The London Perl and Raku Workshop takes place on 26th Oct 2024. If your company depends on Perl, please consider sponsoring and/or attending.

NAME

Bencher - A benchmark framework

VERSION

This document describes version 1.062.4 of Bencher (from Perl distribution Bencher), released on 2024-02-19.

SYNOPSIS

See bencher CLI.

DESCRIPTION

Bencher is a benchmark framework. The main feature of Bencher is permuting list of Perl codes with list of arguments into benchmark items, and then benchmark them. You can run only some of the items as well as filter codes and arguments to use. You can also permute multiple perls and multiple module versions.

TERMINOLOGY

Scenario. A hash data structure that lists participants, datasets, and other things. The bencher CLI can accept a scenario from a module (under Bencher::Scenario::* namespace or Acme::CPANModules::* namespace), a script, or from command-line option. See "SCENARIO".

Participant. What to run or benchmark. Usually a Perl code or code template, or a command or command template. See "participants".

Dataset. Arguments or parameters to permute with a participant. See "datasets".

(Benchmark) item. Participant that has been permuted with dataset into code ready to run. Usually a scenario does not contain items directly, but only participants and datasets, and lets Bencher permute these into items.

SCENARIO

The core data structure that you need to prepare is the scenario. It is a DefHash (i.e. just a regular Perl hash). The two most important keys of this hash are: participants and datasets.

An example scenario (from Bench::Scenario::Example):

package Bencher::Scenario::Example;
our $scenario = {
    participants => [
        {fcall_template => q[Text::Wrap::wrap('', '', <text>)]},
    ],
    datasets => [
        { name=>"foobar x100",   args => {text=>"foobar " x 100} },
        { name=>"foobar x1000",  args => {text=>"foobar " x 1000} },
        { name=>"foobar x10000", args => {text=>"foobar " x 10000} },
    ],
};
1;

participants

participants (array) lists Perl codes (or external commands) that we want to benchmark.

Specifying participant's code

There are several kinds of code you can specify:

  • module

    First, you can just specify module (str, a Perl module name). This is useful when running scenario in "Running benchmark in module startup mode" in module_startup mode. Also useful to instruct Bencher to load the module. When not in module startup mode, there is no code in this participant to run. In addition to module you can also specify import_args which can be a string or an arrayref.

  • modules

    You can also specify modules (an array of Perl module names) if you want to benchmark several modules together. Similarly, this is only useful for running in module startup mode. To specify import arguments for each of these modules, use import_args_array.

  • code

    You can specify code (a coderef) which contains the code to benchmark. However, the point of Bencher is to use fcall_template or at least code_template to be able to easily permute the code with datasets (see below). So you should only specify code when you cannot specify fcall_template or code_template or the other way.

    Caveat: if you run bencher in multiperl or multi-module-version mode, the current implementaiton will fill the code templates and dump the scenario into a temporary script and run that script with separate perl processes. The dumped coderef will lack 'use' statement or code in BEGIN block. Avoid using that in your code.

  • fcall_template

    You can specify fcall_template, and this is the recommended way whenever possible. It is a string containing a function call code, in the form of:

    MODULENAME::FUNCTIONAME(ARG, ...)

    or

    CLASSNAME->FUNCTIONAME(ARG, ...)

    For example:

    Text::Levenshtein::fastdistance(<word1>, <word2>)

    Another example:

    Module::CoreList->is_code(<module>)

    It can be used to benchmark a function call or a method call. From this format, Bencher can easily extract the module name so user can also run in module startup mode.

    By using a template, Bencher can generate actual codes from this template by combining it with datasets. The words enclosed in <...> will be replaced with actual arguments specified in "datasets". The arguments are automatically encoded as Perl value, so it's safe to use arrayref or complex structures as argument values (however, you can use <...:raw> to avoid this automatic encoding).

