c - 使用 MPI_Cart_create、MPICH 获取 NULL 通信器

Question

我为我的家庭作业(并行编程)编写了这个程序，但我在运行时遇到了一些错误。

这是来源:

/****************************************
 * Program:                 *
 *  ALL_TO_ALL_BROADCAST_MESH       *
 *                      *
 * Author:                  *
 *  -------------------         *
 *  -------------------         *
 ****************************************/

/*
 * Program's using MPI_Send, MPI_Recv and cartesian topology functions
 *
 * Compile using: mpicc -o all_to_all_bc all_to_all_bc.c
 */

#include <stdlib.h>
#include <stdio.h>
#include <math.h>
#include <string.h>
#include "mpi.h"

/* Number of dimensions */
#define NUMB_DIMS 2

/* Max value for the rand() function */
#define MAX_NUMB 128

#define TRUE 1
#define FALSE 0

/* Checks if square root of n can be calculated     *
 * Returns the square root on success or -1 on failure  */
int is_sqrt(int n);

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

    int my_rank;            /* Rank of the current process          */
    int size;           /* Number of processes in the mesh      */
    int mesh_size;          /* Size of a single mesh row            */
    int my_numb;            /* Message                  */
    int *all_numbs;         /* All numbers owned by the current process */
    int numb_numbs;         /* Number of numbers in the all_numbs array */
    int left, right, up, down;  /* Neighbourghs of the process as rank numbers  */

    /* Neighbourghs as cartesian coords */
    int *my_coords, *left_coords, *right_coords, *up_coords, *down_coords;

    /* Allocate memory for cartesian coords */
    my_coords = (int *)malloc(NUMB_DIMS*sizeof(int));
    left_coords = (int *)malloc(NUMB_DIMS*sizeof(int));
    right_coords = (int *)malloc(NUMB_DIMS*sizeof(int));
    up_coords = (int *)malloc(NUMB_DIMS*sizeof(int));
    down_coords = (int *)malloc(NUMB_DIMS*sizeof(int));

    MPI_Comm grid_comm; /* Mesh Communicator */

    int *dim_size = (int *)malloc(NUMB_DIMS*sizeof(int));   /* Sizes of mesh dimensions     */
    int *peroids = (int *)malloc(NUMB_DIMS*sizeof(int));    /* True if the column wraps around  */
    int reorder;                        /* I have no idea what this does ;) */

    int i,j; /* counters */

    /* MPI initialization */
    MPI_Init(&argc,&argv);
    MPI_Comm_size(MPI_COMM_WORLD,&size);
    MPI_Comm_rank(MPI_COMM_WORLD,&my_rank);

    /* Check if p^2=N */
    if (mesh_size=is_sqrt(size)<0) {
        perror("Square root of the number of processes can't be calculated!");
        MPI_Finalize();
        return -1;
    }

    /* Configure parameters for MPI_Cart_Create() */
    for (i=0;i<NUMB_DIMS;i++) {
        dim_size[i]=mesh_size;
        peroids[i]=TRUE;
    }
    reorder=FALSE;

    /* create new communicator */
    if (MPI_Cart_create(MPI_COMM_WORLD,NUMB_DIMS,dim_size,peroids,reorder,&grid_comm)<0) {
        perror("Unable to create new communicator!");
        MPI_Finalize();
        return -1;
    }

    /* calculate the coords and ranks */
    MPI_Cart_coords(grid_comm,my_rank,NUMB_DIMS,my_coords);

    /* !!this block is for 2D grid only!! */
        memcpy(left_coords,my_coords,NUMB_DIMS*sizeof(int));
        memcpy(right_coords,my_coords,NUMB_DIMS*sizeof(int));
        memcpy(up_coords,my_coords,NUMB_DIMS*sizeof(int));
        memcpy(down_coords,my_coords,NUMB_DIMS*sizeof(int));

        left_coords[0]=(left_coords[0]-1)%mesh_size;
        right_coords[0]=(right_coords[0]+1)%mesh_size;
        up_coords[1]=(up_coords[1]+1)%mesh_size;
        down_coords[1]=(down_coords[1]-1)%mesh_size;

        MPI_Cart_rank(grid_comm,left_coords,&left);
        MPI_Cart_rank(grid_comm,right_coords,&right);
        MPI_Cart_rank(grid_comm,up_coords,&up);
        MPI_Cart_rank(grid_comm,down_coords,&down);
    /* !!this block is for 2D grid only!! */

    /* get a random number for this process */
    my_numb = rand()%MAX_NUMB;

    all_numbs = (int *)malloc(size*sizeof(int));

    MPI_Status status;

    /****************
     * BROADCASTING *
     ****************/

    printf("I'm process (%d,%d),\nMy number is %d\n\n",my_coords[0],my_coords[1],my_numb);

    all_numbs[0]=my_numb;
    numb_numbs=1;

