10 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --num=100000 --db="/dev/shm/merge-random-10-operands-10K-10KB" --cache_size=1000000000 --use_existing_db --disable_auto_compactions
[FullMergeV2]
readseq : 24.325 micros/op 41109 ops/sec; 402.1 MB/s
readseq : 1.470 micros/op 680272 ops/sec; 6653.7 MB/s
readseq : 1.231 micros/op 812347 ops/sec; 7945.5 MB/s
readseq : 1.091 micros/op 916590 ops/sec; 8965.1 MB/s
readseq : 1.109 micros/op 901713 ops/sec; 8819.6 MB/s
[master]
readseq : 27.257 micros/op 36687 ops/sec; 358.8 MB/s
readseq : 4.443 micros/op 225073 ops/sec; 2201.4 MB/s
readseq : 5.830 micros/op 171526 ops/sec; 1677.7 MB/s
readseq : 4.173 micros/op 239635 ops/sec; 2343.8 MB/s
readseq : 4.150 micros/op 240963 ops/sec; 2356.8 MB/s
```
Test Plan: COMPILE_WITH_ASAN=1 make check -j64
Reviewers: yhchiang, andrewkr, sdong
Reviewed By: sdong
Subscribers: lovro, andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D57075
9 years ago
|
|
|
}
|
|
|
|
|
|
|
|
if (tailing_iter) {
|
|
|
|
// Tailing iterator
|
|
|
|
int iter_cnt = 0;
|
|
|
|
ReadOptions ro;
|
|
|
|
ro.tailing = true;
|
|
|
|
ro.total_order_seek = true;
|
|
|
|
Iterator* iter = db_->NewIterator(ro);
|
Introduce FullMergeV2 (eliminate memcpy from merge operators)
Summary:
This diff update the code to pin the merge operator operands while the merge operation is done, so that we can eliminate the memcpy cost, to do that we need a new public API for FullMerge that replace the std::deque<std::string> with std::vector<Slice>
This diff is stacked on top of D56493 and D56511
In this diff we
- Update FullMergeV2 arguments to be encapsulated in MergeOperationInput and MergeOperationOutput which will make it easier to add new arguments in the future
- Replace std::deque<std::string> with std::vector<Slice> to pass operands
- Replace MergeContext std::deque with std::vector (based on a simple benchmark I ran https://gist.github.com/IslamAbdelRahman/78fc86c9ab9f52b1df791e58943fb187)
- Allow FullMergeV2 output to be an existing operand
```
[Everything in Memtable | 10K operands | 10 KB each | 1 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --merge_keys=10000 --num=10000 --disable_auto_compactions --value_size=10240 --write_buffer_size=1000000000
[FullMergeV2]
readseq : 0.607 micros/op 1648235 ops/sec; 16121.2 MB/s
readseq : 0.478 micros/op 2091546 ops/sec; 20457.2 MB/s
readseq : 0.252 micros/op 3972081 ops/sec; 38850.5 MB/s
readseq : 0.237 micros/op 4218328 ops/sec; 41259.0 MB/s
readseq : 0.247 micros/op 4043927 ops/sec; 39553.2 MB/s
[master]
readseq : 3.935 micros/op 254140 ops/sec; 2485.7 MB/s
readseq : 3.722 micros/op 268657 ops/sec; 2627.7 MB/s
readseq : 3.149 micros/op 317605 ops/sec; 3106.5 MB/s
readseq : 3.125 micros/op 320024 ops/sec; 3130.1 MB/s
readseq : 4.075 micros/op 245374 ops/sec; 2400.0 MB/s
```
```
[Everything in Memtable | 10K operands | 10 KB each | 10 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --merge_keys=1000 --num=10000 --disable_auto_compactions --value_size=10240 --write_buffer_size=1000000000
[FullMergeV2]
readseq : 3.472 micros/op 288018 ops/sec; 2817.1 MB/s
readseq : 2.304 micros/op 434027 ops/sec; 4245.2 MB/s
readseq : 1.163 micros/op 859845 ops/sec; 8410.0 MB/s
readseq : 1.192 micros/op 838926 ops/sec; 8205.4 MB/s
readseq : 1.250 micros/op 800000 ops/sec; 7824.7 MB/s
[master]
readseq : 24.025 micros/op 41623 ops/sec; 407.1 MB/s
readseq : 18.489 micros/op 54086 ops/sec; 529.0 MB/s
readseq : 18.693 micros/op 53495 ops/sec; 523.2 MB/s
readseq : 23.621 micros/op 42335 ops/sec; 414.1 MB/s
readseq : 18.775 micros/op 53262 ops/sec; 521.0 MB/s
```
```
[Everything in Block cache | 10K operands | 10 KB each | 1 operand per key]
[FullMergeV2]
$ DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --num=100000 --db="/dev/shm/merge-random-10K-10KB" --cache_size=1000000000 --use_existing_db --disable_auto_compactions
readseq : 14.741 micros/op 67837 ops/sec; 663.5 MB/s
readseq : 1.029 micros/op 971446 ops/sec; 9501.6 MB/s
readseq : 0.974 micros/op 1026229 ops/sec; 10037.4 MB/s
readseq : 0.965 micros/op 1036080 ops/sec; 10133.8 MB/s
readseq : 0.943 micros/op 1060657 ops/sec; 10374.2 MB/s
[master]
readseq : 16.735 micros/op 59755 ops/sec; 584.5 MB/s
readseq : 3.029 micros/op 330151 ops/sec; 3229.2 MB/s
readseq : 3.136 micros/op 318883 ops/sec; 3119.0 MB/s
readseq : 3.065 micros/op 326245 ops/sec; 3191.0 MB/s
readseq : 3.014 micros/op 331813 ops/sec; 3245.4 MB/s
```
```
