在 While 循环中嵌套 For 循环的 OpenMP 程序偶尔会挂起

Question

我尝试使用 OpenMP 在 while 循环内并行化 for 循环，并遇到程序间歇性挂起的问题，特别是当

condition

变量接近 1 时。下面是简化的代码片段：

#include <omp.h>
#include <stdio.h>

void task(int thread_id, int condition) {
    printf("Hello from thread %d with condition %d\n", thread_id, condition);
}

int main() {
    int condition = 5;
    #pragma omp parallel num_threads(4) shared(condition)
    {
        while(condition) {
            #pragma omp master
            {
                // Update condition in master thread
                condition--;
                #pragma omp flush(condition)
            }
            #pragma omp barrier // Wait for master to update condition
    
            #pragma omp for
            for(int i = 0; i < 4; ++i) {
                task(omp_get_thread_num(), condition);
            }
        }
    }
    return 0;
}

输出示例（当

condition

为 1 时挂起。理想情况下，程序应在

condition

等于 0 时停止。）：

Hello from thread 0 with condition 4
Hello from thread 1 with condition 4
Hello from thread 2 with condition 4
Hello from thread 3 with condition 4
Hello from thread 3 with condition 3
Hello from thread 1 with condition 3
Hello from thread 0 with condition 3
Hello from thread 2 with condition 3
Hello from thread 2 with condition 2
Hello from thread 3 with condition 2
Hello from thread 1 with condition 2
Hello from thread 0 with condition 2
Hello from thread 1 with condition 1
Hello from thread 3 with condition 1
Hello from thread 0 with condition 1
Hello from thread 2 with condition 1
...

尽管使用

#pragma omp flush(condition)

来确保

condition

变量的可见性并使用

#pragma omp barrier

来同步线程，但程序有时会停止，看似死锁，特别是当

condition

达到 1 时。 while 循环没有按预期退出。

我还尝试在 while 循环末尾添加一个额外的屏障，但问题仍然存在。这可能是由于死锁或其他同步问题造成的吗？任何见解或解决方案将不胜感激。预先感谢您！

Answer 1

编译如下代码会报告代码中的数据争用：

$ clang -O3 -g -fopenmp -fsanitize=thread so-78260058.c
$ ./a.out
WARNING: ThreadSanitizer: data race (pid=80729)
  Read of size 4 at 0x7ffee9bf9638 by thread T2:
    #0 main.omp_outlined_debug__ so-78260058.c:12:15 (a.out+0xe60b8)
    #1 main.omp_outlined so-78260058.c:10:5 (a.out+0xe6345)
    #2 __kmp_invoke_microtask <null> (libomp.so+0xbcbd2)
    #3 main so-78260058.c:10:5 (a.out+0xe6058)

  Previous write of size 4 at 0x7ffee9bf9638 by main thread:
    #0 main.omp_outlined_debug__ so-78260058.c:16:26 (a.out+0xe6101)
    #1 main.omp_outlined so-78260058.c:10:5 (a.out+0xe6345)
    #2 __kmp_invoke_microtask <null> (libomp.so+0xbcbd2)
    #3 main so-78260058.c:10:5 (a.out+0xe6058)

  Location is stack of main thread.

  Location is global '??' at 0x7ffee9bdb000 ([stack]+0x1e638)

  Thread T2 (tid=80732, running) created by main thread at:
    #0 pthread_create llvm-project/compiler-rt/lib/tsan/rtl/tsan_interceptors_posix.cpp:1020:234 (a.out+0x5f1bb)
    #1 __kmp_create_worker <null> (libomp.so+0x9c096)

SUMMARY: ThreadSanitizer: data race so-78260058.c:12:15 in main.omp_outlined_debug__

该报告是关于第 12 行中的读取 (

while(condition

) 和第 16 行中的写入 (

condition--;

)。该列还清楚地标识了条件变量。正如评论中所讨论的，修复数据争用的最小更改是在

master

构造之前添加一个屏障，刷新是多余的，并且在

master

构造之后的屏障也是不必要的，因为在

for

区域的末端。

下面的代码更容易阅读：

#include <omp.h>
#include <stdio.h>

void task(int thread_id, int condition) {
  printf("Hello from thread %d with condition %d\n", thread_id, condition);
}

int main() {
  int condition = 5;
  while (condition) {
    condition--;

#pragma omp parallel for num_threads(4)
    for (int i = 0; i < 4; ++i) {
      task(omp_get_thread_num(), condition);
    }
  }
  return 0;
}

由于大多数 OpenMP 实现都维护线程池，因此生成并行区域的开销不会明显高于两个必要的屏障。 （如果您使用嵌套并行性，请检查

KMP_HOT_TEAMS_MAX_LEVEL

环境变量）