rocksdb/table/block_based/block_builder.cc

233 lines
8.6 KiB
C++
Raw Normal View History

// Copyright (c) 2011-present, Facebook, Inc. All rights reserved.
// This source code is licensed under both the GPLv2 (found in the
// COPYING file in the root directory) and Apache 2.0 License
// (found in the LICENSE.Apache file in the root directory).
//
// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file. See the AUTHORS file for names of contributors.
//
// BlockBuilder generates blocks where keys are prefix-compressed:
//
// When we store a key, we drop the prefix shared with the previous
// string. This helps reduce the space requirement significantly.
// Furthermore, once every K keys, we do not apply the prefix
// compression and store the entire key. We call this a "restart
// point". The tail end of the block stores the offsets of all of the
// restart points, and can be used to do a binary search when looking
// for a particular key. Values are stored as-is (without compression)
// immediately following the corresponding key.
//
// An entry for a particular key-value pair has the form:
// shared_bytes: varint32
// unshared_bytes: varint32
// value_length: varint32
// key_delta: char[unshared_bytes]
// value: char[value_length]
// shared_bytes == 0 for restart points.
//
// The trailer of the block has the form:
// restarts: uint32[num_restarts]
// num_restarts: uint32
// restarts[i] contains the offset within the block of the ith restart point.
#include "table/block_based/block_builder.h"
#include <assert.h>
#include <algorithm>
#include "db/dbformat.h"
#include "rocksdb/comparator.h"
#include "table/block_based/data_block_footer.h"
#include "util/coding.h"
namespace ROCKSDB_NAMESPACE {
BlockBuilder::BlockBuilder(
int block_restart_interval, bool use_delta_encoding,
bool use_value_delta_encoding,
BlockBasedTableOptions::DataBlockIndexType index_type,
double data_block_hash_table_util_ratio)
: block_restart_interval_(block_restart_interval),
Introduce ReadOptions::pin_data (support zero copy for keys) Summary: This patch update the Iterator API to introduce new functions that allow users to keep the Slices returned by key() valid as long as the Iterator is not deleted ReadOptions::pin_data : If true keep loaded blocks in memory as long as the iterator is not deleted Iterator::IsKeyPinned() : If true, this mean that the Slice returned by key() is valid as long as the iterator is not deleted Also add a new option BlockBasedTableOptions::use_delta_encoding to allow users to disable delta_encoding if needed. Benchmark results (using https://phabricator.fb.com/P20083553) ``` // $ du -h /home/tec/local/normal.4K.Snappy/db10077 // 6.1G /home/tec/local/normal.4K.Snappy/db10077 // $ du -h /home/tec/local/zero.8K.LZ4/db10077 // 6.4G /home/tec/local/zero.8K.LZ4/db10077 // Benchmarks for shard db10077 // _build/opt/rocks/benchmark/rocks_copy_benchmark \ // --normal_db_path="/home/tec/local/normal.4K.Snappy/db10077" \ // --zero_db_path="/home/tec/local/zero.8K.LZ4/db10077" // First run // ============================================================================ // rocks/benchmark/RocksCopyBenchmark.cpp relative time/iter iters/s // ============================================================================ // BM_StringCopy 1.73s 576.97m // BM_StringPiece 103.74% 1.67s 598.55m // ============================================================================ // Match rate : 1000000 / 1000000 // Second run // ============================================================================ // rocks/benchmark/RocksCopyBenchmark.cpp relative time/iter iters/s // ============================================================================ // BM_StringCopy 611.99ms 1.63 // BM_StringPiece 203.76% 300.35ms 3.33 // ============================================================================ // Match rate : 1000000 / 1000000 ``` Test Plan: Unit tests Reviewers: sdong, igor, anthony, yhchiang, rven Reviewed By: rven Subscribers: dhruba, lovro, adsharma Differential Revision: https://reviews.facebook.net/D48999
2015-12-16 21:08:30 +01:00
use_delta_encoding_(use_delta_encoding),
use_value_delta_encoding_(use_value_delta_encoding),
restarts_(1, 0), // First restart point is at offset 0
counter_(0),
finished_(false) {
