bencho

A command-line benchmarking tool.

:book: This is an early version of the package.

Bencho is inspired by hyperfine, but it measures explicit markers emitted by the benchmarked process instead of only measuring the process runtime. This is useful when you need to benchmark a specific operation inside a larger program or compare application-level metrics.

Markers are plain text lines written to stdout or stderr. They can represent time, memory, or any numeric value. You can emit any number of markers from any language, framework, logger, or tracing tool.

See ./examples/min.js for a complete example.

$ bencho 'node --expose-gc ./examples/min.js {mode}' -n 'min.js {mode}' -r 1 -l mode=math,math-spread,loop,sort,reduce --reporter benchmark --allow-multiple

Marker        Type    Command             Mean       StdDev       StdDev, %  Ratio   Lowest
------------  ------  ------------------  ---------  -----------  ---------  ------  ------
time          time    min.js math           0.148ms      0.017ms    11.744%    1.00     yes
time          time    min.js math-spread    0.152ms      0.027ms    17.487%    1.03
time          time    min.js loop           0.241ms      0.341ms   141.105%    1.63
time          time    min.js sort          18.774ms      0.528ms     2.815%  126.58
time          time    min.js reduce         1.246ms      0.066ms     5.270%    8.40

memory        memory  min.js math         4.180 MiB  130.722 kiB     3.054%    1.00     yes
memory        memory  min.js math-spread  4.251 MiB  104.033 kiB     2.390%    1.02
memory        memory  min.js loop         4.358 MiB  415.165 kiB     9.304%    1.04
memory        memory  min.js sort         6.026 MiB  163.173 kiB     2.644%    1.44
memory        memory  min.js reduce       4.411 MiB  132.157 kiB     2.926%    1.06

max           value   min.js math             99999            0     0.000%    1.00     yes
max           value   min.js math-spread      99999            0     0.000%    1.00     yes
max           value   min.js loop             99999            0     0.000%    1.00     yes
max           value   min.js sort             99999            0     0.000%    1.00     yes
max           value   min.js reduce           99999            0     0.000%    1.00     yes

process.time  time    min.js math            4.018s          0ms     0.000%    1.00
process.time  time    min.js math-spread     4.011s          0ms     0.000%    1.00     yes
process.time  time    min.js loop            4.028s          0ms     0.000%    1.00
process.time  time    min.js sort            5.873s          0ms     0.000%    1.46
process.time  time    min.js reduce          4.124s          0ms     0.000%    1.03

How does it work?

Add markers to the code you want to benchmark. A marker is a line with this format:

bencho_marker:<marker_label>;<marker_type>;<marker_value>

Examples:

bencho_marker:time;time;12.5
bencho_marker:memory;memory;1048576
bencho_marker:counter;value;42

Bencho starts your command, reads stdout and stderr, extracts marker lines, groups them by label and type, and reports statistics.

:book: Bencho does not transform or instrument your code.

You can print marker lines manually with console.log or console.error, or use the helper functions exported by bencho. If markers should not be present in production builds, gate or remove them in your own build process.

Installation

:book: Right now we only support installation from the npm registry.

npm install --save-dev bencho

Usage

Use bencho --help for the full CLI reference. More examples are available in the ./examples directory.

The examples below use bencho. For a local project install, run it as npx bencho or from an npm script.

Before adding markers:

// ./code.mjs
async function findAllEmptyResolutions() { /* code */ }

await findAllEmptyResolutions();

After adding markers:

// ./code.mjs
import { performance } from 'node:perf_hooks';
import * as bencho from 'bencho';

async function findAllEmptyResolutions() { /* code */ }

const startTime = performance.now();

await findAllEmptyResolutions();

const time = performance.now() - startTime;
const memory = process.memoryUsage().heapUsed;

bencho.time('time', time);
bencho.memory('memory', memory);

Run a fixed number of measured runs:

$ bencho 'node ./code.mjs' --warmup 10 --runs 50

--warmup runs are executed before measurement and are not included in the result. --runs selects exact mode and performs exactly that many measured runs.

When --runs is omitted, Bencho uses adaptive mode. In this mode, the first measured run determines how many measured runs will be collected:

max(floor(3000 / first_process_time_ms), min_runs)

The calculated count includes the first measured run. By default, adaptive mode uses at least 10 measured runs and an internal 3 second target. --min-runs changes the minimum, and --max-runs caps the measured run count.

$ bencho 'node ./code.mjs' --warmup 10 --min-runs 10

With this command:

• if the first measured run takes about 400ms, Bencho runs 10 warmup runs and 10 measured runs;
• if it takes about 100ms, Bencho runs 10 warmup runs and 30 measured runs.

The internal 3 second target is only used to estimate the run count. It is not a timeout and does not limit total benchmark duration. Use --max-runs when fast commands would otherwise run too many times.

--runs cannot be combined with --min-runs or --max-runs.

Results:

Benchmark #1
Command:    node ./code.mjs
            node ./code.mjs

Marker        Type    Mean       StdDev   StdDev, %  Min        Max
------------  ------  ---------  -------  ---------  ---------  ---------
time          time      0.011ms  0.002ms    19.064%    0.008ms    0.017ms
process.time  time     36.135ms  2.163ms     5.985%   33.077ms   43.087ms
memory        memory  5.127 MiB  0.000 B     0.000%  5.127 MiB  5.127 MiB