MPI Asynchoronous发送和接收未按预期工作

时间:2015-10-20 08:41:26

标签: c parallel-processing mpi garbage

我对使用C中的MPI库进行通信有疑问。这是我的示例代码:

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

int main (int argc, char **argv) {

    // Initialize the MPI Environment
    MPI_Init (&argc, &argv);

    int world_size;
    MPI_Comm_size (MPI_COMM_WORLD, &world_size);

    int world_rank;
    MPI_Comm_rank (MPI_COMM_WORLD, &world_rank);

    int workers = world_size - 1;

    if (world_rank == 0) {
    }
    else if (world_rank == 1) {
        int to_recv;
        int to_send = world_rank;
        MPI_Request request1, request2;
        MPI_Irecv (&to_recv, sizeof (int), MPI_INT, world_rank + 1, 0, MPI_COMM_WORLD, &request1);
        MPI_Isend (&to_send, sizeof (int), MPI_INT, world_rank + 1, 0, MPI_COMM_WORLD, &request2);

        MPI_Wait (&request1, MPI_STATUS_IGNORE);
        MPI_Wait (&request2, MPI_STATUS_IGNORE);

        printf("I am %d\n", world_rank);
        printf("I received rank %d\n", to_recv);
    }
    else if (world_rank == workers) {
        int to_recv;
        int to_send = world_rank;
        MPI_Request request1, request2;
        MPI_Irecv (&to_recv, sizeof (int), MPI_INT, world_rank - 1, 0, MPI_COMM_WORLD, &request1);
        MPI_Isend (&to_send, sizeof (int), MPI_INT, world_rank - 1, 0, MPI_COMM_WORLD, &request2);

        MPI_Wait (&request1, MPI_STATUS_IGNORE);
        MPI_Wait (&request2, MPI_STATUS_IGNORE);

        printf("I am %d\n", world_rank);
        printf("I received rank %d\n", to_recv);
    }
    else {
        int to_recvl, to_recvr;
        int to_send1 = world_rank, to_send2 = world_rank;

        MPI_Request request1, request2, request3, request4;

        MPI_Irecv (&to_recvl, sizeof (int), MPI_INT, world_rank - 1, 0, MPI_COMM_WORLD, &request1);
        MPI_Isend (&to_send1, sizeof (int), MPI_INT, world_rank - 1, 0, MPI_COMM_WORLD, &request2);

        MPI_Irecv (&to_recvr, sizeof (int), MPI_INT, world_rank + 1, 0, MPI_COMM_WORLD, &request3);
        MPI_Isend (&to_send2, sizeof (int), MPI_INT, world_rank + 1, 0, MPI_COMM_WORLD, &request4);

        MPI_Wait (&request1, MPI_STATUS_IGNORE);
        MPI_Wait (&request2, MPI_STATUS_IGNORE);

        MPI_Wait (&request3, MPI_STATUS_IGNORE);
        MPI_Wait (&request4, MPI_STATUS_IGNORE);

        printf("I am %d\n", world_rank);
        printf("I received ranks %d and %d\n", to_recvl, to_recvr);
    }

    MPI_Finalize ();

    return 0;
}

所以这就是我基本上想要完成的事情:

假设我生成了11个进程(0表示主进程,进程为1-10个进程)。然后,

  • 进程1将其排名发送到进程2
  • 流程1从流程2获得排名
  • 过程10将其等级发送到过程9
  • 过程10从过程9接收等级
  • 进程i将其等级发送到处理i + 1和i-1(1&lt; i&lt; 10)
  • 处理i从处理i + 1和i-1(1&lt; i&lt; 10)
  • 接收等级

在运行程序时,我会在某些情况下获得垃圾值,如下面添加的那样:

I am 2
I received ranks 1 and 3
I am 3
I received ranks 2 and 4
I am 4
I received ranks 1499161032 and 5
I am 6
I received ranks 1488867784 and 7
I am 9
I received rank 8
I am 32767
I received rank 2
I am 5
I received ranks 1507000776 and 6
I am 7
I received ranks 6 and 8
I am 8
I received ranks 7 and 9

我不知道造成这个问题的原因。任何帮助将不胜感激。

谢谢!

1 个答案:

答案 0 :(得分:5)

一个非常简单的问题,MPI_Isend()MPI_Irecv()中的第二个参数应该是您希望发送或接收的count元素。您在大多数系统中使用sizeof(int) 4,而您只想发送/接收1个元素,因此计数应为1.