2016-02-10 00:12:00 +01:00
|
|
|
// Copyright (c) 2011-present, Facebook, Inc. All rights reserved.
|
2013-10-16 23:59:46 +02:00
|
|
|
// 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.
|
|
|
|
//
|
2013-12-11 04:03:13 +01:00
|
|
|
#include "db/memtable_list.h"
|
2012-11-29 01:42:36 +01:00
|
|
|
|
2014-10-28 19:54:33 +01:00
|
|
|
#ifndef __STDC_FORMAT_MACROS
|
|
|
|
#define __STDC_FORMAT_MACROS
|
|
|
|
#endif
|
|
|
|
|
|
|
|
#include <inttypes.h>
|
2012-10-19 23:00:53 +02:00
|
|
|
#include <string>
|
|
|
|
#include "db/memtable.h"
|
2014-01-24 23:30:28 +01:00
|
|
|
#include "db/version_set.h"
|
2017-04-06 04:02:00 +02:00
|
|
|
#include "monitoring/thread_status_util.h"
|
Fix flush not being commit while writing manifest
Summary:
Fix flush not being commit while writing manifest, which is a recent bug introduced by D60075.
The issue:
# Options.max_background_flushes > 1
# Background thread A pick up a flush job, flush, then commit to manifest. (Note that mutex is released before writing manifest.)
# Background thread B pick up another flush job, flush. When it gets to `MemTableList::InstallMemtableFlushResults`, it notices another thread is commiting, so it quit.
# After the first commit, thread A doesn't double check if there are more flush result need to commit, leaving the second flush uncommitted.
Test Plan: run the test. Also verify the new test hit deadlock without the fix.
Reviewers: sdong, igor, lightmark
Reviewed By: lightmark
Subscribers: andrewkr, omegaga, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D60969
2016-07-21 19:10:41 +02:00
|
|
|
#include "rocksdb/db.h"
|
2013-08-23 17:38:13 +02:00
|
|
|
#include "rocksdb/env.h"
|
|
|
|
#include "rocksdb/iterator.h"
|
2017-02-03 01:38:40 +01:00
|
|
|
#include "table/merging_iterator.h"
|
2012-10-19 23:00:53 +02:00
|
|
|
#include "util/coding.h"
|
2014-04-03 06:49:51 +02:00
|
|
|
#include "util/log_buffer.h"
|
Fix flush not being commit while writing manifest
Summary:
Fix flush not being commit while writing manifest, which is a recent bug introduced by D60075.
The issue:
# Options.max_background_flushes > 1
# Background thread A pick up a flush job, flush, then commit to manifest. (Note that mutex is released before writing manifest.)
# Background thread B pick up another flush job, flush. When it gets to `MemTableList::InstallMemtableFlushResults`, it notices another thread is commiting, so it quit.
# After the first commit, thread A doesn't double check if there are more flush result need to commit, leaving the second flush uncommitted.
Test Plan: run the test. Also verify the new test hit deadlock without the fix.
Reviewers: sdong, igor, lightmark
Reviewed By: lightmark
Subscribers: andrewkr, omegaga, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D60969
2016-07-21 19:10:41 +02:00
|
|
|
#include "util/sync_point.h"
|
2012-10-19 23:00:53 +02:00
|
|
|
|
2013-10-04 06:49:15 +02:00
|
|
|
namespace rocksdb {
|
2012-10-19 23:00:53 +02:00
|
|
|
|
|
|
|
class InternalKeyComparator;
|
|
|
|
class Mutex;
|
|
|
|
class VersionSet;
|
|
|
|
|
2015-08-19 22:32:09 +02:00
|
|
|
void MemTableListVersion::AddMemTable(MemTable* m) {
|
|
|
|
memlist_.push_front(m);
|
|
|
|
*parent_memtable_list_memory_usage_ += m->ApproximateMemoryUsage();
|
|
|
|
}
|
|
|
|
|
|
|
|
void MemTableListVersion::UnrefMemTable(autovector<MemTable*>* to_delete,
|
|
|
|
MemTable* m) {
|
|
|
|
if (m->Unref()) {
|
|
|
|
to_delete->push_back(m);
|
|
|
|
assert(*parent_memtable_list_memory_usage_ >= m->ApproximateMemoryUsage());
|
|
|
|
*parent_memtable_list_memory_usage_ -= m->ApproximateMemoryUsage();
|
|
|
|
} else {
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
MemTableListVersion::MemTableListVersion(
|
|
|
|
size_t* parent_memtable_list_memory_usage, MemTableListVersion* old)
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
: max_write_buffer_number_to_maintain_(
|
2015-08-19 22:32:09 +02:00
|
|
|
old->max_write_buffer_number_to_maintain_),
|
|
|
|
parent_memtable_list_memory_usage_(parent_memtable_list_memory_usage) {
|
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 23:52:08 +01:00
|
|
|
if (old != nullptr) {
|
|
|
|
memlist_ = old->memlist_;
|
|
|
|
for (auto& m : memlist_) {
