cacheline 对 Go 程序的影响

[smallnest]

https://colobu.com/2019/01/24/cacheline-affects-performance-in-go/

[cybbeta]

好文，在java开发中也经常需要处理缓存行填充

[goith]

里面例子我运行了四次, 并不是你说的情况。

[@sjs_19_176 cacheline]$ go test -gcflags "-N -l" -bench .
goos: linux
goarch: amd64
pkg: demo/cacheline
BenchmarkPad_Increase-4          5294812               232 ns/op
BenchmarkNoPad_Increase-4       10684954               112 ns/op
PASS
ok      demo/cacheline  2.777s
[@sjs_19_176 cacheline]$ go test -gcflags "-N -l" -bench .
goos: linux
goarch: amd64
pkg: demo/cacheline
BenchmarkPad_Increase-4         10496142               119 ns/op
BenchmarkNoPad_Increase-4        7455829               160 ns/op
PASS
ok      demo/cacheline  2.726s
[@sjs_19_176 cacheline]$ go test -gcflags "-N -l" -bench .
goos: linux
goarch: amd64
pkg: demo/cacheline
BenchmarkPad_Increase-4         10242184               201 ns/op
BenchmarkNoPad_Increase-4       11318866               109 ns/op
PASS
ok      demo/cacheline  3.524s
[@sjs_19_176 cacheline]$ go test -gcflags "-N -l" -bench .
goos: linux
goarch: amd64
pkg: demo/cacheline
BenchmarkPad_Increase-4         11115954               117 ns/op
BenchmarkNoPad_Increase-4        7965044               158 ns/op
PASS
ok      demo/cacheline  2.826s
[@sjs_19_176 cacheline]$ 

[smallnest]

一个完整的测试 https://play.golang.org/p/hp_MXQWEJvB

@goith 里面例子我运行了四次, 并不是你说的情况。

[@sjs_19_176 cacheline]$ go test -gcflags "-N -l" -bench .
goos: linux
goarch: amd64
pkg: demo/cacheline
BenchmarkPad_Increase-4          5294812               232 ns/op
BenchmarkNoPad_Increase-4       10684954               112 ns/op
PASS
ok      demo/cacheline  2.777s
[@sjs_19_176 cacheline]$ go test -gcflags "-N -l" -bench .
goos: linux
goarch: amd64
pkg: demo/cacheline
BenchmarkPad_Increase-4         10496142               119 ns/op
BenchmarkNoPad_Increase-4        7455829               160 ns/op
PASS
ok      demo/cacheline  2.726s
[@sjs_19_176 cacheline]$ go test -gcflags "-N -l" -bench .
goos: linux
goarch: amd64
pkg: demo/cacheline
BenchmarkPad_Increase-4         10242184               201 ns/op
BenchmarkNoPad_Increase-4       11318866               109 ns/op
PASS
ok      demo/cacheline  3.524s
[@sjs_19_176 cacheline]$ go test -gcflags "-N -l" -bench .
goos: linux
goarch: amd64
pkg: demo/cacheline
BenchmarkPad_Increase-4         11115954               117 ns/op
BenchmarkNoPad_Increase-4        7965044               158 ns/op
PASS
ok      demo/cacheline  2.826s
[@sjs_19_176 cacheline]$ 

[JOJO0527]

type Pad struct {
	a   uint64
	_p1 [cpu.CacheLinePadSize - unsafe.Sizeof(uint64(1))%cpu.CacheLinePadSize]byte
	b   uint64
	_p2 [cpu.CacheLinePadSize - unsafe.Sizeof(uint64(1))%cpu.CacheLinePadSize]byte
	c   uint64
	_p3 [cpu.CacheLinePadSize - unsafe.Sizeof(uint64(1))%cpu.CacheLinePadSize]byte
}

用这种形式获取padding大小更准确，Mac m1的CacheLinePadSize是128，并不是传统x86的64字节。

package cpu

import "runtime"

// cacheLineSize is used to prevent false sharing of cache lines.
// We choose 128 because Apple Silicon, a.k.a. M1, has 128-byte cache line size.
// It doesn't cost much and is much more future-proof.
const cacheLineSize = 128

[JOJO0527]

goos: darwin
goarch: arm64
pkg: go-praitce/go/app/benchmark/cachelinepadding
BenchmarkPad_Increase
BenchmarkPad_Increase-8     	15356199	        73.22 ns/op
BenchmarkNoPad_Increase
BenchmarkNoPad_Increase-8   	13037071	        94.33 ns/op
PASS

附上m1的测试结果