rocksdb/db/memtablelist.cc

263 lines
8.2 KiB
C++
Raw Normal View History

// Copyright (c) 2013, Facebook, Inc. All rights reserved.
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree. An additional grant
// of patent rights can be found in the PATENTS file in the same directory.
//
#include "db/memtablelist.h"
#include <string>
#include "rocksdb/db.h"
#include "db/memtable.h"
#include "rocksdb/env.h"
#include "rocksdb/iterator.h"
#include "util/coding.h"
namespace rocksdb {
class InternalKeyComparator;
class Mutex;
class VersionSet;
MemTableListVersion Summary: MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful. This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team. I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet()) Test Plan: `make check` works. I will also do `make valgrind_check` before commit Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak Reviewed By: kailiu CC: leveldb Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 22:52:08 +00:00
MemTableListVersion::MemTableListVersion(MemTableListVersion* old) {
if (old != nullptr) {
memlist_ = old->memlist_;
size_ = old->size_;
for (auto& m : memlist_) {
m->Ref();
}
}
}
MemTableListVersion Summary: MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful. This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team. I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet()) Test Plan: `make check` works. I will also do `make valgrind_check` before commit Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak Reviewed By: kailiu CC: leveldb Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 22:52:08 +00:00
void MemTableListVersion::Ref() { ++refs_; }
void MemTableListVersion::Unref(std::vector<MemTable*>* to_delete) {
--refs_;
if (refs_ == 0) {
// if to_delete is equal to nullptr it means we're confident
// that refs_ will not be zero
assert(to_delete != nullptr);
for (const auto& m : memlist_) {
MemTable* x = m->Unref();
if (x != nullptr) {
to_delete->push_back(x);
}
}
MemTableListVersion Summary: MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful. This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team. I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet()) Test Plan: `make check` works. I will also do `make valgrind_check` before commit Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak Reviewed By: kailiu CC: leveldb Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 22:52:08 +00:00
delete this;
}
}
MemTableListVersion Summary: MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful. This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team. I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet()) Test Plan: `make check` works. I will also do `make valgrind_check` before commit Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak Reviewed By: kailiu CC: leveldb Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 22:52:08 +00:00
int MemTableListVersion::size() const { return size_; }
// Returns the total number of memtables in the list
MemTableListVersion Summary: MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful. This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team. I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet()) Test Plan: `make check` works. I will also do `make valgrind_check` before commit Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak Reviewed By: kailiu CC: leveldb Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 22:52:08 +00:00
int MemTableList::size() const {
assert(num_flush_not_started_ <= current_->size_);
return current_->size_;
}
// Search all the memtables starting from the most recent one.
// Return the most recent value found, if any.
// Operands stores the list of merge operations to apply, so far.
bool MemTableListVersion::Get(const LookupKey& key, std::string* value,
Status* s, MergeContext& merge_context,
const Options& options) {
for (auto& memtable : memlist_) {
if (memtable->Get(key, value, s, merge_context, options)) {
return true;
}
}
return false;
}
void MemTableListVersion::AddIterators(const ReadOptions& options,
std::vector<Iterator*>* iterator_list) {
for (auto& m : memlist_) {
iterator_list->push_back(m->NewIterator(options));
}
}
void MemTableListVersion::Add(MemTable* m) {
assert(refs_ == 1); // only when refs_ == 1 is MemTableListVersion mutable
m->Ref();
memlist_.push_front(m);
++size_;
}
void MemTableListVersion::Remove(MemTable* m) {
assert(refs_ == 1); // only when refs_ == 1 is MemTableListVersion mutable
MemTable* x __attribute__((unused)) = m->Unref();
assert(x == nullptr); // it still needs to be alive!
memlist_.remove(m);
--size_;
}
// Returns true if there is at least one memtable on which flush has
// not yet started.
