70c7302fb5
Summary: This PR implements cache eviction using reinforcement learning. It includes two implementations: 1. An implementation of Thompson Sampling for the Bernoulli Bandit [1]. 2. An implementation of LinUCB with disjoint linear models [2]. The idea is that a cache uses multiple eviction policies, e.g., MRU, LRU, and LFU. The cache learns which eviction policy is the best and uses it upon a cache miss. Thompson Sampling is contextless and does not include any features. LinUCB includes features such as level, block type, caller, column family id to decide which eviction policy to use. [1] Daniel J. Russo, Benjamin Van Roy, Abbas Kazerouni, Ian Osband, and Zheng Wen. 2018. A Tutorial on Thompson Sampling. Found. Trends Mach. Learn. 11, 1 (July 2018), 1-96. DOI: https://doi.org/10.1561/2200000070 [2] Lihong Li, Wei Chu, John Langford, and Robert E. Schapire. 2010. A contextual-bandit approach to personalized news article recommendation. In Proceedings of the 19th international conference on World wide web (WWW '10). ACM, New York, NY, USA, 661-670. DOI=http://dx.doi.org/10.1145/1772690.1772758 Pull Request resolved: https://github.com/facebook/rocksdb/pull/5610 Differential Revision: D16435067 Pulled By: HaoyuHuang fbshipit-source-id: 6549239ae14115c01cb1e70548af9e46d8dc21bb
386 lines
16 KiB
C++
386 lines
16 KiB
C++
// 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).
|
|
|
|
#pragma once
|
|
|
|
#include <map>
|
|
#include <set>
|
|
#include <vector>
|
|
|
|
#include "db/dbformat.h"
|
|
#include "rocksdb/env.h"
|
|
#include "rocksdb/utilities/sim_cache.h"
|
|
#include "trace_replay/block_cache_tracer.h"
|
|
#include "utilities/simulator_cache/cache_simulator.h"
|
|
|
|
namespace rocksdb {
|
|
|
|
// Statistics of a key refereneced by a Get.
|
|
struct GetKeyInfo {
|
|
uint64_t key_id = 0;
|
|
std::vector<uint64_t> access_sequence_number_timeline;
|
|
std::vector<uint64_t> access_timeline;
|
|
|
|
void AddAccess(const BlockCacheTraceRecord& access,
|
|
uint64_t access_sequnce_number) {
|
|
access_sequence_number_timeline.push_back(access_sequnce_number);
|
|
access_timeline.push_back(access.access_timestamp);
|
|
}
|
|
};
|
|
|
|
// Statistics of a block.
|
|
struct BlockAccessInfo {
|
|
uint64_t block_id = 0;
|
|
uint64_t num_accesses = 0;
|
|
uint64_t block_size = 0;
|
|
uint64_t first_access_time = 0;
|
|
uint64_t last_access_time = 0;
|
|
uint64_t num_keys = 0;
|
|
std::map<std::string, std::map<TableReaderCaller, uint64_t>>
|
|
key_num_access_map; // for keys exist in this block.
|
|
std::map<std::string, std::map<TableReaderCaller, uint64_t>>
|
|
non_exist_key_num_access_map; // for keys do not exist in this block.
|
|
uint64_t num_referenced_key_exist_in_block = 0;
|
|
uint64_t referenced_data_size = 0;
|
|
std::map<TableReaderCaller, uint64_t> caller_num_access_map;
|
|
// caller:timestamp:number_of_accesses. The granularity of the timestamp is
|
|
// seconds.
|
|
std::map<TableReaderCaller, std::map<uint64_t, uint64_t>>
|
|
caller_num_accesses_timeline;
|
|
// Unique blocks since the last access.
|
|
std::set<std::string> unique_blocks_since_last_access;
|
|
// Number of reuses grouped by reuse distance.
|
|
std::map<uint64_t, uint64_t> reuse_distance_count;
|
|
|
|
// The access sequence numbers of this block.
|
|
std::vector<uint64_t> access_sequence_number_timeline;
|
|
std::map<TableReaderCaller, std::vector<uint64_t>>
|
|
caller_access_sequence__number_timeline;
|
|
// The access timestamp in microseconds of this block.
