函数调用开销

使用s_call来测试性能，循环1G次。

2.35,0.00,17,0,0,0,99%
2.34,0.00,13,0,0,0,99%
2.34,0.00,10,0,0,0,100%
2.35,0.00,10,0,0,0,99%
2.34,0.00,14,0,0,0,99%
2.34,0.00,6,0,0,0,99%

统计结果如下：

  • time mean = 2.34
  • time var = 0.000022

每次call的开销为2.34ns，约7个指令周期。当然，这些并没有考虑调用压栈和数据返回。

内核调用开销

使用s_syscall来测试性能，循环1G次。这里特意选用了一个不可能失败的内核函数，getpid，来衡量每次进入getpid的开销。

4.37,0.00,76,0,0,0,99%
4.34,0.00,43,0,0,0,99%
4.37,0.00,124,0,0,0,99%
4.37,0.00,63,0,0,0,99%
4.36,0.00,48,0,0,0,99%
4.36,0.00,47,0,0,0,99%

统计结果如下：

  • time mean = 4.36
  • time var = 0.00011

这里可以看到，纯粹的内核进入开销小到非常惊人，只有4.36ns，约合13个指令周期，而且这里还要进行数据的查询和返回。所以内核调用开销在下面的测试中全部忽略不计。