mitata

Install

bun add mitata

npm install mitata

try mitata in browser with ai assistant at https://bolt.new/~/mitata

Recommendations

• use dedicated hardware for running benchmarks
• read writing good benchmarks & LLVM benchmarking tips
• run with manual garbage collection enabled (e.g. node --expose-gc ...)
• install optional hardware counters extension to see cpu stats like IPC (instructions per cycle)
• make sure your runtime has high-resolution timers and other relevant options/permissions enabled

Quick Start

1import { run, bench, boxplot, summary } from 'mitata';
2
3function fibonacci(n) {
4  if (n <= 1) return n;
5  return fibonacci(n - 1) + fibonacci(n - 2);
6}
7
8bench('fibonacci(40)', () => fibonacci(40));
9
10boxplot(() => {
11  summary(() => {
12    bench('Array.from($size)', function* (state) {
13      const size = state.get('size');
14      yield () => Array.from({ length: size });
15    }).range('size', 1, 1024);
16  });
17});
18
19await run();

1#include "src/mitata.hpp"
2
3int fibonacci(int n) {
4  if (n <= 1) return n;
5  return fibonacci(n - 1) + fibonacci(n - 2);
6}
7
8int main() {
9  mitata::runner runner;
10  runner.bench("noop", []() { });
11
12  runner.summary([&]() {
13    runner.bench("empty fn", []() { });
14    runner.bench("fibonacci", []() { fibonacci(20); });
15  });
16
17  auto stats = runner.run();
18}

configure your experience

1import { run } from 'mitata';
2
3await run({ format: 'json' }) // output json
4await run({ filter: /new Array.*/ }) // only run benchmarks that match regex filter
5await run({ throw: true }); // will immediately throw instead of handling error quietly
6await run({ format: { mitata: { name: 'fixed' } } }); // benchmarks name column is fixed length
7
8// c++
9auto stats = runner.run({ .colors = true, .format = "json", .filter = std::regex(".*") });

garbage collection

By default, on runtimes with exposed manual gc (like bun or node with --expose-gc), mitata runs garbage collection once after each benchmark warmup.

This behavior can be customized using gc(mode) method on benchmarks:

1bench('lots of allocations', () => {
2  Array.from({ length: 1024 }, () => Array.from({ length: 1024 }, () => new Array(1024)));
3})
4  // mode: false | 'once' (default) | 'inner'
5
6  // once mode runs gc after warmup
7  // inner mode runs gc after warmup and before each (batch-)iteration
8  .gc('inner');

gc impact and memory usage

For runtimes that provide manual garbage collection or offer access to javscript vm heap usage metrics, additional row will be shown with garbage collection timings or/and estimated heap usage.

1------------------------------------------- -------------------------------
2new Array(512)               509.42 ns/iter 536.53 ns  ▅▃█      ▂          
3                    (449.52 ns … 632.54 ns) 609.34 ns  ███   ▃▅▆█▇         
4                    (  0.00  b …  24.00 kb)   1.61 kb ▆████▅▄██████▅▅▅█▅▄▂▂
5
6Array.from(512)                1.29 µs/iter   1.30 µs  ▂▆█                 
7                        (1.27 µs … 1.48 µs)   1.40 µs ▂███▇▆▃▃▂▁▁▂▁▁▁▁▁▁▁▁▁
8                  gc(457.25 µs … 760.54 µs) 512.32  b (  0.00  b… 84.00 kb)

universal compatibility

Out of box mitata can detect engine/runtime it's running on and fall back to using alternative non-standard I/O functions. If your engine or runtime is missing support, open an issue or pr requesting for support.

how to use mitata with engine CLIs like d8, jsc, graaljs, spidermonkey

1$ xs bench.mjs
2$ quickjs bench.mjs
3$ d8 --expose-gc bench.mjs
4$ spidermonkey -m bench.mjs
5$ graaljs --js.timer-resolution=1 bench.mjs
6$ /System/Library/Frameworks/JavaScriptCore.framework/Versions/Current/Helpers/jsc bench.mjs

1// bench.mjs
2
3import { print } from './src/lib.mjs';
4import { run, bench } from './src/main.mjs'; // git clone
5import { run, bench } from './node_modules/mitata/src/main.mjs'; // npm install
6
7print('hello world'); // works on every engine

adding arguments and parameters to your benchmarks has never been so easy

With other benchmarking libraries, often it's quite hard to easily make benchmarks that go over a range or run the same function with different arguments without writing spaghetti code, but now with mitata converting your benchmark to use arguments is just a function call away.