|
|
std::vector<uint64_t> access_timeline;
|
|
std::map<TableReaderCaller, std::vector<uint64_t>> caller_access_timeline;
|
|
|
|
void AddAccess(const BlockCacheTraceRecord& access,
|
|
uint64_t access_sequnce_number) {
|
|
if (block_size != 0 && access.block_size != 0) {
|
|
assert(block_size == access.block_size);
|
|
}
|
|
if (num_keys != 0 && access.num_keys_in_block != 0) {
|
|
assert(num_keys == access.num_keys_in_block);
|
|
}
|
|
if (first_access_time == 0) {
|
|
first_access_time = access.access_timestamp;
|
|
}
|
|
last_access_time = access.access_timestamp;
|
|
block_size = access.block_size;
|
|
caller_num_access_map[access.caller]++;
|
|
num_accesses++;
|
|
// access.access_timestamp is in microsecond.
|
|
const uint64_t timestamp_in_seconds =
|
|
access.access_timestamp / kMicrosInSecond;
|
|
caller_num_accesses_timeline[access.caller][timestamp_in_seconds] += 1;
|
|
// Populate the feature vectors.
|
|
access_sequence_number_timeline.push_back(access_sequnce_number);
|
|
caller_access_sequence__number_timeline[access.caller].push_back(
|
|
access_sequnce_number);
|
|
access_timeline.push_back(access.access_timestamp);
|
|
caller_access_timeline[access.caller].push_back(access.access_timestamp);
|
|
if (BlockCacheTraceHelper::IsGetOrMultiGetOnDataBlock(access.block_type,
|
|
access.caller)) {
|
|
num_keys = access.num_keys_in_block;
|
|
if (access.referenced_key_exist_in_block == Boolean::kTrue) {
|
|
if (key_num_access_map.find(access.referenced_key) ==
|
|
key_num_access_map.end()) {
|
|
referenced_data_size += access.referenced_data_size;
|
|
}
|
|
key_num_access_map[access.referenced_key][access.caller]++;
|
|
num_referenced_key_exist_in_block++;
|
|
if (referenced_data_size > block_size && block_size != 0) {
|
|
ParsedInternalKey internal_key;
|
|
ParseInternalKey(access.referenced_key, &internal_key);
|
|
}
|
|
} else {
|
|
non_exist_key_num_access_map[access.referenced_key][access.caller]++;
|
|
}
|
|
}
|
|
}
|
|
};
|
|
|
|
// Aggregates stats of a block given a block type.
|
|
struct BlockTypeAccessInfoAggregate {
|
|
std::map<std::string, BlockAccessInfo> block_access_info_map;
|
|
};
|
|
|
|
// Aggregates BlockTypeAggregate given a SST file.
|
|
struct SSTFileAccessInfoAggregate {
|
|
uint32_t level;
|
|
std::map<TraceType, BlockTypeAccessInfoAggregate> block_type_aggregates_map;
|
|
};
|
|
|
|
// Aggregates SSTFileAggregate given a column family.
|
|
struct ColumnFamilyAccessInfoAggregate {
|
|
std::map<uint64_t, SSTFileAccessInfoAggregate> fd_aggregates_map;
|
|
};
|
|
|
|
struct Features {
|
|
std::vector<uint64_t> elapsed_time_since_last_access;
|
|
std::vector<uint64_t> num_accesses_since_last_access;
|
|
std::vector<uint64_t> num_past_accesses;
|
|
};
|
|
|
|
struct Predictions {
|
|
std::vector<uint64_t> elapsed_time_till_next_access;
|
|
std::vector<uint64_t> num_accesses_till_next_access;
|
|
};
|
|
|
|
class BlockCacheTraceAnalyzer {
|
|
public:
|
|
BlockCacheTraceAnalyzer(
|
|
const std::string& trace_file_path, const std::string& output_dir,
|
|
const std::string& human_readable_trace_file_path,
|
|
bool compute_reuse_distance, bool mrc_only,
|
|
std::unique_ptr<BlockCacheTraceSimulator>&& cache_simulator);
|
|
~BlockCacheTraceAnalyzer() = default;
|
|
// No copy and move.
