.TH HYPERFINE 1
.SH NAME
hyperfine \- command\-line benchmarking tool
.SH SYNOPSIS
.B hyperfine
.RB [ \-hVNS ]
.RB [ \-\-ignore\-failure[=MODE] ]
.RB [ \-\-warmup
.IR NUM ]
.RB [ \-\-min\-runs
.IR NUM ]
.RB [ \-\-max\-runs
.IR NUM ]
.RB [ \-\-runs
.IR NUM ]
.RB [ \-\-setup
.IR CMD ]
.RB [ \-\-prepare
.IR CMD ]
.RB [ \-\-conclude
.IR CMD ]
.RB [ \-\-cleanup
.IR CMD ]
.RB [ \-\-parameter\-scan
.IR VAR
.IR MIN
.IR MAX ]
.RB [ \-\-parameter\-step\-size
.IR DELTA ]
.RB [ \-\-parameter\-list
.IR VAR
.IR VALUES ]
.RB [ \-\-shell
.IR SHELL ]
.RB [ \-\-style
.IR TYPE ]
.RB [ \-\-metrics
.IR METRIC[:UNIT],...|default|time|all ]
.RB [ \-\-export\-asciidoc
.IR FILE ]
.RB [ \-\-export\-csv
.IR FILE ]
.RB [ \-\-export\-json
.IR FILE ]
.RB [ \-\-export\-markdown
.IR FILE ]
.RB [ \-\-export\-orgmode
.IR FILE ]
.RB [ \-\-output
.IR WHERE ]
.RB [ \-\-input
.IR WHERE ]
.RB [ \-\-command\-name
.IR NAME ]
.RI [ COMMAND... ]
.SH DESCRIPTION
A command\-line benchmarking tool which includes:
.LP
.RS
* Statistical analysis across multiple runs
.RE
.RS
* Support for arbitrary shell commands
.RE
.RS
* Constant feedback about the benchmark progress and current estimates
.RE
.RS
* Warmup runs can be executed before the actual benchmark
.RE
.RS
* Cache-clearing commands can be set up before each benchmark run
.RE
.RS
* Statistical outlier detection to detect interference from other programs and caching effects
.RE
.RS
* Export results to various formats: CSV, JSON, Markdown, AsciiDoc
.RE
.RS
* Parameterized benchmarks (e.g. vary the number of threads)
.RE
.LP
Commands run and appear in input order. The first benchmark after parameter
expansion is always the reference for comparisons. Each command's results are
printed when it finishes, with one row per selected metric showing mean,
standard deviation, and range. Standard deviation is unavailable for a
single run.
Subsequent commands show the percentage change of their mean from the reference.
A positive change means more; a negative change means less. These changes do not
indicate statistical significance. A zero reference produces an undefined change
(displayed as N/A). All exports preserve input order.
.SH OPTIONS
.HP
\fB\-w\fR, \fB\-\-warmup\fR \fINUM\fP
.IP
Perform \fINUM\fP warmup runs before the actual benchmark. This can be used
to fill (disk) caches for I/O\-heavy programs.
.HP
\fB\-m\fR, \fB\-\-min\-runs\fR \fINUM\fP
.IP
Perform at least \fINUM\fP runs for each command. Default: 10.
.HP
\fB\-M\fR, \fB\-\-max\-runs\fR \fINUM\fP
.IP
Perform at most \fINUM\fP runs for each command. By default, there is no
limit.
.HP
\fB\-r\fR, \fB\-\-runs\fR \fINUM\fP
.IP
Perform exactly \fINUM\fP runs for each command. If this option is not specified,
\fBhyperfine\fR automatically determines the number of runs.
.HP
\fB\-s\fR, \fB\-\-setup\fR \fICMD...\fP
.IP
Execute \fICMD\fP once before each set of benchmark runs. This is useful
for compiling your software with the provided parameters, or to do any
other work that should happen once before a series of benchmark runs,
not every time as would happen with the \fB\-\-prepare\fR option.
.HP
\fB\-p\fR, \fB\-\-prepare\fR \fICMD...\fP
.IP
Execute \fICMD\fP before each benchmark run. This is useful for clearing disk caches,
for example.
The \fB\-\-prepare\fR option can be specified once for all commands or multiple times,
once for each command. In the latter case, each preparation command will be
run prior to the corresponding benchmark command.
