You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
rocksdb/tools/trace_analyzer_tool.cc

1926 lines
67 KiB

RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
// 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).
//
#ifndef ROCKSDB_LITE
#ifdef GFLAGS
#ifdef NUMA
#include <numa.h>
#endif
#ifndef OS_WIN
#include <unistd.h>
#endif
#include <cinttypes>
#include <cmath>
#include <cstdio>
#include <cstdlib>
#include <memory>
#include <sstream>
#include <stdexcept>
#include "db/db_impl/db_impl.h"
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
#include "db/memtable.h"
#include "db/write_batch_internal.h"
Introduce a new storage specific Env API (#5761) Summary: The current Env API encompasses both storage/file operations, as well as OS related operations. Most of the APIs return a Status, which does not have enough metadata about an error, such as whether its retry-able or not, scope (i.e fault domain) of the error etc., that may be required in order to properly handle a storage error. The file APIs also do not provide enough control over the IO SLA, such as timeout, prioritization, hinting about placement and redundancy etc. This PR separates out the file/storage APIs from Env into a new FileSystem class. The APIs are updated to return an IOStatus with metadata about the error, as well as to take an IOOptions structure as input in order to allow more control over the IO. The user can set both ```options.env``` and ```options.file_system``` to specify that RocksDB should use the former for OS related operations and the latter for storage operations. Internally, a ```CompositeEnvWrapper``` has been introduced that inherits from ```Env``` and redirects individual methods to either an ```Env``` implementation or the ```FileSystem``` as appropriate. When options are sanitized during ```DB::Open```, ```options.env``` is replaced with a newly allocated ```CompositeEnvWrapper``` instance if both env and file_system have been specified. This way, the rest of the RocksDB code can continue to function as before. This PR also ports PosixEnv to the new API by splitting it into two - PosixEnv and PosixFileSystem. PosixEnv is defined as a sub-class of CompositeEnvWrapper, and threading/time functions are overridden with Posix specific implementations in order to avoid an extra level of indirection. The ```CompositeEnvWrapper``` translates ```IOStatus``` return code to ```Status```, and sets the severity to ```kSoftError``` if the io_status is retryable. The error handling code in RocksDB can then recover the DB automatically. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5761 Differential Revision: D18868376 Pulled By: anand1976 fbshipit-source-id: 39efe18a162ea746fabac6360ff529baba48486f
5 years ago
#include "env/composite_env_wrapper.h"
#include "file/line_file_reader.h"
#include "file/writable_file_writer.h"
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
#include "options/cf_options.h"
#include "rocksdb/db.h"
#include "rocksdb/env.h"
#include "rocksdb/iterator.h"
#include "rocksdb/slice.h"
#include "rocksdb/slice_transform.h"
#include "rocksdb/status.h"
#include "rocksdb/table_properties.h"
#include "rocksdb/utilities/ldb_cmd.h"
#include "rocksdb/write_batch.h"
#include "table/meta_blocks.h"
#include "table/table_reader.h"
#include "tools/trace_analyzer_tool.h"
#include "trace_replay/trace_replay.h"
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
#include "util/coding.h"
#include "util/compression.h"
#include "util/gflags_compat.h"
#include "util/random.h"
#include "util/string_util.h"
using GFLAGS_NAMESPACE::ParseCommandLineFlags;
DEFINE_string(trace_path, "", "The trace file path.");
DEFINE_string(output_dir, "", "The directory to store the output files.");
DEFINE_string(output_prefix, "trace",
"The prefix used for all the output files.");
DEFINE_bool(output_key_stats, false,
"Output the key access count statistics to file\n"
"for accessed keys:\n"
"file name: <prefix>-<query_type>-<cf_id>-accessed_key_stats.txt\n"
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
"Format:[cf_id value_size access_keyid access_count]\n"
"for the whole key space keys:\n"
"File name: <prefix>-<query_type>-<cf_id>-whole_key_stats.txt\n"
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
"Format:[whole_key_space_keyid access_count]");
DEFINE_bool(output_access_count_stats, false,
"Output the access count distribution statistics to file.\n"
"File name: <prefix>-<query_type>-<cf_id>-accessed_"
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
"key_count_distribution.txt \n"
"Format:[access_count number_of_access_count]");
DEFINE_bool(output_time_series, false,
"Output the access time in second of each key, "
"such that we can have the time series data of the queries \n"
"File name: <prefix>-<query_type>-<cf_id>-time_series.txt\n"
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
"Format:[type_id time_in_sec access_keyid].");
DEFINE_bool(try_process_corrupted_trace, false,
"In default, trace_analyzer will exit if the trace file is "
"corrupted due to the unexpected tracing cases. If this option "
"is enabled, trace_analyzer will stop reading the trace file, "
"and start analyzing the read-in data.");
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
DEFINE_int32(output_prefix_cut, 0,
"The number of bytes as prefix to cut the keys.\n"
"If it is enabled, it will generate the following:\n"
"For accessed keys:\n"
"File name: <prefix>-<query_type>-<cf_id>-"
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
"accessed_key_prefix_cut.txt \n"
"Format:[acessed_keyid access_count_of_prefix "
"number_of_keys_in_prefix average_key_access "
"prefix_succ_ratio prefix]\n"
"For whole key space keys:\n"
"File name: <prefix>-<query_type>-<cf_id>"
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
"-whole_key_prefix_cut.txt\n"
"Format:[start_keyid_in_whole_keyspace prefix]\n"
"if 'output_qps_stats' and 'top_k' are enabled, it will output:\n"
"File name: <prefix>-<query_type>-<cf_id>"
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
"-accessed_top_k_qps_prefix_cut.txt\n"
"Format:[the_top_ith_qps_time QPS], [prefix qps_of_this_second].");
DEFINE_bool(convert_to_human_readable_trace, false,
"Convert the binary trace file to a human readable txt file "
"for further processing. "
"This file will be extremely large "
"(similar size as the original binary trace file). "
"You can specify 'no_key' to reduce the size, if key is not "
"needed in the next step.\n"
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
"File name: <prefix>_human_readable_trace.txt\n"
"Format:[type_id cf_id value_size time_in_micorsec <key>].");
DEFINE_bool(output_qps_stats, false,
"Output the query per second(qps) statistics \n"
"For the overall qps, it will contain all qps of each query type. "
"The time is started from the first trace record\n"
"File name: <prefix>_qps_stats.txt\n"
"Format: [qps_type_1 qps_type_2 ...... overall_qps]\n"
"For each cf and query, it will have its own qps output.\n"
"File name: <prefix>-<query_type>-<cf_id>_qps_stats.txt \n"
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
"Format:[query_count_in_this_second].");
DEFINE_bool(no_print, false, "Do not print out any result");
DEFINE_string(
print_correlation, "",
"intput format: [correlation pairs][.,.]\n"
"Output the query correlations between the pairs of query types "
"listed in the parameter, input should select the operations from:\n"
"get, put, delete, single_delete, rangle_delete, merge. No space "
"between the pairs separated by commar. Example: =[get,get]... "
"It will print out the number of pairs of 'A after B' and "
"the average time interval between the two query.");
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
DEFINE_string(key_space_dir, "",
"<the directory stores full key space files> \n"
"The key space files should be: <column family id>.txt");
DEFINE_bool(analyze_get, false, "Analyze the Get query.");
DEFINE_bool(analyze_put, false, "Analyze the Put query.");
DEFINE_bool(analyze_delete, false, "Analyze the Delete query.");
DEFINE_bool(analyze_single_delete, false, "Analyze the SingleDelete query.");
DEFINE_bool(analyze_range_delete, false, "Analyze the DeleteRange query.");
DEFINE_bool(analyze_merge, false, "Analyze the Merge query.");
DEFINE_bool(analyze_iterator, false,
" Analyze the iterate query like Seek() and SeekForPrev().");
DEFINE_bool(analyze_multiget, false,
" Analyze the MultiGet query. NOTE: for"
" MultiGet, we analyze each KV-pair read in one MultiGet query. "
"Therefore, the total queries and QPS are calculated based on "
"the number of KV-pairs being accessed not the number of MultiGet."