|
|
BlockCacheTraceAnalyzer(const BlockCacheTraceAnalyzer&) = delete;
|
|
BlockCacheTraceAnalyzer& operator=(const BlockCacheTraceAnalyzer&) = delete;
|
|
BlockCacheTraceAnalyzer(BlockCacheTraceAnalyzer&&) = delete;
|
|
BlockCacheTraceAnalyzer& operator=(BlockCacheTraceAnalyzer&&) = delete;
|
|
|
|
// Read all access records in the given trace_file, maintains the stats of
|
|
// a block, and aggregates the information by block type, sst file, and column
|
|
// family. Subsequently, the caller may call Print* functions to print
|
|
// statistics.
|
|
Status Analyze();
|
|
|
|
// Print a summary of statistics of the trace, e.g.,
|
|
// Number of files: 2 Number of blocks: 50 Number of accesses: 50
|
|
// Number of Index blocks: 10
|
|
// Number of Filter blocks: 10
|
|
// Number of Data blocks: 10
|
|
// Number of UncompressionDict blocks: 10
|
|
// Number of RangeDeletion blocks: 10
|
|
// ***************************************************************
|
|
// Caller Get: Number of accesses 10
|
|
// Caller Get: Number of accesses per level break down
|
|
// Level 0: Number of accesses: 10
|
|
// Caller Get: Number of accesses per block type break down
|
|
// Block Type Index: Number of accesses: 2
|
|
// Block Type Filter: Number of accesses: 2
|
|
// Block Type Data: Number of accesses: 2
|
|
// Block Type UncompressionDict: Number of accesses: 2
|
|
// Block Type RangeDeletion: Number of accesses: 2
|
|
void PrintStatsSummary() const;
|
|
|
|
// Print block size distribution and the distribution break down by block type
|
|
// and column family.
|
|
void PrintBlockSizeStats() const;
|
|
|
|
// Print access count distribution and the distribution break down by block
|
|
// type and column family.
|
|
void PrintAccessCountStats(bool user_access_only, uint32_t bottom_k,
|
|
uint32_t top_k) const;
|
|
|
|
// Print data block accesses by user Get and Multi-Get.
|
|
// It prints out 1) A histogram on the percentage of keys accessed in a data
|
|
// block break down by if a referenced key exists in the data block andthe
|
|
// histogram break down by column family. 2) A histogram on the percentage of
|
|
// accesses on keys exist in a data block and its break down by column family.
|
|
void PrintDataBlockAccessStats() const;
|
|
|
|
// Write the percentage of accesses break down by column family into a csv
|
|
// file saved in 'output_dir'.
|
|
//
|
|
// The file is named "percentage_of_accesses_summary". The file format is
|
|
// caller,cf_0,cf_1,...,cf_n where the cf_i is the column family name found in
|
|
// the trace.
|
|
void WritePercentAccessSummaryStats() const;
|
|
|
|
// Write the percentage of accesses for the given caller break down by column
|
|
// family, level, and block type into a csv file saved in 'output_dir'.
|
|
//
|
|
// It generates two files: 1) caller_level_percentage_of_accesses_summary and
|
|
// 2) caller_bt_percentage_of_accesses_summary which break down by the level
|
|
// and block type, respectively. The file format is
|
|
// level/bt,cf_0,cf_1,...,cf_n where cf_i is the column family name found in
|
|
// the trace.
|
|
void WriteDetailedPercentAccessSummaryStats(TableReaderCaller caller) const;
|
|
|
|
// Write the access count summary into a csv file saved in 'output_dir'.
|
|
// It groups blocks by their access count.
|
|
//
|
|
// It generates two files: 1) cf_access_count_summary and 2)
|
|
// bt_access_count_summary which break down the access count by column family
|
|
// and block type, respectively. The file format is
|
|
// cf/bt,bucket_0,bucket_1,...,bucket_N.
|
|
void WriteAccessCountSummaryStats(
|
|
const std::vector<uint64_t>& access_count_buckets,
|
|
bool user_access_only) const;
|
|
|
|
// Write miss ratio curves of simulated cache configurations into a csv file
|
|
// named "mrc" saved in 'output_dir'.