  • code_template

    Aside from fcall_template, you can also use code_template (a string containing arbitrary code), in the cases where the code you want to benchmark cannot be expressed as a simple function/method call, for example (taken from Bencher::Scenario::ComparisonOps):

    participants => [
        {name=>'1k-numeq'      , code_template=>'my $val =     1; for (1..1000) { if ($val ==     1) {} if ($val ==     2) {} }'},
        {name=>'1k-streq-len1' , code_template=>'my $val = "a"  ; for (1..1000) { if ($val eq "a"  ) {} if ($val eq "b"  ) {} }'},
        {name=>'1k-streq-len3' , code_template=>'my $val = "foo"; for (1..1000) { if ($val eq "foo") {} if ($val eq "bar") {} }'},
        {name=>'1k-streq-len10', code_template=>'my $val = "abcdefghij"; for (1..1000) { if ($val eq "abcdefghij") {} if ($val eq "klmnopqrst") {} }'},
    ],

    Like in fcall_template, words enclosed in <...> will be replaced with actual data. When generating actual code, Bencher will enclose the code template with sub { .. }.

    Caveat: if you run bencher in multiperl or multi-module-version mode, the current implementaiton will fill the code templates and dump the scenario into a temporary script and run that script with separate perl processes. The dumped coderef will lack 'use' statement or code in BEGIN block. Avoid using that in your code.

  • cmdline, cmdline_template, perl_cmdline, perl_cmdline_templates

    Or, if you are benchmarking external commands, you specify cmdline (array or strings, or strings) or cmdline_template (array/str) or perl_cmdline or perl_cmdline_template instead. An array cmdline will not use shell, while the string version will use shell. perl_cmdline* are the same as cmdline* except the first implicit argument/prefix is perl. When the cmdline template is filled with the arguments, the values will automatically be shell-escaped (unless you use the <...:raw> syntax).

    When using code template, code will be generated and eval-ed in the main package.

Specifying participant's name

By default, Bencher will attempt to figure out the name for a participant (a sequence number starting from 1, a module name or module name followed by function name, etc). You can also specify name for a participant explicitly so you can refer to it more easily later, e.g.:

participants => [
    {name=>'pp', fcall_template=>'List::MoreUtils::PP::uniq(@{<array>})'},
    {name=>'xs', fcall_template=>'List::MoreUtils::XS::uniq(@{<array>})'},
],

List of properties for a participant

This is a reference section.

  • name (str)

    From DefHash.

  • summary (str)

    From DefHash.

  • description (str)

    From DefHash.

  • tags (array of str)

    From DefHash. Define tag(s) for this participant. Can be used to include/exclude groups of participants having the same tags.

  • module (str)

  • modules (array of str)

    These are modules that are to be benchmarked. When running the benchmark scenario, these modules will be require'd first, so you don't have to require them again manually in your code/code template.

  • helper_modules (array of str)

    These are helper modules that are required when running the participant code, but are not the main subject to be benchmarked. They will also be require'd when running the benchmark scenario so you don't have to require them manually in your code/code template.

    The difference with the modules property is: the modules specified in modules should be specified as phase=x_benchmarks, rel=x_benchmarks prerequisites while modules specified in helper_modules should not. But all modules can be specified as phase=x_benchmarks, rel=requires prerequisites.

  • function (str)

  • fcall_template (str)

  • code_template (str)

  • code (code)

  • cmdline (str|array of str)

  • cmdline_template (str|array of str)

  • perl_cmdline (str|array of str)

  • perl_cmdline_template (str|array of str)

  • result_is_list (bool, default 0)

    This is useful when dumping item's codes, so Bencher will use a list context when receiving result.

  • include_by_default (bool, default 1)

    Can be set to false if you want to exclude participant by default when running benchmark, unless the participant is explicitly included e.g. using --include-participant command-line option.

datasets

datasets (array) lists the function inputs (or command-line arguments). You can name each dataset too, to be able to refer to it more easily.

Other properties you can add to a dataset: include_by_default (bool, default true, can be set to false if you want to exclude dataset by default when running benchmark, unless the dataset is explicitly included).

  • name (str)

    From DefHash.

  • summary (str)

    From DefHash.

  • description (str)

    From DefHash.

  • tags (array of str)

    From DefHash. Define tag(s) for this dataset. Can be used to include/exclude groups of datasets having the same tags.

