将矩阵与 MPI 相乘（分段错误）

Question

我最近开始学习 MPI，您可能已经猜到了，我已经遇到了一个我自己无法解决的错误！

我想编写一个将两个矩阵相乘的程序。但是 我还没有走那么远，事实上，我一开始就坚持广播矩阵。

#define MASTER 0
if (rank == MASTER) {
        A = (double *) malloc(N * N * sizeof(double));
        B = (double *) malloc(N * N * sizeof(double));
        matFillRand(N, A);
        matFillRand(N, B);
    }

if (rank == MASTER) {
        P = (double *) malloc(N * N * sizeof(double));
    }

matMulMPI(N, A, B, P);

if (rank == MASTER) {
        printMatrix(N, P);
}

（理论上）进行数学运算的函数如下：

void matMulMPI(long N, double *a, double *b, double *c) {
    long i, j, k;   
    int rank, size;
    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
    MPI_Comm_size(MPI_COMM_WORLD, &size);

    MPI_Bcast(&N, 1, MPI_LONG, MASTER, MPI_COMM_WORLD);

    MPI_Bcast(b, N*N, MPI_DOUBLE, MASTER, MPI_COMM_WORLD);

    printMatrix(N, b);

    //TO-DO: Broadcast A
    //TO-DO: Do Math
}

而且这个广播不起作用。我收到以下消息：

* Process received signal Signal: Segmentation fault (11) Signal code: Invalid permissions (2) Failing at address: 0x401560 Signal: Segmentation fault (11) Signal code: Invalid permissions (2) Failing at address: 0x401560 [ 0] /lib/x86_64-linux-gnu/libpthread.so.0(+0x10340) [0x7fc3ede6b340] [ 1] /lib/x86_64-linux-gnu/libc.so.6(+0x981c0) [0x7fc3edb2e1c0] [ 2] /usr/lib/libmpi.so.1(opal_convertor_unpack+0x105) [0x7fc3ee1788d5] [ 3] /usr/lib/openmpi/lib/openmpi/mca_pml_ob1.so(mca_pml_ob1_recv_frag_callback_match+0x460) [0x7fc3e6587630] [ 4] /usr/lib/openmpi/lib/openmpi/mca_btl_sm.so(mca_btl_sm_component_progress+0x487) [0x7fc3e572a137] [ 5] /usr/lib/libmpi.so.1(opal_progress+0x5a) [0x7fc3ee1849ea] [ 6] /usr/lib/libmpi.so.1(ompi_request_default_wait+0x16d) [0x7fc3ee0d1c0d] [ 7] /usr/lib/openmpi/lib/openmpi/mca_coll_tuned.so(ompi_coll_tuned_bcast_intra_generic+0x49e) [0x7fc3e486da9e] [ 8] /usr/lib/openmpi/lib/openmpi/mca_coll_tuned.so(ompi_coll_tuned_bcast_intra_binomial+0xb7) [0x7fc3e486df27] [ 9] /usr/lib/openmpi/lib/openmpi/mca_coll_tuned.so(ompi_coll_tuned_bcast_intra_dec_fixed+0xcc) [0x7fc3e486573c] [10] /usr/lib/openmpi/lib/openmpi/mca_coll_sync.so(mca_coll_sync_bcast+0x64) [0x7fc3e4a7d6a4] [11] /usr/lib/libmpi.so.1(MPI_Bcast+0x13d) [0x7fc3ee0df78d] [12] ./matMul() [0x4011a9] [13] ./matMul() [0x401458] [14] /lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0xf5) [0x7fc3edab7ec5] [15] ./matMul() [0x400b49] End of error message [ 0] /lib/x86_64-linux-gnu/libpthread.so.0(+0x10340) [0x7fa4a1fe5340] [ 1] /lib/x86_64-linux-gnu/libc.so.6(+0x981c0) [0x7fa4a1ca81c0] [ 2] /usr/lib/libmpi.so.1(opal_convertor_unpack+0x105) [0x7fa4a22f28d5] [ 3] /usr/lib/openmpi/lib/openmpi/mca_pml_ob1.so(mca_pml_ob1_recv_frag_callback_match+0x460) [0x7fa49a701630] [ 4] /usr/lib/openmpi/lib/openmpi/mca_btl_sm.so(mca_btl_sm_component_progress+0x487) [0x7fa4998a4137] [ 5] /usr/lib/libmpi.so.1(opal_progress+0x5a) [0x7fa4a22fe9ea] [ 6] /usr/lib/libmpi.so.1(ompi_request_default_wait+0x16d) [0x7fa4a224bc0d] [ 7] /usr/lib/openmpi/lib/openmpi/mca_coll_tuned.so(ompi_coll_tuned_bcast_intra_generic+0x4e0) [0x7fa4989e7ae0] [ 8] /usr/lib/openmpi/lib/openmpi/mca_coll_tuned.so(ompi_coll_tuned_bcast_intra_binomial+0xb7) [0x7fa4989e7f27] [ 9] /usr/lib/openmpi/lib/openmpi/mca_coll_tuned.so(ompi_coll_tuned_bcast_intra_dec_fixed+0xcc) [0x7fa4989df73c] [10] /usr/lib/openmpi/lib/openmpi/mca_coll_sync.so(mca_coll_sync_bcast+0x64) [0x7fa498bf76a4] [11] /usr/lib/libmpi.so.1(MPI_Bcast+0x13d) [0x7fa4a225978d] [12] ./matMul() [0x4011a9] [13] ./matMul() [0x401458] [14] /lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0xf5) [0x7fa4a1c31ec5] [15] ./matMul() [0x400b49] End of error message * -------------------------------------------------------------------------- mpirun noticed that process rank 2 with PID 12466 on node rtidev5.etf.bg.ac.rs exited on signal 11 (Segmentation fault). -------------------------------------------------------------------------- 2 total processes killed (some possibly by mpirun during cleanup)

Answer 1

我已经弄明白了。所有进程（不仅仅是master）都需要先分配内存

所以缺少的行是

void matMulMPI(long N, double *a, double *b, double *c) {
    ...

    MPI_Bcast(&N, 1, MPI_LONG, MASTER, MPI_COMM_WORLD);

    b = (double *) malloc(N * N * sizeof(double));

    MPI_Bcast(b, N*N, MPI_DOUBLE, MASTER, MPI_COMM_WORLD);

    ...
}

将矩阵与 MPI 相乘（分段错误）

Multiplying matrices with MPI (segmentation fault)

matrix

mpi

segmentation-fault