"It can be improved in the future if needed");
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
DEFINE_bool(no_key, false,
" Does not output the key to the result files to make smaller.");
DEFINE_bool(print_overall_stats, true,
" Print the stats of the whole trace, "
"like total requests, keys, and etc.");
DEFINE_bool(output_key_distribution, false, "Print the key size distribution.");
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
DEFINE_bool(
output_value_distribution, false,
"Out put the value size distribution, only available for Put and Merge.\n"
"File name: <prefix>-<query_type>-<cf_id>"
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
"-accessed_value_size_distribution.txt\n"
"Format:[Number_of_value_size_between x and "
"x+value_interval is: <the count>]");
DEFINE_int32(print_top_k_access, 1,
"<top K of the variables to be printed> "
"Print the top k accessed keys, top k accessed prefix "
"and etc.");
DEFINE_int32(output_ignore_count, 0,
"<threshold>, ignores the access count <= this value, "
"it will shorter the output.");
DEFINE_int32(value_interval, 8,
"To output the value distribution, we need to set the value "
"intervals and make the statistic of the value size distribution "
"in different intervals. The default is 8.");
DEFINE_double(sample_ratio, 1.0,
"If the trace size is extremely huge or user want to sample "
"the trace when analyzing, sample ratio can be set (0, 1.0]");
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
namespace ROCKSDB_NAMESPACE {
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
const size_t kShadowValueSize = 10;
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
std::map<std::string, int> taOptToIndex = {
{"get", 0}, {"put", 1},
{"delete", 2}, {"single_delete", 3},
{"range_delete", 4}, {"merge", 5},
{"iterator_Seek", 6}, {"iterator_SeekForPrev", 7},
{"multiget", 8}};
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
std::map<int, std::string> taIndexToOpt = {
{0, "get"}, {1, "put"},
{2, "delete"}, {3, "single_delete"},
{4, "range_delete"}, {5, "merge"},
{6, "iterator_Seek"}, {7, "iterator_SeekForPrev"},
{8, "multiget"}};
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
namespace {
uint64_t MultiplyCheckOverflow(uint64_t op1, uint64_t op2) {
if (op1 == 0 || op2 == 0) {
return 0;
}
if (std::numeric_limits<uint64_t>::max() / op1 < op2) {
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
return op1;
}
return (op1 * op2);
}
} // namespace
// The default constructor of AnalyzerOptions
AnalyzerOptions::AnalyzerOptions()
: correlation_map(kTaTypeNum, std::vector<int>(kTaTypeNum, -1)) {}
AnalyzerOptions::~AnalyzerOptions() {}
void AnalyzerOptions::SparseCorrelationInput(const std::string& in_str) {
std::string cur = in_str;
if (cur.size() == 0) {
return;
}
while (!cur.empty()) {
if (cur.compare(0, 1, "[") != 0) {
fprintf(stderr, "Invalid correlation input: %s\n", in_str.c_str());
exit(1);
}
std::string opt1, opt2;
std::size_t split = cur.find_first_of(",");
if (split != std::string::npos) {
opt1 = cur.substr(1, split - 1);
} else {
fprintf(stderr, "Invalid correlation input: %s\n", in_str.c_str());
exit(1);
}
std::size_t end = cur.find_first_of("]");
if (end != std::string::npos) {
opt2 = cur.substr(split + 1, end - split - 1);
} else {
fprintf(stderr, "Invalid correlation input: %s\n", in_str.c_str());
exit(1);
}
cur = cur.substr(end + 1);
if (taOptToIndex.find(opt1) != taOptToIndex.end() &&
taOptToIndex.find(opt2) != taOptToIndex.end()) {
correlation_list.push_back(
std::make_pair(taOptToIndex[opt1], taOptToIndex[opt2]));
} else {
fprintf(stderr, "Invalid correlation input: %s\n", in_str.c_str());
exit(1);
}
}
int sequence = 0;
for (auto& it : correlation_list) {
correlation_map[it.first][it.second] = sequence;
sequence++;
}
return;
}
// The trace statistic struct constructor
TraceStats::TraceStats() {
cf_id = 0;
cf_name = "0";
a_count = 0;
a_key_id = 0;
a_key_size_sqsum = 0;
a_key_size_sum = 0;
a_key_mid = 0;
a_value_size_sqsum = 0;
a_value_size_sum = 0;
a_value_mid = 0;
a_peak_qps = 0;
a_ave_qps = 0.0;
}
TraceStats::~TraceStats() {}
// The trace analyzer constructor
TraceAnalyzer::TraceAnalyzer(std::string& trace_path, std::string& output_path,
AnalyzerOptions _analyzer_opts)
: write_batch_ts_(0),
trace_name_(trace_path),
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
output_path_(output_path),
analyzer_opts_(_analyzer_opts) {
ROCKSDB_NAMESPACE::EnvOptions env_options;
env_ = ROCKSDB_NAMESPACE::Env::Default();
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
offset_ = 0;
total_requests_ = 0;
total_access_keys_ = 0;
total_gets_ = 0;
total_writes_ = 0;
total_seeks_ = 0;
total_seek_prevs_ = 0;
total_multigets_ = 0;
trace_create_time_ = 0;
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
begin_time_ = 0;
end_time_ = 0;
time_series_start_ = 0;
cur_time_sec_ = 0;
if (FLAGS_sample_ratio > 1.0 || FLAGS_sample_ratio <= 0) {
sample_max_ = 1;
} else {
sample_max_ = static_cast<uint32_t>(1.0 / FLAGS_sample_ratio);
}
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
ta_.resize(kTaTypeNum);
ta_[0].type_name = "get";
if (FLAGS_analyze_get) {
ta_[0].enabled = true;
} else {
ta_[0].enabled = false;
}
ta_[1].type_name = "put";
if (FLAGS_analyze_put) {
ta_[1].enabled = true;
} else {
ta_[1].enabled = false;
}
ta_[2].type_name = "delete";
if (FLAGS_analyze_delete) {
ta_[2].enabled = true;
} else {
ta_[2].enabled = false;
}
ta_[3].type_name = "single_delete";
if (FLAGS_analyze_single_delete) {
ta_[3].enabled = true;
} else {
ta_[3].enabled = false;
}
ta_[4].type_name = "range_delete";
if (FLAGS_analyze_range_delete) {
ta_[4].enabled = true;
} else {
ta_[4].enabled = false;
}
ta_[5].type_name = "merge";
if (FLAGS_analyze_merge) {
ta_[5].enabled = true;
} else {
ta_[5].enabled = false;
}
ta_[6].type_name = "iterator_Seek";
if (FLAGS_analyze_iterator) {
ta_[6].enabled = true;
} else {
ta_[6].enabled = false;
}
ta_[7].type_name = "iterator_SeekForPrev";
if (FLAGS_analyze_iterator) {
ta_[7].enabled = true;
} else {
ta_[7].enabled = false;
}
ta_[8].type_name = "multiget";
if (FLAGS_analyze_multiget) {
ta_[8].enabled = true;
} else {
ta_[8].enabled = false;
}
for (int i = 0; i < kTaTypeNum; i++) {
ta_[i].sample_count = 0;
}
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
}
TraceAnalyzer::~TraceAnalyzer() {}
// Prepare the processing
// Initiate the global trace reader and writer here
Status TraceAnalyzer::PrepareProcessing() {
Status s;
// Prepare the trace reader
if (trace_reader_ == nullptr) {
s = NewFileTraceReader(env_, env_options_, trace_name_, &trace_reader_);
} else {
s = trace_reader_->Reset();
}
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
if (!s.ok()) {
return s;
}
// Prepare and open the trace sequence file writer if needed
if (FLAGS_convert_to_human_readable_trace) {
std::string trace_sequence_name;
trace_sequence_name =
output_path_ + "/" + FLAGS_output_prefix + "-human_readable_trace.txt";
s = env_->NewWritableFile(trace_sequence_name, &trace_sequence_f_,
env_options_);
if (!s.ok()) {
return s;
}
}
// prepare the general QPS file writer
if (FLAGS_output_qps_stats) {
std::string qps_stats_name;
qps_stats_name =
output_path_ + "/" + FLAGS_output_prefix + "-qps_stats.txt";
s = env_->NewWritableFile(qps_stats_name, &qps_f_, env_options_);
if (!s.ok()) {
return s;
}
qps_stats_name =
output_path_ + "/" + FLAGS_output_prefix + "-cf_qps_stats.txt";
s = env_->NewWritableFile(qps_stats_name, &cf_qps_f_, env_options_);
if (!s.ok()) {
return s;
}
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
}
return Status::OK();
}
Status TraceAnalyzer::ReadTraceHeader(Trace* header) {
assert(header != nullptr);
std::string encoded_trace;
// Read the trace head
Status s = trace_reader_->Read(&encoded_trace);
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
if (!s.ok()) {
return s;
}
s = TracerHelper::DecodeTrace(encoded_trace, header);
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
if (header->type != kTraceBegin) {
return Status::Corruption("Corrupted trace file. Incorrect header.");
}
if (header->payload.substr(0, kTraceMagic.length()) != kTraceMagic) {
return Status::Corruption("Corrupted trace file. Incorrect magic.");
}
return s;
}
Status TraceAnalyzer::ReadTraceFooter(Trace* footer) {
assert(footer != nullptr);
Status s = ReadTraceRecord(footer);
if (!s.ok()) {
return s;
}
if (footer->type != kTraceEnd) {
return Status::Corruption("Corrupted trace file. Incorrect footer.");
}
return s;
}
Status TraceAnalyzer::ReadTraceRecord(Trace* trace) {
assert(trace != nullptr);
std::string encoded_trace;
Status s = trace_reader_->Read(&encoded_trace);
if (!s.ok()) {
return s;
}
return TracerHelper::DecodeTrace(encoded_trace, trace);
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
}
// process the trace itself and redirect the trace content
// to different operation type handler. With different race
// format, this function can be changed
Status TraceAnalyzer::StartProcessing() {
Status s;
Trace header;
s = ReadTraceHeader(&header);
if (!s.ok()) {
fprintf(stderr, "Cannot read the header\n");
return s;
}
// Set the default trace file version as version 0.2
int trace_file_version = 2;
s = TracerHelper::ParseTraceHeader(header, &trace_file_version, &db_version_);
if (!s.ok()) {
return s;
}
trace_create_time_ = header.ts;
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
if (FLAGS_output_time_series) {
time_series_start_ = header.ts;
}
Trace trace;
std::unique_ptr<TraceRecord> record;
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
while (s.ok()) {
trace.reset();
s = ReadTraceRecord(&trace);
if (!s.ok()) {
break;
}
end_time_ = trace.ts;
if (trace.type == kTraceEnd) {
break;
}
// Do not count TraceEnd (if there is one)
total_requests_++;
s = TracerHelper::DecodeTraceRecord(&trace, trace_file_version, &record);
if (s.IsNotSupported()) {
continue;
}
if (!s.ok()) {
return s;
}
s = record->Accept(this, nullptr);
if (!s.ok()) {
fprintf(stderr, "Cannot process the TraceRecord\n");
return s;
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
}
}
if (s.IsIncomplete()) {
// Fix it: Reaching eof returns Incomplete status at the moment.
return Status::OK();
}
return s;
}
// After the trace is processed by StartProcessing, the statistic data
// is stored in the map or other in memory data structures. To get the
// other statistic result such as key size distribution, value size
// distribution, these data structures are re-processed here.
