bench.pl - Compare the performance of perl code snippets across multiple perls.
# Basic: run the tests in t/perf/benchmarks against two or # more perls bench.pl [options] perlA[=labelA] perlB[=labelB] ... # run the tests against the same perl twice, with varying options bench.pl [options] perlA=bigint --args='-Mbigint' perlA=plain # Run bench on blead, saving results to file; then modify the blead # binary, and benchmark again, comparing against the saved results bench.pl [options] --write=blead.time ./perl=blead # ... hack hack hack, updating ./perl ... bench.pl --read=blead.time ./perl=hacked # You can also combine --read with --write and new benchmark runs bench.pl --read=blead.time --write=last.time -- ./perl=hacked
By default, bench.pl will run code snippets found in t/perf/benchmarks (or similar) under cachegrind, in order to calculate how many instruction reads, data writes, branches, cache misses, etc. that one execution of the snippet uses. Usually it will run them against two or more perl executables and show how much each test has gotten better or worse.
It is modelled on the perlbench tool, but since it measures instruction reads etc., rather than timings, it is much more precise and reproducible. It is also considerably faster, and is capable of running tests in parallel (with
-j). Rather than displaying a single relative percentage per test/perl combination, it displays values for 13 different measurements, such as instruction reads, conditional branch misses etc.
There are options to write the raw data to a file, and to read it back. This means that you can view the same run data in different views with different selection and sort options. You can also use this mechanism to save the results of timing one perl, and then read it back while timing a modification, so that you don't have rerun the same tests on the same perl over and over, or have two perl executables built at the same time.
=label after each perl executable is used in the display output. If you are doing a two step benchmark then you should provide a label for at least the "base" perl. If a label isn't specified, it defaults to the name of the perl executable. Labels must be unique across all current executables, plus any previous ones obtained via --read.
In its most general form, the specification of a perl executable is:
path/perl=+mylabel --args='-foo -bar' --args='-baz' \ --env='A=a' --env='B=b'
This defines how to run the executable path/perl. It has a label, which due to the
+, is appended to the binary name to give a label of
path/perl=+mylabel (without the
+, the label would be just
It can be optionally followed by one or more
--env switches, which specify extra command line arguments or environment variables to use when invoking that executable. Each
--env switch should be of the form
--arg values are concatenated to the eventual command line, along with the global
--perlargs value if any. The above would cause a system() call looking something like:
PERL_HASH_SEED=0 A=a B=b valgrind --tool=cachegrind \ path/perl -foo -bar -baz ....
What action to perform. The default is grind, which runs the benchmarks using cachegrind as the back end. The only other action at the moment is selftest, which runs some basic sanity checks and produces TAP output.
Enable debugging output.
Display basic usage information.
Display progress information.
Specify a subset of tests to run (or in the case of
--read, to read). It may be either a comma-separated list of test names, or a regular expression. For example
-r file --read=file
Read in saved data from a previous
--writerun from the specified file. If
--testsis present too, then only tests matching those conditions are read from the file.
--readmay be specified multiple times, in which case the results across all files are aggregated. The list of test names from each file (after filtering by
--tests) must be identical across all files.
This list of tests is used instead of that obtained from the normal benchmark file (or
--benchfile) for any benchmarks that are run.
The perl labels must be unique across all read in test results.
JSON::PPto be available.
Benchmarks will be run for all perls specified on the command line. These options can be used to modify the benchmarking behavior:
Generate a unique label for every executable which doesn't have an explicit
=label. Works by stripping out common prefixes and suffixes from the executable names, then for any non-unique names, appending
-1, etc. text directly surrounding the unique part which look like version numbers (i.e. which match
/[0-9\.]+/) aren't stripped. For example,
perl-5.20.0-threaded perl-5.22.0-threaded perl-5.24.0-threaded
stripped to unique parts would be:
20 22 24
but is actually only stripped down to:
5.20.0 5.22.0 5.24.0
If the final results are plain integers, they are prefixed with "p" to avoid looking like column numbers to switches like
The path of the file which contains the benchmarks (t/perf/benchmarks by default).
Optional command-line arguments to pass to all cachegrind invocations.
-j N --jobs=N
Run N jobs in parallel (default 1). This determines how many cachegrind process will run at a time, and should generally be set to the number of CPUs available.
Optional command-line arguments to pass to every perl executable. This may optionaly be combined with
--argsswitches following individual perls. For example:
bench.pl --perlargs='-Ilib -It/lib' .... \ perlA --args='-Mstrict' \ perlB --args='-Mwarnings'
would cause the invocations
perlA -Ilib -It/lib -Mstrict perlB -Ilib -It/lib -Mwarnings
Any results accumulated via --read or by running benchmarks can be output in any or all of these three ways:
-w file --write=file
Save the raw data to the specified file. It can be read back later with
--read. If combined with
--readthen the output file will be the merge of the file read and any additional perls added on the command line.
JSON::PPto be available.
Exit with a zero status if the named field is in the specified range; exit with 1 otherwise. It will complain if more than one test or perl has been specified. It is intended to be called as part of a bisect run, to determine when something changed. For example,
bench.pl -j 8 --tests=foo --bisect=Ir,100,105 --perlargs=-Ilib \ ./miniperl
might be called from bisect to find when the number of instruction reads for test foo falls outside the range 100..105.
Display the results to stdout in human-readable form. This is enabled by default, except with --write and --bisect. The following sub-options alter how --show behaves.
Only display the overall average, rather than the results for each individual test.
Display the results for a single perl executable in a compact form. Which perl to display is specified in the same manner as
Display only the specified fields; for example,
If only one field is selected, the output is in more compact form.
Specify which perl column in the output to treat as the 100% norm. It may be:
* a column number (0..N-1),
* a negative column number (-1..-N) which counts from the right (so -1 is the right-most column),
* or a perl executable name,
* or a perl executable label.
It defaults to the leftmost column.
Display raw data counts rather than percentages in the outputs. This allows you to see the exact number of intruction reads, branch misses etc. for each test/perl combination. It also causes the
AVERAGEdisplay per field to be calculated based on the average of each tests's count rather than average of each percentage. This means that tests with very high counts will dominate.
Order the tests in the output based on the value of field in the column perl. The perl value is as per
--norm. For example
bench.pl --sort=Dw:perl-5.20.0 \ perl-5.16.0 perl-5.18.0 perl-5.20.0