switch (index_type) {
case BlockBasedTableOptions::kDataBlockBinarySearch:
break;
case BlockBasedTableOptions::kDataBlockBinaryAndHash:
data_block_hash_index_builder_.Initialize(
data_block_hash_table_util_ratio);
break;
default:
assert(0);
}
assert(block_restart_interval_ >= 1);
Miscellaneous performance improvements Summary: I was investigating performance issues in the SstFileWriter and found all of the following: - The SstFileWriter::Add() function created a local InternalKey every time it was called generating a allocation and free each time. Changed to have an InternalKey member variable that can be reset with the new InternalKey::Set() function. - In SstFileWriter::Add() the smallest_key and largest_key values were assigned the result of a ToString() call, but it is simpler to just assign them directly from the user's key. - The Slice class had no move constructor so each time one was returned from a function a new one had to be allocated, the old data copied to the new, and the old one was freed. I added the move constructor which also required a copy constructor and assignment operator. - The BlockBuilder::CurrentSizeEstimate() function calculates the current estimate size, but was being called 2 or 3 times for each key added. I changed the class to maintain a running estimate (equal to the original calculation) so that the function can return an already calculated value. - The code in BlockBuilder::Add() that calculated the shared bytes between the last key and the new key duplicated what Slice::difference_offset does, so I replaced it with the standard function. - BlockBuilder::Add() had code to copy just the changed portion into the last key value (and asserted that it now matched the new key). It is more efficient just to copy the whole new key over. - Moved this same code up into the 'if (use_delta_encoding_)' since the last key value is only needed when delta encoding is on. - FlushBlockBySizePolicy::BlockAlmostFull calculated a standard deviation value each time it was called, but this information would only change if block_size of block_size_deviation changed, so I created a member variable to hold the value to avoid the calculation each time. - Each PutVarint??() function has a buffer and calls std::string::append(). Two or three calls in a row could share a buffer and a single call to std::string::append(). Some of these will be helpful outside of the SstFileWriter. I'm not 100% the addition of the move constructor is appropriate as I wonder why this wasn't done before - maybe because of compiler compatibility? I tried it on gcc 4.8 and 4.9. Test Plan: The changes should not affect the results so the existing tests should all still work and no new tests were added. The value of the changes was seen by manually testing the SstFileWriter class through MyRocks and adding timing code to identify problem areas. Reviewers: sdong, IslamAbdelRahman Reviewed By: IslamAbdelRahman Subscribers: andrewkr, dhruba Differential Revision: https://reviews.facebook.net/D59607
2016-06-13 18:57:43 +02:00
estimate_ = sizeof(uint32_t) + sizeof(uint32_t);
}
void BlockBuilder::Reset() {
buffer_.clear();
restarts_.resize(1); // First restart point is at offset 0
assert(restarts_[0] == 0);
Miscellaneous performance improvements Summary: I was investigating performance issues in the SstFileWriter and found all of the following: - The SstFileWriter::Add() function created a local InternalKey every time it was called generating a allocation and free each time. Changed to have an InternalKey member variable that can be reset with the new InternalKey::Set() function. - In SstFileWriter::Add() the smallest_key and largest_key values were assigned the result of a ToString() call, but it is simpler to just assign them directly from the user's key. - The Slice class had no move constructor so each time one was returned from a function a new one had to be allocated, the old data copied to the new, and the old one was freed. I added the move constructor which also required a copy constructor and assignment operator. - The BlockBuilder::CurrentSizeEstimate() function calculates the current estimate size, but was being called 2 or 3 times for each key added. I changed the class to maintain a running estimate (equal to the