Status TraceAnalyzer::MakeStatistics() {
int ret;
Status s;
for (int type = 0; type < kTaTypeNum; type++) {
if (!ta_[type].enabled) {
continue;
}
for (auto& stat : ta_[type].stats) {
stat.second.a_key_id = 0;
for (auto& record : stat.second.a_key_stats) {
record.second.key_id = stat.second.a_key_id;
stat.second.a_key_id++;
if (record.second.access_count <=
static_cast<uint64_t>(FLAGS_output_ignore_count)) {
continue;
}
// Generate the key access count distribution data
if (FLAGS_output_access_count_stats) {
if (stat.second.a_count_stats.find(record.second.access_count) ==
stat.second.a_count_stats.end()) {
stat.second.a_count_stats[record.second.access_count] = 1;
} else {
stat.second.a_count_stats[record.second.access_count]++;
}
}
// Generate the key size distribution data
if (FLAGS_output_key_distribution) {
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
if (stat.second.a_key_size_stats.find(record.first.size()) ==
stat.second.a_key_size_stats.end()) {
stat.second.a_key_size_stats[record.first.size()] = 1;
} else {
stat.second.a_key_size_stats[record.first.size()]++;
}
}
if (!FLAGS_print_correlation.empty()) {
s = MakeStatisticCorrelation(stat.second, record.second);
if (!s.ok()) {
return s;
}
}
}
// Output the prefix cut or the whole content of the accessed key space
if (FLAGS_output_key_stats || FLAGS_output_prefix_cut > 0) {
s = MakeStatisticKeyStatsOrPrefix(stat.second);
if (!s.ok()) {
return s;
}
}
// output the access count distribution
if (FLAGS_output_access_count_stats && stat.second.a_count_dist_f) {
for (auto& record : stat.second.a_count_stats) {
ret = snprintf(buffer_, sizeof(buffer_),
"access_count: %" PRIu64 " num: %" PRIu64 "\n",
record.first, record.second);
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
if (ret < 0) {
return Status::IOError("Format the output failed");
}
std::string printout(buffer_);
s = stat.second.a_count_dist_f->Append(printout);
if (!s.ok()) {
fprintf(stderr, "Write access count distribution file failed\n");
return s;
}
}
}
// find the medium of the key size
uint64_t k_count = 0;
bool get_mid = false;
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
for (auto& record : stat.second.a_key_size_stats) {
k_count += record.second;
if (!get_mid && k_count >= stat.second.a_key_mid) {
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
stat.second.a_key_mid = record.first;
get_mid = true;
}
if (FLAGS_output_key_distribution && stat.second.a_key_size_f) {
ret = snprintf(buffer_, sizeof(buffer_), "%" PRIu64 " %" PRIu64 "\n",
record.first, record.second);
if (ret < 0) {
return Status::IOError("Format output failed");
}
std::string printout(buffer_);
s = stat.second.a_key_size_f->Append(printout);
if (!s.ok()) {
fprintf(stderr, "Write key size distribution file failed\n");
return s;
}
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
}
}
// output the value size distribution
uint64_t v_begin = 0, v_end = 0, v_count = 0;
get_mid = false;
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
for (auto& record : stat.second.a_value_size_stats) {
v_begin = v_end;
v_end = (record.first + 1) * FLAGS_value_interval;
v_count += record.second;
if (!get_mid && v_count >= stat.second.a_count / 2) {
stat.second.a_value_mid = (v_begin + v_end) / 2;
get_mid = true;
}
if (FLAGS_output_value_distribution && stat.second.a_value_size_f &&
(type == TraceOperationType::kPut ||
type == TraceOperationType::kMerge)) {
ret = snprintf(buffer_, sizeof(buffer_),
"Number_of_value_size_between %" PRIu64 " and %" PRIu64
" is: %" PRIu64 "\n",
v_begin, v_end, record.second);
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
if (ret < 0) {
return Status::IOError("Format output failed");
}
std::string printout(buffer_);
s = stat.second.a_value_size_f->Append(printout);
if (!s.ok()) {
fprintf(stderr, "Write value size distribution file failed\n");
return s;
}
}
}
}
}
// Make the QPS statistics
if (FLAGS_output_qps_stats) {
s = MakeStatisticQPS();
if (!s.ok()) {
return s;
}
}
return Status::OK();
}
// Process the statistics of the key access and
// prefix of the accessed keys if required
Status TraceAnalyzer::MakeStatisticKeyStatsOrPrefix(TraceStats& stats) {
int ret;
Status s;
std::string prefix = "0";
uint64_t prefix_access = 0;
uint64_t prefix_count = 0;
uint64_t prefix_succ_access = 0;
double prefix_ave_access = 0.0;
stats.a_succ_count = 0;
for (auto& record : stats.a_key_stats) {
// write the key access statistic file
if (!stats.a_key_f) {
return Status::IOError("Failed to open accessed_key_stats file.");
}
stats.a_succ_count += record.second.succ_count;
double succ_ratio = 0.0;
if (record.second.access_count > 0) {
succ_ratio = (static_cast<double>(record.second.succ_count)) /
record.second.access_count;
}
ret = snprintf(buffer_, sizeof(buffer_),
"%u %zu %" PRIu64 " %" PRIu64 " %f\n", record.second.cf_id,
record.second.value_size, record.second.key_id,
record.second.access_count, succ_ratio);
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
if (ret < 0) {
return Status::IOError("Format output failed");
}
std::string printout(buffer_);
s = stats.a_key_f->Append(printout);
if (!s.ok()) {
fprintf(stderr, "Write key access file failed\n");
return s;
}
// write the prefix cut of the accessed keys
if (FLAGS_output_prefix_cut > 0 && stats.a_prefix_cut_f) {
if (record.first.compare(0, FLAGS_output_prefix_cut, prefix) != 0) {
std::string prefix_out =
ROCKSDB_NAMESPACE::LDBCommand::StringToHex(prefix);
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
if (prefix_count == 0) {
prefix_ave_access = 0.0;
} else {
prefix_ave_access =
(static_cast<double>(prefix_access)) / prefix_count;
}
double prefix_succ_ratio = 0.0;
if (prefix_access > 0) {
prefix_succ_ratio =
(static_cast<double>(prefix_succ_access)) / prefix_access;
}
ret =
snprintf(buffer_, sizeof(buffer_),
"%" PRIu64 " %" PRIu64 " %" PRIu64 " %f %f %s\n",
record.second.key_id, prefix_access, prefix_count,
prefix_ave_access, prefix_succ_ratio, prefix_out.c_str());
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
if (ret < 0) {
return Status::IOError("Format output failed");
}
std::string pout(buffer_);
s = stats.a_prefix_cut_f->Append(pout);
if (!s.ok()) {
fprintf(stderr, "Write accessed key prefix file failed\n");
return s;
}
// make the top k statistic for the prefix
if (static_cast<int32_t>(stats.top_k_prefix_access.size()) <
FLAGS_print_top_k_access) {
stats.top_k_prefix_access.push(
std::make_pair(prefix_access, prefix_out));
} else {
if (prefix_access > stats.top_k_prefix_access.top().first) {
stats.top_k_prefix_access.pop();
stats.top_k_prefix_access.push(
std::make_pair(prefix_access, prefix_out));
}
}
if (static_cast<int32_t>(stats.top_k_prefix_ave.size()) <
FLAGS_print_top_k_access) {
stats.top_k_prefix_ave.push(
std::make_pair(prefix_ave_access, prefix_out));
} else {
if (prefix_ave_access > stats.top_k_prefix_ave.top().first) {
stats.top_k_prefix_ave.pop();
stats.top_k_prefix_ave.push(
std::make_pair(prefix_ave_access, prefix_out));
}
}
prefix = record.first.substr(0, FLAGS_output_prefix_cut);
prefix_access = 0;
prefix_count = 0;
prefix_succ_access = 0;
}
prefix_access += record.second.access_count;
prefix_count += 1;
prefix_succ_access += record.second.succ_count;
}
}
return Status::OK();
}
// Process the statistics of different query type
// correlations
Status TraceAnalyzer::MakeStatisticCorrelation(TraceStats& stats,
StatsUnit& unit) {
if (stats.correlation_output.size() !=
analyzer_opts_.correlation_list.size()) {
return Status::Corruption("Cannot make the statistic of correlation.");
}
for (int i = 0; i < static_cast<int>(analyzer_opts_.correlation_list.size());
i++) {
if (i >= static_cast<int>(stats.correlation_output.size()) ||
i >= static_cast<int>(unit.v_correlation.size())) {
break;
}
stats.correlation_output[i].first += unit.v_correlation[i].count;
stats.correlation_output[i].second += unit.v_correlation[i].total_ts;
}
return Status::OK();
}
// Process the statistics of QPS
Status TraceAnalyzer::MakeStatisticQPS() {
if (begin_time_ == 0) {
begin_time_ = trace_create_time_;
}
uint32_t duration =
static_cast<uint32_t>((end_time_ - begin_time_) / 1000000);
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
int ret;
Status s;
std::vector<std::vector<uint32_t>> type_qps(
duration, std::vector<uint32_t>(kTaTypeNum + 1, 0));
std::vector<uint64_t> qps_sum(kTaTypeNum + 1, 0);
std::vector<uint32_t> qps_peak(kTaTypeNum + 1, 0);
qps_ave_.resize(kTaTypeNum + 1);
for (int type = 0; type < kTaTypeNum; type++) {
if (!ta_[type].enabled) {
continue;
}
for (auto& stat : ta_[type].stats) {
uint32_t time_line = 0;
uint64_t cf_qps_sum = 0;
for (auto& time_it : stat.second.a_qps_stats) {
if (time_it.first >= duration) {
continue;
}
type_qps[time_it.first][kTaTypeNum] += time_it.second;
type_qps[time_it.first][type] += time_it.second;
cf_qps_sum += time_it.second;
if (time_it.second > stat.second.a_peak_qps) {