|
|
|
|
m->Ref();
|
|
|
|
}
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
|
|
|
|
memlist_history_ = old->memlist_history_;
|
|
|
|
for (auto& m : memlist_history_) {
|
|
|
|
m->Ref();
|
|
|
|
}
|
2012-10-19 23:00:53 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
MemTableListVersion::MemTableListVersion(
|
2015-08-19 22:32:09 +02:00
|
|
|
size_t* parent_memtable_list_memory_usage,
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
int max_write_buffer_number_to_maintain)
|
2015-08-19 22:32:09 +02:00
|
|
|
: max_write_buffer_number_to_maintain_(max_write_buffer_number_to_maintain),
|
|
|
|
parent_memtable_list_memory_usage_(parent_memtable_list_memory_usage) {}
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
|
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 23:52:08 +01:00
|
|
|
void MemTableListVersion::Ref() { ++refs_; }
|
|
|
|
|
2015-08-19 22:32:09 +02:00
|
|
|
// called by superversion::clean()
|
2014-01-28 19:35:48 +01:00
|
|
|
void MemTableListVersion::Unref(autovector<MemTable*>* to_delete) {
|
2014-01-25 22:50:30 +01:00
|
|
|
assert(refs_ >= 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 23:52:08 +01:00
|
|
|
--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_) {
|
2015-08-19 22:32:09 +02:00
|
|
|
UnrefMemTable(to_delete, m);
|
2013-11-25 20:55:36 +01:00
|
|
|
}
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
for (const auto& m : memlist_history_) {
|
2015-08-19 22:32:09 +02:00
|
|
|
UnrefMemTable(to_delete, m);
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
}
|
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 23:52:08 +01:00
|
|
|
delete this;
|
2012-10-19 23:00:53 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
int MemTableList::NumNotFlushed() const {
|
|
|
|
int size = static_cast<int>(current_->memlist_.size());
|
|
|
|
assert(num_flush_not_started_ <= size);
|
|
|
|
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 23:52:08 +01:00
|
|
|
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
int MemTableList::NumFlushed() const {
|
|
|
|
return static_cast<int>(current_->memlist_history_.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 23:52:08 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
// 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.
|
2017-01-08 23:08:51 +01:00
|
|
|
bool MemTableListVersion::Get(const LookupKey& key, std::string* value,
|
2015-05-29 23:36:35 +02:00
|
|
|
Status* s, MergeContext* merge_context,
|
2016-11-04 02:40:23 +01:00
|
|
|
RangeDelAggregator* range_del_agg,
|
|
|
|
SequenceNumber* seq,
|
|
|
|
const ReadOptions& read_opts) {
|
2017-01-08 23:08:51 +01:00
|
|
|
return GetFromList(&memlist_, key, value, s, merge_context, range_del_agg,
|
2016-11-04 02:40:23 +01:00
|
|
|
seq, read_opts);
|
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 23:52:08 +01:00
|
|
|
}
|
|
|
|
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
bool MemTableListVersion::GetFromHistory(const LookupKey& key,
|
|
|
|
std::string* value, Status* s,
|
2015-05-29 23:36:35 +02:00
|
|
|
MergeContext* merge_context,
|
2016-11-04 02:40:23 +01:00
|
|
|
RangeDelAggregator* range_del_agg,
|
|
|
|
SequenceNumber* seq,
|
|
|
|
const ReadOptions& read_opts) {
|
2017-01-08 23:08:51 +01:00
|
|
|
return GetFromList(&memlist_history_, key, value, s, merge_context,
|
|
|
|
range_del_agg, seq, read_opts);
|
2015-05-29 23:36:35 +02:00
|
|
|
}
|
|
|
|
|
2017-01-08 23:08:51 +01:00
|
|
|
bool MemTableListVersion::GetFromList(std::list<MemTable*>* list,
|
|
|
|
const LookupKey& key, std::string* value,
|
|
|
|
Status* s, MergeContext* merge_context,
|
|
|
|
RangeDelAggregator* range_del_agg,
|
|
|
|
SequenceNumber* seq,
|
|
|
|
const ReadOptions& read_opts) {
|
2015-05-29 23:36:35 +02:00
|
|
|
*seq = kMaxSequenceNumber;
|
|
|
|
|
|
|
|
for (auto& memtable : *list) {
|
|
|
|
SequenceNumber current_seq = kMaxSequenceNumber;
|
|
|
|
|
2017-01-08 23:08:51 +01:00
|
|
|
bool done = memtable->Get(key, value, s, merge_context, range_del_agg,
|
2016-11-04 02:40:23 +01:00
|
|
|
¤t_seq, read_opts);
|
2015-05-29 23:36:35 +02:00
|
|
|
if (*seq == kMaxSequenceNumber) {
|
|
|
|
// Store the most recent sequence number of any operation on this key.
