t/perf/README

   1 Git performance tests
   2 =====================
   3
   4 This directory holds performance testing scripts for git tools.  The
   5 first part of this document describes the various ways in which you
   6 can run them.
   7
   8 When fixing the tools or adding enhancements, you are strongly
   9 encouraged to add tests in this directory to cover what you are
  10 trying to fix or enhance.  The later part of this short document
  11 describes how your test scripts should be organized.
  12
  13
  14 Running Tests
  15 -------------
  16
  17 The easiest way to run tests is to say "make".  This runs all
  18 the tests on the current git repository.
  19
  20     === Running 2 tests in this tree ===
  21     [...]
  22     Test                                     this tree
  23     ---------------------------------------------------------
  24     0001.1: rev-list --all                   0.54(0.51+0.02)
  25     0001.2: rev-list --all --objects         6.14(5.99+0.11)
  26     7810.1: grep worktree, cheap regex       0.16(0.16+0.35)
  27     7810.2: grep worktree, expensive regex   7.90(29.75+0.37)
  28     7810.3: grep --cached, cheap regex       3.07(3.02+0.25)
  29     7810.4: grep --cached, expensive regex   9.39(30.57+0.24)
  30
  31 You can compare multiple repositories and even git revisions with the
  32 'run' script:
  33
  34     $ ./run . origin/next /path/to/git-tree p0001-rev-list.sh
  35
  36 where . stands for the current git tree.  The full invocation is
  37
  38     ./run [<revision|directory>...] [--] [<test-script>...]
  39
  40 A '.' argument is implied if you do not pass any other
  41 revisions/directories.
  42
  43 You can also manually test this or another git build tree, and then
  44 call the aggregation script to summarize the results:
  45
  46     $ ./p0001-rev-list.sh
  47     [...]
  48     $ ./run /path/to/other/git -- ./p0001-rev-list.sh
  49     [...]
  50     $ ./aggregate.perl . /path/to/other/git ./p0001-rev-list.sh
  51
  52 aggregate.perl has the same invocation as 'run', it just does not run
  53 anything beforehand.
  54
  55 You can set the following variables (also in your config.mak):
  56
  57     GIT_PERF_REPEAT_COUNT
  58         Number of times a test should be repeated for best-of-N
  59         measurements.  Defaults to 3.
  60
  61     GIT_PERF_MAKE_OPTS
  62         Options to use when automatically building a git tree for
  63         performance testing. E.g., -j6 would be useful. Passed
  64         directly to make as "make $GIT_PERF_MAKE_OPTS".
  65
  66     GIT_PERF_MAKE_COMMAND
  67         An arbitrary command that'll be run in place of the make
  68         command, if set the GIT_PERF_MAKE_OPTS variable is
  69         ignored. Useful in cases where source tree changes might
  70         require issuing a different make command to different
  71         revisions.
  72
  73         This can be (ab)used to monkeypatch or otherwise change the
  74         tree about to be built. Note that the build directory can be
  75         re-used for subsequent runs so the make command might get
  76         executed multiple times on the same tree, but don't count on
  77         any of that, that's an implementation detail that might change
  78         in the future.
  79
  80     GIT_PERF_REPO
  81     GIT_PERF_LARGE_REPO
  82         Repositories to copy for the performance tests.  The normal
  83         repo should be at least git.git size.  The large repo should
  84         probably be about linux.git size for optimal results.
  85         Both default to the git.git you are running from.
  86
  87     GIT_PERF_EXTRA
  88         Boolean to enable additional tests. Most test scripts are
  89         written to detect regressions between two versions of Git, and
  90         the output will compare timings for individual tests between
  91         those versions. Some scripts have additional tests which are not
  92         run by default, that show patterns within a single version of
  93         Git (e.g., performance of index-pack as the number of threads
  94         changes). These can be enabled with GIT_PERF_EXTRA.
  95
  96 You can also pass the options taken by ordinary git tests; the most
  97 useful one is:
  98
  99 --root=<directory>::
 100         Create "trash" directories used to store all temporary data during
 101         testing under <directory>, instead of the t/ directory.
 102         Using this option with a RAM-based filesystem (such as tmpfs)
 103         can massively speed up the test suite.
 104
 105
 106 Naming Tests
 107 ------------
 108
 109 The performance test files are named as:
 110
 111         pNNNN-commandname-details.sh
 112
 113 where N is a decimal digit.  The same conventions for choosing NNNN as
 114 for normal tests apply.
 115
 116
 117 Writing Tests
 118 -------------
 119
 120 The perf script starts much like a normal test script, except it
 121 sources perf-lib.sh:
 122
 123         #!/bin/sh
 124         #
 125         # Copyright (c) 2005 Junio C Hamano
 126         #
 127
 128         test_description='xxx performance test'
 129         . ./perf-lib.sh
 130
 131 After that you will want to use some of the following:
 132
 133         test_perf_fresh_repo    # sets up an empty repository
 134         test_perf_default_repo  # sets up a "normal" repository
 135         test_perf_large_repo    # sets up a "large" repository
 136
 137         test_perf_default_repo sub  # ditto, in a subdir "sub"
 138
 139         test_checkout_worktree  # if you need the worktree too
 140
 141 At least one of the first two is required!
 142
 143 You can use test_expect_success as usual. In both test_expect_success
 144 and in test_perf, running "git" points to the version that is being
 145 perf-tested. The $MODERN_GIT variable points to the git wrapper for the
 146 currently checked-out version (i.e., the one that matches the t/perf
 147 scripts you are running).  This is useful if your setup uses commands
 148 that only work with newer versions of git than what you might want to
 149 test (but obviously your new commands must still create a state that can
 150 be used by the older version of git you are testing).
 151
 152 For actual performance tests, use
 153
 154         test_perf 'descriptive string' '
 155                 command1 &&
 156                 command2
 157         '
 158
 159 test_perf spawns a subshell, for lack of better options.  This means
 160 that
 161
 162 * you _must_ export all variables that you need in the subshell
 163
 164 * you _must_ flag all variables that you want to persist from the
 165   subshell with 'test_export':
 166
 167         test_perf 'descriptive string' '
 168                 foo=$(git rev-parse HEAD) &&
 169                 test_export foo
 170         '
 171
 172   The so-exported variables are automatically marked for export in the
 173   shell executing the perf test.  For your convenience, test_export is
 174   the same as export in the main shell.
 175
 176   This feature relies on a bit of magic using 'set' and 'source'.
 177   While we have tried to make sure that it can cope with embedded
 178   whitespace and other special characters, it will not work with
 179   multi-line data.
 180
 181 Rather than tracking the performance by run-time as `test_perf` does, you
 182 may also track output size by using `test_size`. The stdout of the
 183 function should be a single numeric value, which will be captured and
 184 shown in the aggregated output. For example:
 185
 186         test_perf 'time foo' '
 187                 ./foo >foo.out
 188         '
 189
 190         test_size 'output size'
 191                 wc -c <foo.out
 192         '
 193
 194 might produce output like:
 195
 196         Test                origin           HEAD
 197         -------------------------------------------------------------
 198         1234.1 time foo     0.37(0.79+0.02)  0.26(0.51+0.02) -29.7%
 199         1234.2 output size             4.3M             3.6M -14.7%
 200
 201 The item being measured (and its units) is up to the test; the context
 202 and the test title should make it clear to the user whether bigger or
 203 smaller numbers are better. Unlike test_perf, the test code will only be
 204 run once, since output sizes tend to be more deterministic than timings.