定长遍历数组的访问时间答案

【问题标题】：The accessing time of traversal a array via a constant length定长遍历数组的访问时间
【发布时间】：2013-04-27 07:51:48
【问题描述】：

我想测量缓存的速度（我问过question），所以我写了这段代码：

#include<cstdio>
#include<ctime>
#include<windows.h>
const int mod = 100000007;
volatile int f[100000007];
volatile int *p;
int main()
{
    FILE *f1;
    f1=fopen("CacheTest.txt","w");
    for(int l=1;l<=100000;)
    {
        clock_t st,ed;
        st=clock();
        int now=0;
        for(int i=0;i<mod;i++)
        {
            now+=l;//note:mod is a prime and that will traversal the array
            if (now>=mod) now-=mod;
        }
        ed=clock();
        double extime=(double)(ed-st)/CLOCKS_PER_SEC;
        st=clock();
        now=0;
        for(int i=0;i<mod;i++)
        {
            p=&f[now];
            *p=1;
            now+=l;
            if (now>=mod) now-=mod;
        }
        ed=clock();
        double cost=(double)(ed-st)/CLOCKS_PER_SEC;
        fprintf(f1,"length %d costs %f seconds access time:%f\n",l,cost,cost-extime);
        printf("length %d costs %f seconds access time:%f\n",l,cost,cost-extime);
        if (l<50) l++;
        else if (l<100) l+=2;
        else if (l<1000) l+=20;
        else if (l<10000) l+=200;
        else if (l<100000) l+=2000;
        else l+=20000;
    }
}

结果如下：

length 1 costs 0.994000 seconds access time:0.153000
length 2 costs 0.651000 seconds access time:-0.048000
length 3 costs 1.034000 seconds access time:0.414000
length 4 costs 1.151000 seconds access time:0.018000
length 5 costs 1.221000 seconds access time:0.086000
length 6 costs 1.313000 seconds access time:0.179000
length 7 costs 1.475000 seconds access time:0.442000
length 8 costs 1.584000 seconds access time:0.437000
length 9 costs 1.700000 seconds access time:0.567000
length 10 costs 1.672000 seconds access time:0.665000
length 11 costs 1.578000 seconds access time:0.573000
length 12 costs 2.166000 seconds access time:1.280000
length 13 costs 2.070000 seconds access time:0.988000
length 14 costs 2.430000 seconds access time:1.402000
length 15 costs 2.564000 seconds access time:1.428000
length 16 costs 2.651000 seconds access time:1.519000
length 17 costs 2.721000 seconds access time:1.654000
length 18 costs 2.648000 seconds access time:1.518000
length 19 costs 2.775000 seconds access time:1.638000
length 20 costs 2.757000 seconds access time:1.636000
length 21 costs 2.816000 seconds access time:1.684000
length 22 costs 2.842000 seconds access time:1.706000
length 23 costs 2.856000 seconds access time:1.751000
length 24 costs 2.702000 seconds access time:1.568000
length 25 costs 2.624000 seconds access time:1.618000
length 26 costs 2.912000 seconds access time:1.777000
length 27 costs 2.617000 seconds access time:1.468000
length 28 costs 2.910000 seconds access time:1.778000
length 29 costs 2.779000 seconds access time:1.649000
length 30 costs 2.803000 seconds access time:1.727000
length 31 costs 2.596000 seconds access time:1.465000
length 32 costs 2.068000 seconds access time:1.188000
length 33 costs 2.012000 seconds access time:1.294000
length 34 costs 3.258000 seconds access time:2.381000
length 35 costs 3.538000 seconds access time:2.406000
length 36 costs 3.619000 seconds access time:2.556000
length 37 costs 3.847000 seconds access time:2.717000
length 38 costs 3.958000 seconds access time:2.827000
length 39 costs 3.910000 seconds access time:2.841000
length 40 costs 3.722000 seconds access time:2.592000
length 41 costs 4.095000 seconds access time:2.960000
length 42 costs 3.440000 seconds access time:2.420000
length 43 costs 4.217000 seconds access time:3.085000
length 44 costs 4.413000 seconds access time:3.457000
length 45 costs 4.637000 seconds access time:3.507000
length 46 costs 4.248000 seconds access time:3.115000
length 47 costs 4.924000 seconds access time:3.856000
length 48 costs 5.072000 seconds access time:3.942000
length 49 costs 4.619000 seconds access time:3.615000
length 50 costs 5.025000 seconds access time:4.136000
length 52 costs 5.196000 seconds access time:4.059000
length 54 costs 4.937000 seconds access time:3.806000
length 56 costs 5.264000 seconds access time:4.132000
length 58 costs 5.137000 seconds access time:4.004000
length 60 costs 5.113000 seconds access time:4.107000
length 62 costs 4.665000 seconds access time:3.537000
length 64 costs 5.030000 seconds access time:3.900000
length 66 costs 4.969000 seconds access time:4.016000
length 68 costs 5.201000 seconds access time:4.096000
length 70 costs 5.066000 seconds access time:3.931000
...
for the length 70~10000 , the access time is around 4s.