1import { bench } from 'mitata';
2
3bench(function* look_mom_no_spaghetti(state) {
4  const len = state.get('len');
5  const len2 = state.get('len2');
6  yield () => new Array(len * len2);
7})
8
9.args('len', [1, 2, 3])
10.range('len', 1, 1024) // 1, 8, 64, 512...
11.dense_range('len', 1, 100) // 1, 2, 3 ... 99, 100
12.args({ len: [1, 2, 3], len2: ['4', '5', '6'] }) // every possible combination

computed parameters

For cases where you need unique copy of value for each iteration, mitata supports creating computed parameters that do not count towards benchmark results (note: there is no guarantee of recompute time, order, or call count):

1bench('deleting $keys from object', function* (state) {
2  const keys = state.get('keys');
3
4  const obj = {};
5  for (let i = 0; i < keys; i++) obj[i] = i;
6
7  yield {
8    [0]() {
9      return { ...obj };
10    },
11
12    bench(p0) {
13      for (let i = 0; i < keys; i++) delete p0[i];
14    },
15  };
16}).args('keys', [1, 10, 100]);

concurrency

concurrency option enables transparent concurrent execution of asynchronous benchmark, providing insights into:

• scalability of async functions
• potential bottlenecks in parallel code
• performance under different levels of concurrency

(note: concurrent benchmarks may have higher variance due to scheduling, contention, event loop and async overhead)

1bench('sleepAsync(1000) x $concurrency', function* () {
2  // concurrency inherited from arguments
3  yield async () => await sleepAsync(1000);
4}).args('concurrency', [1, 5, 10]);
5
6bench('sleepAsync(1000) x 5', function* () {
7  yield {
8    // concurrency is set manually
9    concurrency: 5,
10
11    async bench() {
12      await sleepAsync(1000);
13    },
14  };
15});

hardware counters

bun add @mitata/counters

npm install @mitata/counters

supported on: macos (apple silicon) | linux (amd64, aarch64)

linux:

• /proc/sys/kernel/perf_event_paranoid has to be set to 2 or lower
• on some vm systems pmu is disabled by hypervisor (usually when cpu core is shared across vms)

macos:

• Apple Silicon CPU optimization guide/handbook
• Xcode must be installed for complete cpu counters support
• Instruments.app (CPU Counters) has to be closed during benchmarking
• Corrupted install of Xcode/Command Line Tools can result in kernel panic (requires Xcode/Command Line Tools reinstall)

By installing @mitata/counters package you can enable collection and displaying of hardware counters for benchmarks.

1------------------------------------------- -------------------------------
2new Array(1024)              332.67 ns/iter 337.90 ns   █                  
3                    (295.63 ns … 507.93 ns) 455.66 ns ▂██▇▄▂▂▂▁▂▁▃▃▃▂▂▁▁▁▁▁
4                  2.41 ipc ( 48.66% stalls)  37.89% L1 data cache
5          1.11k cycles   2.69k instructions  33.09% retired LD/ST ( 888.96)
6
7new URL(google.com)          246.40 ns/iter 245.10 ns       █▃             
8                    (206.01 ns … 841.23 ns) 302.39 ns ▁▁▁▁▂███▇▃▂▂▂▂▂▂▂▁▁▁▁
9                  4.12 ipc (  1.05% stalls)  98.88% L1 data cache
10         856.49 cycles   3.53k instructions  28.65% retired LD/ST (  1.01k)

helpful warnings

For those who love doing micro-benchmarks, mitata can automatically detect and inform you about optimization passes like dead code elimination without requiring any special engine flags.