[Everything in Block cache | 10K operands | 10 KB each | 10 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --num=100000 --db="/dev/shm/merge-random-10-operands-10K-10KB" --cache_size=1000000000 --use_existing_db --disable_auto_compactions
[FullMergeV2]
readseq : 24.325 micros/op 41109 ops/sec; 402.1 MB/s
readseq : 1.470 micros/op 680272 ops/sec; 6653.7 MB/s
readseq : 1.231 micros/op 812347 ops/sec; 7945.5 MB/s
readseq : 1.091 micros/op 916590 ops/sec; 8965.1 MB/s
readseq : 1.109 micros/op 901713 ops/sec; 8819.6 MB/s
[master]
readseq : 27.257 micros/op 36687 ops/sec; 358.8 MB/s
readseq : 4.443 micros/op 225073 ops/sec; 2201.4 MB/s
readseq : 5.830 micros/op 171526 ops/sec; 1677.7 MB/s
readseq : 4.173 micros/op 239635 ops/sec; 2343.8 MB/s
readseq : 4.150 micros/op 240963 ops/sec; 2356.8 MB/s
```
Test Plan: COMPILE_WITH_ASAN=1 make check -j64
Reviewers: yhchiang, andrewkr, sdong
Reviewed By: sdong
Subscribers: lovro, andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D57075
9 years ago
|
|
|
|
|
|
|
// Verify ForwardIterator::Next()
|
|
|
|
iter_cnt = 0;
|
|
|
|
auto data_iter = true_data.begin();
|
|
|
|
for (iter->SeekToFirst(); iter->Valid(); iter->Next(), data_iter++) {
|
|
|
|
ASSERT_EQ(iter->key().ToString(), data_iter->first);
|
|
|
|
ASSERT_EQ(iter->value().ToString(), data_iter->second);
|
|
|
|
iter_cnt++;
|
|
|
|
total_reads++;
|
|
|
|
}
|
|
|
|
ASSERT_EQ(data_iter, true_data.end())
|
|
|
|
<< iter_cnt << " / " << true_data.size();
|
|
|
|
|
|
|
|
// Verify ForwardIterator::Seek()
|
|
|
|
for (auto kv : true_data) {
|
|
|
|
iter->Seek(kv.first);
|
|
|
|
ASSERT_EQ(kv.first, iter->key().ToString());
|
|
|
|
ASSERT_EQ(kv.second, iter->value().ToString());
|
|
|
|
total_reads++;
|
|
|
|
}
|
|
|
|
|
|
|
|
delete iter;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (total_reads_res) {
|
|
|
|
*total_reads_res = total_reads;
|
|
|
|
}
|
Introduce FullMergeV2 (eliminate memcpy from merge operators)
Summary:
This diff update the code to pin the merge operator operands while the merge operation is done, so that we can eliminate the memcpy cost, to do that we need a new public API for FullMerge that replace the std::deque<std::string> with std::vector<Slice>
This diff is stacked on top of D56493 and D56511
In this diff we
- Update FullMergeV2 arguments to be encapsulated in MergeOperationInput and MergeOperationOutput which will make it easier to add new arguments in the future
- Replace std::deque<std::string> with std::vector<Slice> to pass operands
- Replace MergeContext std::deque with std::vector (based on a simple benchmark I ran https://gist.github.com/IslamAbdelRahman/78fc86c9ab9f52b1df791e58943fb187)
- Allow FullMergeV2 output to be an existing operand
```
[Everything in Memtable | 10K operands | 10 KB each | 1 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --merge_keys=10000 --num=10000 --disable_auto_compactions --value_size=10240 --write_buffer_size=1000000000
[FullMergeV2]
readseq : 0.607 micros/op 1648235 ops/sec; 16121.2 MB/s
readseq : 0.478 micros/op 2091546 ops/sec; 20457.2 MB/s
readseq : 0.252 micros/op 3972081 ops/sec; 38850.5 MB/s
readseq : 0.237 micros/op 4218328 ops/sec; 41259.0 MB/s
readseq : 0.247 micros/op 4043927 ops/sec; 39553.2 MB/s
[master]
readseq : 3.935 micros/op 254140 ops/sec; 2485.7 MB/s
readseq : 3.722 micros/op 268657 ops/sec; 2627.7 MB/s
readseq : 3.149 micros/op 317605 ops/sec; 3106.5 MB/s
readseq : 3.125 micros/op 320024 ops/sec; 3130.1 MB/s
readseq : 4.075 micros/op 245374 ops/sec; 2400.0 MB/s
```
```
[Everything in Memtable | 10K operands | 10 KB each | 10 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="mergerandom,readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --merge_keys=1000 --num=10000 --disable_auto_compactions --value_size=10240 --write_buffer_size=1000000000
[FullMergeV2]
readseq : 3.472 micros/op 288018 ops/sec; 2817.1 MB/s
readseq : 2.304 micros/op 434027 ops/sec; 4245.2 MB/s
readseq : 1.163 micros/op 859845 ops/sec; 8410.0 MB/s
readseq : 1.192 micros/op 838926 ops/sec; 8205.4 MB/s
readseq : 1.250 micros/op 800000 ops/sec; 7824.7 MB/s
[master]
readseq : 24.025 micros/op 41623 ops/sec; 407.1 MB/s
readseq : 18.489 micros/op 54086 ops/sec; 529.0 MB/s
readseq : 18.693 micros/op 53495 ops/sec; 523.2 MB/s
readseq : 23.621 micros/op 42335 ops/sec; 414.1 MB/s
readseq : 18.775 micros/op 53262 ops/sec; 521.0 MB/s
```
```
[Everything in Block cache | 10K operands | 10 KB each | 1 operand per key]
[FullMergeV2]
$ DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --num=100000 --db="/dev/shm/merge-random-10K-10KB" --cache_size=1000000000 --use_existing_db --disable_auto_compactions
readseq : 14.741 micros/op 67837 ops/sec; 663.5 MB/s
readseq : 1.029 micros/op 971446 ops/sec; 9501.6 MB/s
readseq : 0.974 micros/op 1026229 ops/sec; 10037.4 MB/s
readseq : 0.965 micros/op 1036080 ops/sec; 10133.8 MB/s
readseq : 0.943 micros/op 1060657 ops/sec; 10374.2 MB/s
[master]
readseq : 16.735 micros/op 59755 ops/sec; 584.5 MB/s
readseq : 3.029 micros/op 330151 ops/sec; 3229.2 MB/s
readseq : 3.136 micros/op 318883 ops/sec; 3119.0 MB/s
readseq : 3.065 micros/op 326245 ops/sec; 3191.0 MB/s
readseq : 3.014 micros/op 331813 ops/sec; 3245.4 MB/s
```
```
[Everything in Block cache | 10K operands | 10 KB each | 10 operand per key]
DEBUG_LEVEL=0 make db_bench -j64 && ./db_bench --benchmarks="readseq,readseq,readseq,readseq,readseq" --merge_operator="max" --num=100000 --db="/dev/shm/merge-random-10-operands-10K-10KB" --cache_size=1000000000 --use_existing_db --disable_auto_compactions
[FullMergeV2]
readseq : 24.325 micros/op 41109 ops/sec; 402.1 MB/s
readseq : 1.470 micros/op 680272 ops/sec; 6653.7 MB/s
readseq : 1.231 micros/op 812347 ops/sec; 7945.5 MB/s
readseq : 1.091 micros/op 916590 ops/sec; 8965.1 MB/s
readseq : 1.109 micros/op 901713 ops/sec; 8819.6 MB/s
[master]
readseq : 27.257 micros/op 36687 ops/sec; 358.8 MB/s
readseq : 4.443 micros/op 225073 ops/sec; 2201.4 MB/s
readseq : 5.830 micros/op 171526 ops/sec; 1677.7 MB/s
readseq : 4.173 micros/op 239635 ops/sec; 2343.8 MB/s
readseq : 4.150 micros/op 240963 ops/sec; 2356.8 MB/s
```
Test Plan: COMPILE_WITH_ASAN=1 make check -j64
Reviewers: yhchiang, andrewkr, sdong
Reviewed By: sdong
Subscribers: lovro, andrewkr, dhruba
Differential Revision: https://reviews.facebook.net/D57075
9 years ago
|
|
|
}
|
|
|
|
|
|
|
|
void DBTestBase::VerifyDBInternal(
|
|
|
|
std::vector<std::pair<std::string, std::string>> true_data) {
|
|
|
|
Arena arena;
|
|
|
|
InternalKeyComparator icmp(last_options_.comparator);
|
|
|
|
ReadOptions read_options;
|
Skip swaths of range tombstone covered keys in merging iterator (2022 edition) (#10449)
Summary:
Delete range logic is moved from `DBIter` to `MergingIterator`, and `MergingIterator` will seek to the end of a range deletion if possible instead of scanning through each key and check with `RangeDelAggregator`.
With the invariant that a key in level L (consider memtable as the first level, each immutable and L0 as a separate level) has a larger sequence number than all keys in any level >L, a range tombstone `[start, end)` from level L covers all keys in its range in any level >L. This property motivates optimizations in iterator:
- in `Seek(target)`, if level L has a range tombstone `[start, end)` that covers `target.UserKey`, then for all levels > L, we can do Seek() on `end` instead of `target` to skip some range tombstone covered keys.
- in `Next()/Prev()`, if the current key is covered by a range tombstone `[start, end)` from level L, we can do `Seek` to `end` for all levels > L.
This PR implements the above optimizations in `MergingIterator`. As all range tombstone covered keys are now skipped in `MergingIterator`, the range tombstone logic is removed from `DBIter`. The idea in this PR is similar to https://github.com/facebook/rocksdb/issues/7317, but this PR leaves `InternalIterator` interface mostly unchanged. **Credit**: the cascading seek optimization and the sentinel key (discussed below) are inspired by [Pebble](https://github.com/cockroachdb/pebble/blob/master/merging_iter.go) and suggested by ajkr in https://github.com/facebook/rocksdb/issues/7317. The two optimizations are mostly implemented in `SeekImpl()/SeekForPrevImpl()` and `IsNextDeleted()/IsPrevDeleted()` in `merging_iterator.cc`. See comments for each method for more detail.
One notable change is that the minHeap/maxHeap used by `MergingIterator` now contains range tombstone end keys besides point key iterators. This helps to reduce the number of key comparisons. For example, for a range tombstone `[start, end)`, a `start` and an `end` `HeapItem` are inserted into the heap. When a `HeapItem` for range tombstone start key is popped from the minHeap, we know this range tombstone becomes "active" in the sense that, before the range tombstone's end key is popped from the minHeap, all the keys popped from this heap is covered by the range tombstone's internal key range `[start, end)`.