|
|
|
|
// Since we only care about the most recent change, we only need to
|
|
|
|
// return the first operation found when searching memtables in
|
|
|
|
// reverse-chronological order.
|
|
|
|
*seq = current_seq;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (done) {
|
|
|
|
assert(*seq != kMaxSequenceNumber);
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
return true;
|
|
|
|
}
|
2016-11-04 19:53:38 +01:00
|
|
|
if (!done && !s->ok() && !s->IsMergeInProgress() && !s->IsNotFound()) {
|
|
|
|
return false;
|
|
|
|
}
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
}
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2016-11-04 19:53:38 +01:00
|
|
|
Status MemTableListVersion::AddRangeTombstoneIterators(
|
|
|
|
const ReadOptions& read_opts, Arena* arena,
|
|
|
|
RangeDelAggregator* range_del_agg) {
|
|
|
|
assert(range_del_agg != nullptr);
|
|
|
|
for (auto& m : memlist_) {
|
2016-11-19 23:14:35 +01:00
|
|
|
std::unique_ptr<InternalIterator> range_del_iter(
|
|
|
|
m->NewRangeTombstoneIterator(read_opts));
|
2016-11-04 19:53:38 +01:00
|
|
|
Status s = range_del_agg->AddTombstones(std::move(range_del_iter));
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
2015-10-13 00:06:38 +02:00
|
|
|
void MemTableListVersion::AddIterators(
|
|
|
|
const ReadOptions& options, std::vector<InternalIterator*>* iterator_list,
|
|
|
|
Arena* arena) {
|
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 23:52:08 +01:00
|
|
|
for (auto& m : memlist_) {
|
2014-09-05 02:40:41 +02:00
|
|
|
iterator_list->push_back(m->NewIterator(options, arena));
|
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 23:52:08 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
In DB::NewIterator(), try to allocate the whole iterator tree in an arena
Summary:
In this patch, try to allocate the whole iterator tree starting from DBIter from an arena
1. ArenaWrappedDBIter is created when serves as the entry point of an iterator tree, with an arena in it.
2. Add an option to create iterator from arena for following iterators: DBIter, MergingIterator, MemtableIterator, all mem table's iterators, all table reader's iterators and two level iterator.
3. MergeIteratorBuilder is created to incrementally build the tree of internal iterators. It is passed to mem table list and version set and add iterators to it.
Limitations:
(1) Only DB::NewIterator() without tailing uses the arena. Other cases, including readonly DB and compactions are still from malloc
(2) Two level iterator itself is allocated in arena, but not iterators inside it.
Test Plan: make all check
Reviewers: ljin, haobo
Reviewed By: haobo
Subscribers: leveldb, dhruba, yhchiang, igor
Differential Revision: https://reviews.facebook.net/D18513
2014-06-03 01:38:00 +02:00
|
|
|
void MemTableListVersion::AddIterators(
|
|
|
|
const ReadOptions& options, MergeIteratorBuilder* merge_iter_builder) {
|
|
|
|
for (auto& m : memlist_) {
|
|
|
|
merge_iter_builder->AddIterator(
|
|
|
|
m->NewIterator(options, merge_iter_builder->GetArena()));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2014-04-23 02:17:33 +02:00
|
|
|
uint64_t MemTableListVersion::GetTotalNumEntries() const {
|
|
|
|
uint64_t total_num = 0;
|
|
|
|
for (auto& m : memlist_) {
|
2015-03-19 00:11:02 +01:00
|
|
|
total_num += m->num_entries();
|
|
|
|
}
|
|
|
|
return total_num;
|
|
|
|
}
|
|
|
|
|
2017-02-06 23:42:38 +01:00
|
|
|
MemTable::MemTableStats MemTableListVersion::ApproximateStats(
|
|
|
|
const Slice& start_ikey, const Slice& end_ikey) {
|
|
|
|
MemTable::MemTableStats total_stats = {0, 0};
|
2015-06-13 03:04:30 +02:00
|
|
|
