MPI中的进程数是否有限制? [英] Does the number of processes in MPI have a limit?

问题描述

我正在阅读使用MPI"并尝试自己执行代码.在6.3章中有一个网格分解代码.它可以编译，没有任何警告或错误，并且可以在少量进程中运行，但是在笔记本电脑上却可以以较大的数字(例如30)运行.我的笔记本电脑是4核，超线程和8G RAM.两种版本的la_grid_2d_new均不起作用，但是第一个版本可以容忍更大的数字(例如35)，但不能运行40个进程.我不知道为什么.请问你能帮帮我吗?非常感谢.

I am reading "Using MPI" and try to execute the code myself. There is a grid decomposition code in Chapter 6.3. It compiles with no warnings or errors, and runs with small number processes, but fails with larger numbers, say 30, on my laptop. My laptop is 4 core, hyperthreaded, and 8G RAM. Both versions of la_grid_2d_new do not work, but the first one tolerate a little larger number, say 35, but fails for 40 processes. I am not sure why. Could you help me please? Thanks a lot.

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

typedef struct
{
    int P, Q;
    int p, q;
    MPI_Comm grid_comm;
    MPI_Comm row_comm;
    MPI_Comm col_comm;
} LA_Grid_2d;

LA_Grid_2d *la_grid_2d_new(MPI_Comm comm, int P, int Q)
{
    LA_Grid_2d *grid;
    MPI_Comm row, col;
    int my_rank, p, q;
    MPI_Comm_rank(comm, &my_rank);
    p=my_rank/Q;
    q=my_rank%Q;
    MPI_Comm_split(comm, p, q, &row);
    MPI_Comm_split(comm, q, p, &col);
    grid=(LA_Grid_2d *)malloc(sizeof(LA_Grid_2d));
    grid->grid_comm=comm;
    grid->row_comm=row;
    grid->col_comm=col;
    grid->P=P;
    grid->Q=Q;
    grid->p=p;
    grid->q=q;
    return grid;
}

LA_Grid_2d *la_grid_2d_new_II(MPI_Comm comm, int P, int Q)
{
    LA_Grid_2d *grid;
    MPI_Comm comm_2d, row, col;
    int my_rank, p, q;
    int dims[2]={P,Q}, local[2], period[2]={0,0}, remain_dims[2];

    MPI_Cart_create(comm, 2, dims, period, 1, &comm_2d);
    MPI_Comm_rank(comm, &my_rank);
    MPI_Cart_coords(comm_2d, my_rank, 2, local);
    p=local[0];
    q=local[1];
    remain_dims[0]=0;
    remain_dims[1]=1;
    MPI_Cart_sub(comm_2d, remain_dims, &row);
    remain_dims[0]=1;
    remain_dims[1]=0;
    MPI_Cart_sub(comm_2d, remain_dims, &col);
    grid=(LA_Grid_2d *)malloc(sizeof(LA_Grid_2d));
    grid->grid_comm=comm;
    grid->row_comm=row;
    grid->col_comm=col;
    grid->P=P;
    grid->Q=Q;
    grid->p=p;
    grid->q=q;
    return grid;
}

void la_grid_2d_delete(LA_Grid_2d *grid)
{
    free(grid);
}
int main(int argc, char **argv)
{
    LA_Grid_2d *pgrid;
    int size, rank, dims[2]={0,0}, row, col;
    MPI_Init(&argc, &argv);
    MPI_Comm_size(MPI_COMM_WORLD, &size);
    MPI_Comm_rank(MPI_COMM_WORLD, &rank);
    if(rank==0)
        printf("size=%d rank=%d\n", size, rank);
    MPI_Dims_create(size, 2, dims);
//  pgrid=la_grid_2d_new(MPI_COMM_WORLD, dims[0], dims[1]);
    pgrid=la_grid_2d_new_II(MPI_COMM_WORLD, dims[0], dims[1]);
    if(rank==0)
        printf("dims[0]=%d dims[1]=%d\n", dims[0], dims[1]);
    MPI_Reduce(&rank, &row, 1, MPI_INT, MPI_SUM, 0, pgrid->row_comm);
    MPI_Reduce(&rank, &col, 1, MPI_INT, MPI_SUM, 0, pgrid->col_comm);
    la_grid_2d_delete(pgrid);
    MPI_Finalize();
    if(rank==0)
        printf("row=%d col=%d\n", row, col);
    return 0;
}

错误消息是:

shuang@phoebe:~/usingMPI$ mpiexec -n 20 ./grid
size=20 rank=0
dims[0]=5 dims[1]=4
row=6 col=40

shuang@phoebe:~/usingMPI$ mpiexec -n 30 ./grid
size=30 rank=0
dims[0]=6 dims[1]=5
[phoebe:14939] *** Process received signal ***
[phoebe:14939] Signal: Floating point exception (8)
[phoebe:14939] Signal code: Integer divide-by-zero (1)
[phoebe:14939] Failing at address: 0x7fb1e599e6f7
[phoebe:14939] [ 0] /lib/x86_64-linux-gnu/libpthread.so.0(+0xfcb0) [0x7fb1e5714cb0]
[phoebe:14939] [ 1] /usr/lib/libmpi.so.0(mca_topo_base_cart_coords+0x57) [0x7fb1e599e6f7]
[phoebe:14939] [ 2] /usr/lib/libmpi.so.0(mca_topo_base_cart_sub+0x166) [0x7fb1e599ec36]
[phoebe:14939] [ 3] /usr/lib/libmpi.so.0(PMPI_Cart_sub+0xba) [0x7fb1e596f34a]
[phoebe:14939] [ 4] ./grid(la_grid_2d_new_II+0xd6) [0x400df6]
[phoebe:14939] [ 5] ./grid(main+0x98) [0x400f07]
[phoebe:14939] [ 6] /lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0xed) [0x7fb1e536776d]
[phoebe:14939] [ 7] ./grid() [0x400b99]
[phoebe:14939] *** End of error message ***
--------------------------------------------------------------------------
mpiexec noticed that process rank 22 with PID 14939 on node phoebe exited on signal 8 (Floating point exception).
--------------------------------------------------------------------------

推荐答案

@Sean如果您想尝试另一个OpenMPI，通常可以下载它并使用类似的东西进行编译

@Sean If you want to try another OpenMPI you can normally download it and compile with something like

./configure --prefix=/opt/ompi-[version]
make
sudo make install

由于这将安装到非标准位置(以便以后删除)，因此需要设置LD_LIBRARY_PATH =/opt/ompi- [version]/lib并指定mpicc和mpirun的完整路径，以确保您调用正确的版本.在构建过程中的某个地方，它将提醒您有关设置LD_LIBRARY_PATH的信息.

Since this will install to a non-standard location(for easy removal later) you will need to set LD_LIBRARY_PATH=/opt/ompi-[version]/lib and specify the full path to mpicc and mpirun to ensure you call the right version. Somewhere in the build process it will remind you about setting LD_LIBRARY_PATH.

这篇关于MPI中的进程数是否有限制?的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！