缓存行对齐优化不会减少缓存未命中

Cache line alignment optimization not reducing cache miss

我得到了这段代码,演示了缓存行对齐优化如何通过从 http://blog.kongfy.com/2016/10/cache-coherence-sequential-consistency-and-memory-barrier/

减少 'false sharing' 来工作

代码:

/*
 * Demo program for showing the drawback of "false sharing"
 *
 * Use it with perf!
 *
 * Compile: g++ -O2 -o false_share false_share.cpp -lpthread
 * Usage: perf stat -e cache-misses ./false_share <loopcount> <is_aligned>
 */

#include <pthread.h>
#include <stdio.h>
#include <stdlib.h>
#include <time.h>
#include <sys/time.h>
#include <sys/resource.h>

#define CACHE_ALIGN_SIZE 64
#define CACHE_ALIGNED __attribute__((aligned(CACHE_ALIGN_SIZE)))

int gLoopCount;

inline int64_t current_time()
{
  struct timeval t;
  if (gettimeofday(&t, NULL) < 0) {
  }
  return (static_cast<int64_t>(t.tv_sec) * static_cast<int64_t>(1000000) + static_cast<int64_t>(t.tv_usec));
}

struct value {
  int64_t val;
};
value data[2] CACHE_ALIGNED;

struct aligned_value {
  int64_t val;
} CACHE_ALIGNED;
aligned_value aligned_data[2] CACHE_ALIGNED;

void* worker1(int64_t *val)
{
  printf("worker1 start...\n");

  volatile int64_t &v = *val;
  for (int i = 0; i < gLoopCount; ++i) {
    v += 1;
  }

  printf("worker1 exit...\n");
}

// duplicate worker function for perf report
void* worker2(int64_t *val)
{
  printf("worker2 start...\n");

  volatile int64_t &v = *val;
  for (int i = 0; i < gLoopCount; ++i) {
    v += 1;
  }

  printf("worker2 exit...\n");
}

int main(int argc, char *argv[])
{
  pthread_t race_thread_1;
  pthread_t race_thread_2;

  bool is_aligned;

  /* Check arguments to program*/
  if(argc != 3) {
    fprintf(stderr, "USAGE: %s <loopcount> <is_aligned>\n", argv[0]);
    exit(1);
  }

  /* Parse argument */
  gLoopCount = atoi(argv[1]); /* Don't bother with format checking */
  is_aligned = atoi(argv[2]); /* Don't bother with format checking */

  printf("size of unaligned data : %d\n", sizeof(data));
  printf("size of aligned data   : %d\n", sizeof(aligned_data));

  void *val_0, *val_1;
  if (is_aligned) {
    val_0 = (void *)&aligned_data[0].val;
    val_1 = (void *)&aligned_data[1].val;
  } else {
    val_0 = (void *)&data[0].val;
    val_1 = (void *)&data[1].val;
  }

  int64_t start_time = current_time();

  /* Start the threads */
  pthread_create(&race_thread_1, NULL, (void* (*)(void*))worker1, val_0);
  pthread_create(&race_thread_2, NULL, (void* (*)(void*))worker2, val_1);

  /* Wait for the threads to end */
  pthread_join(race_thread_1, NULL);
  pthread_join(race_thread_2, NULL);

  int64_t end_time = current_time();

  printf("time : %d us\n", end_time - start_time);

  return 0;
}

预期性能结果:

[jingyan.kfy@OceanBase224006 work]$ perf stat -e cache-misses ./false_share 100000000 0
size of unaligned data : 16
size of aligned data   : 128
worker2 start...
worker1 start...
worker1 exit...
worker2 exit...
time : 452451 us

 Performance counter stats for './false_share 100000000 0':

         3,105,245 cache-misses

       0.455033803 seconds time elapsed

[jingyan.kfy@OceanBase224006 work]$ perf stat -e cache-misses ./false_share 100000000 1
size of unaligned data : 16
size of aligned data   : 128
worker1 start...
worker2 start...
worker1 exit...
worker2 exit...
time : 326994 us

 Performance counter stats for './false_share 100000000 1':

            27,735 cache-misses

       0.329737667 seconds time elapsed

然而,我 运行 自己编写代码并得到非常接近 运行 的时间,未对齐时缓存未命中计数甚至更低:

我的结果:

$ perf stat -e cache-misses ./false_share 100000000 0
size of unaligned data : 16
size of aligned data   : 128
worker1 start...
worker2 start...
worker2 exit...
worker1 exit...
time : 169465 us

 Performance counter stats for './false_share 100000000 0':

            37,698      cache-misses:u                                              

       0.171625603 seconds time elapsed

       0.334919000 seconds user
       0.001988000 seconds sys


$ perf stat -e cache-misses ./false_share 100000000 1
size of unaligned data : 16
size of aligned data   : 128
worker2 start...
worker1 start...
worker2 exit...
worker1 exit...
time : 118798 us

 Performance counter stats for './false_share 100000000 1':

            38,375      cache-misses:u                                              

       0.121072715 seconds time elapsed

       0.230043000 seconds user
       0.001973000 seconds sys

我应该如何理解这种不一致?