for (auto& m : memlist_) {
|
2017-02-06 23:42:38 +01:00
|
|
|
auto mStats = m->ApproximateStats(start_ikey, end_ikey);
|
|
|
|
total_stats.size += mStats.size;
|
|
|
|
total_stats.count += mStats.count;
|
2015-06-13 03:04:30 +02:00
|
|
|
}
|
2017-02-06 23:42:38 +01:00
|
|
|
return total_stats;
|
2015-06-13 03:04:30 +02:00
|
|
|
}
|
|
|
|
|
2015-03-19 00:11:02 +01:00
|
|
|
uint64_t MemTableListVersion::GetTotalNumDeletes() const {
|
|
|
|
uint64_t total_num = 0;
|
|
|
|
for (auto& m : memlist_) {
|
|
|
|
total_num += m->num_deletes();
|
2014-04-23 02:17:33 +02:00
|
|
|
}
|
|
|
|
return total_num;
|
|
|
|
}
|
|
|
|
|
2015-05-29 23:36:35 +02:00
|
|
|
SequenceNumber MemTableListVersion::GetEarliestSequenceNumber(
|
|
|
|
bool include_history) const {
|
|
|
|
if (include_history && !memlist_history_.empty()) {
|
|
|
|
return memlist_history_.back()->GetEarliestSequenceNumber();
|
|
|
|
} else if (!memlist_.empty()) {
|
|
|
|
return memlist_.back()->GetEarliestSequenceNumber();
|
|
|
|
} else {
|
|
|
|
return kMaxSequenceNumber;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2014-01-27 02:40:43 +01:00
|
|
|
// caller is responsible for referencing m
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
void MemTableListVersion::Add(MemTable* m, autovector<MemTable*>* to_delete) {
|
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 23:52:08 +01:00
|
|
|
assert(refs_ == 1); // only when refs_ == 1 is MemTableListVersion mutable
|
2015-08-19 22:32:09 +02:00
|
|
|
AddMemTable(m);
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
|
|
|
|
TrimHistory(to_delete);
|
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 23:52:08 +01:00
|
|
|
}
|
|
|
|
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
// Removes m from list of memtables not flushed. Caller should NOT Unref m.
|
|
|
|
void MemTableListVersion::Remove(MemTable* m,
|
|
|
|
autovector<MemTable*>* to_delete) {
|
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 23:52:08 +01:00
|
|
|
assert(refs_ == 1); // only when refs_ == 1 is MemTableListVersion mutable
|
|
|
|
memlist_.remove(m);
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
|
|
|
|
if (max_write_buffer_number_to_maintain_ > 0) {
|
|
|
|
memlist_history_.push_front(m);
|
|
|
|
TrimHistory(to_delete);
|
|
|
|
} else {
|
2015-08-19 22:32:09 +02:00
|
|
|
UnrefMemTable(to_delete, m);
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Make sure we don't use up too much space in history
|
|
|
|
void MemTableListVersion::TrimHistory(autovector<MemTable*>* to_delete) {
|
|
|
|
while (memlist_.size() + memlist_history_.size() >
|
|
|
|
static_cast<size_t>(max_write_buffer_number_to_maintain_) &&
|
|
|
|
!memlist_history_.empty()) {
|
|
|
|
MemTable* x = memlist_history_.back();
|
|
|
|
memlist_history_.pop_back();
|
|
|
|
|
2015-08-19 22:32:09 +02:00
|
|
|
UnrefMemTable(to_delete, x);
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
}
|
2012-10-19 23:00:53 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
// Returns true if there is at least one memtable on which flush has
|
|
|
|
// not yet started.
|
2014-03-18 20:37:42 +01:00
|
|
|
bool MemTableList::IsFlushPending() const {
|
2013-09-05 02:24:35 +02:00
|
|
|
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 23:52:08 +01:00
|
|
|
(num_flush_not_started_ >= min_write_buffer_number_to_merge_)) {
|
2014-10-27 22:50:21 +01:00
|
|
|
assert(imm_flush_needed.load(std::memory_order_relaxed));
|
2012-10-19 23:00:53 +02:00
|
|
|
return true;
|
|
|
|
}
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2013-06-11 23:23:58 +02:00
|
|
|
// Returns the memtables that need to be flushed.