Another major change, *delete range sentinel key*, is made to `LevelIterator`. Before this PR, when all point keys in an SST file are iterated through in `MergingIterator`, a level iterator would advance to the next SST file in its level. In the case when an SST file has a range tombstone that covers keys beyond the SST file's last point key, advancing to the next SST file would lose this range tombstone. Consequently, `MergingIterator` could return keys that should have been deleted by some range tombstone. We prevent this by pretending that file boundaries in each SST file are sentinel keys. A `LevelIterator` now only advance the file iterator once the sentinel key is processed.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10449
Test Plan:
- Added many unit tests in db_range_del_test
- Stress test: `./db_stress --readpercent=5 --prefixpercent=19 --writepercent=20 -delpercent=10 --iterpercent=44 --delrangepercent=2`
- Additional iterator stress test is added to verify against iterators against expected state: https://github.com/facebook/rocksdb/issues/10538. This is based on ajkr's previous attempt https://github.com/facebook/rocksdb/pull/5506#issuecomment-506021913.
```
python3 ./tools/db_crashtest.py blackbox --simple --write_buffer_size=524288 --target_file_size_base=524288 --max_bytes_for_level_base=2097152 --compression_type=none --max_background_compactions=8 --value_size_mult=33 --max_key=5000000 --interval=10 --duration=7200 --delrangepercent=3 --delpercent=9 --iterpercent=25 --writepercent=60 --readpercent=3 --prefixpercent=0 --num_iterations=1000 --range_deletion_width=100 --verify_iterator_with_expected_state_one_in=1
```
- Performance benchmark: I used a similar setup as in the blog [post](http://rocksdb.org/blog/2018/11/21/delete-range.html) that introduced DeleteRange, "a database with 5 million data keys, and 10000 range tombstones (ignoring those dropped during compaction) that were written in regular intervals after 4.5 million data keys were written". As expected, the performance with this PR depends on the range tombstone width.
```
# Setup:
TEST_TMPDIR=/dev/shm ./db_bench_main --benchmarks=fillrandom --writes=4500000 --num=5000000
TEST_TMPDIR=/dev/shm ./db_bench_main --benchmarks=overwrite --writes=500000 --num=5000000 --use_existing_db=true --writes_per_range_tombstone=50
# Scan entire DB
TEST_TMPDIR=/dev/shm ./db_bench_main --benchmarks=readseq[-X5] --use_existing_db=true --num=5000000 --disable_auto_compactions=true
# Short range scan (10 Next())
TEST_TMPDIR=/dev/shm/width-100/ ./db_bench_main --benchmarks=seekrandom[-X5] --use_existing_db=true --num=500000 --reads=100000 --seek_nexts=10 --disable_auto_compactions=true
# Long range scan(1000 Next())
TEST_TMPDIR=/dev/shm/width-100/ ./db_bench_main --benchmarks=seekrandom[-X5] --use_existing_db=true --num=500000 --reads=2500 --seek_nexts=1000 --disable_auto_compactions=true
```
Avg over of 10 runs (some slower tests had fews runs):
For the first column (tombstone), 0 means no range tombstone, 100-10000 means width of the 10k range tombstones, and 1 means there is a single range tombstone in the entire DB (width is 1000). The 1 tombstone case is to test regression when there's very few range tombstones in the DB, as no range tombstone is likely to take a different code path than with range tombstones.
- Scan entire DB
| tombstone width | Pre-PR ops/sec | Post-PR ops/sec | ±% |
| ------------- | ------------- | ------------- | ------------- |
| 0 range tombstone |2525600 (± 43564) |2486917 (± 33698) |-1.53% |
| 100 |1853835 (± 24736) |2073884 (± 32176) |+11.87% |
| 1000 |422415 (± 7466) |1115801 (± 22781) |+164.15% |
| 10000 |22384 (± 227) |227919 (± 6647) |+918.22% |
| 1 range tombstone |2176540 (± 39050) |2434954 (± 24563) |+11.87% |
- Short range scan
| tombstone width | Pre-PR ops/sec | Post-PR ops/sec | ±% |
| ------------- | ------------- | ------------- | ------------- |
| 0 range tombstone |35398 (± 533) |35338 (± 569) |-0.17% |
| 100 |28276 (± 664) |31684 (± 331) |+12.05% |
| 1000 |7637 (± 77) |25422 (± 277) |+232.88% |
| 10000 |1367 |28667 |+1997.07% |
| 1 range tombstone |32618 (± 581) |32748 (± 506) |+0.4% |
- Long range scan
| tombstone width | Pre-PR ops/sec | Post-PR ops/sec | ±% |
| ------------- | ------------- | ------------- | ------------- |
| 0 range tombstone |2262 (± 33) |2353 (± 20) |+4.02% |
| 100 |1696 (± 26) |1926 (± 18) |+13.56% |
| 1000 |410 (± 6) |1255 (± 29) |+206.1% |
| 10000 |25 |414 |+1556.0% |
| 1 range tombstone |1957 (± 30) |2185 (± 44) |+11.65% |
- Microbench does not show significant regression: https://gist.github.com/cbi42/59f280f85a59b678e7e5d8561e693b61
Reviewed By: ajkr
Differential Revision: D38450331
Pulled By: cbi42
fbshipit-source-id: b5ef12e8d8c289ed2e163ccdf277f5039b511fca
2 years ago
|
|
|
auto iter =
|
|
|
|
dbfull()->NewInternalIterator(read_options, &arena, kMaxSequenceNumber);
|
|
|
|
iter->SeekToFirst();
|
|
|
|
for (auto p : true_data) {
|
|
|
|
ASSERT_TRUE(iter->Valid());
|
|
|
|
ParsedInternalKey ikey;
|
|
|
|
ASSERT_OK(ParseInternalKey(iter->key(), &ikey, true /* log_err_key */));
|
|
|
|
ASSERT_EQ(p.first, ikey.user_key);
|
|
|
|
ASSERT_EQ(p.second, iter->value());
|
|
|
|
iter->Next();
|
|
|
|
};
|
|
|
|
ASSERT_FALSE(iter->Valid());