original calculation) so that the function can return an already calculated value. - The code in BlockBuilder::Add() that calculated the shared bytes between the last key and the new key duplicated what Slice::difference_offset does, so I replaced it with the standard function. - BlockBuilder::Add() had code to copy just the changed portion into the last key value (and asserted that it now matched the new key). It is more efficient just to copy the whole new key over. - Moved this same code up into the 'if (use_delta_encoding_)' since the last key value is only needed when delta encoding is on. - FlushBlockBySizePolicy::BlockAlmostFull calculated a standard deviation value each time it was called, but this information would only change if block_size of block_size_deviation changed, so I created a member variable to hold the value to avoid the calculation each time. - Each PutVarint??() function has a buffer and calls std::string::append(). Two or three calls in a row could share a buffer and a single call to std::string::append(). Some of these will be helpful outside of the SstFileWriter. I'm not 100% the addition of the move constructor is appropriate as I wonder why this wasn't done before - maybe because of compiler compatibility? I tried it on gcc 4.8 and 4.9. Test Plan: The changes should not affect the results so the existing tests should all still work and no new tests were added. The value of the changes was seen by manually testing the SstFileWriter class through MyRocks and adding timing code to identify problem areas. Reviewers: sdong, IslamAbdelRahman Reviewed By: IslamAbdelRahman Subscribers: andrewkr, dhruba Differential Revision: https://reviews.facebook.net/D59607
2016-06-13 18:57:43 +02:00
estimate_ = sizeof(uint32_t) + sizeof(uint32_t);
counter_ = 0;
finished_ = false;
last_key_.clear();
if (data_block_hash_index_builder_.Valid()) {
data_block_hash_index_builder_.Reset();
}
#ifndef NDEBUG
add_with_last_key_called_ = false;
#endif
}
void BlockBuilder::SwapAndReset(std::string& buffer) {
std::swap(buffer_, buffer);
Reset();
}
size_t BlockBuilder::EstimateSizeAfterKV(const Slice& key,
const Slice& value) const {
size_t estimate = CurrentSizeEstimate();
// Note: this is an imprecise estimate as it accounts for the whole key size
// instead of non-shared key size.
estimate += key.size();
// In value delta encoding we estimate the value delta size as half the full
// value size since only the size field of block handle is encoded.
estimate +=
!use_value_delta_encoding_ || (counter_ >= block_restart_interval_)
? value.size()
: value.size() / 2;
if (counter_ >= block_restart_interval_) {
estimate += sizeof(uint32_t); // a new restart entry.
}
estimate += sizeof(int32_t); // varint for shared prefix length.
// Note: this is an imprecise estimate as we will have to encoded size, one
// for shared key and one for non-shared key.
estimate += VarintLength(key.size()); // varint for key length.
if (!use_value_delta_encoding_ || (counter_ >= block_restart_interval_)) {
estimate += VarintLength(value.size()); // varint for value length.
}
return estimate;
}
Slice BlockBuilder::Finish() {
// Append restart array
for (size_t i = 0; i < restarts_.size(); i++) {
PutFixed32(&buffer_, restarts_[i]);
}
uint32_t num_restarts = static_cast<uint32_t>(restarts_.size());
BlockBasedTableOptions::DataBlockIndexType index_type =
BlockBasedTableOptions::kDataBlockBinarySearch;
if (data_block_hash_index_builder_.Valid() &&
CurrentSizeEstimate() <= kMaxBlockSizeSupportedByHashIndex) {
data_block_hash_index_builder_.Finish(buffer_);
index_type = BlockBasedTableOptions::kDataBlockBinaryAndHash;
}
// footer is a packed format of data_block_index_type and num_restarts
uint32_t block_footer = PackIndexTypeAndNumRestarts(index_type, num_restarts);
PutFixed32(&buffer_, block_footer);
finished_ = true;
return Slice(buffer_);
}
void BlockBuilder::Add(const Slice& key, const Slice& value,
const Slice* const delta_value) {
// Ensure no unsafe mixing of Add and AddWithLastKey
assert(!add_with_last_key_called_);
AddWithLastKeyImpl(key, value, last_key_, delta_value, buffer_.size());
if (use_delta_encoding_) {
// Update state
// We used to just copy the changed data, but it appears to be
// faster to just copy the whole thing.
