The go-labs's discuss from bg5sbk

go-labs's Issues

链表遍历可能需要考虑cache miss

测试环境的数据分布在连续内存中：

func init() {
        for i := 0; i < 1000; i++ {
                var newData = new(BigStruct)
                newData.C30 = i
                newData.next = data
                data = newData
        }
}

真实环境中，玩家数据在内存中的分布可能不是连续的

func init_cache_miss() {
        for i := 0; i < 1800; i++ {
                var newData = new(BigStruct)
                if i%5 >= 3 {
                        newData.C30 = i
                        newData.next = otherUserData
                        otherUserData = newData
                } else {
                        newData.C30 = i
                        newData.next = data
                        data = newData
                }
        }
}

原始版本：

PASS
Benchmark_Loop1-8     300000          5156 ns/op
Benchmark_Loop2-8     200000          5891 ns/op
Benchmark_Loop3-8     100000         14506 ns/op
Benchmark_Loop4-8      20000         95709 ns/op
Benchmark_Loop5-8       3000        447557 ns/op
Benchmark_Loop6-8     200000          8395 ns/op
Benchmark_Loop7-8     200000          9301 ns/op
ok      go-labs/labs07  12.440s

模拟 cache miss 情况

PASS
Benchmark_Loop1-8     200000          9960 ns/op
Benchmark_Loop2-8     200000         10466 ns/op
Benchmark_Loop3-8     100000         18221 ns/op
Benchmark_Loop4-8      10000        107809 ns/op
Benchmark_Loop5-8       3000        507910 ns/op
Benchmark_Loop6-8     200000         11533 ns/op
Benchmark_Loop7-8     200000         11954 ns/op
ok      go-labs/labs07  13.934s

~~我觉得还是 slice 效率更优。因为 slice 可以保证同一个玩家的数据在一段连续内存。~~

labs09_test.go
testing: warning: no tests to run
PASS
Benchmark_NormalFuncCall 2000000000 0.68 ns/op
Benchmark_VarFuncCall1 1000000000 2.37 ns/op
Benchmark_VarFuncCall2 1000000000 2.83 ns/op
ok github.com/realint/labs/src/labs09 7.195s

Recommend Projects