MemTableListVersion Summary: MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful. This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team. I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet()) Test Plan: `make check` works. I will also do `make valgrind_check` before commit Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak Reviewed By: kailiu CC: leveldb Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 22:52:08 +00:00
bool MemTableList::IsFlushPending() {
if ((flush_requested_ && num_flush_not_started_ >= 1) ||
MemTableListVersion Summary: MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful. This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team. I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet()) Test Plan: `make check` works. I will also do `make valgrind_check` before commit Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak Reviewed By: kailiu CC: leveldb Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 22:52:08 +00:00
(num_flush_not_started_ >= min_write_buffer_number_to_merge_)) {
assert(imm_flush_needed.NoBarrier_Load() != nullptr);
return true;
}
return false;
}
// Returns the memtables that need to be flushed.
void MemTableList::PickMemtablesToFlush(std::vector<MemTable*>* ret) {
MemTableListVersion Summary: MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful. This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team. I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet()) Test Plan: `make check` works. I will also do `make valgrind_check` before commit Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak Reviewed By: kailiu CC: leveldb Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 22:52:08 +00:00
const auto& memlist = current_->memlist_;
for (auto it = memlist.rbegin(); it != memlist.rend(); ++it) {
MemTable* m = *it;
if (!m->flush_in_progress_) {
assert(!m->flush_completed_);
num_flush_not_started_--;
if (num_flush_not_started_ == 0) {
imm_flush_needed.Release_Store(nullptr);
}
m->flush_in_progress_ = true; // flushing will start very soon
ret->push_back(m);
}
}
flush_requested_ = false; // start-flush request is complete
}
// Record a successful flush in the manifest file
Status MemTableList::InstallMemtableFlushResults(
const std::vector<MemTable*> &mems,
VersionSet* vset, Status flushStatus,
port::Mutex* mu, Logger* info_log,
uint64_t file_number,
std::set<uint64_t>& pending_outputs,
std::vector<MemTable*>* to_delete) {
mu->AssertHeld();
// If the flush was not successful, then just reset state.
// Maybe a suceeding attempt to flush will be successful.
if (!flushStatus.ok()) {
for (MemTable* m : mems) {
assert(m->flush_in_progress_);
assert(m->file_number_ == 0);
m->flush_in_progress_ = false;
m->flush_completed_ = false;
m->edit_.Clear();
num_flush_not_started_++;
imm_flush_needed.Release_Store((void *)1);
pending_outputs.erase(file_number);
}
return flushStatus;
}
// flush was sucessful
for (size_t i = 0; i < mems.size(); ++i) {
// All the edits are associated with the first memtable of this batch.
assert(i == 0 || mems[i]->GetEdits()->NumEntries() == 0);
mems[i]->flush_completed_ = true;
mems[i]->file_number_ = file_number;
}
// if some other thread is already commiting, then return
Status s;
if (commit_in_progress_) {
return s;
}
// Only a single thread can be executing this piece of code
commit_in_progress_ = true;
// scan all memtables from the earliest, and commit those
// (in that order) that have finished flushing. Memetables
// are always committed in the order that they were created.
MemTableListVersion Summary: MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful. This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team. I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet()) Test Plan: `make check` works. I will also do `make valgrind_check` before commit Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak Reviewed By: kailiu CC: leveldb Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 22:52:08 +00:00
while (!current_->memlist_.empty() && s.ok()) {
MemTable* m = current_->memlist_.back(); // get the last element
if (!m->flush_completed_) {
break;
}
Log(info_log,
"Level-0 commit table #%lu started",
(unsigned long)m->file_number_);
// this can release and reacquire the mutex.
s = vset->LogAndApply(&m->edit_, mu);
MemTableListVersion Summary: MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful. This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team. I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet()) Test Plan: `make check` works. I will also do `make valgrind_check` before commit Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak Reviewed By: kailiu CC: leveldb Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 22:52:08 +00:00
// we will be changing the version in the next code path,
// so we better create a new one, since versions are immutable
InstallNewVersion();
// All the later memtables that have the same filenum
// are part of the same batch. They can be committed now.
uint64_t mem_id = 1; // how many memtables has been flushed.
do {
if (s.ok()) { // commit new state
Log(info_log,
"Level-0 commit table #%lu: memtable #%lu done",
(unsigned long)m->file_number_,
(unsigned long)mem_id);
MemTableListVersion Summary: MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful. This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team. I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet()) Test Plan: `make check` works. I will also do `make valgrind_check` before commit Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak Reviewed By: kailiu CC: leveldb Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 22:52:08 +00:00
current_->Remove(m);
assert(m->file_number_ > 0);
// pending_outputs can be cleared only after the newly created file
// has been written to a committed version so that other concurrently
// executing compaction threads do not mistakenly assume that this
// file is not live.
pending_outputs.erase(m->file_number_);
if (m->Unref() != nullptr) {
to_delete->push_back(m);
}
} else {
//commit failed. setup state so that we can flush again.