|
|
|
|
iter->~InternalIterator();
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
uint64_t DBTestBase::GetNumberOfSstFilesForColumnFamily(
|
|
|
|
DB* db, std::string column_family_name) {
|
|
|
|
std::vector<LiveFileMetaData> metadata;
|
|
|
|
db->GetLiveFilesMetaData(&metadata);
|
|
|
|
uint64_t result = 0;
|
|
|
|
for (auto& fileMetadata : metadata) {
|
|
|
|
result += (fileMetadata.column_family_name == column_family_name);
|
|
|
|
}
|
|
|
|
return result;
|
|
|
|
}
|
|
|
|
|
|
|
|
uint64_t DBTestBase::GetSstSizeHelper(Temperature temperature) {
|
|
|
|
std::string prop;
|
|
|
|
EXPECT_TRUE(dbfull()->GetProperty(
|
|
|
|
DB::Properties::kLiveSstFilesSizeAtTemperature +
|
|
|
|
std::to_string(static_cast<uint8_t>(temperature)),
|
|
|
|
&prop));
|
|
|
|
return static_cast<uint64_t>(std::atoi(prop.c_str()));
|
|
|
|
}
|
|
|
|
|
Experimental support for SST unique IDs (#8990)
Summary:
* New public header unique_id.h and function GetUniqueIdFromTableProperties
which computes a universally unique identifier based on table properties
of table files from recent RocksDB versions.
* Generation of DB session IDs is refactored so that they are
guaranteed unique in the lifetime of a process running RocksDB.
(SemiStructuredUniqueIdGen, new test included.) Along with file numbers,
this enables SST unique IDs to be guaranteed unique among SSTs generated
in a single process, and "better than random" between processes.
See https://github.com/pdillinger/unique_id
* In addition to public API producing 'external' unique IDs, there is a function
for producing 'internal' unique IDs, with functions for converting between the
two. In short, the external ID is "safe" for things people might do with it, and
the internal ID enables more "power user" features for the future. Specifically,
the external ID goes through a hashing layer so that any subset of bits in the
external ID can be used as a hash of the full ID, while also preserving
uniqueness guarantees in the first 128 bits (bijective both on first 128 bits
and on full 192 bits).
Intended follow-up:
* Use the internal unique IDs in cache keys. (Avoid conflicts with https://github.com/facebook/rocksdb/issues/8912) (The file offset can be XORed into
the third 64-bit value of the unique ID.)
* Publish the external unique IDs in FileStorageInfo (https://github.com/facebook/rocksdb/issues/8968)
Pull Request resolved: https://github.com/facebook/rocksdb/pull/8990
Test Plan:
Unit tests added, and checking of unique ids in stress test.
NOTE in stress test we do not generate nearly enough files to thoroughly
stress uniqueness, but the test trims off pieces of the ID to check for
uniqueness so that we can infer (with some assumptions) stronger
properties in the aggregate.
Reviewed By: zhichao-cao, mrambacher
Differential Revision: D31582865
Pulled By: pdillinger
fbshipit-source-id: 1f620c4c86af9abe2a8d177b9ccf2ad2b9f48243
3 years ago
|
|
|
void VerifySstUniqueIds(const TablePropertiesCollection& props) {
|
|
|
|
ASSERT_FALSE(props.empty()); // suspicious test if empty
|
|
|
|
std::unordered_set<std::string> seen;
|
|
|
|
for (auto& pair : props) {
|
|
|
|
std::string id;
|
|
|
|
ASSERT_OK(GetUniqueIdFromTableProperties(*pair.second, &id));
|
|
|
|
ASSERT_TRUE(seen.insert(id).second);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
template <CacheEntryRole R>
|
|
|
|
TargetCacheChargeTrackingCache<R>::TargetCacheChargeTrackingCache(
|
|
|
|
std::shared_ptr<Cache> target)
|
|
|
|
: CacheWrapper(std::move(target)),
|
|
|
|
cur_cache_charge_(0),
|
|
|
|
cache_charge_peak_(0),
|
|
|
|
cache_charge_increment_(0),
|
|
|
|
last_peak_tracked_(false),
|
|
|
|
cache_charge_increments_sum_(0) {}
|
|
|
|
|
|
|
|
template <CacheEntryRole R>
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
Status TargetCacheChargeTrackingCache<R>::Insert(const Slice& key,
|
|
|
|
ObjectPtr value,
|
|
|
|
const CacheItemHelper* helper,
|
|
|
|
size_t charge, Handle** handle,
|
|
|
|
Priority priority) {
|
|
|
|
Status s = target_->Insert(key, value, helper, charge, handle, priority);
|
|
|
|
if (helper == kCrmHelper) {
|
|
|
|
if (last_peak_tracked_) {
|
|
|
|
cache_charge_peak_ = 0;
|
|
|
|
cache_charge_increment_ = 0;
|
|
|
|
last_peak_tracked_ = false;
|
|
|
|
}
|
|
|
|
if (s.ok()) {
|
|
|
|
cur_cache_charge_ += charge;
|
|
|
|
}
|
|
|
|
cache_charge_peak_ = std::max(cache_charge_peak_, cur_cache_charge_);
|
|
|
|
cache_charge_increment_ += charge;
|
|
|
|
}
|
|
|
|
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
template <CacheEntryRole R>
|
|
|
|
bool TargetCacheChargeTrackingCache<R>::Release(Handle* handle,
|
|
|
|
bool erase_if_last_ref) {
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
auto helper = GetCacheItemHelper(handle);
|
|
|
|
if (helper == kCrmHelper) {
|
|
|
|
if (!last_peak_tracked_) {
|
|
|
|
cache_charge_peaks_.push_back(cache_charge_peak_);
|
|
|
|
cache_charge_increments_sum_ += cache_charge_increment_;
|
|
|
|
last_peak_tracked_ = true;
|
|
|
|
}
|
|
|
|
cur_cache_charge_ -= GetCharge(handle);
|
|
|
|
}
|
|
|
|
bool is_successful = target_->Release(handle, erase_if_last_ref);
|
|
|
|
return is_successful;
|
|
|
|
}
|
|
|
|
|
|
|
|
template <CacheEntryRole R>
|
Major Cache refactoring, CPU efficiency improvement (#10975)
