go 性能测试

例子

源码文件

//fib.go
package main

//
func FibRaw(x int) int {
	if x == 0 {
		return 0
	} else if x <= 2 {
		return 1
	} else {
		return FibRaw(x-2) + FibRaw(x-1)
	}
}

//
var cache = make(map[int]int)

func FibCache(x int) int {
	if val, found := cache[x]; found {
		return val
	}

	var ret int
	if x == 0 {
		ret = 0
	} else if x <= 2 {
		ret = 1
	} else {
		ret = FibCache(x-2) + FibCache(x-1)
	}

	cache[x] = ret
	return ret
}

测试文件

//fib_test.go
package main
//bench
func BenchmarkFibRaw(b *testing.B) {
	for i := 0; i < b.N; i++ {
		FibRaw(i % 40)
	}
}

func BenchmarkFibCache(b *testing.B) {
	for i := 0; i < b.N; i++ {
		FibCache(i % 40)
	}
}

//prof
func fibCacheLoop(times int) {
	for i := 0; i < times; i++ {
		FibCache(40)
	}
}
func BenchmarkFibPro(b *testing.B) {

	FibRaw(40)
	fibCacheLoop(30000000)
}

性能测试

运行全部测试用例,用 . 模式匹配所有性能测试用例

sfw@ux ~/src/fib/output  $ go test -run=xxx -bench=. ../.
goos: linux
goarch: amd64
BenchmarkFibRaw-2            100          99909281 ns/op
BenchmarkFibCache-2     20000000               116 ns/op
BenchmarkFibPro-2              1        6975007074 ns/op
PASS
ok      _/home/sfw/src/fib      19.446s

运行某一类测试用例,比如 FibCache

sfw@ux ~/src/fib/output  $ go test -run=xxx -bench=FibCache ../.
goos: linux
goarch: amd64
BenchmarkFibCache-2     20000000               116 ns/op
PASS
ok      _/home/sfw/src/fib      2.457s

性能剖析

涉及的工具

• pprof ( go 自带 )
• graphviz ( 第三方 )
• go-torch ( 第三方 )

cpu 性能剖析

-cpurofile 参数生成性能文件

sfw@ux ~/src/fib/output  $ go test -run=xxx -bench=FibPro ../. -cpuprofile=cpu.prof
goos: linux
goarch: amd64
BenchmarkFibPro-2              1        6215340183 ns/op
PASS
ok      _/home/sfw/src/fib      6.428s
-rwxrwxrwx 1 root root 5.1K 5月  29 16:44 cpu.prof*

1, -run=xxx 这里 “xxx” 模式匹配不到任何功能测试函数,所以不执行功能测试.还可写成 -run=NONE, -run=“none” 等匹配不到但又能自解释的模式.

2, Test 和 Benchmark 测试用例中不用加入 runtime.pprof 包中的性能测函数,已经默认打开了.只需指定输出文件即可 -cpuprofile=cpu.prof. 这对要对某个函数进行性能剖析非常方便.

3, Test 函数一般不导出性能剖析文件,因为他们的用途是验证功能的正确与否. Bench 函数是专门用来剖析性能的,正适合导出性能剖析文件,进行性能刨析.

pprof 工具解析性能文件

sfw@ux ~/src/fib/output  $ go tool pprof fib.test cpu.prof
File: fib.test
Type: cpu
Time: May 29, 2019 at 5:23pm (CST)
Duration: 8.76s, Total samples = 8.50s (97.00%)
Entering interactive mode (type "help" for commands, "o" for options)
(pprof) top  // <--- top 指令
Showing nodes accounting for 8440ms, 99.29% of 8500ms total
Dropped 14 nodes (cum <= 42.50ms)
      flat  flat%   sum%        cum   cum%
    3530ms 41.53% 41.53%     3530ms 41.53%  _/home/sfw/src/fib.FibRaw
    3320ms 39.06% 80.59%     4020ms 47.29%  runtime.mapaccess2_fast64
     570ms  6.71% 87.29%      570ms  6.71%  runtime.aeshash64
     520ms  6.12% 93.41%     4540ms 53.41%  _/home/sfw/src/fib.FibCache
     410ms  4.82% 98.24%     4950ms 58.24%  _/home/sfw/src/fib.fibCacheLoop
      90ms  1.06% 99.29%       90ms  1.06%  runtime.add
         0     0% 99.29%     8480ms 99.76%  _/home/sfw/src/fib.BenchmarkFibPro
         0     0% 99.29%     8480ms 99.76%  testing.(*B).run1.func1
         0     0% 99.29%     8480ms 99.76%  testing.(*B).runN
(pprof) list FibCache // <--- list 指令
Total: 8.50s
ROUTINE ======================== _/home/sfw/src/fib.FibCache in /home/sfw/src/fib/fib.go
     520ms      4.54s (flat, cum) 53.41% of Total
         .          .     12:}
         .          .     13:
         .          .     14://
         .          .     15:var cache = make(map[int]int)
         .          .     16:
     150ms      150ms     17:func FibCache(x int) int {
     240ms      4.26s     18:   if val, found := cache[x]; found {
     130ms      130ms     19:           return val
         .          .     20:   }
         .          .     21:
         .          .     22:   var ret int
         .          .     23:   if x == 0 {
         .          .     24:           ret = 0     

graphviz 的 svg 指令生成 svg 文件

(pprof) svg	// <--- svg 指令
Generating report in profile001.svg

1,从图中可以看出 FibRaw 和 fibCacheLoop 的耗时比是 2:3, fibCacheLoop 调用了 30000000 次 FibCache, 所以 FibRaw 的 cpu 耗时是 FibCache 的 20000000 倍(忽略函数调,用缓存等).

2,如果要对比的两个函数性能相差太多时,不作放大处理(如 FibCache),值小的那个可能会显示0,就是被忽略了,那样就没法量化比较了,需要放大小的一方.

go-torch 生成 svg 图

sfw@ux ~/src/fib/output  $ go-torch cpu.prof ../.
INFO[17:38:48] Run pprof command: go tool pprof -raw -seconds 30 cpu.prof ../.
INFO[17:38:48] Writing svg to torch.svg