Log(info_log,
"Level-0 commit table #%lu: memtable #%lu failed",
(unsigned long)m->file_number_,
(unsigned long)mem_id);
m->flush_completed_ = false;
m->flush_in_progress_ = false;
m->edit_.Clear();
num_flush_not_started_++;
pending_outputs.erase(m->file_number_);
m->file_number_ = 0;
imm_flush_needed.Release_Store((void *)1);
s = Status::IOError("Unable to commit flushed memtable");
}
++mem_id;
MemTableListVersion Summary: MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful. This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team. I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet()) Test Plan: `make check` works. I will also do `make valgrind_check` before commit Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak Reviewed By: kailiu CC: leveldb Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 22:52:08 +00:00
} while (!current_->memlist_.empty() && (m = current_->memlist_.back()) &&
m->file_number_ == file_number);
}
commit_in_progress_ = false;
return s;
}
// New memtables are inserted at the front of the list.
void MemTableList::Add(MemTable* m) {
MemTableListVersion Summary: MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful. This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team. I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet()) Test Plan: `make check` works. I will also do `make valgrind_check` before commit Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak Reviewed By: kailiu CC: leveldb Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 22:52:08 +00:00
assert(current_->size_ >= num_flush_not_started_);
InstallNewVersion();
current_->Add(m);
m->MarkImmutable();
num_flush_not_started_++;
if (num_flush_not_started_ == 1) {
imm_flush_needed.Release_Store((void *)1);
}
}
// Returns an estimate of the number of bytes of data in use.
size_t MemTableList::ApproximateMemoryUsage() {
size_t size = 0;
MemTableListVersion Summary: MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful. This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team. I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet()) Test Plan: `make check` works. I will also do `make valgrind_check` before commit Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak Reviewed By: kailiu CC: leveldb Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 22:52:08 +00:00
for (auto& memtable : current_->memlist_) {
size += memtable->ApproximateMemoryUsage();
}
return size;
}
MemTableListVersion Summary: MemTableListVersion is to MemTableList what Version is to VersionSet. I took almost the same ideas to develop MemTableListVersion. The reason is to have copying std::list done in background, while flushing, rather than in foreground (MultiGet() and NewIterator()) under a mutex! Also, whenever we copied MemTableList, we copied also some MemTableList metadata (flush_requested_, commit_in_progress_, etc.), which was wasteful. This diff avoids std::list copy under a mutex in both MultiGet() and NewIterator(). I created a small database with some number of immutable memtables, and creating 100.000 iterators in a single-thread (!) decreased from {188739, 215703, 198028} to {154352, 164035, 159817}. A lot of the savings come from code under a mutex, so we should see much higher savings with multiple threads. Creating new iterator is very important to LogDevice team. I also think this diff will make SuperVersion obsolete for performance reasons. I will try it in the next diff. SuperVersion gave us huge savings on Get() code path, but I think that most of the savings came from copying MemTableList under a mutex. If we had MemTableListVersion, we would never need to copy the entire object (like we still do in NewIterator() and MultiGet()) Test Plan: `make check` works. I will also do `make valgrind_check` before commit Reviewers: dhruba, haobo, kailiu, sdong, emayanke, tnovak Reviewed By: kailiu CC: leveldb Differential Revision: https://reviews.facebook.net/D15255
2014-01-24 22:52:08 +00:00
void MemTableList::InstallNewVersion() {
if (current_->refs_ == 1) {
// we're the only one using the version, just keep using it
} else {
// somebody else holds the current version, we need to create new one
MemTableListVersion* version = current_;
current_ = new MemTableListVersion(current_);
version->Unref();
}
}
} // namespace rocksdb