Summary:
This is several refactorings bundled into one to avoid having to incrementally re-modify uses of Cache several times. Overall, there are breaking changes to Cache class, and it becomes more of low-level interface for implementing caches, especially block cache. New internal APIs make using Cache cleaner than before, and more insulated from block cache evolution. Hopefully, this is the last really big block cache refactoring, because of rather effectively decoupling the implementations from the uses. This change also removes the EXPERIMENTAL designation on the SecondaryCache support in Cache. It seems reasonably mature at this point but still subject to change/evolution (as I warn in the API docs for Cache).
The high-level motivation for this refactoring is to minimize code duplication / compounding complexity in adding SecondaryCache support to HyperClockCache (in a later PR). Other benefits listed below.
* static_cast lines of code +29 -35 (net removed 6)
* reinterpret_cast lines of code +6 -32 (net removed 26)
## cache.h and secondary_cache.h
* Always use CacheItemHelper with entries instead of just a Deleter. There are several motivations / justifications:
* Simpler for implementations to deal with just one Insert and one Lookup.
* Simpler and more efficient implementation because we don't have to track which entries are using helpers and which are using deleters
* Gets rid of hack to classify cache entries by their deleter. Instead, the CacheItemHelper includes a CacheEntryRole. This simplifies a lot of code (cache_entry_roles.h almost eliminated). Fixes https://github.com/facebook/rocksdb/issues/9428.
* Makes it trivial to adjust SecondaryCache behavior based on kind of block (e.g. don't re-compress filter blocks).
* It is arguably less convenient for many direct users of Cache, but direct users of Cache are now rare with introduction of typed_cache.h (below).
* I considered and rejected an alternative approach in which we reduce customizability by assuming each secondary cache compatible value starts with a Slice referencing the uncompressed block contents (already true or mostly true), but we apparently intend to stack secondary caches. Saving an entry from a compressed secondary to a lower tier requires custom handling offered by SaveToCallback, etc.
* Make CreateCallback part of the helper and introduce CreateContext to work with it (alternative to https://github.com/facebook/rocksdb/issues/10562). This cleans up the interface while still allowing context to be provided for loading/parsing values into primary cache. This model works for async lookup in BlockBasedTable reader (reader owns a CreateContext) under the assumption that it always waits on secondary cache operations to finish. (Otherwise, the CreateContext could be destroyed while async operation depending on it continues.) This likely contributes most to the observed performance improvement because it saves an std::function backed by a heap allocation.
* Use char* for serialized data, e.g. in SaveToCallback, where void* was confusingly used. (We use `char*` for serialized byte data all over RocksDB, with many advantages over `void*`. `memcpy` etc. are legacy APIs that should not be mimicked.)
* Add a type alias Cache::ObjectPtr = void*, so that we can better indicate the intent of the void* when it is to be the object associated with a Cache entry. Related: started (but did not complete) a refactoring to move away from "value" of a cache entry toward "object" or "obj". (It is confusing to call Cache a key-value store (like DB) when it is really storing arbitrary in-memory objects, not byte strings.)
* Remove unnecessary key param from DeleterFn. This is good for efficiency in HyperClockCache, which does not directly store the cache key in memory. (Alternative to https://github.com/facebook/rocksdb/issues/10774)
* Add allocator to Cache DeleterFn. This is a kind of future-proofing change in case we get more serious about using the Cache allocator for memory tracked by the Cache. Right now, only the uncompressed block contents are allocated using the allocator, and a pointer to that allocator is saved as part of the cached object so that the deleter can use it. (See CacheAllocationPtr.) If in the future we are able to "flatten out" our Cache objects some more, it would be good not to have to track the allocator as part of each object.
* Removes legacy `ApplyToAllCacheEntries` and changes `ApplyToAllEntries` signature for Deleter->CacheItemHelper change.
## typed_cache.h
Adds various "typed" interfaces to the Cache as internal APIs, so that most uses of Cache can use simple type safe code without casting and without explicit deleters, etc. Almost all of the non-test, non-glue code uses of Cache have been migrated. (Follow-up work: CompressedSecondaryCache deserves deeper attention to migrate.) This change expands RocksDB's internal usage of metaprogramming and SFINAE (https://en.cppreference.com/w/cpp/language/sfinae).
The existing usages of Cache are divided up at a high level into these new interfaces. See updated existing uses of Cache for examples of how these are used.
* PlaceholderCacheInterface - Used for making cache reservations, with entries that have a charge but no value.