由于您引用的博客是中文的,因此很难提供帮助。不过,我注意到第一个图似乎显示了多路架构。所以我做了一些实验。

a) 我的 PC,Intel(R) Core(TM) i7-2600K CPU @ 3.40GHz,单插槽,两个内核,每个内核两个 threes:

0:

time : 195389 us

 Performance counter stats for './a.out 100000000 0':

             8 980      cache-misses:u                                              

       0,198584628 seconds time elapsed

       0,391694000 seconds user
       0,000000000 seconds sys

和 1:

time : 191413 us

 Performance counter stats for './a.out 100000000 1':

             9 020      cache-misses:u                                              

       0,192953853 seconds time elapsed

       0,378434000 seconds user
       0,000000000 seconds sys

差别不大

b) 现在是 2 路工作站

Thread(s) per core: 2
Core(s) per socket: 12
Socket(s): 2
NUMA node(s): 2
Model name: Intel(R) Xeon(R) CPU E5-2650 v4 @ 2.20GHz

0:

time : 454679 us

 Performance counter stats for './a.out 100000000 0':

         5,644,133      cache-misses                                                

       0.456665966 seconds time elapsed

       0.738173000 seconds user

1:

time : 346871 us

 Performance counter stats for './a.out 100000000 1':

            42,217      cache-misses                                                

       0.348814583 seconds time elapsed

       0.539676000 seconds user
       0.000000000 seconds sys

差别很大


最后一点。你写:

the cache miss count is even lower when NOT ALIGNED

不,不是。除了您的程序之外,您的处理器还有 运行 各种任务。此外,您是 运行 2 个线程,它们可能以不同的时间顺序访问缓存。所有这些都可能影响缓存利用率。您需要多次重复测量并进行比较。就个人而言,当我看到任何性能结果的差异小于 10% 时,我认为它们无法区分。


更新

我还对您的代码进行了扩展到 3 个线程的实验,因此其中一些肯定必须 运行 在不同的内核上,因此,仅共享 L3 缓存。

我查看了 How to catch the L3-cache hits and misses by perf tool in Linux 并得到了这个命令:

 perf stat -e cache-misses,cache-references,LLC-loads,LLC-stores,L1-dcache-load-misses,L1-dcache-prefetch-misses,L1-dcache-store-misses ./a.out 100000000 0

0:

time : 214253 us

 Performance counter stats for './a.out 100000000 0':

             4 765      cache-misses:u            #    0,018 % of all cache refs      (57,39%)
        25 992 887      cache-references:u                                            (57,56%)
        17 430 736      LLC-loads:u                                                   (57,56%)
         8 591 378      LLC-stores:u                                                  (57,56%)
        28 110 342      L1-dcache-load-misses:u                                       (57,40%)
        14 661 378      L1-dcache-prefetch-misses:u                                     (57,80%)
            32 269      L1-dcache-store-misses:u                                      (57,49%)

       0,215484922 seconds time elapsed

       0,627426000 seconds user
       0,006635000 seconds sys

1:

time : 194253 us

 Performance counter stats for './a.out 100000000 1':

             4 509      cache-misses:u            #   30,715 % of all cache refs      (57,15%)
            14 680      cache-references:u                                            (57,45%)
             7 954      LLC-loads:u                                                   (57,49%)
             1 565      LLC-stores:u                                                  (57,92%)
             4 442      L1-dcache-load-misses:u                                       (57,91%)
               836      L1-dcache-prefetch-misses:u                                     (57,02%)
               984      L1-dcache-store-misses:u                                      (56,85%)

       0,195145645 seconds time elapsed

       0,569986000 seconds user
       0,000000000 seconds sys

因此:

  • 对齐(3 线程)版本系统地(有点)比未对齐(我重复测试了几次)更快,即使在单路机器上也是如此。
  • 不太清楚“缓存未命中”选项实际报告的是什么
  • L1 缓存、LLC 缓存和缓存引用数量中的“虚假数据共享”会受到巨大的(数字)惩罚。
  • 请记住,这些是基于硬件的统计数据:如果其他进程是 运行,它们会将自己的贡献添加到这些结果中