Tyarel8 483974311d
feat(std): further bench improvements (#15856)
# Description

Following #15843, I have tinkered more with it and realized that there
are plenty of features from
[hyperfine](https://github.com/sharkdp/hyperfine) that could be
implemented pretty easily.

- `--warmup` flag to do `n` runs without benchmarking first, useful to
fill disk cache
```nu
@example "use --warmup to fill the disk cache before benchmarking" { bench { fd } { jwalk . -k } -w 1 -n 10 }
```
- `--setup`, `--prepare`, `--cleanup`, `--conclude` flags to run code
before/after benchmarks
```nu
@example "use `--setup` to compile before benchmarking" { bench { ./target/release/foo } --setup { cargo build --release } }
@example "use `--prepare` to benchmark rust compilation speed" { bench { cargo build --release } --prepare { cargo clean } }
```
- `--ignore-errors` to ignore any errors in the benchmarked commands
- benchmarked commands are now `| ignore` so that externals don't fill
the screen
2025-06-02 22:32:44 +02:00

174 lines
5.9 KiB
Plaintext

# run a piece of `nushell` code multiple times and measure the time of execution.
#
# this command returns a benchmark report in the form of a table/record, or a string if using `--pretty`
#
# if multiple commands are passed, it will show a comparison of their runtimes.
@example "measure the performance of simple addition" { bench { 1 + 2 } } --result {
mean: 2308ns,
min: 2000ns,
max: 8500ns,
std: 895ns
}
@example "do 10 runs and show the time of each" { bench { 1 + 2 } -n 10 --verbose } --result {
mean: 3170ns,
min: 2200ns,
max: 9800ns,
std: 2228ns,
times: [
9800ns,
3100ns,
2800ns,
2300ns,
2500ns,
2200ns,
2300ns,
2300ns,
2200ns,
2200ns
]
}
@example "get a pretty benchmark report" { bench { 1 + 2 } --pretty } --result "3µs 125ns +/- 2µs 408ns"
@example "compare multiple commands" { bench { 2 + 4 } { 2 ** 4 } } --result [
[
code,
mean,
min,
max,
std,
ratio
];
[
"{ 2 + 4 }",
2406ns,
2100ns,
9400ns,
1012ns,
1.02732707087959
],
[
"{ 2 ** 4 }",
2342ns,
2100ns,
5300ns,
610ns,
1.0
]
]
@example "compare multiple commands with pretty report" { bench { 2 + 4 } { 2 ** 4 } --pretty } --result "
Benchmark 1: { 2 + 4 }
2µs 494ns +/- 1µs 105ns
Benchmark 2: { 2 ** 4 }
2µs 348ns +/- 565ns
{ 2 + 4 } ran
1 times faster than { 2 ** 4 }"
@example "use `--setup` to compile before benchmarking" { bench { ./target/release/foo } --setup { cargo build --release } }
@example "use `--prepare` to benchmark rust compilation speed" { bench { cargo build --release } --prepare { cargo clean } }
@example "use `--warmup` to fill the disk cache before benchmarking" { bench { fd } { jwalk . -k } -w 1 -n 10 }
export def main [
...commands: closure # the piece(s) of `nushell` code to measure the performance of
--rounds (-n): int = 50 # the number of benchmark rounds (hopefully the more rounds the less variance)
--warmup (-w): int = 0 # the number of warmup rounds (not timed) to do before the benchmark, useful for filling the disk cache in I/O-heavy programs
--setup (-s): closure # command to run before all benchmarks
--cleanup (-c): closure # command to run after all benchmarks
--prepare (-S): closure # command to run before each timing round (for multiple commands, the index is passed to this closure so you can do different behaviour)
--conclude (-C): closure # command to run after each timing round (for multiple commands, the index is passed to this closure so you can do different behaviour)
--ignore-errors (-i) # ignore errors in the command
--verbose (-v) # show individual times (has no effect if used with `--pretty`)
--progress (-P) # prints the progress
--pretty (-p) # shows the results in human-readable format: "<mean> +/- <stddev>"
]: [
nothing -> record<mean: duration, std: duration, times: list<duration>>
nothing -> record<mean: duration, std: duration>
nothing -> table<code: string, mean: duration, std: duration, ratio: float, times: list<duration>>
nothing -> table<code: string, mean: duration, std: duration, ratio: float>
nothing -> string
] {
if $setup != null { do $setup | ignore }
let results = (
$commands | enumerate | each {|x|
let code = $x.item
let idx = $x.index
let bench_num = if ($commands | length) > 1 { $" ($idx + 1)"}
seq 1 $warmup | each {|i|
if $progress { print -n $"Warmup($bench_num): ($i) / ($warmup)\r" }
if $prepare != null { $idx | do $prepare $idx | ignore }
do --ignore-errors=$ignore_errors $code | ignore
if $conclude != null { $idx | do $conclude $idx | ignore }
}
if $progress and $warmup > 0 { print $"Warmup($bench_num): ($warmup) / ($warmup)" }
let times: list<duration> = (
seq 1 $rounds | each {|i|
if $progress { print -n $"Benchmark($bench_num): ($i) / ($rounds)\r" }
if $prepare != null { $idx | do $prepare $idx | ignore }
let time = timeit { do --ignore-errors=$ignore_errors $code | ignore }
if $conclude != null { $idx | do $conclude $idx | ignore }
$time
}
)
if $progress { print $"Benchmark($bench_num): ($rounds) / ($rounds)" }
{
mean: ($times | math avg)
min: ($times | math min)
max: ($times | math max)
std: ($times | into int | into float | math stddev | into int | into duration)
}
| if $verbose { merge { times: $times }} else {}
}
)
if $cleanup != null { do $cleanup | ignore }
# One benchmark
if ($results | length) == 1 {
let report = $results | first
if $pretty {
return $"($report.mean) +/- ($report.std)"
} else {
return $report
}
}
# Multiple benchmarks
let min_mean = $results | get mean | math min
let results = (
$commands
| each { view source $in | nu-highlight }
| wrap code
| merge $results
| insert ratio { $in.mean / $min_mean }
)
if $pretty {
$results
| enumerate
| each {|x|
let report = $x.item
print $"Benchmark ($x.index + 1): ($report.code)\n\t($report.mean) +/- ($report.std)"
}
let results = $results | sort-by ratio
print $"\n($results.0.code) ran"
$results
| skip
| each {|report|
print $"\t(ansi green)($report.ratio | math round -p 2)(ansi reset) times faster than ($report.code)"
}
ignore
} else {
$results
}
}