github.com/loov/hrtime

hrtime

Package hrtime implements high-resolution timing functions and benchmarking utilities.

hrtime relies on using the best timing mechanism on a particular system. At the moment, for Windows it is using Performance Counters and on other platforms standard time.Now (since it's good enough).

Package also supports using hardware time stamp counters (TSC). They offer better accuracy and on some platforms correspond to the processor cycles. However, they are not supported on all platforms.

For example measuring time.Sleep on Mac and Windows.

Example

package main

import (
    "fmt"
    "time"

    "github.com/loov/hrtime"
)

func main() {
    start := hrtime.Now()
    time.Sleep(1000 * time.Nanosecond)
    fmt.Println(hrtime.Since(start))

    const numberOfExperiments = 4096

    bench := hrtime.NewBenchmark(numberOfExperiments)
    for bench.Next() {
        time.Sleep(1000 * time.Nanosecond)
    }
    fmt.Println(bench.Histogram(10))
}

Output on Mac:

12µs
  avg 14.5µs;  min 2µs;  p50 12µs;  max 74µs;
  p90 22µs;  p99 44µs;  p999 69µs;  p9999 74µs;
        2µs [ 229] ██▌
       10µs [3239] ████████████████████████████████████████
       20µs [ 483] ██████
       30µs [  80] █
       40µs [  39] ▌
       50µs [  17] ▌
       60µs [   6]
       70µs [   3]
       80µs [   0]
       90µs [   0]

Output on Windows:

1.5155ms
  avg 1.49ms;  min 576µs;  p50 1.17ms;  max 2.47ms;
  p90 2.02ms;  p99 2.3ms;  p999 2.37ms;  p9999 2.47ms;
      577µs [   1]
      600µs [  57] █▌
      800µs [ 599] █████████████████
        1ms [1399] ████████████████████████████████████████
      1.2ms [  35] █
      1.4ms [   7]
      1.6ms [  91] ██▌
      1.8ms [ 995] ████████████████████████████
        2ms [ 778] ██████████████████████
      2.2ms [ 134] ███▌

A full explanation why it outputs this is out of the scope of this document. However, all sleep instructions have a specified granularity and time.Sleep actual sleeping time is requested time ± sleep granularity. There are also other explanations to that behavior.

Benchmarking

hrtime/hrtesting can be used to supplement existing benchmarks with more details:

package hrtesting_test

import (
	"fmt"
	"runtime"
	"testing"

	"github.com/loov/hrtime/hrtesting"
)

func BenchmarkReport(b *testing.B) {
	bench := hrtesting.NewBenchmark(b)
	defer bench.Report()

	for bench.Next() {
		r := fmt.Sprintf("hello, world %d", 123)
		runtime.KeepAlive(r)
	}
}