.HP
\fB\-C\fR, \fB\-\-conclude\fR \fICMD...\fP
.IP
Execute \fICMD\fP after each benchmark run. This is useful for clearing disk caches,
for example.
The \fB\-\-conclude\fR option can be specified once for all commands or multiple times,
once for each command. In the latter case, each conclusion command will be
run after the corresponding benchmark command.
.HP
\fB\-c\fR, \fB\-\-cleanup\fR \fICMD...\fP
.IP
Execute \fICMD\fP after the completion of all benchmarking runs for each individual
command to be benchmarked. This is useful if the commands to be benchmarked
produce artifacts that need to be cleaned up. It only runs once per series of
benchmark runs, as opposed to the \fB\-\-conclude\fR option which runs after
every run.
.HP
\fB\-P\fR, \fB\-\-parameter\-scan\fR \fIVAR\fP \fIMIN\fP \fIMAX\fP
.IP
Perform benchmark runs for each value in the range \fIMIN..MAX\fP. Replaces the
string '{\fIVAR\fP}' in each command by the current parameter value.
.IP
.RS
Example:
.RS
\fBhyperfine\fR \fB\-\-prepare\fR 'make clean' \fB\-P\fR threads 1 8 'make \-j {threads}'
.RE
.RE
.IP
This performs benchmarks for 'make \-j 1', 'make \-j 2', ..., 'make \-j 8'.
.IP
To have the value increase following different patterns, use shell
arithmetic.
.IP
.RS
Example:
.RS
\fBhyperfine\fR \fB\-\-shell=bash\fR \fB\-P\fR size 0 3 'sleep $((2**{size}))'
.RE
.RE
.IP
This performs benchmarks with power of 2 increases: 'sleep 1', 'sleep
2', 'sleep 4', ...
.IP
The exact syntax may vary depending on your shell and OS.
.HP
\fB\-D\fR, \fB\-\-parameter\-step\-size\fR \fIDELTA\fP
.IP
This argument requires \fB\-\-parameter\-scan\fR to be specified as well. Traverse the
range \fIMIN..MAX\fP in steps of \fIDELTA\fP, which must be positive.
.IP
.RS
Example:
.RS
\fBhyperfine\fR \fB\-P\fR delay 0.3 0.7 \fB\-D\fR 0.2 'sleep {delay}'
.RE
.RE
.IP
This performs benchmarks for 'sleep 0.3', 'sleep 0.5' and 'sleep 0.7'.
.HP
\fB\-L\fR, \fB\-\-parameter\-list\fR \fIVAR\fP \fIVALUES\fP
.IP
Perform benchmark runs for each value in the comma\-separated list of \fIVALUES\fP.
Replaces the string '{\fIVAR\fP}' in each command by the current parameter value.
.IP
.RS
Example:
.RS
\fBhyperfine\fR \fB\-L\fR compiler g++,clang++ '{compiler} \-O2 main.cpp'
.RE
.RE
.IP
This performs benchmarks for 'g++ \-O2 main.cpp' and 'clang++ \-O2 main.cpp'.
.IP
The option can be specified multiple times to run benchmarks for all
possible parameter combinations.
.HP
\fB\-\-shell\fR \fISHELL\fP
.IP
Set the shell to use for executing commands (default: none). This can be
the name or the path to the shell executable, or a full command line
like "bash \fB\-\-norc\fR". It can also be set to "default" to explicitly
select the platform shell (sh on Unix, cmd.exe on Windows). Finally, this can also be
set to "none" to disable the shell. In this case, commands will be
executed directly. They can still have arguments, but more complex
things like "sleep 0.1; sleep 0.2" are not possible without a shell.
This option also applies to setup, prepare, conclude, and cleanup commands.
.HP
\fB\-S\fR
.IP
An alias for '\-\-shell=default' (sh on Unix, cmd.exe on Windows).
.HP
\fB\-N\fR
.IP
An alias for '\-\-shell=none'.
.HP
\fB\-i\fR, \fB\-\-ignore\-failure\fR[=\fIMODE\fP]
.IP
Ignore failures of the benchmarked programs. Without a value or with
\fBall\-non\-zero\fR, all non\-zero exit codes are ignored.