1-------------------------------------- -------------------------------
21 + 1                   318.63 ps/iter 325.37 ps        ▇  █           !
3                (267.92 ps … 14.28 ns) 382.81 ps ▁▁▁▁▁▁▁█▁▁█▁▁▁▁▁▁▁▁▁▁
4empty function          319.36 ps/iter 325.37 ps          █ ▅          !
5                (248.62 ps … 46.61 ns) 382.81 ps ▁▁▁▁▁▁▃▁▁█▁█▇▁▁▁▁▁▁▁▁
6
7! = benchmark was likely optimized out (dead code elimination)

powerful visualizations right in your terminal

With mitata’s ascii rendering capabilities, now you can easily visualize samples in barplots, boxplots, lineplots, histograms, and get clear summaries without any additional tools or dependencies.

1import { summary, barplot, boxplot, lineplot } from 'mitata';
2
3// wrap bench() calls in visualization scope
4barplot(() => {
5  bench(...)
6});
7
8                        ┌                                            ┐
9                  1 + 1 ┤■ 318.11 ps 
10             Date.now() ┤■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■ 27.69 ns 
11                        └                                            ┘
12
13// scopes can be async
14await boxplot(async () => {
15  // ...
16});
17
18                        ┌                                            ┐
19                                        ╷┌─┬─┐                       ╷
20            Bubble Sort                 ├┤ │ ├───────────────────────┤
21                                        ╵└─┴─┘                       ╵
22                        ┬   ╷
23             Quick Sort │───┤
24                        ┴   ╵
25                        ┬
26            Native Sort │
27                        ┴
28                        └                                            ┘
29                        90.88 µs            2.43 ms            4.77 ms
30
31// can combine multiple visualizations
32lineplot(() => {
33  summary(() => {
34    // ...
35  });
36
37  // bench() calls here wont be part of summary
38});
39
40summary
41  new Array($len)
42   5.42…8.33x faster than Array.from($len)
43
44                        ┌                                            ┐
45      Array.from($size)                                            ⢠⠊
46       new Array($size)                                          ⢀⠔⠁ 
47                                                                ⡠⠃   
48                                                              ⢀⠎     
49                                                             ⡔⠁      
50                                                           ⡠⠊        
51                                                         ⢀⠜          
52                                                        ⡠⠃           
53                                                       ⡔⠁            
54                                                     ⢀⠎              
55                                                    ⡠⠃               
56                                                  ⢀⠜                 
57                                                 ⢠⠊             ⣀⣀⠤⠤⠒
58                                                ⡰⠁       ⣀⡠⠤⠔⠒⠊⠉     
59                                           ⣀⣀⣀⠤⠜   ⣀⡠⠤⠒⠊⠉            
60                         ⣤⣤⣤⣤⣤⣤⣤⣤⣤⣤⣤⣤⣔⣒⣒⣊⣉⠭⠤⠤⠤⠤⠤⠒⠊⠉               
61                        └                                            ┘

give your own code power of mitata

In case you don’t need all the fluff that comes with mitata or just need raw results, mitata exports its fundamental building blocks to allow you to easily build your own tooling and wrappers without losing any core benefits of using mitata.

1#include "src/mitata.hpp"
2
3int main() {
4  auto stats = mitata::lib::fn([]() { /***/ })
5}

1import { B, measure } from 'mitata';
2
3// lowest level for power users
4const stats = await measure(function* (state) {
5  const size = state.get('x');
6
7  yield {
8    [0]() {
9      return size;
10    },
11
12    bench(size) {
13      return new Array(size);
14    },
15  };
16}, {
17  args: { x: 1 },
18  batch_samples: 5 * 1024,
19  min_cpu_time: 1000 * 1e6,
20});
21
22// explore how magic happens
23console.log(stats.debug) // -> jit optimized source code of benchmark
24
25// higher level api that includes mitata's argument and range features
26const b = new B('new Array($x)', function* (state) {
27  const size = state.get('x');
28  yield () => new Array(size);
29}).args('x', [1, 5, 10]);
30
31const trial = await b.run();

accuracy down to picoseconds

By leveraging the power of javascript JIT compilation, mitata is able to generate zero-overhead measurement loops that provide picoseconds precision in timing measurements. These loops are so precise that they can even be reused to provide additional features like CPU clock frequency estimation and dead code elimination detection, all while staying inside javascript vm sandbox.