|
|
//
|
|
// The file format is
|
|
// "cache_name,num_shard_bits,capacity,miss_ratio,total_accesses".
|
|
void WriteMissRatioCurves() const;
|
|
|
|
// Write miss ratio timeline of simulated cache configurations into several
|
|
// csv files, one per cache capacity saved in 'output_dir'.
|
|
//
|
|
// The file format is
|
|
// "time,label_1_access_per_second,label_2_access_per_second,...,label_N_access_per_second"
|
|
// where N is the number of unique cache names
|
|
// (cache_name+num_shard_bits+ghost_capacity).
|
|
void WriteMissRatioTimeline(uint64_t time_unit) const;
|
|
|
|
// Write misses timeline of simulated cache configurations into several
|
|
// csv files, one per cache capacity saved in 'output_dir'.
|
|
//
|
|
// The file format is
|
|
// "time,label_1_access_per_second,label_2_access_per_second,...,label_N_access_per_second"
|
|
// where N is the number of unique cache names
|
|
// (cache_name+num_shard_bits+ghost_capacity).
|
|
void WriteMissTimeline(uint64_t time_unit) const;
|
|
|
|
// Write the access timeline into a csv file saved in 'output_dir'.
|
|
//
|
|
// The file is named "label_access_timeline".The file format is
|
|
// "time,label_1_access_per_second,label_2_access_per_second,...,label_N_access_per_second"
|
|
// where N is the number of unique labels found in the trace.
|
|
void WriteAccessTimeline(const std::string& label, uint64_t time_unit,
|
|
bool user_access_only) const;
|
|
|
|
// Write the reuse distance into a csv file saved in 'output_dir'. Reuse
|
|
// distance is defined as the cumulated size of unique blocks read between two
|
|
// consective accesses on the same block.
|
|
//
|
|
// The file is named "label_reuse_distance". The file format is
|
|
// bucket,label_1,label_2,...,label_N.
|
|
void WriteReuseDistance(const std::string& label_str,
|
|
const std::vector<uint64_t>& distance_buckets) const;
|
|
|
|
// Write the reuse interval into a csv file saved in 'output_dir'. Reuse
|
|
// interval is defined as the time between two consecutive accesses on the
|
|
// same block.
|
|
//
|
|
// The file is named "label_reuse_interval". The file format is
|
|
// bucket,label_1,label_2,...,label_N.
|
|
void WriteReuseInterval(const std::string& label_str,
|
|
const std::vector<uint64_t>& time_buckets) const;
|
|
|
|
// Write the reuse lifetime into a csv file saved in 'output_dir'. Reuse
|
|
// lifetime is defined as the time interval between the first access of a
|
|
// block and its last access.
|
|
//
|
|
// The file is named "label_reuse_lifetime". The file format is
|
|
// bucket,label_1,label_2,...,label_N.
|
|
void WriteReuseLifetime(const std::string& label_str,
|
|
const std::vector<uint64_t>& time_buckets) const;
|
|
|
|
// Write the reuse timeline into a csv file saved in 'output_dir'.
|
|
//
|
|
// The file is named
|
|
// "block_type_user_access_only_reuse_window_reuse_timeline". The file format
|
|
// is start_time,0,1,...,N where N equals trace_duration / reuse_window.
|
|
void WriteBlockReuseTimeline(uint64_t reuse_window, bool user_access_only,
|
|
TraceType block_type) const;
|
|
|
|
// Write the Get spatical locality into csv files saved in 'output_dir'.
|
|
//
|
|
// It generates three csv files. label_percent_ref_keys,
|
|
// label_percent_accesses_on_ref_keys, and
|
|
// label_percent_data_size_on_ref_keys.