前几行访问时间很低，不知道为什么。

然后程序就稳定下来了，访问时间在0.4s左右，我认为原因是大部分访问都命中了L1级缓存。

随着长度的增加，时间变为 1.5s 左右。也许 L1 缓存不起作用，但 L2 缓存可以。

那么时间到了4s，可能根本没有命中缓存？

上面的结果似乎总体上符合我的预期。

但是！发生了一些意想不到的事情，我很惊讶。

当长度超过20000时，访问时间变为：

length 20000 costs 1.828000 seconds access time:1.453000
length 22000 costs 1.867000 seconds access time:1.492000
length 24000 costs 1.959000 seconds access time:1.584000
length 26000 costs 1.977000 seconds access time:1.603000
length 28000 costs 1.987000 seconds access time:1.613000
length 30000 costs 2.045000 seconds access time:1.669000
length 32000 costs 1.951000 seconds access time:1.575000
length 34000 costs 1.980000 seconds access time:1.611000
length 36000 costs 1.988000 seconds access time:1.613000
length 38000 costs 1.993000 seconds access time:1.619000
length 40000 costs 1.850000 seconds access time:1.473000
length 42000 costs 2.010000 seconds access time:1.634000
length 44000 costs 1.991000 seconds access time:1.615000
length 46000 costs 1.977000 seconds access time:1.601000
length 48000 costs 1.977000 seconds access time:1.601000
length 50000 costs 1.927000 seconds access time:1.552000
length 52000 costs 1.907000 seconds access time:1.531000
length 54000 costs 2.013000 seconds access time:1.637000
length 56000 costs 2.014000 seconds access time:1.639000
length 58000 costs 2.011000 seconds access time:1.640000
length 60000 costs 1.978000 seconds access time:1.606000
length 62000 costs 2.046000 seconds access time:1.672000
length 64000 costs 2.076000 seconds access time:1.700000
length 66000 costs 2.068000 seconds access time:1.693000
length 68000 costs 2.014000 seconds access time:1.639000
length 70000 costs 2.123000 seconds access time:1.748000
length 72000 costs 2.032000 seconds access time:1.646000
length 74000 costs 2.015000 seconds access time:1.638000
length 76000 costs 2.023000 seconds access time:1.646000
length 78000 costs 1.873000 seconds access time:1.497000
length 80000 costs 1.655000 seconds access time:1.278000

谁能解释为什么会发生这些奇怪的事情？我现在被卡住了，对此一无所知。

提前致谢。

【问题讨论】：

我无法解释，但我会分享我系统的数据。我把素数减少到 10000019，因为我没有你那么多的记忆。对于小长度，它逐渐增长到 1.6 访问时间（长度 8），然后保持稳定直到长度 2400，然后增长到 2.1 访问时间（长度 4200）并保持稳定，除了长度 4800、6400 下降到 1.7， 9600, 16000, 20000, 32000, 40000, 50000, 64000, 80000, 100000。这就是你所期望的，我想。您确定您的系统未被使用吗？
你是对的，我再次测试我的代码，现在这些事情不再发生了。
只是好奇，你有什么样的 CPU/缓存级别？
我的CPU是INTEL I5 3317U

标签： c caching time hardware

【解决方案1】：

当你测试这样的东西时，你应该让你的系统处于空闲状态。

我在电脑上测试了3次，长度>5000时保持3s。

【讨论】：