MPI_BARRIER 不工作

Posted 2023-03-17

【中文标题】MPI_BARRIER 不工作

【英文标题】：MPI_BARRIER not working

【发布时间】：2011-06-21 08:23:10

【问题描述】：

为什么这里的屏障不起作用？如果我使用它，程序会被阻塞，否则我会以奇怪的顺序得到输出：

Number of worker tasks = 4
   sending 1-th element q=0.011000 to task 1
   sending 2-th element q=0.012000 to task 2
   received 1-th element q=0.011000 in task 1
   processed 1-th element q=6.105000 in task 1
   sending 3-th element q=0.013000 to task 3
   received 2-th element q=0.012000 in task 2
   processed 2-th element q=13.320000 in task 2
   sending 4-th element q=0.014000 to task 4
   received 3-th element q=0.013000 in task 3
   processed 3-th element q=21.645000 in task 3
 starting to get data in MASTER
   MASTER received 0-th element q=6.105000 from task 1
   MASTER received 0-th element q=13.320000 from task 2
   received 4-th element q=0.014000 in task 4
   processed 4-th element q=31.080000 in task 4
   MASTER received 0-th element q=21.645000 from task 3
   MASTER received 0-th element q=31.080000 from task 4
 end

代码：

#include "mpi.h"               /* required MPI library */
#include <stdio.h>
#include <math.h>

#define NRRR 16               /* number of rows in matrix A */
#define NLLL 16                 /* number of columns in matrix A */
#define MASTER 0               /* taskid of first task */
#define FROM_MASTER 1          /* setting a message type */
#define FROM_WORKER 2          /* setting a message type */

int main(argc,argv)
int argc;
char *argv[];

    int numtasks,              /* number of tasks in partition */
    taskid,                /* a task identifier */
    numworkers,            /* number of worker tasks */
    source,                /* task id of message source */
    dest,                  /* task id of message destination */
    mtype,
    i,j,
    rc;                 /* message type */
    double  qr[NRRR],
    ql[NLLL],
    element_r[NRRR][3],
    element_l[NLLL][3];           


    MPI_Status status;  
    rc = MPI_Init(&argc,&argv);
    rc|= MPI_Comm_size(MPI_COMM_WORLD,&numtasks);   
    rc|= MPI_Comm_rank(MPI_COMM_WORLD,&taskid); 

    if (rc != 0)
        printf ("error initializing MPI and obtaining task ID info\n");

    numworkers = numtasks-1;

    // MASTER
    if (taskid == MASTER)
    
        printf("\n\n\n\nNumber of worker tasks = %d\n",numworkers);

        // init element_r and element_l
        for(j=0;j<NRRR;j++)
            element_r[j][0]=j;
            element_r[j][1]=j+1;
            element_r[j][2]=j+2;
            qr[j] = j*1e-4+1e-3;
        

        for(i=0;i<NLLL;i++)
            element_l[i][0]=12000+i;
            element_l[i][1]=12000+i+1;
            element_l[i][2]=12000+i+2;
            ql[i] = i*1e-3 +1e-2 ;
        

        mtype = FROM_MASTER;
        for (dest=1; dest<=numworkers; dest++)
        
            printf("   sending %d-th element q=%f to task %d\n",dest,ql[dest],dest);
            MPI_Send(&ql[dest], 1, MPI_DOUBLE, dest, mtype, MPI_COMM_WORLD);                        
        


        mtype = FROM_WORKER;
        printf(" starting to get data in MASTER\n");
        for (i=1; i<=numworkers; i++)
        
            source = i;
            MPI_Recv(&ql[source], 1, MPI_DOUBLE, source, mtype, MPI_COMM_WORLD, &status);
         
        MPI_Barrier(MPI_COMM_WORLD); 
        for (i=1; i<=numworkers; i++)
        
            source = i;
            printf("   MASTER received %d-th element q=%f from task %d\n",taskid,ql[source],source);            
         
        printf(" end\n");

    

    // WORKER
    if (taskid > MASTER)
    
        mtype = FROM_MASTER;        
        MPI_Recv(&ql, 1, MPI_DOUBLE, MASTER, mtype, MPI_COMM_WORLD, &status);
        printf("   received %d-th element q=%f in task %d\n",taskid,ql[0],taskid);
        ql[0]=ql[0]*555*taskid;
        printf("   processed %d-th element q=%f in task %d\n",taskid,ql[0],taskid);

        mtype = FROM_WORKER;
        MPI_Send(&ql, 1, MPI_DOUBLE, MASTER, mtype, MPI_COMM_WORLD);
    

    MPI_Finalize();

【参考方案1】：

MPI_Barrier 是一个集体通信调用：它将阻塞，直到参数中提供的通信器中的所有进程都调用它。由于您的工作进程从不调用 MPI_Barrier，因此主进程会无限期阻塞。

【讨论】：

好的，然后我在woker任务的末尾添加了MPI_BARRIER，现在一切似乎都正常了。现在我想知道使用屏障时是否会受到性能损失

使用屏障时会有一些性能损失，但如果您需要同步进程以获得正确的行为，它们是必要的。优化并行算法的一个方面是设计您的代码，以便不需要同步。

@Werner：的确，Barrier 只会损害性能，而不会改善它。另一方面，它可以让你作为开发人员更方便。在其他时候，在其他进程完成某些任务之前，进程可能无法执行任何有意义的事情，因此屏障是必不可少的。您可以自行评估是否需要设置障碍，或者是否可以容忍性能损失。

你能看看这个问题***.com/questions/71247968/…谢谢。