|
|
void WriteGetSpatialLocality(
|
|
const std::string& label_str,
|
|
const std::vector<uint64_t>& percent_buckets) const;
|
|
|
|
void WriteCorrelationFeatures(const std::string& label_str,
|
|
uint32_t max_number_of_values) const;
|
|
|
|
void WriteCorrelationFeaturesForGet(uint32_t max_number_of_values) const;
|
|
|
|
const std::map<std::string, ColumnFamilyAccessInfoAggregate>&
|
|
TEST_cf_aggregates_map() const {
|
|
return cf_aggregates_map_;
|
|
}
|
|
|
|
private:
|
|
std::set<std::string> ParseLabelStr(const std::string& label_str) const;
|
|
|
|
std::string BuildLabel(const std::set<std::string>& labels,
|
|
const std::string& cf_name, uint64_t fd,
|
|
uint32_t level, TraceType type,
|
|
TableReaderCaller caller, uint64_t block_key) const;
|
|
|
|
void ComputeReuseDistance(BlockAccessInfo* info) const;
|
|
|
|
Status RecordAccess(const BlockCacheTraceRecord& access);
|
|
|
|
void UpdateReuseIntervalStats(
|
|
const std::string& label, const std::vector<uint64_t>& time_buckets,
|
|
const std::map<uint64_t, uint64_t> timeline,
|
|
std::map<std::string, std::map<uint64_t, uint64_t>>*
|
|
label_time_num_reuses,
|
|
uint64_t* total_num_reuses) const;
|
|
|
|
std::string OutputPercentAccessStats(
|
|
uint64_t total_accesses,
|
|
const std::map<std::string, uint64_t>& cf_access_count) const;
|
|
|
|
void WriteStatsToFile(
|
|
const std::string& label_str, const std::vector<uint64_t>& time_buckets,
|
|
const std::string& filename_suffix,
|
|
const std::map<std::string, std::map<uint64_t, uint64_t>>& label_data,
|
|
uint64_t ntotal) const;
|
|
|
|
void TraverseBlocks(
|
|
std::function<void(const std::string& /*cf_name*/, uint64_t /*fd*/,
|
|
uint32_t /*level*/, TraceType /*block_type*/,
|
|
const std::string& /*block_key*/,
|
|
uint64_t /*block_key_id*/,
|
|
const BlockAccessInfo& /*block_access_info*/)>
|
|
block_callback) const;
|
|
|
|
void UpdateFeatureVectors(
|
|
const std::vector<uint64_t>& access_sequence_number_timeline,
|
|
const std::vector<uint64_t>& access_timeline, const std::string& label,
|
|
std::map<std::string, Features>* label_features,
|
|
std::map<std::string, Predictions>* label_predictions) const;
|
|
|
|
void WriteCorrelationFeaturesToFile(
|
|
const std::string& label,
|
|
const std::map<std::string, Features>& label_features,
|
|
const std::map<std::string, Predictions>& label_predictions,
|
|
uint32_t max_number_of_values) const;
|
|
|
|
Status WriteHumanReadableTraceRecord(const BlockCacheTraceRecord& access,
|
|
uint64_t block_id, uint64_t get_key_id);
|
|
|
|
rocksdb::Env* env_;
|
|
const std::string trace_file_path_;
|
|
const std::string output_dir_;
|
|
std::string human_readable_trace_file_path_;
|
|
const bool compute_reuse_distance_;
|
|
const bool mrc_only_;
|
|
|
|
BlockCacheTraceHeader header_;
|
|
std::unique_ptr<BlockCacheTraceSimulator> cache_simulator_;
|
|
std::map<std::string, ColumnFamilyAccessInfoAggregate> cf_aggregates_map_;
|
|
std::map<std::string, BlockAccessInfo*> block_info_map_;
|
|
std::unordered_map<std::string, GetKeyInfo> get_key_info_map_;
|
|
uint64_t access_sequence_number_ = 0;
|
|
uint64_t trace_start_timestamp_in_seconds_ = 0;
|
|
uint64_t trace_end_timestamp_in_seconds_ = 0;
|
|
MissRatioStats miss_ratio_stats_;
|
|
uint64_t unique_block_id_ = 1;
|
|
uint64_t unique_get_key_id_ = 1;
|
|
char trace_record_buffer_[1024 * 1024];
|
|
std::unique_ptr<rocksdb::WritableFile> human_readable_trace_file_writer_;
|
|
};
|
|
|
|
int block_cache_trace_analyzer_tool(int argc, char** argv);
|
|
|
|
} // namespace rocksdb
|