Go 实现Benchmark函数

背景

go 必须要 test 才能跑benchmark，导致一些情况下想要在main函数中测试benchmark会麻烦一些，因此我实现了一个简单的且没有开销的benchmark函数，方便使用！其次也方便大家学习下如何实现一个零开销的benchmark框架!

benchamrk 实现

对于有timeout的benchamrk，每次都去比较 time.Before(timeout) 开销非常的大，而且 benchmark 的对外函数也不能是一个空函数一定要带 count ，因为函数调用大概会劣化 ns 级别！

所以一般的benchmark算法都是梯度benchmark，即 1，10，100，1000，10000，100000，1000000 ... 的数量级去benchmark，好处就是避免了大量的time.Since 计算开销，因为time.Since单次在 30ns 左右（Linux环境下），开销非常大的！

go 复制代码

package pprof

import (
    "fmt"
    "sync"
    "sync/atomic"
    "time"
)

func ParallelBenchmark(name string, thread int, duration time.Duration, execute func(count int)) {
    wg := sync.WaitGroup{}
    wg.Add(thread)
    totalCount := uint64(0)
    totalSpend := uint64(0)
    for i := 0; i < thread; i++ {
       go func() {
          defer wg.Done()
          spend, count := Benchmark(duration, execute)
          atomic.AddUint64(&totalSpend, uint64(spend))
          atomic.AddUint64(&totalCount, uint64(count))
       }()
    }
    wg.Wait()
    fmt.Printf("name=%s thread=%d duration=%s total=%d avg=%s\n", name, thread, duration, totalCount, Avg(time.Duration(totalSpend), int(totalCount)))
}

func Avg(spend time.Duration, count int) string {
    avg := float64(spend) / float64(count)
    if avg > 100 {
       return time.Duration(avg).String()
    }
    return fmt.Sprintf("%.4fns", avg)
}

func Benchmark(duration time.Duration, bench func(count int)) (time.Duration, int) {
    const maxTotalCount = 1000000000 // 10E
    count := 1
    totalSpend := time.Duration(0)
    totalCount := 0
    for {
       start := time.Now()
       bench(count)
       spend := time.Since(start)

       totalSpend = totalSpend + spend
       totalCount = totalCount + count

       if totalCount >= maxTotalCount {
          break
       }
       subSpend := duration - totalSpend
       if subSpend <= 0 {
          break
       }
       count = totalCount*10 - totalCount
       if subCount := int(float64(subSpend) / (float64(totalSpend) / float64(totalCount))); count > subCount {
          count = subCount
       }
    }
    return totalSpend, totalCount
}

profile 实现

go 复制代码

package pprof

import (
    "net/http"
    _ "net/http/pprof"
    "os"
    "runtime"
    "runtime/pprof"
)

// InitPProf
// go InitPProf()
func InitPProf() {
    err := http.ListenAndServe(":12345", http.DefaultServeMux)
    if err != nil {
       panic(err)
    }
}

func StartCPUProfile(fileName string) (stop func()) {
    f, err := os.Create(fileName)
    if err != nil {
       panic(err)
    }
    if err := pprof.StartCPUProfile(f); err != nil {
       if err := f.Close(); err != nil {
          panic(err)
       }
       panic(err)
    }
    return func() {
       pprof.StopCPUProfile()
       if err := f.Close(); err != nil {
          panic(err)
       }
    }
}

func StartMemProfile(fileName string) (stop func()) {
    f, err := os.Create(fileName)
    if err != nil {
       panic(err)
    }
    return func() {
       defer func() {
          if err := f.Close(); err != nil {
             panic(err)
          }
       }()
       runtime.GC() // get up-to-date statistics
       if err := pprof.WriteHeapProfile(f); err != nil {
          panic(err)
       }
    }
}

例子

go 复制代码

package main

import (
	"github.com/anthony-dong/golang/pkg/pprof"
	"sync"
	"time"
)

func main() {
	// 记录 cup pprof
	//stop := pprof.StartCPUProfile("cpu.out")
	//defer stop()

	// 并发测试 sync map的性能
	mm := sync.Map{}
	pprof.ParallelBenchmark("test1", 64, time.Second, func(count int) {
		for i := 0; i < count; i++ {
			mm.Store(i%10000, 1)
		}
	})
	// name=test1 thread=32 duration=1s total=6708009 avg=4.772µs
	// name=test1 thread=64 duration=1s total=6883456 avg=9.3µs
}
=