c - MPI_Barrier 不在循环内工作

Question

我对 MPI 函数运行了一些测试以了解它是如何工作的，并且使用 MPI_Barrier 得到了一个奇怪的结果:如果我在代码中使用它，它会做每个人都期望的事情

int main(int argc, char *argv[])
{
  <some code>
  MPI_Barrier(MPI_COMM_WORLD);
  <more code>
  MPI_Barrier(MPI_COMM_WORLD);
  <...>
}

但是当我从循环内部调用它时，我会得到随机结果。具体来说，我有以下测试代码:

#include "mpi.h" 
#include <stdio.h>

int main(int argc, char *argv[]) 
{ 
  int i, rb, rank, nprocs;

  MPI_Init(&argc,&argv); 
  MPI_Comm_size(MPI_COMM_WORLD,&nprocs); 
  MPI_Comm_rank(MPI_COMM_WORLD,&rank);

  i=0;
  while(i<5)
  {
    rb=MPI_Barrier(MPI_COMM_WORLD);
    printf("Itartion %d.  I am %d of %d. MPIBarrierRes: %d\n", i, rank, nprocs, rb);
    i++;
  } 
  MPI_Finalize(); 
  return 0; 
} 

当我用 3 个任务运行它时，我随机得到:

Itartion 0.  I am 0 of 3. MPIBarrierRes: 0
Itartion 0.  I am 2 of 3. MPIBarrierRes: 0
Itartion 0.  I am 1 of 3. MPIBarrierRes: 0
Itartion 1.  I am 0 of 3. MPIBarrierRes: 0
Itartion 1.  I am 1 of 3. MPIBarrierRes: 0
Itartion 1.  I am 2 of 3. MPIBarrierRes: 0
Itartion 2.  I am 0 of 3. MPIBarrierRes: 0
Itartion 2.  I am 1 of 3. MPIBarrierRes: 0
Itartion 2.  I am 2 of 3. MPIBarrierRes: 0
Itartion 3.  I am 0 of 3. MPIBarrierRes: 0
Itartion 3.  I am 1 of 3. MPIBarrierRes: 0
Itartion 3.  I am 2 of 3. MPIBarrierRes: 0
Itartion 4.  I am 0 of 3. MPIBarrierRes: 0
Itartion 4.  I am 1 of 3. MPIBarrierRes: 0
Itartion 4.  I am 2 of 3. MPIBarrierRes: 0

这是我所期望的，或者只是相反:

Itartion 0.  I am 2 of 3. MPIBarrierRes: 0
Itartion 1.  I am 2 of 3. MPIBarrierRes: 0
Itartion 2.  I am 2 of 3. MPIBarrierRes: 0
Itartion 3.  I am 2 of 3. MPIBarrierRes: 0
Itartion 4.  I am 2 of 3. MPIBarrierRes: 0
Itartion 0.  I am 0 of 3. MPIBarrierRes: 0
Itartion 1.  I am 0 of 3. MPIBarrierRes: 0
Itartion 2.  I am 0 of 3. MPIBarrierRes: 0
Itartion 3.  I am 0 of 3. MPIBarrierRes: 0
Itartion 4.  I am 0 of 3. MPIBarrierRes: 0
Itartion 0.  I am 1 of 3. MPIBarrierRes: 0
Itartion 1.  I am 1 of 3. MPIBarrierRes: 0
Itartion 2.  I am 1 of 3. MPIBarrierRes: 0
Itartion 3.  I am 1 of 3. MPIBarrierRes: 0
Itartion 4.  I am 1 of 3. MPIBarrierRes: 0

或者介于两者之间的东西，比如

Itartion 0.  I am 1 of 3. MPIBarrierRes: 0
Itartion 0.  I am 0 of 3. MPIBarrierRes: 0
Itartion 1.  I am 0 of 3. MPIBarrierRes: 0
Itartion 0.  I am 2 of 3. MPIBarrierRes: 0
Itartion 1.  I am 1 of 3. MPIBarrierRes: 0
Itartion 2.  I am 0 of 3. MPIBarrierRes: 0
Itartion 1.  I am 2 of 3. MPIBarrierRes: 0
Itartion 2.  I am 1 of 3. MPIBarrierRes: 0
Itartion 3.  I am 0 of 3. MPIBarrierRes: 0
Itartion 2.  I am 2 of 3. MPIBarrierRes: 0
Itartion 3.  I am 1 of 3. MPIBarrierRes: 0
Itartion 4.  I am 0 of 3. MPIBarrierRes: 0
Itartion 3.  I am 2 of 3. MPIBarrierRes: 0
Itartion 4.  I am 1 of 3. MPIBarrierRes: 0
Itartion 4.  I am 2 of 3. MPIBarrierRes: 0

谁能告诉我 MPI_Barrier 和循环之间是否存在冲突？ (我只发现了在不同任务中使用不同大小的循环避免死锁的警告。)
如果有，我可以做些什么来强制任务在开始新的循环迭代之前相互等待？
如果没有，这段代码有什么问题？

谢谢!

Answer 1

它与您的循环无关。 MPI 没有暗示 IO 的顺序化。如果您需要按顺序打印出来，则必须明确地将它们发送到一个等级，并在那里全部打印出来。