With computed parameters and garbage collection tuning, you can tap into mitata's code generation capabilities to further refine the accuracy of your benchmarks. Using computed parameters ensures that parameters computation is moved outside the benchmark, thereby preventing the javascript JIT from performing loop invariant code motion optimization.

1// node --expose-gc --allow-natives-syntax tools/compare.mjs
2clk: ~2.71 GHz
3cpu: Apple M2 Pro
4runtime: node 23.3.0 (arm64-darwin)
5
6benchmark                   avg (min … max) p75   p99    (min … top 1%)
7------------------------------------------- -------------------------------
8a / b                          4.59 ns/iter   4.44 ns █                    
9                       (4.33 ns … 25.86 ns)   6.91 ns ██▂▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁
10                  6.70 ipc (  2.17% stalls)    NaN% L1 data cache
11          16.80 cycles  112.52 instructions   0.00% retired LD/ST (   0.00)
12
13a / b (computed)               4.23 ns/iter   4.10 ns ▇█                   
14                       (3.88 ns … 30.03 ns)   7.26 ns ██▅▂▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁
15                  6.40 ipc (  2.10% stalls)    NaN% L1 data cache
16          15.70 cycles  100.53 instructions   0.00% retired LD/ST (   0.00)
174.59 ns/iter - https://npmjs.com/mitata
18
19// vs other libraries
20
21a / b x 90,954,882 ops/sec ±2.13% (92 runs sampled)
2210.99 ns/iter - https://npmjs.com/benchmark
23
24┌─────────┬───────────┬──────────────────────┬─────────────────────┬────────────────────────────┬───────────────────────────┬──────────┐
25│ (index) │ Task name │ Latency average (ns) │ Latency median (ns) │ Throughput average (ops/s) │ Throughput median (ops/s) │ Samples  │
26├─────────┼───────────┼──────────────────────┼─────────────────────┼────────────────────────────┼───────────────────────────┼──────────┤
27│ 0       │ 'a / b'   │ '27.71 ± 0.09%'      │ '41.00'             │ '28239766 ± 0.01%'         │ '24390243'                │ 36092096 │
28└─────────┴───────────┴──────────────────────┴─────────────────────┴────────────────────────────┴───────────────────────────┴──────────┘
2927.71 ns/iter - vitest bench / https://npmjs.com/tinybench
30
31a / b x 86,937,932 ops/sec (11 runs sampled) v8-never-optimize=true min..max=(11.32ns...11.62ns)
3211.51 ns/iter - https://npmjs.com/bench-node
33
34╔══════════════╤═════════╤════════════════════╤═══════════╗
35║ Slower tests │ Samples │             Result │ Tolerance ║
36╟──────────────┼─────────┼────────────────────┼───────────╢
37║ Fastest test │ Samples │             Result │ Tolerance ║
38╟──────────────┼─────────┼────────────────────┼───────────╢
39║ a / b        │   10000 │ 14449822.99 op/sec │  ± 4.04 % ║
40╚══════════════╧═════════╧════════════════════╧═══════════╝
4169.20 ns/iter - https://npmjs.com/cronometro

