c - MPI_Gatherv，MPI_Gatherv : Pending request (no error), 错误堆栈中的 fatal error :

Question

编辑#1:

所以解决方案是:

线

MPI_Gatherv(buffer, rank, MPI_INT, buffer, receive_counts, receive_displacements, MPI_INT, 0, MPI_COMM_WORLD);

必须改为

MPI_Gatherv(buffer, receive_counts[rank], MPI_INT, buffer, receive_counts, receive_displacements, MPI_INT, 0, MPI_COMM_WORLD);

再次感谢您的帮助

---

原帖:

我的代码来自DeinoMPI

当我运行 mpiexec -localonly 4 skusamGatherv.exe 时，一切都ok。

如果我换行

int receive_counts[4] = { 0, 1, 2, 3 };

到

int receive_counts[4] = { 0, 1, 2, 1 };

编译仍然可以，但是当我运行 mpiexec -localonly 4 skusamGatherv.exe 时，我会得到错误

我认为它应该工作

感谢帮助

---

我会得到错误:

Fatal error in MPI_Gatherv: Message truncated, error stack:
MPI_Gatherv(363)........................: MPI_Gatherv failed(sbuf=0012FF4C, scou
nt=0, MPI_INT, rbuf=0012FF2C, rcnts=0012FEF0, displs=0012FED8, MPI_INT, root=0,
MPI_COMM_WORLD) failed
MPIDI_CH3_PktHandler_EagerShortSend(351): Message from rank 3 and tag 4 truncate
d; 12 bytes received but buffer size is 4
unable to read the cmd header on the pmi context, Error = -1
.
0. [0][0][0][0][0][0] , [0][0][0][0][0][0]
Error posting readv, An existing connection was forcibly closed by the remote ho
st.(10054)
unable to read the cmd header on the pmi context, Error = -1
.
Error posting readv, An existing connection was forcibly closed by the remote ho
st.(10054)
1. [1][1][1][1][1][1] , [0][0][0][0][0][0]
unable to read the cmd header on the pmi context, Error = -1
.
Error posting readv, An existing connection was forcibly closed by the remote ho
st.(10054)
2. [2][2][2][2][2][2] , [0][0][0][0][0][0]
unable to read the cmd header on the pmi context, Error = -1
.
Error posting readv, An existing connection was forcibly closed by the remote ho
st.(10054)
3. [3][3][3][3][3][3] , [0][0][0][0][0][0]

job aborted:
rank: node: exit code[: error message]
0: jan-pc-nb: 1: Fatal error in MPI_Gatherv: Message truncated, error stack:
MPI_Gatherv(363)........................: MPI_Gatherv failed(sbuf=0012FF4C, scou
nt=0, MPI_INT, rbuf=0012FF2C, rcnts=0012FEF0, displs=0012FED8, MPI_INT, root=0,
MPI_COMM_WORLD) failed
MPIDI_CH3_PktHandler_EagerShortSend(351): Message from rank 3 and tag 4 truncate
d; 12 bytes received but buffer size is 4
1: jan-pc-nb: 1
2: jan-pc-nb: 1
3: jan-pc-nb: 1
Press any key to continue . . .

我的代码:

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

int main(int argc, char *argv[])
{
    int buffer[6];
    int rank, size, i;
    int receive_counts[4] = { 0, 1, 2, 3 };
    int receive_displacements[4] = { 0, 0, 1, 3 };

    MPI_Init(&argc, &argv);
    MPI_Comm_size(MPI_COMM_WORLD, &size);
    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
    if (size != 4)
    {
        if (rank == 0)
        {
            printf("Please run with 4 processes\n");fflush(stdout);
        }
        MPI_Finalize();
        return 0;
    }
    for (i=0; i<rank; i++)
    {
        buffer[i] = rank;
    }
    MPI_Gatherv(buffer, rank, MPI_INT, buffer, receive_counts, receive_displacements, MPI_INT, 0, MPI_COMM_WORLD);
    if (rank == 0)
    {
        for (i=0; i<6; i++)
        {
            printf("[%d]", buffer[i]);
        }
        printf("\n");
        fflush(stdout);
    }
    MPI_Finalize();
    return 0;
}

Answer 1

退后一步，考虑一下 MPI_Gatherv 在做什么:它是一个 MPI_Gather(在本例中为 0 级)，每个处理器可以在其中发送不同数量的数据。

在您的示例中，等级 0 发送 0 个整数，等级 1 发送 1 个整数，等级 2 发送 2 个整数，等级 3 发送 3 个整数。

MPIDI_CH3_PktHandler_EagerShortSend(351): Message from rank 3 and tag 4 truncated; 12 bytes received but buffer size is 4

它隐藏在许多其他信息中，但它表示第 3 级发送了 3 个整数(12 字节)，但第 0 级只有 1 个整数的空间。

查看 gatherv 的前三个参数:'buffer, rank, MPI_INT'。无论您将 receive 设置为什么，等级 3 总是会发送 3 个整数。

请注意，您可以填充缓冲区不足(例如，您可以将 receive_counts 中的最后一项设置为 100)，但您告诉 MPI 库使用更小的 receive_counts[3] 只期望 1 个 int，即使你发送了 3 个。