stat.second.a_peak_qps = time_it.second;
}
if (stat.second.a_qps_f) {
while (time_line < time_it.first) {
ret = snprintf(buffer_, sizeof(buffer_), "%u\n", 0);
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
if (ret < 0) {
return Status::IOError("Format the output failed");
}
std::string printout(buffer_);
s = stat.second.a_qps_f->Append(printout);
if (!s.ok()) {
fprintf(stderr, "Write QPS file failed\n");
return s;
}
time_line++;
}
ret = snprintf(buffer_, sizeof(buffer_), "%u\n", time_it.second);
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
if (ret < 0) {
return Status::IOError("Format the output failed");
}
std::string printout(buffer_);
s = stat.second.a_qps_f->Append(printout);
if (!s.ok()) {
fprintf(stderr, "Write QPS file failed\n");
return s;
}
if (time_line == time_it.first) {
time_line++;
}
}
// Process the top k QPS peaks
if (FLAGS_output_prefix_cut > 0) {
if (static_cast<int32_t>(stat.second.top_k_qps_sec.size()) <
FLAGS_print_top_k_access) {
stat.second.top_k_qps_sec.push(
std::make_pair(time_it.second, time_it.first));
} else {
if (stat.second.top_k_qps_sec.size() > 0 &&
stat.second.top_k_qps_sec.top().first < time_it.second) {
stat.second.top_k_qps_sec.pop();
stat.second.top_k_qps_sec.push(
std::make_pair(time_it.second, time_it.first));
}
}
}
}
if (duration == 0) {
stat.second.a_ave_qps = 0;
} else {
stat.second.a_ave_qps = (static_cast<double>(cf_qps_sum)) / duration;
}
// Output the accessed unique key number change overtime
if (stat.second.a_key_num_f) {
uint64_t cur_uni_key =
static_cast<uint64_t>(stat.second.a_key_stats.size());
double cur_ratio = 0.0;
uint64_t cur_num = 0;
for (uint32_t i = 0; i < duration; i++) {
auto find_time = stat.second.uni_key_num.find(i);
if (find_time != stat.second.uni_key_num.end()) {
cur_ratio = (static_cast<double>(find_time->second)) / cur_uni_key;
cur_num = find_time->second;
}
ret = snprintf(buffer_, sizeof(buffer_), "%" PRIu64 " %.12f\n",
cur_num, cur_ratio);
if (ret < 0) {
return Status::IOError("Format the output failed");
}
std::string printout(buffer_);
s = stat.second.a_key_num_f->Append(printout);
if (!s.ok()) {
fprintf(stderr,
"Write accessed unique key number change file failed\n");
return s;
}
}
}
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
// output the prefix of top k access peak
if (FLAGS_output_prefix_cut > 0 && stat.second.a_top_qps_prefix_f) {
while (!stat.second.top_k_qps_sec.empty()) {
ret = snprintf(buffer_, sizeof(buffer_), "At time: %u with QPS: %u\n",
stat.second.top_k_qps_sec.top().second,
stat.second.top_k_qps_sec.top().first);
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
if (ret < 0) {
return Status::IOError("Format the output failed");
}
std::string printout(buffer_);
s = stat.second.a_top_qps_prefix_f->Append(printout);
if (!s.ok()) {
fprintf(stderr, "Write prefix QPS top K file failed\n");
return s;
}
uint32_t qps_time = stat.second.top_k_qps_sec.top().second;
stat.second.top_k_qps_sec.pop();
if (stat.second.a_qps_prefix_stats.find(qps_time) !=
stat.second.a_qps_prefix_stats.end()) {
for (auto& qps_prefix : stat.second.a_qps_prefix_stats[qps_time]) {
std::string qps_prefix_out =
ROCKSDB_NAMESPACE::LDBCommand::StringToHex(qps_prefix.first);
ret = snprintf(buffer_, sizeof(buffer_),
"The prefix: %s Access count: %u\n",
qps_prefix_out.c_str(), qps_prefix.second);
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
if (ret < 0) {
return Status::IOError("Format the output failed");
}
std::string pout(buffer_);
s = stat.second.a_top_qps_prefix_f->Append(pout);
if (!s.ok()) {
fprintf(stderr, "Write prefix QPS top K file failed\n");
return s;
}
}
}
}
}
}
}
if (qps_f_) {
for (uint32_t i = 0; i < duration; i++) {
for (int type = 0; type <= kTaTypeNum; type++) {
if (type < kTaTypeNum) {
ret = snprintf(buffer_, sizeof(buffer_), "%u ", type_qps[i][type]);
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
} else {
ret = snprintf(buffer_, sizeof(buffer_), "%u\n", type_qps[i][type]);
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
}
if (ret < 0) {
return Status::IOError("Format the output failed");
}
std::string printout(buffer_);
s = qps_f_->Append(printout);
if (!s.ok()) {
return s;
}
qps_sum[type] += type_qps[i][type];
if (type_qps[i][type] > qps_peak[type]) {
qps_peak[type] = type_qps[i][type];
}
}
}
}
if (cf_qps_f_) {
int cfs_size = static_cast<uint32_t>(cfs_.size());
uint32_t v;
for (uint32_t i = 0; i < duration; i++) {
for (int cf = 0; cf < cfs_size; cf++) {
if (cfs_[cf].cf_qps.find(i) != cfs_[cf].cf_qps.end()) {
v = cfs_[cf].cf_qps[i];
} else {
v = 0;
}
if (cf < cfs_size - 1) {
ret = snprintf(buffer_, sizeof(buffer_), "%u ", v);
} else {
ret = snprintf(buffer_, sizeof(buffer_), "%u\n", v);
}
if (ret < 0) {
return Status::IOError("Format the output failed");
}
std::string printout(buffer_);
s = cf_qps_f_->Append(printout);
if (!s.ok()) {
return s;
}
}
}
}
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
qps_peak_ = qps_peak;
for (int type = 0; type <= kTaTypeNum; type++) {
if (duration == 0) {
qps_ave_[type] = 0;
} else {
qps_ave_[type] = (static_cast<double>(qps_sum[type])) / duration;
}
}
return Status::OK();
}
// In reprocessing, if we have the whole key space
// we can output the access count of all keys in a cf
// we can make some statistics of the whole key space
// also, we output the top k accessed keys here
Status TraceAnalyzer::ReProcessing() {
int ret;
Status s;
for (auto& cf_it : cfs_) {
uint32_t cf_id = cf_it.first;
// output the time series;
if (FLAGS_output_time_series) {
for (int type = 0; type < kTaTypeNum; type++) {
if (!ta_[type].enabled ||
ta_[type].stats.find(cf_id) == ta_[type].stats.end()) {
continue;
}
TraceStats& stat = ta_[type].stats[cf_id];
if (!stat.time_series_f) {
fprintf(stderr, "Cannot write time_series of '%s' in '%u'\n",
ta_[type].type_name.c_str(), cf_id);
continue;
}
while (!stat.time_series.empty()) {
uint64_t key_id = 0;
auto found = stat.a_key_stats.find(stat.time_series.front().key);
if (found != stat.a_key_stats.end()) {
key_id = found->second.key_id;
}
ret =
snprintf(buffer_, sizeof(buffer_), "%u %" PRIu64 " %" PRIu64 "\n",
stat.time_series.front().type,
stat.time_series.front().ts, key_id);
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
if (ret < 0) {
return Status::IOError("Format the output failed");
}
std::string printout(buffer_);
s = stat.time_series_f->Append(printout);
if (!s.ok()) {
fprintf(stderr, "Write time series file failed\n");
return s;
}
stat.time_series.pop_front();
}
}
}
// process the whole key space if needed
if (!FLAGS_key_space_dir.empty()) {
std::string whole_key_path =
FLAGS_key_space_dir + "/" + std::to_string(cf_id) + ".txt";
std::string input_key, get_key;
std::vector<std::string> prefix(kTaTypeNum);
std::unique_ptr<FSSequentialFile> file;
Introduce a new storage specific Env API (#5761) Summary: The current Env API encompasses both storage/file operations, as well as OS related operations. Most of the APIs return a Status, which does not have enough metadata about an error, such as whether its retry-able or not, scope (i.e fault domain) of the error etc., that may be required in order to properly handle a storage error. The file APIs also do not provide enough control over the IO SLA, such as timeout, prioritization, hinting about placement and redundancy etc. This PR separates out the file/storage APIs from Env into a new FileSystem class. The APIs are updated to return an IOStatus with metadata about the error, as well as to take an IOOptions structure as input in order to allow more control over the IO. The user can set both ```options.env``` and ```options.file_system``` to specify that RocksDB should use the former for OS related operations and the latter for storage operations. Internally, a ```CompositeEnvWrapper``` has been introduced that inherits from ```Env``` and redirects individual methods to either an ```Env``` implementation or the ```FileSystem``` as appropriate. When options are sanitized during ```DB::Open```, ```options.env``` is replaced with a newly allocated ```CompositeEnvWrapper``` instance if both env and file_system have been specified. This way, the rest of the RocksDB code can continue to function as before. This PR also ports PosixEnv to the new API by splitting it into two - PosixEnv and PosixFileSystem. PosixEnv is defined as a sub-class of CompositeEnvWrapper, and threading/time functions are overridden with Posix specific implementations in order to avoid an extra level of indirection. The ```CompositeEnvWrapper``` translates ```IOStatus``` return code to ```Status```, and sets the severity to ```kSoftError``` if the io_status is retryable. The error handling code in RocksDB can then recover the DB automatically. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5761 Differential Revision: D18868376 Pulled By: anand1976 fbshipit-source-id: 39efe18a162ea746fabac6360ff529baba48486f
5 years ago