|
2014-01-02 20:26:57 +01:00
|
|
|
void MemTableList::PickMemtablesToFlush(autovector<MemTable*>* ret) {
|
2015-03-13 18:45:40 +01:00
|
|
|
AutoThreadOperationStageUpdater stage_updater(
|
|
|
|
ThreadStatus::STAGE_PICK_MEMTABLES_TO_FLUSH);
|
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 23:52:08 +01:00
|
|
|
const auto& memlist = current_->memlist_;
|
|
|
|
for (auto it = memlist.rbegin(); it != memlist.rend(); ++it) {
|
2012-10-19 23:00:53 +02:00
|
|
|
MemTable* m = *it;
|
|
|
|
if (!m->flush_in_progress_) {
|
|
|
|
assert(!m->flush_completed_);
|
2012-11-29 01:42:36 +01:00
|
|
|
num_flush_not_started_--;
|
2012-10-19 23:00:53 +02:00
|
|
|
if (num_flush_not_started_ == 0) {
|
2014-10-27 22:50:21 +01:00
|
|
|
imm_flush_needed.store(false, std::memory_order_release);
|
2012-10-19 23:00:53 +02:00
|
|
|
}
|
2014-01-28 19:35:48 +01:00
|
|
|
m->flush_in_progress_ = true; // flushing will start very soon
|
2013-06-11 23:23:58 +02:00
|
|
|
ret->push_back(m);
|
2012-10-19 23:00:53 +02:00
|
|
|
}
|
|
|
|
}
|
2014-01-28 19:35:48 +01:00
|
|
|
flush_requested_ = false; // start-flush request is complete
|
2012-10-19 23:00:53 +02:00
|
|
|
}
|
|
|
|
|
2014-02-12 20:42:54 +01:00
|
|
|
void MemTableList::RollbackMemtableFlush(const autovector<MemTable*>& mems,
|
2014-11-07 20:50:34 +01:00
|
|
|
uint64_t file_number) {
|
2015-03-13 18:45:40 +01:00
|
|
|
AutoThreadOperationStageUpdater stage_updater(
|
|
|
|
ThreadStatus::STAGE_MEMTABLE_ROLLBACK);
|
2014-02-12 20:42:54 +01:00
|
|
|
assert(!mems.empty());
|
|
|
|
|
|
|
|
// If the flush was not successful, then just reset state.
|
2015-04-25 11:14:27 +02:00
|
|
|
// Maybe a succeeding attempt to flush will be successful.
|
2014-02-12 20:42:54 +01:00
|
|
|
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_++;
|
|
|
|
}
|
2014-10-27 22:50:21 +01:00
|
|
|
imm_flush_needed.store(true, std::memory_order_release);
|
2014-02-12 20:42:54 +01:00
|
|
|
}
|
|
|
|
|
2012-10-19 23:00:53 +02:00
|
|
|
// Record a successful flush in the manifest file
|
2013-06-11 23:23:58 +02:00
|
|
|
Status MemTableList::InstallMemtableFlushResults(
|
2014-10-02 01:19:16 +02:00
|
|
|
ColumnFamilyData* cfd, const MutableCFOptions& mutable_cf_options,
|
2015-02-05 06:39:45 +01:00
|
|
|
const autovector<MemTable*>& mems, VersionSet* vset, InstrumentedMutex* mu,
|
2014-11-07 20:50:34 +01:00
|
|
|
uint64_t file_number, autovector<MemTable*>* to_delete,
|
|
|
|
Directory* db_directory, LogBuffer* log_buffer) {
|
2015-03-13 18:45:40 +01:00
|
|
|
AutoThreadOperationStageUpdater stage_updater(
|
|
|
|
ThreadStatus::STAGE_MEMTABLE_INSTALL_FLUSH_RESULTS);
|
2012-10-19 23:00:53 +02:00
|
|
|
mu->AssertHeld();
|
|
|
|
|
2015-04-25 11:14:27 +02:00
|
|
|
// flush was successful
|
2013-10-28 06:55:46 +01:00
|
|
|
for (size_t i = 0; i < mems.size(); ++i) {
|
2013-06-11 23:23:58 +02:00
|
|
|
// All the edits are associated with the first memtable of this batch.
|
2013-10-28 07:08:54 +01:00
|
|
|
assert(i == 0 || mems[i]->GetEdits()->NumEntries() == 0);
|
2013-06-11 23:23:58 +02:00
|
|
|
|
2013-10-28 06:55:46 +01:00
|
|
|
mems[i]->flush_completed_ = true;
|
|
|
|
mems[i]->file_number_ = file_number;
|
2013-06-11 23:23:58 +02:00
|
|
|
}
|
2012-10-19 23:00:53 +02:00
|
|
|
|
2015-04-25 11:14:27 +02:00
|
|
|
// if some other thread is already committing, then return
|
2012-10-19 23:00:53 +02:00
|
|
|
Status s;
|
|
|
|
if (commit_in_progress_) {
|
Fix flush not being commit while writing manifest
Summary:
Fix flush not being commit while writing manifest, which is a recent bug introduced by D60075.
The issue:
# Options.max_background_flushes > 1
# Background thread A pick up a flush job, flush, then commit to manifest. (Note that mutex is released before writing manifest.)