last_key_.assign(key.data(), key.size());
}
}
void BlockBuilder::AddWithLastKey(const Slice& key, const Slice& value,
const Slice& last_key_param,
const Slice* const delta_value) {
// Ensure no unsafe mixing of Add and AddWithLastKey
assert(last_key_.empty());
#ifndef NDEBUG
add_with_last_key_called_ = false;
#endif
// Here we make sure to use an empty `last_key` on first call after creation
// or Reset. This is more convenient for the caller and we can be more
// clever inside BlockBuilder. On this hot code path, we want to avoid
// conditional jumps like `buffer_.empty() ? ... : ...` so we can use a
// fast min operation instead, with an assertion to be sure our logic is
// sound.
size_t buffer_size = buffer_.size();
size_t last_key_size = last_key_param.size();
assert(buffer_size == 0 || buffer_size >= last_key_size);
Slice last_key(last_key_param.data(), std::min(buffer_size, last_key_size));
AddWithLastKeyImpl(key, value, last_key, delta_value, buffer_size);
}
inline void BlockBuilder::AddWithLastKeyImpl(const Slice& key,
const Slice& value,
const Slice& last_key,
const Slice* const delta_value,
size_t buffer_size) {
assert(!finished_);
assert(counter_ <= block_restart_interval_);
assert(!use_value_delta_encoding_ || delta_value);
Introduce ReadOptions::pin_data (support zero copy for keys) Summary: This patch update the Iterator API to introduce new functions that allow users to keep the Slices returned by key() valid as long as the Iterator is not deleted ReadOptions::pin_data : If true keep loaded blocks in memory as long as the iterator is not deleted Iterator::IsKeyPinned() : If true, this mean that the Slice returned by key() is valid as long as the iterator is not deleted Also add a new option BlockBasedTableOptions::use_delta_encoding to allow users to disable delta_encoding if needed. Benchmark results (using https://phabricator.fb.com/P20083553) ``` // $ du -h /home/tec/local/normal.4K.Snappy/db10077 // 6.1G /home/tec/local/normal.4K.Snappy/db10077 // $ du -h /home/tec/local/zero.8K.LZ4/db10077 // 6.4G /home/tec/local/zero.8K.LZ4/db10077 // Benchmarks for shard db10077 // _build/opt/rocks/benchmark/rocks_copy_benchmark \ // --normal_db_path="/home/tec/local/normal.4K.Snappy/db10077" \ // --zero_db_path="/home/tec/local/zero.8K.LZ4/db10077" // First run // ============================================================================ // rocks/benchmark/RocksCopyBenchmark.cpp relative time/iter iters/s // ============================================================================ // BM_StringCopy 1.73s 576.97m // BM_StringPiece 103.74% 1.67s 598.55m // ============================================================================ // Match rate : 1000000 / 1000000 // Second run // ============================================================================ // rocks/benchmark/RocksCopyBenchmark.cpp relative time/iter iters/s // ============================================================================ // BM_StringCopy 611.99ms 1.63 // BM_StringPiece 203.76% 300.35ms 3.33 // ============================================================================ // Match rate : 1000000 / 1000000 ``` Test Plan: Unit tests Reviewers: sdong, igor, anthony, yhchiang, rven Reviewed By: rven Subscribers: dhruba, lovro, adsharma Differential Revision: https://reviews.facebook.net/D48999
2015-12-16 21:08:30 +01:00
size_t shared = 0; // number of bytes shared with prev key
if (counter_ >= block_restart_interval_) {
// Restart compression
restarts_.push_back(static_cast<uint32_t>(buffer_size));