s = env_->GetFileSystem()->NewSequentialFile(
whole_key_path, FileOptions(env_options_), &file, nullptr);
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
if (!s.ok()) {
fprintf(stderr, "Cannot open the whole key space file of CF: %u\n",
cf_id);
file.reset();
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
}
if (file) {
size_t kTraceFileReadaheadSize = 2 * 1024 * 1024;
LineFileReader lf_reader(
std::move(file), whole_key_path,
kTraceFileReadaheadSize /* filereadahead_size */);
Support read rate-limiting in SequentialFileReader (#9973) Summary: Added rate limiter and read rate-limiting support to SequentialFileReader. I've updated call sites to SequentialFileReader::Read with appropriate IO priority (or left a TODO and specified IO_TOTAL for now). The PR is separated into four commits: the first one added the rate-limiting support, but with some fixes in the unit test since the number of request bytes from rate limiter in SequentialFileReader are not accurate (there is overcharge at EOF). The second commit fixed this by allowing SequentialFileReader to check file size and determine how many bytes are left in the file to read. The third commit added benchmark related code. The fourth commit moved the logic of using file size to avoid overcharging the rate limiter into backup engine (the main user of SequentialFileReader). Pull Request resolved: https://github.com/facebook/rocksdb/pull/9973 Test Plan: - `make check`, backup_engine_test covers usage of SequentialFileReader with rate limiter. - Run db_bench to check if rate limiting is throttling as expected: Verified that reads and writes are together throttled at 2MB/s, and at 0.2MB chunks that are 100ms apart. - Set up: `./db_bench --benchmarks=fillrandom -db=/dev/shm/test_rocksdb` - Benchmark: ``` strace -ttfe read,write ./db_bench --benchmarks=backup -db=/dev/shm/test_rocksdb --backup_rate_limit=2097152 --use_existing_db strace -ttfe read,write ./db_bench --benchmarks=restore -db=/dev/shm/test_rocksdb --restore_rate_limit=2097152 --use_existing_db ``` - db bench on backup and restore to ensure no performance regression. - backup (avg over 50 runs): pre-change: 1.90443e+06 micros/op; post-change: 1.8993e+06 micros/op (improve by 0.2%) - restore (avg over 50 runs): pre-change: 1.79105e+06 micros/op; post-change: 1.78192e+06 micros/op (improve by 0.5%) ``` # Set up ./db_bench --benchmarks=fillrandom -db=/tmp/test_rocksdb -num=10000000 # benchmark TEST_TMPDIR=/tmp/test_rocksdb NUM_RUN=50 for ((j=0;j<$NUM_RUN;j++)) do ./db_bench -db=$TEST_TMPDIR -num=10000000 -benchmarks=backup -use_existing_db | egrep 'backup' # Restore #./db_bench -db=$TEST_TMPDIR -num=10000000 -benchmarks=restore -use_existing_db done > rate_limit.txt && awk -v NUM_RUN=$NUM_RUN '{sum+=$3;sum_sqrt+=$3^2}END{print sum/NUM_RUN, sqrt(sum_sqrt/NUM_RUN-(sum/NUM_RUN)^2)}' rate_limit.txt >> rate_limit_2.txt ``` Reviewed By: hx235 Differential Revision: D36327418 Pulled By: cbi42 fbshipit-source-id: e75d4307cff815945482df5ba630c1e88d064691
3 years ago
for (cfs_[cf_id].w_count = 0; lf_reader.ReadLine(
&get_key, Env::IO_TOTAL /* rate_limiter_priority */);
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
++cfs_[cf_id].w_count) {
input_key = ROCKSDB_NAMESPACE::LDBCommand::HexToString(get_key);
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
for (int type = 0; type < kTaTypeNum; type++) {
if (!ta_[type].enabled) {
continue;
}
TraceStats& stat = ta_[type].stats[cf_id];
if (stat.w_key_f) {
if (stat.a_key_stats.find(input_key) != stat.a_key_stats.end()) {
ret = snprintf(buffer_, sizeof(buffer_),
"%" PRIu64 " %" PRIu64 "\n", cfs_[cf_id].w_count,
stat.a_key_stats[input_key].access_count);
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
if (ret < 0) {
return Status::IOError("Format the output failed");
}
std::string printout(buffer_);
s = stat.w_key_f->Append(printout);
if (!s.ok()) {
fprintf(stderr, "Write whole key space access file failed\n");
return s;
}
}
}
// Output the prefix cut file of the whole key space
if (FLAGS_output_prefix_cut > 0 && stat.w_prefix_cut_f) {
if (input_key.compare(0, FLAGS_output_prefix_cut, prefix[type]) !=
0) {
prefix[type] = input_key.substr(0, FLAGS_output_prefix_cut);
std::string prefix_out =
ROCKSDB_NAMESPACE::LDBCommand::StringToHex(prefix[type]);
ret = snprintf(buffer_, sizeof(buffer_), "%" PRIu64 " %s\n",
cfs_[cf_id].w_count, prefix_out.c_str());
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
if (ret < 0) {
return Status::IOError("Format the output failed");
}
std::string printout(buffer_);
s = stat.w_prefix_cut_f->Append(printout);
if (!s.ok()) {
fprintf(stderr,
"Write whole key space prefix cut file failed\n");
return s;
}
}
}
}
// Make the statistics fo the key size distribution
if (FLAGS_output_key_distribution) {
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
if (cfs_[cf_id].w_key_size_stats.find(input_key.size()) ==
cfs_[cf_id].w_key_size_stats.end()) {
cfs_[cf_id].w_key_size_stats[input_key.size()] = 1;
} else {
cfs_[cf_id].w_key_size_stats[input_key.size()]++;
}
}
}
s = lf_reader.GetStatus();
if (!s.ok()) {
fprintf(stderr, "Read whole key space file failed\n");
return s;
}
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
}
}
// process the top k accessed keys
if (FLAGS_print_top_k_access > 0) {
for (int type = 0; type < kTaTypeNum; type++) {
if (!ta_[type].enabled ||
ta_[type].stats.find(cf_id) == ta_[type].stats.end()) {
continue;
}
TraceStats& stat = ta_[type].stats[cf_id];
for (auto& record : stat.a_key_stats) {
if (static_cast<int32_t>(stat.top_k_queue.size()) <
FLAGS_print_top_k_access) {
stat.top_k_queue.push(
std::make_pair(record.second.access_count, record.first));
} else {
if (record.second.access_count > stat.top_k_queue.top().first) {
stat.top_k_queue.pop();
stat.top_k_queue.push(
std::make_pair(record.second.access_count, record.first));
}
}
}
}
}
}
return Status::OK();
}
// End the processing, print the requested results
Status TraceAnalyzer::EndProcessing() {
Status s;
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
if (trace_sequence_f_) {
s = trace_sequence_f_->Close();
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
}
if (FLAGS_no_print) {
return s;
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
}
PrintStatistics();
if (s.ok()) {
s = CloseOutputFiles();
}
return s;
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
}
// Insert the corresponding key statistics to the correct type
// and correct CF, output the time-series file if needed
Status TraceAnalyzer::KeyStatsInsertion(const uint32_t& type,
const uint32_t& cf_id,
const std::string& key,
const size_t value_size,
const uint64_t ts) {
Status s;
StatsUnit unit;
unit.key_id = 0;
unit.cf_id = cf_id;
unit.value_size = value_size;
unit.access_count = 1;
unit.latest_ts = ts;
if ((type != TraceOperationType::kGet &&
type != TraceOperationType::kMultiGet) ||
value_size > 0) {
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
unit.succ_count = 1;
} else {
unit.succ_count = 0;
}
unit.v_correlation.resize(analyzer_opts_.correlation_list.size());
for (int i = 0;
i < (static_cast<int>(analyzer_opts_.correlation_list.size())); i++) {
unit.v_correlation[i].count = 0;
unit.v_correlation[i].total_ts = 0;
}
std::string prefix;
if (FLAGS_output_prefix_cut > 0) {
prefix = key.substr(0, FLAGS_output_prefix_cut);
}
if (begin_time_ == 0) {
begin_time_ = ts;
}
uint32_t time_in_sec;
if (ts < begin_time_) {
time_in_sec = 0;
} else {
time_in_sec = static_cast<uint32_t>((ts - begin_time_) / 1000000);
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
}
uint64_t dist_value_size = value_size / FLAGS_value_interval;
auto found_stats = ta_[type].stats.find(cf_id);
if (found_stats == ta_[type].stats.end()) {
ta_[type].stats[cf_id].cf_id = cf_id;
ta_[type].stats[cf_id].cf_name = std::to_string(cf_id);
ta_[type].stats[cf_id].a_count = 1;
ta_[type].stats[cf_id].a_key_id = 0;
ta_[type].stats[cf_id].a_key_size_sqsum = MultiplyCheckOverflow(
static_cast<uint64_t>(key.size()), static_cast<uint64_t>(key.size()));
ta_[type].stats[cf_id].a_key_size_sum = key.size();
ta_[type].stats[cf_id].a_value_size_sqsum = MultiplyCheckOverflow(
static_cast<uint64_t>(value_size), static_cast<uint64_t>(value_size));
ta_[type].stats[cf_id].a_value_size_sum = value_size;
s = OpenStatsOutputFiles(ta_[type].type_name, ta_[type].stats[cf_id]);
if (!FLAGS_print_correlation.empty()) {
s = StatsUnitCorrelationUpdate(unit, type, ts, key);
}
ta_[type].stats[cf_id].a_key_stats[key] = unit;
ta_[type].stats[cf_id].a_value_size_stats[dist_value_size] = 1;
ta_[type].stats[cf_id].a_qps_stats[time_in_sec] = 1;
ta_[type].stats[cf_id].correlation_output.resize(
analyzer_opts_.correlation_list.size());
if (FLAGS_output_prefix_cut > 0) {
std::map<std::string, uint32_t> tmp_qps_map;
tmp_qps_map[prefix] = 1;
ta_[type].stats[cf_id].a_qps_prefix_stats[time_in_sec] = tmp_qps_map;
}
if (time_in_sec != cur_time_sec_) {