  • args (hash)

    Example:

    {filename=>"ujang.txt", size=>10}

    You can supply multiple argument values by adding @ suffix to the argument name. You then supply an array for the values, example:

    {filename=>"ujang.txt", 'size@'=>[10, 100, 1000]}

    This means, for each participant mentioning size, three benchmark items will be generated, one for each value of size.

    Aside from array, you can also use hash for the multiple values. This has a nice effect of showing nicer names (in the hash keys) for the argument value, e.g.:

    {filename=>"ujang.txt", 'size@'=>{"1M"=>1024*2, "1G"=>1024**3, "1T"=>1024**4}}
  • argv (array)

  • include_by_default (bool, default 1)

  • include_participant_tags (array of str)

    Only include participants having one of these tags. For example:

    ['a', 'b']

    will include all participants having either a or b in their tags. To only include participants which have all of a and b in their tags, use:

    ['a & b']
  • exclude_participant_tags (array of str)

    Exclude participants having any of these tags. For example:

    ['a', 'b']

    will exclude all participants having either a or b in their tags. To only exclude participants which have all of a and b in their tags, use:

    ['a & b']
  • result (any)

    Specify result that any participant must return. This can be used to verify that participant code is correct (returns the desired result) before benchmarking. If a participant fails to return this result, benchmarking will be aborted.

Other properties

Other known scenario properties (keys):

  • name

    From DefHash, scenario name (usually short and one word).

  • summary

    From DefHash, a one-line plaintext summary.

  • description (str)

    From DefHash, longer description in Markdown.

  • test (bool, default 1)

    By default, participant code is run once first for testing (e.g. whether it dies or return the correct result) before benchmarking. If your code runs for many seconds, you might want to skip this test and set this to 0.

  • module_startup (bool)

  • code_startup (bool)

  • precision (float)

    Precision to pass to Benchmark::Dumb. Default is 0. Can be overriden via --precision (CLI). takes precedence over default_precision.

  • default_precision (float)

    (DEPRECATED)

    Precision to pass to Benchmark::Dumb. Default is 0. Can be overriden via --precision (CLI).

  • module_startup_precision (float)

    Precision to pass to Benchmark::Dumb, when running benchmark in module_startup mode. Default is from precision or default_precision. Can be overriden with --precision (CLI).

  • result (any)

    Like per-dataset result property, which you normally should use instead of this. But this property can be used when a scenario does not have any dataset.

  • with_args_size (bool)

    Show the size of the item code's arguments. Size is measured using Devel::Size. The measurement is done once per item when it is testing the code. See also: with_result_size.

  • with_result_size (bool)

    Show the size of the item code's return value. Size is measured using Devel::Size. The measurement is done once per item when it is testing the code. See also: with_args_size.

  • with_process_size (bool)

    Include some memory statistics in each item's result. This currently only works on Linux because the measurement is done by reading /proc/PID/smaps. Also, since this is a per-process information, to get this information each item's code will be run by dumping the code (using B::Deparse) into a temporary file, then running the file (once per item, after the item's code is completed) using a new perl interpreter process. This is done to get a measurement on a clean process that does not load Bencher itself or the other items. This also means that not all code will work: all the caveats in "MULTIPLE PERLS AND MULTIPLE MODULE VERSIONS" apply. In short, all outside data will not be available for the code.

    Also, this information normally does not make sense for external command participants, because what is measured is the memory statistics of the perl process itself, not the external command's processes.

  • capture_stdout (bool)

    Useful for silencing command/code that outputs stuffs to stdout. Note that output capturing might affect timings if your benchmark code outputs a lot of stuffs. See also: capture_stderr.

  • capture_stderr (bool)

    Useful for silencing command/code that outputs stuffs to stderr. Note that output capturing might affect timings if your benchmark code outputs a lot of stuffs. See also: capture_stdout.

  • extra_modules (array of str)

    You can specify extra modules to load here before benchmarking. The modules and their versions will be listed in the result metadata under func.module_versions, for extra information. An example to put here are modules that contain/produce datasets that get benchmarked, because the data might differ from version to version.