Miscellaneous performance improvements Summary: I was investigating performance issues in the SstFileWriter and found all of the following: - The SstFileWriter::Add() function created a local InternalKey every time it was called generating a allocation and free each time. Changed to have an InternalKey member variable that can be reset with the new InternalKey::Set() function. - In SstFileWriter::Add() the smallest_key and largest_key values were assigned the result of a ToString() call, but it is simpler to just assign them directly from the user's key. - The Slice class had no move constructor so each time one was returned from a function a new one had to be allocated, the old data copied to the new, and the old one was freed. I added the move constructor which also required a copy constructor and assignment operator. - The BlockBuilder::CurrentSizeEstimate() function calculates the current estimate size, but was being called 2 or 3 times for each key added. I changed the class to maintain a running estimate (equal to the original calculation) so that the function can return an already calculated value. - The code in BlockBuilder::Add() that calculated the shared bytes between the last key and the new key duplicated what Slice::difference_offset does, so I replaced it with the standard function. - BlockBuilder::Add() had code to copy just the changed portion into the last key value (and asserted that it now matched the new key). It is more efficient just to copy the whole new key over. - Moved this same code up into the 'if (use_delta_encoding_)' since the last key value is only needed when delta encoding is on. - FlushBlockBySizePolicy::BlockAlmostFull calculated a standard deviation value each time it was called, but this information would only change if block_size of block_size_deviation changed, so I created a member variable to hold the value to avoid the calculation each time. - Each PutVarint??() function has a buffer and calls std::string::append(). Two or three calls in a row could share a buffer and a single call to std::string::append(). Some of these will be helpful outside of the SstFileWriter. I'm not 100% the addition of the move constructor is appropriate as I wonder why this wasn't done before - maybe because of compiler compatibility? I tried it on gcc 4.8 and 4.9. Test Plan: The changes should not affect the results so the existing tests should all still work and no new tests were added. The value of the changes was seen by manually testing the SstFileWriter class through MyRocks and adding timing code to identify problem areas. Reviewers: sdong, IslamAbdelRahman Reviewed By: IslamAbdelRahman Subscribers: andrewkr, dhruba Differential Revision: https://reviews.facebook.net/D59607
2016-06-13 18:57:43 +02:00
estimate_ += sizeof(uint32_t);
Introduce ReadOptions::pin_data (support zero copy for keys) Summary: This patch update the Iterator API to introduce new functions that allow users to keep the Slices returned by key() valid as long as the Iterator is not deleted ReadOptions::pin_data : If true keep loaded blocks in memory as long as the iterator is not deleted Iterator::IsKeyPinned() : If true, this mean that the Slice returned by key() is valid as long as the iterator is not deleted Also add a new option BlockBasedTableOptions::use_delta_encoding to allow users to disable delta_encoding if needed. Benchmark results (using https://phabricator.fb.com/P20083553) ``` // $ du -h /home/tec/local/normal.4K.Snappy/db10077 // 6.1G /home/tec/local/normal.4K.Snappy/db10077 // $ du -h /home/tec/local/zero.8K.LZ4/db10077 // 6.4G /home/tec/local/zero.8K.LZ4/db10077 // Benchmarks for shard db10077 // _build/opt/rocks/benchmark/rocks_copy_benchmark \ // --normal_db_path="/home/tec/local/normal.4K.Snappy/db10077" \ // --zero_db_path="/home/tec/local/zero.8K.LZ4/db10077" // First run // ============================================================================ // rocks/benchmark/RocksCopyBenchmark.cpp relative time/iter iters/s // ============================================================================ // BM_StringCopy 1.73s 576.97m // BM_StringPiece 103.74% 1.67s 598.55m // ============================================================================ // Match rate : 1000000 / 1000000 // Second run // ============================================================================ // rocks/benchmark/RocksCopyBenchmark.cpp relative time/iter iters/s // ============================================================================ // BM_StringCopy 611.99ms 1.63 // BM_StringPiece 203.76% 300.35ms 3.33 // ============================================================================ // Match rate : 1000000 / 1000000 ``` Test Plan: Unit tests Reviewers: sdong, igor, anthony, yhchiang, rven Reviewed By: rven Subscribers: dhruba, lovro, adsharma Differential Revision: https://reviews.facebook.net/D48999