ta_[type].stats[cf_id].uni_key_num[cur_time_sec_] =
static_cast<uint64_t>(ta_[type].stats[cf_id].a_key_stats.size());
cur_time_sec_ = time_in_sec;
}
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
} else {
found_stats->second.a_count++;
found_stats->second.a_key_size_sqsum += MultiplyCheckOverflow(
static_cast<uint64_t>(key.size()), static_cast<uint64_t>(key.size()));
found_stats->second.a_key_size_sum += key.size();
found_stats->second.a_value_size_sqsum += MultiplyCheckOverflow(
static_cast<uint64_t>(value_size), static_cast<uint64_t>(value_size));
found_stats->second.a_value_size_sum += value_size;
auto found_key = found_stats->second.a_key_stats.find(key);
if (found_key == found_stats->second.a_key_stats.end()) {
found_stats->second.a_key_stats[key] = unit;
} else {
found_key->second.access_count++;
if (type != TraceOperationType::kGet || value_size > 0) {
found_key->second.succ_count++;
}
if (!FLAGS_print_correlation.empty()) {
s = StatsUnitCorrelationUpdate(found_key->second, type, ts, key);
}
}
if (time_in_sec != cur_time_sec_) {
found_stats->second.uni_key_num[cur_time_sec_] =
static_cast<uint64_t>(found_stats->second.a_key_stats.size());
cur_time_sec_ = time_in_sec;
}
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
auto found_value =
found_stats->second.a_value_size_stats.find(dist_value_size);
if (found_value == found_stats->second.a_value_size_stats.end()) {
found_stats->second.a_value_size_stats[dist_value_size] = 1;
} else {
found_value->second++;
}
auto found_qps = found_stats->second.a_qps_stats.find(time_in_sec);
if (found_qps == found_stats->second.a_qps_stats.end()) {
found_stats->second.a_qps_stats[time_in_sec] = 1;
} else {
found_qps->second++;
}
if (FLAGS_output_prefix_cut > 0) {
auto found_qps_prefix =
found_stats->second.a_qps_prefix_stats.find(time_in_sec);
if (found_qps_prefix == found_stats->second.a_qps_prefix_stats.end()) {
std::map<std::string, uint32_t> tmp_qps_map;
found_stats->second.a_qps_prefix_stats[time_in_sec] = tmp_qps_map;
}
if (found_stats->second.a_qps_prefix_stats[time_in_sec].find(prefix) ==
found_stats->second.a_qps_prefix_stats[time_in_sec].end()) {
found_stats->second.a_qps_prefix_stats[time_in_sec][prefix] = 1;
} else {
found_stats->second.a_qps_prefix_stats[time_in_sec][prefix]++;
}
}
}
if (cfs_.find(cf_id) == cfs_.end()) {
CfUnit cf_unit;
cf_unit.cf_id = cf_id;
cf_unit.w_count = 0;
cf_unit.a_count = 0;
cfs_[cf_id] = cf_unit;
}
if (FLAGS_output_qps_stats) {
cfs_[cf_id].cf_qps[time_in_sec]++;
}
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
if (FLAGS_output_time_series) {
TraceUnit trace_u;
trace_u.type = type;
trace_u.key = key;
trace_u.value_size = value_size;
trace_u.ts = (ts - time_series_start_) / 1000000;
trace_u.cf_id = cf_id;
ta_[type].stats[cf_id].time_series.push_back(trace_u);
}
return s;
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
}
// Update the correlation unit of each key if enabled
Status TraceAnalyzer::StatsUnitCorrelationUpdate(StatsUnit& unit,
const uint32_t& type_second,
const uint64_t& ts,
const std::string& key) {
if (type_second >= kTaTypeNum) {
fprintf(stderr, "Unknown Type Id: %u\n", type_second);
return Status::NotFound();
}
for (int type_first = 0; type_first < kTaTypeNum; type_first++) {
if (type_first >= static_cast<int>(ta_.size()) ||
type_first >= static_cast<int>(analyzer_opts_.correlation_map.size())) {
break;
}
if (analyzer_opts_.correlation_map[type_first][type_second] < 0 ||
ta_[type_first].stats.find(unit.cf_id) == ta_[type_first].stats.end() ||
ta_[type_first].stats[unit.cf_id].a_key_stats.find(key) ==
ta_[type_first].stats[unit.cf_id].a_key_stats.end() ||
ta_[type_first].stats[unit.cf_id].a_key_stats[key].latest_ts == ts) {
continue;
}
int correlation_id =
analyzer_opts_.correlation_map[type_first][type_second];
// after get the x-y operation time or x, update;
if (correlation_id < 0 ||
correlation_id >= static_cast<int>(unit.v_correlation.size())) {
continue;
}
unit.v_correlation[correlation_id].count++;
unit.v_correlation[correlation_id].total_ts +=
(ts - ta_[type_first].stats[unit.cf_id].a_key_stats[key].latest_ts);
}
unit.latest_ts = ts;
return Status::OK();
}
// when a new trace statistic is created, the file handler
// pointers should be initiated if needed according to
// the trace analyzer options
Status TraceAnalyzer::OpenStatsOutputFiles(const std::string& type,
TraceStats& new_stats) {
Status s;
if (FLAGS_output_key_stats) {
s = CreateOutputFile(type, new_stats.cf_name, "accessed_key_stats.txt",
&new_stats.a_key_f);
s = CreateOutputFile(type, new_stats.cf_name,
"accessed_unique_key_num_change.txt",
&new_stats.a_key_num_f);
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
if (!FLAGS_key_space_dir.empty()) {
s = CreateOutputFile(type, new_stats.cf_name, "whole_key_stats.txt",
&new_stats.w_key_f);
}
}
if (FLAGS_output_access_count_stats) {
s = CreateOutputFile(type, new_stats.cf_name,
"accessed_key_count_distribution.txt",
&new_stats.a_count_dist_f);
}
if (FLAGS_output_prefix_cut > 0) {
s = CreateOutputFile(type, new_stats.cf_name, "accessed_key_prefix_cut.txt",
&new_stats.a_prefix_cut_f);
if (!FLAGS_key_space_dir.empty()) {
s = CreateOutputFile(type, new_stats.cf_name, "whole_key_prefix_cut.txt",
&new_stats.w_prefix_cut_f);
}
if (FLAGS_output_qps_stats) {
s = CreateOutputFile(type, new_stats.cf_name,
"accessed_top_k_qps_prefix_cut.txt",
&new_stats.a_top_qps_prefix_f);
}
}
if (FLAGS_output_time_series) {
s = CreateOutputFile(type, new_stats.cf_name, "time_series.txt",
&new_stats.time_series_f);
}
if (FLAGS_output_value_distribution) {
s = CreateOutputFile(type, new_stats.cf_name,
"accessed_value_size_distribution.txt",
&new_stats.a_value_size_f);
}
if (FLAGS_output_key_distribution) {
s = CreateOutputFile(type, new_stats.cf_name,
"accessed_key_size_distribution.txt",
&new_stats.a_key_size_f);
}
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
if (FLAGS_output_qps_stats) {
s = CreateOutputFile(type, new_stats.cf_name, "qps_stats.txt",
&new_stats.a_qps_f);
}
return s;
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
}
// create the output path of the files to be opened
Status TraceAnalyzer::CreateOutputFile(
const std::string& type, const std::string& cf_name,
const std::string& ending,
std::unique_ptr<ROCKSDB_NAMESPACE::WritableFile>* f_ptr) {
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
std::string path;
path = output_path_ + "/" + FLAGS_output_prefix + "-" + type + "-" + cf_name +
"-" + ending;
Status s;
s = env_->NewWritableFile(path, f_ptr, env_options_);
if (!s.ok()) {
fprintf(stderr, "Cannot open file: %s\n", path.c_str());
exit(1);
}
return Status::OK();
}
// Close the output files in the TraceStats if they are opened
Status TraceAnalyzer::CloseOutputFiles() {
Status s;
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
for (int type = 0; type < kTaTypeNum; type++) {
if (!ta_[type].enabled) {
continue;
}
for (auto& stat : ta_[type].stats) {
if (s.ok() && stat.second.time_series_f) {
s = stat.second.time_series_f->Close();
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
}
if (s.ok() && stat.second.a_key_f) {
s = stat.second.a_key_f->Close();
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
}
if (s.ok() && stat.second.a_key_num_f) {
s = stat.second.a_key_num_f->Close();
}
if (s.ok() && stat.second.a_count_dist_f) {
s = stat.second.a_count_dist_f->Close();
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
}
if (s.ok() && stat.second.a_prefix_cut_f) {
s = stat.second.a_prefix_cut_f->Close();
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
}
if (s.ok() && stat.second.a_value_size_f) {
s = stat.second.a_value_size_f->Close();
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
}
if (s.ok() && stat.second.a_key_size_f) {
s = stat.second.a_key_size_f->Close();
}
if (s.ok() && stat.second.a_qps_f) {
s = stat.second.a_qps_f->Close();
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
}
if (s.ok() && stat.second.a_top_qps_prefix_f) {
s = stat.second.a_top_qps_prefix_f->Close();
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
}
if (s.ok() && stat.second.w_key_f) {
s = stat.second.w_key_f->Close();
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
}
if (s.ok() && stat.second.w_prefix_cut_f) {
s = stat.second.w_prefix_cut_f->Close();
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
}
}
}
return s;
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
}
Status TraceAnalyzer::Handle(const WriteQueryTraceRecord& record,
std::unique_ptr<TraceRecordResult>* /*result*/) {
total_writes_++;
// Note that, if the write happens in a transaction,
// 'Write' will be called twice, one for Prepare, one for
// Commit. Thus, in the trace, for the same WriteBatch, there
// will be two records if it is in a transaction. Here, we only
// process the reord that is committed. If write is non-transaction,
// HasBeginPrepare()==false, so we process it normally.