  • env_hashes (array of hash)

    With this property, you can permute multiple sets of environment variables. Suppose you want to benchmark each participant when running under environment variables FOO=0, FOO=1, and FOO=2. You can specify:

    env_hashes => [
        {FOO=>0},
        {FOO=>1},
        {FOO=>2},
    ]
  • runner (str)

    Set which runner to run the benchmark with. Either Benchmark::Dumb (the default), Benchmark (Benchmark.pm, to get Benchmark.pm-style output), or Benchmark::Dumb::SimpleTime (to be able to run a code with only 1 or 2 iterations without warning, suitable if your code that runs for at least a few seconds and you don't want to wait for too long).

  • on_failure (str, "skip"|"die")

    For a command participant, failure means non-zero exit code. For a Perl-code participant, failure means Perl code dies or (if expected result is specified) the result is not equal to the expected result.

    The default is "die". When set to "skip", will first run the code of each item before benchmarking and trap command failure/Perl exception and if that happens, will "skip" the item.

    Can be overriden in the CLI with --on-failure option.

  • on_result_failure (str, "skip"|"die"|"warn")

    This is like on_failure except that it specifically refer to the failure of item's result not being equal to expected result.

    The default is the value of on_failure.

    There is an extra choice of `warn` for this type of failure, which is to print a warning to STDERR and continue.

    Can be overriden in the CLI with --on-result-failure option.

  • before_parse_scenario (code)

    If specified, then this code will be called before parsing scenario. Code will be given hash argument with the following keys: hook_name (str, set to before_gen_items), scenario (hash, unparsed scenario), stash (hash, which you can use to pass data between hooks).

  • before_parse_participants (code)

    If specified, then this code will be called before parsing/normalizing participants from scenario. Code will be given hash argument with the following keys: hook_name, scenario, stash.

    You can use this hook to, e.g.: generate participants dynamically.

  • before_parse_datasets (code)

    If specified, then this code will be called before parsing/normalizing datasets from scenario. Code will be given hash argument with the following keys: hook_name, scenario, stash.

    You can use this hook to, e.g.: generate datasets dynamically.

  • after_parse_scenario (code)

    If specified, then this code will be called after parsing scenario. Code will be given hash argument with the following keys: hook_name, scenario (hash, parsed scenario), stash.

  • before_gen_items (code)

    If specified, then this code will be called before generating items. Code will be given hash argument with the following keys: hook_name, scenario, stash.

    You can use this hook to, e.g.: modify datasets/participants before being permuted into items.

  • before_bench (code)

    If specified, then this code will be called before starting the benchmark. Code will be given hash argument with the following keys: hook_name, scenario, stash.

  • before_test_item (code)

    If specified, then this code will be called before testing each item's code. Code will be given hash argument with the following keys: hook_name, scenario, stash, item (itself a hash containing the following keys: _name, seq, permute).

  • after_test_item (code)

    If specified, then this code will be called after testing each item's code. Code will be given hash argument with the following keys: hook_name, scenario, stash, item (itself a hash containing the following keys: _name, seq, permute, _result).

    Item's code result can be accessed in _result in item hash.

    The hook is expected to return false (0 or '') or a string which will be interpreted as error message. Thus, this hook can be used to force error during an item's test. The hook can be used to check that the item's code has executed correctly by checking other side effects aside from the return in _result.

  • after_bench (code)

    If specified, then this code will be called after completing benchmark. Code will be given hash argument with the following keys: hook_name, scenario, stash, result (array, enveloped result).

    You can use this hook to, e.g.: do some custom formatting/modification to the result.

  • before_return (code)

    If specified, then this code will be called before displaying/returning the result. Code will be given hash argument with the following keys: hook_name, scenario, stash, result.

    You can use this hook to, e.g.: modify the result in some way.

USING THE BENCHER COMMAND-LINE TOOL

Running benchmark

Running benchmark in code startup mode

Code startup mode can be activated either by specifying --code-startup option from the command-line, or by setting code_startup property to true in the scenario.

