Profiling library







  • Switched to ppx.

  • Minor adjustments to the command line of profiler_tool.exe:

    • Make '-%' an alias for '-percentile'

    • Make '-percentile' accept a comma-separated list of numbers

    • Add '-median' argument that is equivalent to '-percentile 50'


  • Changed delta timers and probes so they record the total amount of time/value
    change between each start and pause.


Initial release


Initial release


  • Solved a problem in which OCaml 4.02 was optimizing away benchmarks,
    making them meaningless.


  • fixed legacy format string


  • Added support for saving inline benchmark measurements to tabular
    files for easy loading into Octave.


  • Improved bench.mli's generated docs and added some usage examples.

    This also partly satisfies issue #3.

  • Added the ability to create groups of benchmarks with a common prefix.

    For example, the prefix "Perf" below is created in created using

    let command = Bench.make_command [
      Bench.Test.create ~name:"" (fun () ->
        ignore ( ()));
      Bench.Test.create_group ~name:"Perf" [
        Bench.Test.create ~name:"" ...

    and the output shows:

    Estimated testing time 7s (7 benchmarks x 1s). Change using -quota SECS.
    │ Name                                      │ Time/Run │ mWd/Run │ Percentage │
    │                                  │  41.38ns │   2.00w │     16.72% │
    │ Calibrator.calibrate                      │ 247.42ns │  32.00w │    100.00% │
    │ Perf/                              │   7.84ns │         │      3.17% │
    │ Perf/TSC.to_time                          │   9.35ns │   2.00w │      3.78% │
    │ Perf/TSC.to_time ( ())             │  13.22ns │   2.00w │      5.34% │
    │ Perf/TSC.to_nanos_since_epoch             │  10.83ns │         │      4.38% │
    │ Perf/TSC.to_nanos_since_epoch( ()) │  14.86ns │         │      6.00% │


  • Fixed a bug in Core_bench where the linear regression was
    sometimes supplied with spurious data.

    This showed up when doing custom regressions that allow for a non-zero


  • Exposed an extensible form of make_command so that
    inline-benchmarking and the other tools can add more commandline

  • A significant rewrite of Core_bench.

    The rewrite provides largely the same functionality as the older
    version. The most visible external change is that the API makes it
    clear that Core_bench performs linear regressions to come up with
    its numbers. Further, it allows running user-specified multivariate
    regressions in addition to the built in ones.

    The underlying code has been cleaned up in many ways, some of which
    are aimed at improving the implementation of inline benchmarking
    (the BENCH syntax, which has not yet been released).


  • Columns that have a + prefix are now always displayed, whereas
    columns that don't are displayed only if they have meaningful data.

  • Added the ability to reload saved metrics (benchmark test data) so
    that bench can re-analyze them.


  • Added support for additional predictors like minor/major GCs and
    compactions, using multi-variable linear regression.

    Replaced linear regression with multi-variable linear regression.
    The original algorithm estimated the cost of a function f by using
    a linear regression of the time taken to run f vs the number of
    runs. The new version adds the ability to include additional
    predictors such as minor GCs, compactions etc.

    This allows a more fine-grained split-up of the running costs of a
    function, distinguishing between the time spent actually running f
    and the time spent doing minor GCs, major GCs or compactions.

  • Added a forking option that allows benchmarks to be run in separate

    This avoids any influence (e.g. polluting the cache, size of live
    heap words) they might otherwise have on each other.


  • Changed -save to output compaction information.

  • Added indexed tests.

    These are benchmarks of the form int -> unit -> unit, which can be
    profiled for a list of user specified ints.


  • Report compaction stats


  • Added R^2 error estimation.

    Adding this metric should give us a sense of how closely the given
    values fit a line. Even dots that are fairly scattered can give
    tight confidence intervals. We would like to have to number to have
    a sense of how much noise we have.