WriteBatch batch(record.GetWriteBatchRep().ToString());
if (batch.Count() == 0 || (batch.HasBeginPrepare() && !batch.HasCommit())) {
return Status::OK();
}
write_batch_ts_ = record.GetTimestamp();
// write_result_ will be updated in batch's handler during iteration.
Status s = batch.Iterate(this);
write_batch_ts_ = 0;
if (!s.ok()) {
fprintf(stderr, "Cannot process the write batch in the trace\n");
return s;
}
return Status::OK();
}
Status TraceAnalyzer::Handle(const GetQueryTraceRecord& record,
std::unique_ptr<TraceRecordResult>* /*result*/) {
total_gets_++;
return OutputAnalysisResult(TraceOperationType::kGet, record.GetTimestamp(),
record.GetColumnFamilyID(),
std::move(record.GetKey()), 0);
}
Status TraceAnalyzer::Handle(const IteratorSeekQueryTraceRecord& record,
std::unique_ptr<TraceRecordResult>* /*result*/) {
TraceOperationType op_type;
if (record.GetSeekType() == IteratorSeekQueryTraceRecord::kSeek) {
op_type = TraceOperationType::kIteratorSeek;
total_seeks_++;
} else {
op_type = TraceOperationType::kIteratorSeekForPrev;
total_seek_prevs_++;
}
// To do: shall we add lower/upper bounds?
return OutputAnalysisResult(op_type, record.GetTimestamp(),
record.GetColumnFamilyID(),
std::move(record.GetKey()), 0);
}
Status TraceAnalyzer::Handle(const MultiGetQueryTraceRecord& record,
std::unique_ptr<TraceRecordResult>* /*result*/) {
total_multigets_++;
std::vector<uint32_t> cf_ids = record.GetColumnFamilyIDs();
std::vector<Slice> keys = record.GetKeys();
std::vector<size_t> value_sizes;
// If the size does not match is not the error of tracing and anayzing, we
// just report it to the user. The analyzing continues.
if (cf_ids.size() > keys.size()) {
printf("The CF ID vector size does not match the keys vector size!\n");
// Make the sure the 2 vectors are of the same (smaller) size.
cf_ids.resize(keys.size());
} else if (cf_ids.size() < keys.size()) {
printf("The CF ID vector size does not match the keys vector size!\n");
// Make the sure the 2 vectors are of the same (smaller) size.
keys.resize(cf_ids.size());
}
// Now the 2 vectors must be of the same size.
value_sizes.resize(cf_ids.size(), 0);
return OutputAnalysisResult(TraceOperationType::kMultiGet,
record.GetTimestamp(), std::move(cf_ids),
std::move(keys), std::move(value_sizes));
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
}
// Handle the Put request in the write batch of the trace
Status TraceAnalyzer::PutCF(uint32_t column_family_id, const Slice& key,
const Slice& value) {
return OutputAnalysisResult(TraceOperationType::kPut, write_batch_ts_,
column_family_id, key, value.size());
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
}
// Handle the Delete request in the write batch of the trace
Status TraceAnalyzer::DeleteCF(uint32_t column_family_id, const Slice& key) {
return OutputAnalysisResult(TraceOperationType::kDelete, write_batch_ts_,
column_family_id, key, 0);
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
}
// Handle the SingleDelete request in the write batch of the trace
Status TraceAnalyzer::SingleDeleteCF(uint32_t column_family_id,
const Slice& key) {
return OutputAnalysisResult(TraceOperationType::kSingleDelete,
write_batch_ts_, column_family_id, key, 0);
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
}
// Handle the DeleteRange request in the write batch of the trace
Status TraceAnalyzer::DeleteRangeCF(uint32_t column_family_id,
const Slice& begin_key,
const Slice& end_key) {
return OutputAnalysisResult(TraceOperationType::kRangeDelete, write_batch_ts_,
{column_family_id, column_family_id},
{begin_key, end_key}, {0, 0});
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
}
// Handle the Merge request in the write batch of the trace
Status TraceAnalyzer::MergeCF(uint32_t column_family_id, const Slice& key,
const Slice& value) {
return OutputAnalysisResult(TraceOperationType::kMerge, write_batch_ts_,
column_family_id, key, value.size());
}
Status TraceAnalyzer::OutputAnalysisResult(TraceOperationType op_type,
uint64_t timestamp,
std::vector<uint32_t> cf_ids,
std::vector<Slice> keys,
std::vector<size_t> value_sizes) {
assert(!cf_ids.empty());
assert(cf_ids.size() == keys.size());
assert(cf_ids.size() == value_sizes.size());
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
Status s;
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
if (FLAGS_convert_to_human_readable_trace && trace_sequence_f_) {
// DeleteRane only writes the begin_key.
size_t cnt =
op_type == TraceOperationType::kRangeDelete ? 1 : cf_ids.size();
for (size_t i = 0; i < cnt; i++) {
s = WriteTraceSequence(op_type, cf_ids[i], keys[i], value_sizes[i],
timestamp);
if (!s.ok()) {
return Status::Corruption("Failed to write the trace sequence to file");
}
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
}
}
if (ta_[op_type].sample_count >= sample_max_) {
ta_[op_type].sample_count = 0;
}
if (ta_[op_type].sample_count > 0) {
ta_[op_type].sample_count++;
return Status::OK();
}
ta_[op_type].sample_count++;
if (!ta_[op_type].enabled) {
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
return Status::OK();
}
for (size_t i = 0; i < cf_ids.size(); i++) {
// Get query does not have value part, just give a fixed value 10 for easy
// calculation.
s = KeyStatsInsertion(
op_type, cf_ids[i], keys[i].ToString(),
value_sizes[i] == 0 ? kShadowValueSize : value_sizes[i], timestamp);
if (!s.ok()) {
return Status::Corruption("Failed to insert key statistics");
}
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
}
return Status::OK();
}
Status TraceAnalyzer::OutputAnalysisResult(TraceOperationType op_type,
uint64_t timestamp, uint32_t cf_id,
const Slice& key,
size_t value_size) {
return OutputAnalysisResult(
op_type, timestamp, std::vector<uint32_t>({cf_id}),
std::vector<Slice>({key}), std::vector<size_t>({value_size}));
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
}
// Before the analyzer is closed, the requested general statistic results are
// printed out here. In current stage, these information are not output to
// the files.
