异步MPI_Irecv的MPI_Waitall错误

Question

我对两个MPI_Irecv都使用了2个MPI_Irecv，然后是2个MPI_Send，然后是MPI_Waitall，如下所示。经过几次计算后，我再次编写了相同的代码块。但似乎MPI进程在第一块代码本身就失败了。

我的通信是这样的，一个矩阵被水平拆分为MPI进程的no，并且通信只发生在矩阵边界之间，矩阵网格下面发送'start‘/第一行到矩阵网格上面，矩阵网格上面发送’end‘/最后一行到矩阵网格下面。

MPI_Request request[2];
MPI_Status status[2];
double grid[size];
double grida[size];
.
.
.
<Calculation for grid2[][]>
...

MPI_Barrier(MPI_COMM_WORLD);
    if (world_rank != 0){
    MPI_Irecv(&grid, size, MPI_DOUBLE, world_rank-1, 0, MPI_COMM_WORLD, &request[1]);
    printf("1 MPI_Irecv");
    }
    if (world_rank != world_size-1){
    MPI_Irecv(&grida, size, MPI_DOUBLE, world_rank+1, 1, MPI_COMM_WORLD, &request[0]);
    printf("2 MPI_Irecv");
    }
    if (world_rank != world_size-1){
    MPI_Send(grid2[end], size, MPI_DOUBLE, world_rank+1, 0, MPI_COMM_WORLD);
    printf("1 MPI_Send");
    }
    if (world_rank != 0){
    MPI_Send(grid2[start], size, MPI_DOUBLE, world_rank-1, 1, MPI_COMM_WORLD);
    printf("2 MPI_Send");
    }
    MPI_Waitall(2, request, status); 

MPI_Barrier(MPI_COMM_WORLD);

.
.
.

<Again the above code but without the initialization of MPI_Request and MPI_Status>

但在这方面，我得到了错误：

 *** Process received signal ***
 Signal: Bus error: 10 (10)
 Signal code: Non-existant physical address (2)
 Failing at address: 0x108bc91e3
 [ 0] 0   libsystem_platform.dylib            0x00007fff50b65f5a _sigtramp + 26
 [ 1] 0   ???                                 0x000000010c61523d 0x0 + 4502671933
 [ 2] 0   libmpi.20.dylib                     0x0000000108bc8e4a MPI_Waitall + 154
 [ 3] 0   dist-jacobi                         0x0000000104b55770 Work + 1488
 [ 4] 0   dist-jacobi                         0x0000000104b54f01 main + 561
 [ 5] 0   libdyld.dylib                       0x00007fff508e5145 start + 1
 [ 6] 0   ???                                 0x0000000000000003 0x0 + 3
 *** End of error message ***
 *** An error occurred in MPI_Waitall
 *** reported by process [1969881089,3]
 *** on communicator MPI_COMM_WORLD
 *** MPI_ERR_REQUEST: invalid request
 *** MPI_ERRORS_ARE_FATAL (processes in this communicator will now abort,
 ***    and potentially your MPI job)
--------------------------------------------------------------------------
mpirun noticed that process rank 0 with PID 0 on node dhcp-10 exited on signal 10 (Bus error: 10).
--------------------------------------------------------------------------

为什么Waitall抛出错误，为什么printf("1 MPI_Irecv");没有被打印出来。这个print语句之前的所有内容都被正确打印出来了。

代码使用MPI_Wait()和MPI_Isend()的方式如下所示：

 // insert barrier 
    MPI_Barrier(MPI_COMM_WORLD);
    if (world_rank != 0){
    MPI_Irecv(&grid, size*2, MPI_DOUBLE, world_rank-1, 0, MPI_COMM_WORLD, &request[0]);
    printf("1 MPI_Irecv");
    }
    if (world_rank != world_size-1){
    MPI_Irecv(&grida, size*2, MPI_DOUBLE, world_rank+1, 1, MPI_COMM_WORLD, &request[1]);
    printf("2 MPI_Irecv");
    }
    if (world_rank != world_size-1){
    MPI_Isend(grid2[end], size*2, MPI_DOUBLE, world_rank+1, 0, MPI_COMM_WORLD, &request[0]);
    printf("1 MPI_Send");
    }
    if (world_rank != 0){
    MPI_Isend(grid2[start], size*2, MPI_DOUBLE, world_rank-1, 1, MPI_COMM_WORLD,  &request[1]);
    printf("2 MPI_Send");
    }
    //MPI_Waitall(2, request, status); 
    MPI_Wait(&request[0], &status[0]);
    MPI_Wait(&request[1], &status[1]);

Answer 1

在最后一列未初始化的情况下使用request[0]，在第一列未初始化的情况下使用request[1]。

一种可能的修复方法是静态地初始化request数组(假设代码中的其他地方没有使用它)。

MPI_Request request[2] = {MPI_REQUEST_NULL, MPI_REQUEST_NULL};

顺便说一句，您可能想要考虑将request重命名为requests，将status重命名为statuses，以清楚地表明这是数组而不是标量。