Core-local statistics

[ajkr]

This diff changes StatisticsImpl from a thread-local approach to a core-local one. The goal is to perform faster aggregations, particularly for applications that have many threads. There should be no behavior change.

Test Plan:

• sysbench with 4k concurrent transaction threads: stats aggregations 6.3x faster compared to thread-local (thread-local: 28ms, core-local: 4.4ms)
• db_bench readwhilewriting, 32 threads, 100M keys, on SSD: stats updates 1.2x slower compared to thread-local

thread-local:

readwhilewriting : 3.072 micros/op 325566 ops/sec; 22.8 MB/s (631365 of 1000000 found)

+ 0.96% 0.11% db_bench.master db_bench.master [.] rocksdb::StatisticsImpl::measureTime
+ 0.89% 0.32% db_bench.master db_bench.master [.] rocksdb::StatisticsImpl::recordTick

core-local:

readwhilewriting : 3.050 micros/op 327826 ops/sec; 22.9 MB/s (631365 of 1000000 found)

+ 1.14% 0.38% db_bench.core-l db_bench.core-local.getcpu.bitshift [.] rocksdb::StatisticsImpl::recordTick
+ 1.09% 0.17% db_bench.core-l db_bench.core-local.getcpu.bitshift [.] rocksdb::StatisticsImpl::measureTime

• db_bench readrandom, 32 threads, 100M keys, in-memory: stats updates 1.2x slower compared to thread-local

thread-local:

+ 0.88% 0.07% db_bench.master db_bench.master [.] rocksdb::StatisticsImpl::measureTime
0.55% 0.18% db_bench.master db_bench.master [.] rocksdb::StatisticsImpl::recordTick

readrandom : 2.049 micros/op 487965 ops/sec; 34.7 MB/s (1000000 of 1000000 found)

core-local:

+ 0.97% 0.13% db_bench.core-l db_bench.core-local.getcpu.bitshift [.] rocksdb::StatisticsImpl::measureTime
+ 0.74% 0.24% db_bench.core-l db_bench.core-local.getcpu.bitshift [.] rocksdb::StatisticsImpl::recordTick

readrandom : 2.086 micros/op 479325 ops/sec; 34.1 MB/s (1000000 of 1000000 found)

[facebook-github-bot]

@ajkr updated the pull request - view changes

[facebook-github-bot]

@ajkr has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

[facebook-github-bot]

@ajkr updated the pull request - view changes - changes since last import

[facebook-github-bot]

@ajkr has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

[facebook-github-bot]

@ajkr updated the pull request - view changes - changes since last import

[facebook-github-bot]

@ajkr has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

[siying]

CLANG Analyze failure with:

internal_repo_rocksdb/repo/monitoring/statistics.h:71:9: error: no matching function for call to 'atomic_init'
std::atomic_init(&tickers_[i], static_cast<uint64_t>(0));
^~~~~~~~~~~~~~~~

[facebook-github-bot]

@ajkr updated the pull request - view changes - changes since last import

[ajkr]

Thanks, @siying, I was confused about the array initialization but fixed it now.

[facebook-github-bot]

@ajkr has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

[ajkr]

can somebody review in time for the June release?

[siying]

Please make sure it works in Windows. Looks like it is failing.

[facebook-github-bot]

@ajkr updated the pull request - view changes - changes since last import

[facebook-github-bot]

@ajkr has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.