go thread-safe rand

发表于 更新于

go 全局的 rand 是线程安全的, 通过 mutex 来保证, 但是 go 提供的 rand.NewSource 使用的 rngSource 并不是线程安全的
这里提供一种基于 TLS 的实现

mtrnd.go

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
package mtrnd

import (
	"math/rand"
	"sync"
	"time"

	"github.com/vizee/asm/hack"
)

var (
	rndmu sync.Mutex
	rnds  []*rand.Rand
)

func currnd() *rand.Rand {
	pid := hack.ProcPin()
	if pid >= len(rnds) {
		hack.ProcUnpin()
		rndmu.Lock()
		pid = hack.ProcPin()
		if pid >= len(rnds) {
			t := make([]*rand.Rand, pid+1)
			n := copy(t, rnds)
			for i := n; i < len(t); i++ {
				t[i] = rand.New(rand.NewSource(time.Now().UnixNano() ^ int64(i)))
			}
			rnds = t
		}
		rndmu.Unlock()
	}
	r := rnds[pid]
	hack.ProcUnpin()
	return r
}

func Int() int {
	return currnd().Int()
}

func Intn(n int) int {
	return currnd().Intn(n)
}

func Uint64() uint64 {
	return currnd().Uint64()
}

mtrnd_test.go:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
package mtrnd

import (
	"math/rand"
	"testing"
)

func TestMT(t *testing.T) {
	var n [256]int
	for i := 0; i < 1000000; i++ {
		n[Intn(256)]++
	}
	min := n[0]
	max := n[0]
	for i := 0; i < 256; i++ {
		if n[i] < min {
			min = n[i]
		}
		if n[i] > max {
			max = n[i]
		}
		t.Log(n[i])
	}
	t.Log("diff", max-min)
}

func BenchmarkGlobal(b *testing.B) {
	for i := 0; i < b.N; i++ {
		_ = rand.Int()
	}
}

func BenchmarkMT(b *testing.B) {
	for i := 0; i < b.N; i++ {
		_ = Int()
	}
}

func BenchmarkGlobalP(b *testing.B) {
	b.RunParallel(func(pb *testing.PB) {
		for pb.Next() {
			_ = rand.Int()
		}
	})
}

func BenchmarkMTP(b *testing.B) {
	b.RunParallel(func(pb *testing.PB) {
		for pb.Next() {
			_ = Int()
		}
	})
}

benchmark result (i7 4c8t):

1
2
3
4
BenchmarkGlobal-8    	50000000	        32.5 ns/op
BenchmarkMT-8        	50000000	        29.4 ns/op
BenchmarkGlobalP-8   	10000000	       170 ns/op
BenchmarkMTP-8       	200000000	         7.06 ns/op