41fb118c5c
One of the big ones, so doing this alone. Bug: T166010 Change-Id: I4848e903b0c0a2f7c58ac19e0baeef0085a36bdc |
||
|---|---|---|
| .. | ||
| data | ||
| benchmarkCommentFormatter.php | ||
| benchmarkEval.php | ||
| benchmarkHooks.php | ||
| benchmarkJSMinPlus.php | ||
| benchmarkJsonCodec.php | ||
| benchmarkLruHash.php | ||
| benchmarkParse.php | ||
| benchmarkPurge.php | ||
| benchmarkSanitizer.php | ||
| benchmarkSettings.php | ||
| benchmarkTidy.php | ||
| benchmarkTitleValue.php | ||
| benchmarkTruncate.php | ||
| README.md | ||
This directory hold several benchmarking scripts used track performances of MediaWiki and/or PHP.
Consistency
To gain greater precision than the time elapsed as reported by the benchmark
itself, one can use the perf_events tool on Linux to count the number of
CPU instructions, in addition to measuring how long it took to execute them.
This should accurately tell you how much machine code is executed. Note that this does not correctly model the cost of each instruction (especially memory access).
For example:
$ perf stat -e instructions php -d opcache.enable_cli=1 \
maintenance/benchmarks/benchmarkEval.php --code="Html::openElement( 'a', [ 'class' => 'foo' ] )" \
--inner=1000 --count=10000
eval:
count: 10000
rate: 590.1/s
total: 16946.61ms
mean: 1.69ms
...
Performance counter stats for 'php maintenance/benchmarks/..':
83651088078 instructions
17.225255198 seconds time elapsed
...
Fixtures
- data/tidy/australia-untidy.html.gz: Representative input text for benchmarkTidy.php. It needs to be decompressed before use.
- data/CommentFormatter/rc100-2021-07-29.json: Input for Linker::formatComment() from https://en.wikipedia.org/w/api.php?action=query&format=json&list=recentchanges&rcprop=title%7Ccomment&rclimit=100