Alternatively, provide a comma\-separated list of exit codes to ignore,
for example \fB\-\-ignore\-failure=1,2\fR.
An explicit value must be attached to the option with an equals sign.
.HP
\fB\-\-style\fR \fITYPE\fP
.IP
Set output style \fITYPE\fP (default: auto). Set this to 'basic' to disable output
coloring and interactive elements. Set it to 'full' to enable all effects even
if no interactive terminal was detected. Set this to 'nocolor' to keep the
interactive output without any colors. Set this to 'color' to keep the colors
without any interactive output. Set this to 'none' to disable all the output
of the tool.
.HP
\fB\-\-metrics\fR \fIMETRIC[:UNIT],...|default|time|all\fP
.IP
Select the metrics to display, in order, with an optional unit for each.
Default: \fBtime_wall_clock,memory_peak_resident\fR, or \fBtime_wall_clock\fR
on Windows, where peak resident memory is not yet supported.
The first metric is primary: its entire terminal row is bold, and it is the
only metric in non\-JSON exports. All metric rows use colors when enabled.
The first benchmark has no change column; subsequent benchmarks show percentage
changes from the first benchmark. Decreases are green and increases are red
when colors are enabled.
.IP
Use \fBdefault\fR for the platform defaults, \fBtime\fR for
\fBtime_wall_clock,time_cpu,time_user,time_system\fR, or \fBall\fR for all available
metrics in the order listed below. The \fBtime\fR preset works on all
platforms, including with a shell.
All presets keep wall\-clock time primary, select display units automatically,
and must be used alone, without other metrics or unit suffixes.
The \fBall\fR preset skips metrics unsupported by the platform or shell,
or missing from any run of a benchmark. If a metric is unavailable for the
reference, its change displays as N/A.
.IP
Time metrics \fBtime_wall_clock\fR, \fBtime_cpu\fR, \fBtime_user\fR, and \fBtime_system\fR
accept units \fBns\fR, \fBus\fR, \fBms\fR, \fBs\fR, \fBmin\fR, and \fBh\fR.
Total CPU time, \fBtime_cpu\fR, is the sum of user and system time for each run.
It is summed across threads and can exceed wall\-clock time when work runs in parallel.
Peak memory, \fBmemory_peak_resident\fR, accepts \fBB\fR, \fBkB\fR,
\fBMB\fR, \fBGB\fR, \fBTB\fR, \fBKiB\fR, \fBMiB\fR, \fBGiB\fR,
and \fBTiB\fR.
.IP
Counter metrics \fBcpu_cycles\fR, \fBinstructions\fR, \fBcache_references\fR,
\fBcache_misses\fR, and \fBbranch_misses\fR accept \fBcount\fR (unscaled),
\fBk\fR (thousand), \fBM\fR (million), and \fBB\fR (billion).
For memory, \fBB\fR instead means bytes.
.IP
If a unit is omitted, terminal and markup output automatically select a suitable
unit. CSV uses base units unless a unit is explicitly selected. JSON always
uses base units, independently of this option.
.IP
Duplicate metrics, invalid units, and unavailable explicitly selected metrics are errors.
Hardware counters require direct execution (no shell) and platform support.
Automatic run counts continue to use elapsed time, regardless of the metrics.
.RS
Example:
.RS
\fBhyperfine\fR \fB\-\-metrics\fR memory_peak_resident:MiB,time_wall_clock:ms,instructions:B './baseline' './candidate'
.RE
.RE
.HP
\fB\-\-export\-asciidoc\fR \fIFILE\fP 
.IP
Export the primary metric's summary statistics as an AsciiDoc table to the given
\fIFILE\fP. Use its selected unit, or choose a unit automatically when omitted.
.HP
\fB\-\-export\-csv\fR \fIFILE\fP
.IP
Export the primary metric's summary statistics as CSV to the given \fIFILE\fP,
identifying the metric and unit explicitly. Use the explicitly selected unit,
or the base unit (seconds, bytes, or unscaled counts) when omitted.
CSV never scales units automatically. For individual samples, use JSON.
.HP
\fB\-\-export\-json\fR \fIFILE\fP
.IP
Export summary statistics and individual samples for all collected metrics as
JSON to the given \fIFILE\fP. Include the primary metric, and always use base
units (seconds, bytes, or unscaled counts), regardless of display settings.