* BasicTypedCacheInterface<TValue> - Used for primary cache storage of objects of type TValue, which can be cleaned up with std::default_delete<TValue>. The role is provided by TValue::kCacheEntryRole or given in an optional template parameter.
* FullTypedCacheInterface<TValue, TCreateContext> - Used for secondary cache compatible storage of objects of type TValue. In addition to BasicTypedCacheInterface constraints, we require TValue::ContentSlice() to return persistable data. This simplifies usage for the normal case of simple secondary cache compatibility (can give you a Slice to the data already in memory). In addition to TCreateContext performing the role of Cache::CreateContext, it is also expected to provide a factory function for creating TValue.
* For each of these, there's a "Shared" version (e.g. FullTypedSharedCacheInterface) that holds a shared_ptr to the Cache, rather than assuming external ownership by holding only a raw `Cache*`.
These interfaces introduce specific handle types for each interface instantiation, so that it's easy to see what kind of object is controlled by a handle. (Ultimately, this might not be worth the extra complexity, but it seems OK so far.)
Note: I attempted to make the cache 'charge' automatically inferred from the cache object type, such as by expecting an ApproximateMemoryUsage() function, but this is not so clean because there are cases where we need to compute the charge ahead of time and don't want to re-compute it.
## block_cache.h
This header is essentially the replacement for the old block_like_traits.h. It includes various things to support block cache access with typed_cache.h for block-based table.
## block_based_table_reader.cc
Before this change, accessing the block cache here was an awkward mix of static polymorphism (template TBlocklike) and switch-case on a dynamic BlockType value. This change mostly unifies on static polymorphism, relying on minor hacks in block_cache.h to distinguish variants of Block. We still check BlockType in some places (especially for stats, which could be improved in follow-up work) but at least the BlockType is a static constant from the template parameter. (No more awkward partial redundancy between static and dynamic info.) This likely contributes to the overall performance improvement, but hasn't been tested in isolation.
The other key source of simplification here is a more unified system of creating block cache objects: for directly populating from primary cache and for promotion from secondary cache. Both use BlockCreateContext, for context and for factory functions.
## block_based_table_builder.cc, cache_dump_load_impl.cc
Before this change, warming caches was super ugly code. Both of these source files had switch statements to basically transition from the dynamic BlockType world to the static TBlocklike world. None of that mess is needed anymore as there's a new, untyped WarmInCache function that handles all the details just as promotion from SecondaryCache would. (Fixes `TODO akanksha: Dedup below code` in block_based_table_builder.cc.)
## Everything else
Mostly just updating Cache users to use new typed APIs when reasonably possible, or changed Cache APIs when not.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10975
Test Plan:
tests updated
Performance test setup similar to https://github.com/facebook/rocksdb/issues/10626 (by cache size, LRUCache when not "hyper" for HyperClockCache):
34MB 1thread base.hyper -> kops/s: 0.745 io_bytes/op: 2.52504e+06 miss_ratio: 0.140906 max_rss_mb: 76.4844
34MB 1thread new.hyper -> kops/s: 0.751 io_bytes/op: 2.5123e+06 miss_ratio: 0.140161 max_rss_mb: 79.3594
34MB 1thread base -> kops/s: 0.254 io_bytes/op: 1.36073e+07 miss_ratio: 0.918818 max_rss_mb: 45.9297
34MB 1thread new -> kops/s: 0.252 io_bytes/op: 1.36157e+07 miss_ratio: 0.918999 max_rss_mb: 44.1523
34MB 32thread base.hyper -> kops/s: 7.272 io_bytes/op: 2.88323e+06 miss_ratio: 0.162532 max_rss_mb: 516.602
34MB 32thread new.hyper -> kops/s: 7.214 io_bytes/op: 2.99046e+06 miss_ratio: 0.168818 max_rss_mb: 518.293
34MB 32thread base -> kops/s: 3.528 io_bytes/op: 1.35722e+07 miss_ratio: 0.914691 max_rss_mb: 264.926
34MB 32thread new -> kops/s: 3.604 io_bytes/op: 1.35744e+07 miss_ratio: 0.915054 max_rss_mb: 264.488
233MB 1thread base.hyper -> kops/s: 53.909 io_bytes/op: 2552.35 miss_ratio: 0.0440566 max_rss_mb: 241.984
233MB 1thread new.hyper -> kops/s: 62.792 io_bytes/op: 2549.79 miss_ratio: 0.044043 max_rss_mb: 241.922
233MB 1thread base -> kops/s: 1.197 io_bytes/op: 2.75173e+06 miss_ratio: 0.103093 max_rss_mb: 241.559
233MB 1thread new -> kops/s: 1.199 io_bytes/op: 2.73723e+06 miss_ratio: 0.10305 max_rss_mb: 240.93
233MB 32thread base.hyper -> kops/s: 1298.69 io_bytes/op: 2539.12 miss_ratio: 0.0440307 max_rss_mb: 371.418
233MB 32thread new.hyper -> kops/s: 1421.35 io_bytes/op: 2538.75 miss_ratio: 0.0440307 max_rss_mb: 347.273
233MB 32thread base -> kops/s: 9.693 io_bytes/op: 2.77304e+06 miss_ratio: 0.103745 max_rss_mb: 569.691
233MB 32thread new -> kops/s: 9.75 io_bytes/op: 2.77559e+06 miss_ratio: 0.103798 max_rss_mb: 552.82
1597MB 1thread base.hyper -> kops/s: 58.607 io_bytes/op: 1449.14 miss_ratio: 0.0249324 max_rss_mb: 1583.55
1597MB 1thread new.hyper -> kops/s: 69.6 io_bytes/op: 1434.89 miss_ratio: 0.0247167 max_rss_mb: 1584.02
1597MB 1thread base -> kops/s: 60.478 io_bytes/op: 1421.28 miss_ratio: 0.024452 max_rss_mb: 1589.45
1597MB 1thread new -> kops/s: 63.973 io_bytes/op: 1416.07 miss_ratio: 0.0243766 max_rss_mb: 1589.24
1597MB 32thread base.hyper -> kops/s: 1436.2 io_bytes/op: 1357.93 miss_ratio: 0.0235353 max_rss_mb: 1692.92
1597MB 32thread new.hyper -> kops/s: 1605.03 io_bytes/op: 1358.04 miss_ratio: 0.023538 max_rss_mb: 1702.78
1597MB 32thread base -> kops/s: 280.059 io_bytes/op: 1350.34 miss_ratio: 0.023289 max_rss_mb: 1675.36
1597MB 32thread new -> kops/s: 283.125 io_bytes/op: 1351.05 miss_ratio: 0.0232797 max_rss_mb: 1703.83
Almost uniformly improving over base revision, especially for hot paths with HyperClockCache, up to 12% higher throughput seen (1597MB, 32thread, hyper). The improvement for that is likely coming from much simplified code for providing context for secondary cache promotion (CreateCallback/CreateContext), and possibly from less branching in block_based_table_reader. And likely a small improvement from not reconstituting key for DeleterFn.