2015-12-16 21:08:30 +01:00
counter_ = 0;
} else if (use_delta_encoding_) {
// See how much sharing to do with previous string
shared = key.difference_offset(last_key);
}
Miscellaneous performance improvements Summary: I was investigating performance issues in the SstFileWriter and found all of the following: - The SstFileWriter::Add() function created a local InternalKey every time it was called generating a allocation and free each time. Changed to have an InternalKey member variable that can be reset with the new InternalKey::Set() function. - In SstFileWriter::Add() the smallest_key and largest_key values were assigned the result of a ToString() call, but it is simpler to just assign them directly from the user's key. - The Slice class had no move constructor so each time one was returned from a function a new one had to be allocated, the old data copied to the new, and the old one was freed. I added the move constructor which also required a copy constructor and assignment operator. - The BlockBuilder::CurrentSizeEstimate() function calculates the current estimate size, but was being called 2 or 3 times for each key added. I changed the class to maintain a running estimate (equal to the original calculation) so that the function can return an already calculated value. - The code in BlockBuilder::Add() that calculated the shared bytes between the last key and the new key duplicated what Slice::difference_offset does, so I replaced it with the standard function. - BlockBuilder::Add() had code to copy just the changed portion into the last key value (and asserted that it now matched the new key). It is more efficient just to copy the whole new key over. - Moved this same code up into the 'if (use_delta_encoding_)' since the last key value is only needed when delta encoding is on. - FlushBlockBySizePolicy::BlockAlmostFull calculated a standard deviation value each time it was called, but this information would only change if block_size of block_size_deviation changed, so I created a member variable to hold the value to avoid the calculation each time. - Each PutVarint??() function has a buffer and calls std::string::append(). Two or three calls in a row could share a buffer and a single call to std::string::append(). Some of these will be helpful outside of the SstFileWriter. I'm not 100% the addition of the move constructor is appropriate as I wonder why this wasn't done before - maybe because of compiler compatibility? I tried it on gcc 4.8 and 4.9. Test Plan: The changes should not affect the results so the existing tests should all still work and no new tests were added. The value of the changes was seen by manually testing the SstFileWriter class through MyRocks and adding timing code to identify problem areas. Reviewers: sdong, IslamAbdelRahman Reviewed By: IslamAbdelRahman Subscribers: andrewkr, dhruba Differential Revision: https://reviews.facebook.net/D59607
2016-06-13 18:57:43 +02:00
const size_t non_shared = key.size() - shared;
if (use_value_delta_encoding_) {
// Add "<shared><non_shared>" to buffer_
PutVarint32Varint32(&buffer_, static_cast<uint32_t>(shared),
static_cast<uint32_t>(non_shared));
} else {
// Add "<shared><non_shared><value_size>" to buffer_
PutVarint32Varint32Varint32(&buffer_, static_cast<uint32_t>(shared),
static_cast<uint32_t>(non_shared),
static_cast<uint32_t>(value.size()));
}
// Add string delta to buffer_ followed by value
buffer_.append(key.data() + shared, non_shared);
// Use value delta encoding only when the key has shared bytes. This would
// simplify the decoding, where it can figure which decoding to use simply by
// looking at the shared bytes size.
if (shared != 0 && use_value_delta_encoding_) {
buffer_.append(delta_value->data(), delta_value->size());
} else {
buffer_.append(value.data(), value.size());
}
if (data_block_hash_index_builder_.Valid()) {
data_block_hash_index_builder_.Add(ExtractUserKey(key),
restarts_.size() - 1);
}
counter_++;
estimate_ += buffer_.size() - buffer_size;
}
} // namespace ROCKSDB_NAMESPACE