.HP
\fB\-\-export\-markdown\fR \fIFILE\fP
.IP
Export the primary metric's summary statistics as a Markdown table to the given
\fIFILE\fP. Use its selected unit, or choose a unit automatically when omitted.
.HP
\fB\-\-export\-orgmode\fR \fIFILE\fP
.IP
Export the primary metric's summary statistics as an Emacs org\-mode table to
the given \fIFILE\fP. Use its selected unit, or choose a unit automatically
when omitted.
.HP
\fB\-\-show\-output\fR
.IP
Print the stdout and stderr of the benchmark instead of suppressing it. This
will increase the time it takes for benchmarks to run, so it should only be
used for debugging purposes or when trying to benchmark output speed.
.HP
\fB\-\-output\fR \fIWHERE\fP
.IP
Control where the output of the benchmark is redirected. Note that
some programs like 'grep' detect when standard output is \fI\,/dev/null\/\fP and
apply certain optimizations. To avoid that, consider using
\-\-output=pipe.
.IP
\fIWHERE\fP can be:
.RS
.IP null
Redirect output to \fI\,/dev/null\/\fP (the default).
.IP pipe
Feed the output through a pipe before discarding it.
.IP inherit
Don't redirect the output at all (same as \&'\-\-show\-output').
.IP "<FILE>"
Write the output to the given file.
.RE
.IP
This option can be specified once for all commands or multiple times,
once for each command. Note: If you want to log the output of each and
every iteration, you can use a shell redirection and the $HYPERFINE_ITERATION
environment variable (requires \fB\-S\fR / \fB\-\-shell=default\fR):
'my-command > output-${HYPERFINE_ITERATION}.log'
.HP
\fB\-\-input\fR \fIWHERE\fP
.IP
Control where the input of the benchmark comes from.
.IP
\fIWHERE\fP can be:
.RS
.IP null
Read from \fI\,/dev/null\/\fP (the default).
.IP "<FILE>"
Read the input from the given file.
.RE
.HP
\fB\-n\fR, \fB\-\-command\-name\fR \fINAME\fP
.IP
Give a meaningful \fINAME\fP to a command. This can be specified multiple times
if several commands are benchmarked.
.HP
\fB\-h\fR, \fB\-\-help\fR
.IP
Print help
.HP
\fB\-V\fR, \fB\-\-version\fR
.IP
Print version
.SH EXAMPLES
.LP
Basic benchmark of 'find . -name todo.txt':
.RS
.nf
\fBhyperfine\fR 'find . -name todo.txt'
.fi
.RE
.LP
Perform benchmarks for 'sleep 0.2' and 'sleep 3.2' with a minimum 5 runs each:
.RS
.nf
\fBhyperfine\fR \fB\-\-min\-runs\fR 5 'sleep 0.2' 'sleep 3.2'
.fi
.RE
.LP
Perform a benchmark of 'grep' with a warm disk cache by executing 3 runs up front that are not part
of the measurement: 
.RS
.nf
\fBhyperfine\fR \fB\-S\fR \fB\-\-warmup\fR 3 'grep -R TODO *'
.fi
.RE
.LP
Export the results of a parameter scan benchmark to a markdown table: 
.RS
.nf
\fBhyperfine\fR \fB\-\-export\-markdown\fR output.md \fB\-\-parameter-scan\fR time 1 5 'sleep {time}'
.fi
.RE
.LP
Demonstrate when each of \fB\-\-setup\fR, \fB\-\-prepare\fR, \fB\-\-conclude\fR, \fIcmd\fP and \fB\-\-cleanup\fR will run:
.RS
.nf
\fBhyperfine\fR \fB\-L\fR n 1,2 \fB\-r\fR 2 \fB\-\-show-output\fR \\
	\fB\-\-setup\fR 'echo setup n={n}' \\
	\fB\-\-prepare\fR 'echo prepare={n}' \\
	\fB\-\-conclude\fR 'echo conclude={n}' \\
	\fB\-\-cleanup\fR 'echo cleanup n={n}' \\
	'echo command n={n}'
.fi
.RE
.SH AUTHOR
.LP
David Peter <mail@david-peter.de>
.LP
Source, bug tracker, and additional information can be found on GitHub:
.I https://github.com/sharkdp/hyperfine