// -----type
// |__cf_id
// |_statistics
void TraceAnalyzer::PrintStatistics() {
for (int type = 0; type < kTaTypeNum; type++) {
if (!ta_[type].enabled) {
continue;
}
ta_[type].total_keys = 0;
ta_[type].total_access = 0;
ta_[type].total_succ_access = 0;
printf("\n################# Operation Type: %s #####################\n",
ta_[type].type_name.c_str());
if (qps_ave_.size() == kTaTypeNum + 1) {
printf("Peak QPS is: %u Average QPS is: %f\n", qps_peak_[type],
qps_ave_[type]);
}
for (auto& stat_it : ta_[type].stats) {
if (stat_it.second.a_count == 0) {
continue;
}
TraceStats& stat = stat_it.second;
uint64_t total_a_keys = static_cast<uint64_t>(stat.a_key_stats.size());
double key_size_ave = 0.0;
double value_size_ave = 0.0;
double key_size_vari = 0.0;
double value_size_vari = 0.0;
if (stat.a_count > 0) {
key_size_ave =
(static_cast<double>(stat.a_key_size_sum)) / stat.a_count;
value_size_ave =
(static_cast<double>(stat.a_value_size_sum)) / stat.a_count;
key_size_vari = std::sqrt((static_cast<double>(stat.a_key_size_sqsum)) /
stat.a_count -
key_size_ave * key_size_ave);
value_size_vari = std::sqrt(
(static_cast<double>(stat.a_value_size_sqsum)) / stat.a_count -
value_size_ave * value_size_ave);
}
if (value_size_ave == 0.0) {
stat.a_value_mid = 0;
}
cfs_[stat.cf_id].a_count += total_a_keys;
ta_[type].total_keys += total_a_keys;
ta_[type].total_access += stat.a_count;
ta_[type].total_succ_access += stat.a_succ_count;
printf("*********************************************************\n");
printf("colume family id: %u\n", stat.cf_id);
printf("Total number of queries to this cf by %s: %" PRIu64 "\n",
ta_[type].type_name.c_str(), stat.a_count);
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
printf("Total unique keys in this cf: %" PRIu64 "\n", total_a_keys);
printf("Average key size: %f key size medium: %" PRIu64
" Key size Variation: %f\n",
key_size_ave, stat.a_key_mid, key_size_vari);
if (type == kPut || type == kMerge) {
printf("Average value size: %f Value size medium: %" PRIu64
" Value size variation: %f\n",
value_size_ave, stat.a_value_mid, value_size_vari);
}
printf("Peak QPS is: %u Average QPS is: %f\n", stat.a_peak_qps,
stat.a_ave_qps);
// print the top k accessed key and its access count
if (FLAGS_print_top_k_access > 0) {
printf("The Top %d keys that are accessed:\n",
FLAGS_print_top_k_access);
while (!stat.top_k_queue.empty()) {
std::string hex_key = ROCKSDB_NAMESPACE::LDBCommand::StringToHex(
stat.top_k_queue.top().second);
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
printf("Access_count: %" PRIu64 " %s\n", stat.top_k_queue.top().first,
hex_key.c_str());
stat.top_k_queue.pop();
}
}
// print the top k access prefix range and
// top k prefix range with highest average access per key
if (FLAGS_output_prefix_cut > 0) {
printf("The Top %d accessed prefix range:\n", FLAGS_print_top_k_access);
while (!stat.top_k_prefix_access.empty()) {
printf("Prefix: %s Access count: %" PRIu64 "\n",
stat.top_k_prefix_access.top().second.c_str(),
stat.top_k_prefix_access.top().first);
stat.top_k_prefix_access.pop();
}
printf("The Top %d prefix with highest access per key:\n",
FLAGS_print_top_k_access);
while (!stat.top_k_prefix_ave.empty()) {
printf("Prefix: %s access per key: %f\n",
stat.top_k_prefix_ave.top().second.c_str(),
stat.top_k_prefix_ave.top().first);
stat.top_k_prefix_ave.pop();
}
}
// print the operation correlations
if (!FLAGS_print_correlation.empty()) {
for (int correlation = 0;
correlation <
static_cast<int>(analyzer_opts_.correlation_list.size());
correlation++) {
printf(
"The correlation statistics of '%s' after '%s' is:",
taIndexToOpt[analyzer_opts_.correlation_list[correlation].second]
.c_str(),
taIndexToOpt[analyzer_opts_.correlation_list[correlation].first]
.c_str());
double correlation_ave = 0.0;
if (stat.correlation_output[correlation].first > 0) {
correlation_ave =
(static_cast<double>(
stat.correlation_output[correlation].second)) /
(stat.correlation_output[correlation].first * 1000);
}
printf(" total numbers: %" PRIu64 " average time: %f(ms)\n",
stat.correlation_output[correlation].first, correlation_ave);
}
}
}
printf("*********************************************************\n");
printf("Total keys of '%s' is: %" PRIu64 "\n", ta_[type].type_name.c_str(),
ta_[type].total_keys);
printf("Total access is: %" PRIu64 "\n", ta_[type].total_access);
total_access_keys_ += ta_[type].total_keys;
}
// Print the overall statistic information of the trace
printf("\n*********************************************************\n");
printf("*********************************************************\n");
printf("The column family based statistics\n");
for (auto& cf : cfs_) {
printf("The column family id: %u\n", cf.first);
printf("The whole key space key numbers: %" PRIu64 "\n", cf.second.w_count);
printf("The accessed key space key numbers: %" PRIu64 "\n",
cf.second.a_count);
}
if (FLAGS_print_overall_stats) {
printf("\n*********************************************************\n");
printf("*********************************************************\n");
if (qps_peak_.size() == kTaTypeNum + 1) {
printf("Average QPS per second: %f Peak QPS: %u\n", qps_ave_[kTaTypeNum],
qps_peak_[kTaTypeNum]);
}
printf("The statistics related to query number need to times: %u\n",
sample_max_);
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
printf("Total_requests: %" PRIu64 " Total_accessed_keys: %" PRIu64
" Total_gets: %" PRIu64 " Total_write_batches: %" PRIu64
" Total_seeks: %" PRIu64 " Total_seek_for_prevs: %" PRIu64
" Total_multigets: %" PRIu64 "\n",
total_requests_, total_access_keys_, total_gets_, total_writes_,
total_seeks_, total_seek_prevs_, total_multigets_);
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
for (int type = 0; type < kTaTypeNum; type++) {
if (!ta_[type].enabled) {
continue;
}
printf("Operation: '%s' has: %" PRIu64 "\n", ta_[type].type_name.c_str(),
ta_[type].total_access);
}
}
}
// Write the trace sequence to file
Status TraceAnalyzer::WriteTraceSequence(const uint32_t& type,
const uint32_t& cf_id,
const Slice& key,
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
const size_t value_size,
const uint64_t ts) {
std::string hex_key =
ROCKSDB_NAMESPACE::LDBCommand::StringToHex(key.ToString());
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
int ret;
ret = snprintf(buffer_, sizeof(buffer_), "%u %u %zu %" PRIu64 "\n", type,
cf_id, value_size, ts);
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
if (ret < 0) {
return Status::IOError("failed to format the output");
}
std::string printout(buffer_);
if (!FLAGS_no_key) {
printout = hex_key + " " + printout;
}
return trace_sequence_f_->Append(printout);
}
// The entrance function of Trace_Analyzer
int trace_analyzer_tool(int argc, char** argv) {
std::string trace_path;
std::string output_path;
AnalyzerOptions analyzer_opts;
ParseCommandLineFlags(&argc, &argv, true);
if (!FLAGS_print_correlation.empty()) {
analyzer_opts.SparseCorrelationInput(FLAGS_print_correlation);
}
std::unique_ptr<TraceAnalyzer> analyzer(
new TraceAnalyzer(FLAGS_trace_path, FLAGS_output_dir, analyzer_opts));
if (!analyzer) {
fprintf(stderr, "Cannot initiate the trace analyzer\n");
exit(1);
}
ROCKSDB_NAMESPACE::Status s = analyzer->PrepareProcessing();
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
if (!s.ok()) {
fprintf(stderr, "%s\n", s.getState());
fprintf(stderr, "Cannot initiate the trace reader\n");
exit(1);
}
s = analyzer->StartProcessing();
if (!s.ok() && !FLAGS_try_process_corrupted_trace) {
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
fprintf(stderr, "%s\n", s.getState());
fprintf(stderr, "Cannot process the trace\n");
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
exit(1);
}
s = analyzer->MakeStatistics();
if (!s.ok()) {
fprintf(stderr, "%s\n", s.getState());
analyzer->EndProcessing();
fprintf(stderr, "Cannot make the statistics\n");
exit(1);
}
s = analyzer->ReProcessing();
if (!s.ok()) {
fprintf(stderr, "%s\n", s.getState());
fprintf(stderr, "Cannot re-process the trace for more statistics\n");
analyzer->EndProcessing();
exit(1);
}
s = analyzer->EndProcessing();
if (!s.ok()) {
fprintf(stderr, "%s\n", s.getState());
fprintf(stderr, "Cannot close the trace analyzer\n");
exit(1);
}
return 0;
}
} // namespace ROCKSDB_NAMESPACE
RocksDB Trace Analyzer (#4091) Summary: A framework of trace analyzing for RocksDB After collecting the trace by using the tool of [PR #3837](https://github.com/facebook/rocksdb/pull/3837). User can use the Trace Analyzer to interpret, analyze, and characterize the collected workload. **Input:** 1. trace file 2. Whole keys space file **Statistics:** 1. Access count of each operation (Get, Put, Delete, SingleDelete, DeleteRange, Merge) in each column family. 2. Key hotness (access count) of each one 3. Key space separation based on given prefix 4. Key size distribution 5. Value size distribution if appliable 6. Top K accessed keys 7. QPS statistics including the average QPS and peak QPS 8. Top K accessed prefix 9. The query correlation analyzing, output the number of X after Y and the corresponding average time intervals **Output:** 1. key access heat map (either in the accessed key space or whole key space) 2. trace sequence file (interpret the raw trace file to line base text file for future use) 3. Time serial (The key space ID and its access time) 4. Key access count distritbution 5. Key size distribution 6. Value size distribution (in each intervals) 7. whole key space separation by the prefix 8. Accessed key space separation by the prefix 9. QPS of each operation and each column family 10. Top K QPS and their accessed prefix range **Test:** 1. Added the unit test of analyzing Get, Put, Delete, SingleDelete, DeleteRange, Merge 2. Generated the trace and analyze the trace **Implemented but not tested (due to the limitation of trace_replay):** 1. Analyzing Iterator, supporting Seek() and SeekForPrev() analyzing 2. Analyzing the number of Key found by Get **Future Work:** 1. Support execution time analyzing of each requests 2. Support cache hit situation and block read situation of Get Pull Request resolved: https://github.com/facebook/rocksdb/pull/4091 Differential Revision: D9256157 Pulled By: zhichao-cao fbshipit-source-id: f0ceacb7eedbc43a3eee6e85b76087d7832a8fe6
6 years ago
#endif // Endif of Gflag
#endif // RocksDB LITE