# Background thread B pick up another flush job, flush. When it gets to `MemTableList::InstallMemtableFlushResults`, it notices another thread is commiting, so it quit.
# After the first commit, thread A doesn't double check if there are more flush result need to commit, leaving the second flush uncommitted.
Test Plan: run the test. Also verify the new test hit deadlock without the fix.
Reviewers: sdong, igor, lightmark
Reviewed By: lightmark
Subscribers: andrewkr, omegaga, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D60969
2016-07-21 19:10:41 +02:00
|
|
|
TEST_SYNC_POINT("MemTableList::InstallMemtableFlushResults:InProgress");
|
2012-10-19 23:00:53 +02:00
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Only a single thread can be executing this piece of code
|
|
|
|
commit_in_progress_ = true;
|
|
|
|
|
Fix flush not being commit while writing manifest
Summary:
Fix flush not being commit while writing manifest, which is a recent bug introduced by D60075.
The issue:
# Options.max_background_flushes > 1
# Background thread A pick up a flush job, flush, then commit to manifest. (Note that mutex is released before writing manifest.)
# Background thread B pick up another flush job, flush. When it gets to `MemTableList::InstallMemtableFlushResults`, it notices another thread is commiting, so it quit.
# After the first commit, thread A doesn't double check if there are more flush result need to commit, leaving the second flush uncommitted.
Test Plan: run the test. Also verify the new test hit deadlock without the fix.
Reviewers: sdong, igor, lightmark
Reviewed By: lightmark
Subscribers: andrewkr, omegaga, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D60969
2016-07-21 19:10:41 +02:00
|
|
|
// Retry until all completed flushes are committed. New flushes can finish
|
|
|
|
// while the current thread is writing manifest where mutex is released.
|
|
|
|
while (s.ok()) {
|
|
|
|
auto& memlist = current_->memlist_;
|
|
|
|
if (memlist.empty() || !memlist.back()->flush_completed_) {
|
2012-10-19 23:00:53 +02:00
|
|
|
break;
|
|
|
|
}
|
Fix flush not being commit while writing manifest
Summary:
Fix flush not being commit while writing manifest, which is a recent bug introduced by D60075.
The issue:
# Options.max_background_flushes > 1
# Background thread A pick up a flush job, flush, then commit to manifest. (Note that mutex is released before writing manifest.)
# Background thread B pick up another flush job, flush. When it gets to `MemTableList::InstallMemtableFlushResults`, it notices another thread is commiting, so it quit.
# After the first commit, thread A doesn't double check if there are more flush result need to commit, leaving the second flush uncommitted.
Test Plan: run the test. Also verify the new test hit deadlock without the fix.
Reviewers: sdong, igor, lightmark
Reviewed By: lightmark
Subscribers: andrewkr, omegaga, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D60969
2016-07-21 19:10:41 +02:00
|
|
|
// 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.
|
|
|
|
uint64_t batch_file_number = 0;
|
|
|
|
size_t batch_count = 0;
|
|
|
|
autovector<VersionEdit*> edit_list;
|
|
|
|
// enumerate from the last (earliest) element to see how many batch finished
|
|
|
|
for (auto it = memlist.rbegin(); it != memlist.rend(); ++it) {
|
|
|
|
MemTable* m = *it;
|
|
|
|
if (!m->flush_completed_) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
if (it == memlist.rbegin() || batch_file_number != m->file_number_) {
|
|
|
|
batch_file_number = m->file_number_;
|
2017-03-16 03:22:52 +01:00
|
|
|
ROCKS_LOG_BUFFER(log_buffer,
|
|
|
|
"[%s] Level-0 commit table #%" PRIu64 " started",
|
|
|
|
cfd->GetName().c_str(), m->file_number_);
|
Fix flush not being commit while writing manifest
Summary:
Fix flush not being commit while writing manifest, which is a recent bug introduced by D60075.
The issue:
# Options.max_background_flushes > 1
# Background thread A pick up a flush job, flush, then commit to manifest. (Note that mutex is released before writing manifest.)
# Background thread B pick up another flush job, flush. When it gets to `MemTableList::InstallMemtableFlushResults`, it notices another thread is commiting, so it quit.
# After the first commit, thread A doesn't double check if there are more flush result need to commit, leaving the second flush uncommitted.
Test Plan: run the test. Also verify the new test hit deadlock without the fix.
Reviewers: sdong, igor, lightmark
Reviewed By: lightmark
Subscribers: andrewkr, omegaga, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D60969
2016-07-21 19:10:41 +02:00
|
|
|
edit_list.push_back(&m->edit_);
|
|
|
|
}
|
|
|
|
batch_count++;
|
2016-07-06 03:09:59 +02:00
|
|
|
}
|
2013-06-11 23:23:58 +02:00
|
|
|
|
Fix flush not being commit while writing manifest
Summary:
Fix flush not being commit while writing manifest, which is a recent bug introduced by D60075.