1// node --expose-gc --allow-natives-syntax --jitless tools/compare.mjs
2clk: ~0.06 GHz
3cpu: Apple M2 Pro
4runtime: node 23.3.0 (arm64-darwin)
5
6benchmark                   avg (min … max) p75   p99    (min … top 1%)
7------------------------------------------- -------------------------------
8a / b                         74.52 ns/iter  75.53 ns █                    
9                     (71.96 ns … 104.94 ns)  92.01 ns █▅▇▅▅▃▃▂▁▁▁▁▁▁▁▁▁▁▁▁▁
10                  5.78 ipc (  0.51% stalls)    NaN% L1 data cache
11         261.51 cycles   1.51k instructions   0.00% retired LD/ST (   0.00)
12
13a / b (computed)              56.05 ns/iter  57.20 ns █                    
14                      (53.62 ns … 84.69 ns)  73.21 ns █▅▆▅▅▃▃▂▂▁▁▁▁▁▁▁▁▁▁▁▁
15                  5.65 ipc (  0.59% stalls)    NaN% L1 data cache
16         197.74 cycles   1.12k instructions   0.00% retired LD/ST (   0.00)
1774.52 ns/iter - https://npmjs.com/mitata
18
19// vs other libraries
20
21a / b x 11,232,032 ops/sec ±0.50% (99 runs sampled)
2289.03 ns/iter - https://npmjs.com/benchmark
23
24┌─────────┬───────────┬──────────────────────┬─────────────────────┬────────────────────────────┬───────────────────────────┬─────────┐
25│ (index) │ Task name │ Latency average (ns) │ Latency median (ns) │ Throughput average (ops/s) │ Throughput median (ops/s) │ Samples │
26├─────────┼───────────┼──────────────────────┼─────────────────────┼────────────────────────────┼───────────────────────────┼─────────┤
27│ 0       │ 'a / b'   │ '215.53 ± 0.08%'     │ '208.00'            │ '4786095 ± 0.01%'          │ '4807692'                 │ 4639738 │
28└─────────┴───────────┴──────────────────────┴─────────────────────┴────────────────────────────┴───────────────────────────┴─────────┘
29215.53 ns/iter - vitest bench / https://npmjs.com/tinybench
30
31a / b x 10,311,999 ops/sec (11 runs sampled) v8-never-optimize=true min..max=(95.66ns...97.51ns)
3296.86 ns/iter - https://npmjs.com/bench-node
33
34╔══════════════╤═════════╤═══════════════════╤═══════════╗
35║ Slower tests │ Samples │            Result │ Tolerance ║
36╟──────────────┼─────────┼───────────────────┼───────────╢
37║ Fastest test │ Samples │            Result │ Tolerance ║
38╟──────────────┼─────────┼───────────────────┼───────────╢
39║ a / b        │    2000 │ 4664908.00 op/sec │  ± 0.94 % ║
40╚══════════════╧═════════╧═══════════════════╧═══════════╝
41214.37 ns/iter - https://npmjs.com/cronometro

writing good benchmarks

Creating accurate and meaningful benchmarks requires careful attention to how modern JavaScript engines optimize code. This covers essential concepts and best practices to ensure your benchmarks measure actual performance characteristics rather than optimization artifacts.

examples

• readme gif
• cpu cache line size
• holey vs packed arrays

dead code elimination

JIT can detect and eliminate code that has no observable effects. To ensure your benchmark code executes as intended, you must create observable side effects.

1import { do_not_optimize } from 'mitata';
2
3bench(function* () {
4  // ❌ Bad: jit can see that function has zero side-effects
5  yield () => new Array(0);
6  // will get optimized to:
7  /*
8    yield () => {};
9  */
10
11  // ✅ Good: do_not_optimize(value) emits code that causes side-effects
12  yield () => do_not_optimize(new Array(0));
13});

garbage collection pressure

For benchmarks involving significant memory allocations, controlling garbage collection frequency can improve results consistency.

1// ❌ Bad: unpredictable gc pauses
2bench(() => {
3  const bigArray = new Array(1000000);
4});
5
6// ✅ Good: gc before each (batch-)iteration
7bench(() => {
8  const bigArray = new Array(1000000);
9}).gc('inner'); // run gc before each iteration

loop invariant code motion optimization

JavaScript engines can optimize away repeated computations by hoisting them out of loops or caching results. Use computed parameters to prevent loop invariant code motion optimization.

1bench(function* (ctx) {
2  const str = 'abc';
3
4  // ❌ Bad: JIT sees that both str and 'c' search value are constants/comptime-known
5  yield () => str.includes('c');
6  // will get optimized to:
7  /*
8    yield () => true;
9  */
10
11  // ❌ Bad: JIT sees that computation doesn't depend on anything inside loop
12  const substr = ctx.get('substr');
13  yield () => str.includes(substr);
14  // will get optimized to:
15  /*
16    const $0 = str.includes(substr);
17    yield () => $0;
18  */
19
20  // ✅ Good: using computed parameters prevents jit from performing any loop optimizations
21  yield {
22    [0]() {
23      return str;
24    },
25
26    [1]() {
27      return substr;
28    },
29
30    bench(str, substr) {
31      return do_not_optimize(str.includes(substr));
32    },
33  };
34}).args('substr', ['c']);

License

MIT © evanwashere