In this mode, instead of running participant's perl code, it runs a perl command:

perl -e 'YOUR_PERL_CODE'

or if there are modules to be loaded:

perl -MModule1 -MModule2 ... -e 'YOUR_PERL_CODE'

and then comparing the time with the baseline:

perl -e1

and measure the startup overhead of each code.

See also module_startup mode.

Running benchmark in module startup mode

Module startup mode can be activated either by specifying --module-startup option from the command-line, or by setting module_startup property to true in the scenario.

In this mode, instead of running each participant's code, module name will be extracted from each participant and this will be benchmarked instead:

perl -MModule1 -e1
perl -MModule2 -e1
...
perl -e1 ;# the baseline, for comparison

Basically, this mode tries to measure the startup overhead of each module in isolation.

Module name can be extracted from a participant if a participant specifies module or fcall_template or modules. When a participant does not contain any module name, it will be skipped.

MULTIPLE PERLS AND MULTIPLE MODULE VERSIONS

Bencher can be instructed to run benchmark items against multiple perl installations, as well as multiple versions of a module.

Bencher uses perlbrew to get the list of available perl installations, so you need to install perlbrew and brew some perls first.

To run against multiple versions of a module, specify the module name in --multimodver then add one or more library include paths using -I. The include paths need to contain different versions of the module.

Caveats. Here is how benchmarking against multiple perls and module versions currently works. Bencher first prepares a new scenario based on the input scenario. But the new scenario contains benchmark items that has been permuted and where the code template has been converted into actual Perl code (a coderef). The new scenario along with the Perl codes in it will be dumped using Data::Dmp (which can deparse code) into a temporary file. A new Bencher process is then started using the appropriate perl interpreter, runs the scenario, and returns the result as JSON. The original Bencher process then collects and combines the per-interpreter results into the final result.

Due to the above way of working, there are some caveats. First, code that contains closures won't work properly because the original variables that the code can see are no longer available in the new process. Also, some scenarios prepare data in a hook like in the before_bench or before_gen_items hook. This also won't work because the new scenario that gets dumped into temporary file currently has all the hooks stripped first.

So in principle, to enable a benchmark item to be run against multiple perls or module versions, make the code self-sufficient. Do not depend on an outside variable. Instead, only depend on the variables in the dataset.

HOMEPAGE

Please visit the project's homepage at https://metacpan.org/release/Bencher.

SOURCE

Source repository is at https://github.com/perlancar/perl-Bencher.

SEE ALSO

bencher

Bencher::Manual::*

BenchmarkAnything. There are lot of overlaps of goals between Bencher and this project. I hope to reuse or interoperate parts of BenchmarkAnything, e.g. storing Bencher results in a BenchmarkAnything storage backend, sending Bencher results to a BenchmarkAnything HTTP server, and so on.

Benchmark, Benchmark::Dumb (Dumbbench)

Bencher::Scenario::* for examples of scenarios.

An article about Bencher written in 2016, https://perladvent.org/2016/2016-12-03.html

AUTHOR

perlancar <perlancar@cpan.org>

CONTRIBUTING

To contribute, you can send patches by email/via RT, or send pull requests on GitHub.

Most of the time, you don't need to build the distribution yourself. You can simply modify the code, then test via:

% prove -l

If you want to build the distribution (e.g. to try to install it locally on your system), you can install Dist::Zilla, Dist::Zilla::PluginBundle::Author::PERLANCAR, Pod::Weaver::PluginBundle::Author::PERLANCAR, and sometimes one or two other Dist::Zilla- and/or Pod::Weaver plugins. Any additional steps required beyond that are considered a bug and can be reported to me.

COPYRIGHT AND LICENSE

This software is copyright (c) 2024 by perlancar <perlancar@cpan.org>.

This is free software; you can redistribute it and/or modify it under the same terms as the Perl 5 programming language system itself.

BUGS

Please report any bugs or feature requests on the bugtracker website https://rt.cpan.org/Public/Dist/Display.html?Name=Bencher

When submitting a bug or request, please include a test-file or a patch to an existing test-file that illustrates the bug or desired feature.