The issue:
# Options.max_background_flushes > 1
# Background thread A pick up a flush job, flush, then commit to manifest. (Note that mutex is released before writing manifest.)
# Background thread B pick up another flush job, flush. When it gets to `MemTableList::InstallMemtableFlushResults`, it notices another thread is commiting, so it quit.
# After the first commit, thread A doesn't double check if there are more flush result need to commit, leaving the second flush uncommitted.
Test Plan: run the test. Also verify the new test hit deadlock without the fix.
Reviewers: sdong, igor, lightmark
Reviewed By: lightmark
Subscribers: andrewkr, omegaga, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D60969
2016-07-21 19:10:41 +02:00
|
|
|
if (batch_count > 0) {
|
|
|
|
// this can release and reacquire the mutex.
|
|
|
|
s = vset->LogAndApply(cfd, mutable_cf_options, edit_list, mu,
|
|
|
|
db_directory);
|
|
|
|
|
|
|
|
// 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 have been flushed.
|
|
|
|
if (s.ok()) { // commit new state
|
|
|
|
while (batch_count-- > 0) {
|
|
|
|
MemTable* m = current_->memlist_.back();
|
2017-03-16 03:22:52 +01:00
|
|
|
ROCKS_LOG_BUFFER(log_buffer, "[%s] Level-0 commit table #%" PRIu64
|
|
|
|
": memtable #%" PRIu64 " done",
|
|
|
|
cfd->GetName().c_str(), m->file_number_, mem_id);
|
Fix flush not being commit while writing manifest
Summary:
Fix flush not being commit while writing manifest, which is a recent bug introduced by D60075.
The issue:
# Options.max_background_flushes > 1
# Background thread A pick up a flush job, flush, then commit to manifest. (Note that mutex is released before writing manifest.)
# Background thread B pick up another flush job, flush. When it gets to `MemTableList::InstallMemtableFlushResults`, it notices another thread is commiting, so it quit.
# After the first commit, thread A doesn't double check if there are more flush result need to commit, leaving the second flush uncommitted.
Test Plan: run the test. Also verify the new test hit deadlock without the fix.
Reviewers: sdong, igor, lightmark
Reviewed By: lightmark
Subscribers: andrewkr, omegaga, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D60969
2016-07-21 19:10:41 +02:00
|
|
|
assert(m->file_number_ > 0);
|
|
|
|
current_->Remove(m, to_delete);
|
|
|
|
++mem_id;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
for (auto it = current_->memlist_.rbegin(); batch_count-- > 0; it++) {
|
|
|
|
MemTable* m = *it;
|
|
|
|
// commit failed. setup state so that we can flush again.
|
2017-03-16 03:22:52 +01:00
|
|
|
ROCKS_LOG_BUFFER(log_buffer, "Level-0 commit table #%" PRIu64
|
|
|
|
": memtable #%" PRIu64 " failed",
|
|
|
|
m->file_number_, mem_id);
|
Fix flush not being commit while writing manifest
Summary:
Fix flush not being commit while writing manifest, which is a recent bug introduced by D60075.
The issue:
# Options.max_background_flushes > 1
# Background thread A pick up a flush job, flush, then commit to manifest. (Note that mutex is released before writing manifest.)
# Background thread B pick up another flush job, flush. When it gets to `MemTableList::InstallMemtableFlushResults`, it notices another thread is commiting, so it quit.
# After the first commit, thread A doesn't double check if there are more flush result need to commit, leaving the second flush uncommitted.
Test Plan: run the test. Also verify the new test hit deadlock without the fix.
Reviewers: sdong, igor, lightmark
Reviewed By: lightmark
Subscribers: andrewkr, omegaga, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D60969
2016-07-21 19:10:41 +02:00
|
|
|
m->flush_completed_ = false;
|
|
|
|
m->flush_in_progress_ = false;
|
|
|
|
m->edit_.Clear();
|
|
|
|
num_flush_not_started_++;
|
|
|
|
m->file_number_ = 0;
|
|
|
|
imm_flush_needed.store(true, std::memory_order_release);
|
|
|
|
++mem_id;
|
|
|
|
}
|
2013-06-11 23:23:58 +02:00
|
|
|
}
|
2016-07-06 03:09:59 +02:00
|
|
|
}
|
2012-10-19 23:00:53 +02:00
|
|
|
}
|
|
|
|
commit_in_progress_ = false;
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
2012-11-29 01:42:36 +01:00
|
|
|
// New memtables are inserted at the front of the list.