    /* Communication along rows */
    for (i=0;i<mesh_size-1;i++) {

        printf("I'm process (%d,%d),\nI'm sending %d to process (%d,%d)\n\n",my_coords[0],my_coords[1],my_numb,right_coords[0],right_coords[1]);
        MPI_Send(&my_numb,sizeof(int),MPI_INT,right,99,grid_comm);

        printf("I'm process (%d,%d),\nI'm recieving data from process (%d,%d)\n\n",my_coords[0],my_coords[1],left_coords[0],left_coords[1]);
        MPI_Recv(&my_numb,sizeof(int),MPI_INT,left,99,grid_comm,&status);

        all_numbs[numb_numbs]=my_numb;
        numb_numbs++;

        printf("My current numbers are:\n");
        for (j=0;j<numb_numbs;j++)
            printf("%d, ",all_numbs[j]);
        printf("\n\n");
    }

    /* Communication along columns */
    my_numb=all_numbs[0];

    for (i=0;i<mesh_size-1;i++) {

        printf("I'm process (%d,%d),\nI'm sending %d to process (%d,%d)\n\n",my_coords[0],my_coords[1],my_numb,up_coords[0],up_coords[1]);
        MPI_Send(&my_numb,sizeof(int),MPI_INT,up,99,grid_comm);

        printf("I'm process (%d,%d),\nI'm recieving data from process (%d,%d)\n\n",my_coords[0],my_coords[1],down_coords[0],down_coords[1]);
        MPI_Recv(&my_numb,sizeof(int),MPI_INT,down,99,grid_comm,&status);

        all_numbs[numb_numbs]=my_numb;
        numb_numbs++;

        printf("My current numbers are:\n");
        for (j=0;j<numb_numbs;j++)
            printf("%d, ",all_numbs[j]);
        printf("\n\n");
    }

    printf("I'm process %d\nBroacasting performed!\nMy numbers are:\n");
    for (i=0;i<numb_numbs;i++)
        printf("%d, ",all_numbs[i]);
    printf("\n\n");

    MPI_Finalize();
    return 0;
}

int is_sqrt(int n) {

    double a=n;
    double b=sqrt(a);
    double c=b*b;

    int result;

    if (a==c) {
        result=(int)b;
        return result;
    }
    else
        return -1;
}

这是我遇到的错误:

% mpirun -np 4 all_to_all_bc
[cli_1]: [cli_3]: aborting job:
Fatal error in MPI_Cart_coords: Invalid communicator, error stack:
MPI_Cart_coords(130): MPI_Cart_coords(MPI_COMM_NULL, rank=3, maxdims=2, coords=0x6a9010) failed
MPI_Cart_coords(74).: Null communicator
aborting job:
Fatal error in MPI_Cart_coords: Invalid communicator, error stack:
MPI_Cart_coords(130): MPI_Cart_coords(MPI_COMM_NULL, rank=1, maxdims=2, coords=0x6a9010) failed
MPI_Cart_coords(74).: Null communicator
[cli_0]: aborting job:
Fatal error in MPI_Cart_coords: Invalid communicator, error stack:
MPI_Cart_coords(130): MPI_Cart_coords(MPI_COMM_NULL, rank=0, maxdims=2, coords=0x6a9010) failed
MPI_Cart_coords(74).: Null communicator
[cli_2]: aborting job:
Fatal error in MPI_Cart_coords: Invalid communicator, error stack:
MPI_Cart_coords(130): MPI_Cart_coords(MPI_COMM_NULL, rank=2, maxdims=2, coords=0x6a9010) failed
MPI_Cart_coords(74).: Null communicator
rank 3 in job 18  ---host---_58157   caused collective abort of all ranks
  exit status of rank 3: killed by signal 9 
rank 2 in job 18  ---host---_58157   caused collective abort of all ranks
  exit status of rank 2: return code 1 
rank 1 in job 18  ---host---_58157   caused collective abort of all ranks
  exit status of rank 1: killed by signal 9 
rank 0 in job 18  ---host---_58157   caused collective abort of all ranks
  exit status of rank 0: return code 1 

Answer 1

编辑:

查看您的错误消息:

Fatal error in MPI_Cart_coords: Invalid communicator, error stack:
MPI_Cart_coords(130): MPI_Cart_coords(MPI_COMM_NULL, rank=3, maxdims=2, coords=0x6a9010) failed
MPI_Cart_coords(74).: Null communicator

看起来 MPI_Cart_coords 的通信器为空。回头看，这是通过调用 MPI_Cart_create ~5 行以上来初始化的。

来自 MPI_Cart_create 手册页(我的重点):

DESCRIPTION

MPI_Cart_create returns a handle to a new communicator to which the Cartesian topology information is attached. If reorder = false then the rank of each process in the new group is identical to its rank in the old group. Otherwise, the function may reorder the processes (possibly so as to choose a good embedding of the virtual topology onto the physical machine). If the total size of the Cartesian grid is smaller than the size of the group of comm, then some processes are returned MPI_COMM_NULL, in analogy to MPI_Comm_split. The call is erroneous if it specifies a grid that is larger than the group size.

看起来这可能是您的问题。