Reviewed By: anand1976
Differential Revision: D42417818
Pulled By: pdillinger
fbshipit-source-id: f86bfdd584dce27c028b151ba56818ad14f7a432
2 years ago
|
|
|
const Cache::CacheItemHelper* TargetCacheChargeTrackingCache<R>::kCrmHelper =
|
|
|
|
CacheReservationManagerImpl<R>::TEST_GetCacheItemHelperForRole();
|
|
|
|
|
|
|
|
template class TargetCacheChargeTrackingCache<
|
|
|
|
CacheEntryRole::kFilterConstruction>;
|
|
|
|
template class TargetCacheChargeTrackingCache<
|
|
|
|
CacheEntryRole::kBlockBasedTableReader>;
|
Account memory of FileMetaData in global memory limit (#9924)
Summary:
**Context/Summary:**
As revealed by heap profiling, allocation of `FileMetaData` for [newly created file added to a Version](https://github.com/facebook/rocksdb/pull/9924/files#diff-a6aa385940793f95a2c5b39cc670bd440c4547fa54fd44622f756382d5e47e43R774) can consume significant heap memory. This PR is to account that toward our global memory limit based on block cache capacity.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/9924
Test Plan:
- Previous `make check` verified there are only 2 places where the memory of the allocated `FileMetaData` can be released
- New unit test `TEST_P(ChargeFileMetadataTestWithParam, Basic)`
- db bench (CPU cost of `charge_file_metadata` in write and compact)
- **write micros/op: -0.24%** : `TEST_TMPDIR=/dev/shm/testdb ./db_bench -benchmarks=fillseq -db=$TEST_TMPDIR -charge_file_metadata=1 (remove this option for pre-PR) -disable_auto_compactions=1 -write_buffer_size=100000 -num=4000000 | egrep 'fillseq'`
- **compact micros/op -0.87%** : `TEST_TMPDIR=/dev/shm/testdb ./db_bench -benchmarks=fillseq -db=$TEST_TMPDIR -charge_file_metadata=1 -disable_auto_compactions=1 -write_buffer_size=100000 -num=4000000 -numdistinct=1000 && ./db_bench -benchmarks=compact -db=$TEST_TMPDIR -use_existing_db=1 -charge_file_metadata=1 -disable_auto_compactions=1 | egrep 'compact'`
table 1 - write
#-run | (pre-PR) avg micros/op | std micros/op | (post-PR) micros/op | std micros/op | change (%)
-- | -- | -- | -- | -- | --
10 | 3.9711 | 0.264408 | 3.9914 | 0.254563 | 0.5111933721
20 | 3.83905 | 0.0664488 | 3.8251 | 0.0695456 | -0.3633711465
40 | 3.86625 | 0.136669 | 3.8867 | 0.143765 | 0.5289363078
80 | 3.87828 | 0.119007 | 3.86791 | 0.115674 | **-0.2673865734**
160 | 3.87677 | 0.162231 | 3.86739 | 0.16663 | **-0.2419539978**
table 2 - compact
#-run | (pre-PR) avg micros/op | std micros/op | (post-PR) micros/op | std micros/op | change (%)
-- | -- | -- | -- | -- | --
10 | 2,399,650.00 | 96,375.80 | 2,359,537.00 | 53,243.60 | -1.67
20 | 2,410,480.00 | 89,988.00 | 2,433,580.00 | 91,121.20 | 0.96
40 | 2.41E+06 | 121811 | 2.39E+06 | 131525 | **-0.96**
80 | 2.40E+06 | 134503 | 2.39E+06 | 108799 | **-0.78**
- stress test: `python3 tools/db_crashtest.py blackbox --charge_file_metadata=1 --cache_size=1` killed as normal
Reviewed By: ajkr
Differential Revision: D36055583
Pulled By: hx235
fbshipit-source-id: b60eab94707103cb1322cf815f05810ef0232625
3 years ago
|
|
|
template class TargetCacheChargeTrackingCache<CacheEntryRole::kFileMetadata>;
|
|
|
|
|
|
|
|
} // namespace ROCKSDB_NAMESPACE
|