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
void MemTableList::Add(MemTable* m, autovector<MemTable*>* to_delete) {
|
|
|
|
assert(static_cast<int>(current_->memlist_.size()) >= num_flush_not_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 23:52:08 +01:00
|
|
|
InstallNewVersion();
|
2014-01-27 02:40:43 +01:00
|
|
|
// this method is used to move mutable memtable into an immutable list.
|
|
|
|
// since mutable memtable is already refcounted by the DBImpl,
|
|
|
|
// and when moving to the imutable list we don't unref it,
|
|
|
|
// we don't have to ref the memtable here. we just take over the
|
|
|
|
// reference from the DBImpl.
|
Support saving history in memtable_list
Summary:
For transactions, we are using the memtables to validate that there are no write conflicts. But after flushing, we don't have any memtables, and transactions could fail to commit. So we want to someone keep around some extra history to use for conflict checking. In addition, we want to provide a way to increase the size of this history if too many transactions fail to commit.
After chatting with people, it seems like everyone prefers just using Memtables to store this history (instead of a separate history structure). It seems like the best place for this is abstracted inside the memtable_list. I decide to create a separate list in MemtableListVersion as using the same list complicated the flush/installalflushresults logic too much.
This diff adds a new parameter to control how much memtable history to keep around after flushing. However, it sounds like people aren't too fond of adding new parameters. So I am making the default size of flushed+not-flushed memtables be set to max_write_buffers. This should not change the maximum amount of memory used, but make it more likely we're using closer the the limit. (We are now postponing deleting flushed memtables until the max_write_buffer limit is reached). So while we might use more memory on average, we are still obeying the limit set (and you could argue it's better to go ahead and use up memory now instead of waiting for a write stall to happen to test this limit).
However, if people are opposed to this default behavior, we can easily set it to 0 and require this parameter be set in order to use transactions.
Test Plan: Added a xfunc test to play around with setting different values of this parameter in all tests. Added testing in memtablelist_test and planning on adding more testing here.
Reviewers: sdong, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D37443
2015-05-29 01:34:24 +02:00
|
|
|
current_->Add(m, to_delete);
|
2013-08-23 08:10:02 +02:00
|
|
|
m->MarkImmutable();
|
2012-10-19 23:00:53 +02:00
|
|
|
num_flush_not_started_++;
|
|
|
|
if (num_flush_not_started_ == 1) {
|
2014-10-27 22:50:21 +01:00
|
|
|
imm_flush_needed.store(true, std::memory_order_release);
|
2012-10-19 23:00:53 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Returns an estimate of the number of bytes of data in use.
|
2015-08-19 22:32:09 +02:00
|
|
|
size_t MemTableList::ApproximateUnflushedMemTablesMemoryUsage() {
|
2014-11-06 20:14:28 +01:00
|
|
|
size_t total_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 23:52:08 +01:00
|
|
|
for (auto& memtable : current_->memlist_) {
|
2014-11-06 20:14:28 +01:00
|
|
|
total_size += memtable->ApproximateMemoryUsage();
|
2012-10-19 23:00:53 +02:00
|
|
|
}
|
2014-11-06 20:14:28 +01:00
|
|
|
return total_size;
|
2012-10-19 23:00:53 +02:00
|
|
|
}
|
|
|
|
|
2015-08-19 22:32:09 +02:00
|
|
|
size_t MemTableList::ApproximateMemoryUsage() { return current_memory_usage_; }
|
|
|
|
|
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 23:52:08 +01:00
|
|
|
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_;
|
2015-08-19 22:32:09 +02:00
|
|
|
current_ = new MemTableListVersion(¤t_memory_usage_, current_);
|
2014-01-25 22:50:30 +01:00
|
|
|
current_->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 23:52:08 +01:00
|
|
|
version->Unref();
|
2012-10-19 23:00:53 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2016-04-18 20:11:51 +02:00
|
|
|
uint64_t MemTableList::GetMinLogContainingPrepSection() {
|
|
|
|
uint64_t min_log = 0;
|
|
|
|
|
|
|
|
for (auto& m : current_->memlist_) {
|
|
|
|
// this mem has been flushed it no longer
|
|
|
|
// needs to hold on the its prep section
|
|
|
|
if (m->flush_completed_) {
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
auto log = m->GetMinLogContainingPrepSection();
|
|
|
|
|
|
|
|
if (log > 0 && (min_log == 0 || log < min_log)) {
|
|
|
|
min_log = log;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return min_log;
|
|
|
|
}
|
|
|
|
|
2013-10-04 06:49:15 +02:00
|
|
|
} // namespace rocksdb
|