1->n MPI_Scatter MPI_Bcast

最新推荐文章于 2024-10-07 16:44:43 发布

原创最新推荐文章于 2024-10-07 16:44:43 发布 · 5.8k 阅读

1 ·

CC 4.0 BY-SA版权

文章标签：

#float #buffer #each

MPI 专栏收录该内容

35 篇文章

订阅专栏

本文详细介绍了MPI中的两种集合通信操作：MPI_Scatter和MPI_Bcast。MPI_Scatter用于将数据从根进程分散到所有进程中，而MPI_Bcast则用于从根进程广播数据到所有进程。通过示例和执行过程，阐述了这两个函数的使用方法，强调了在MPI程序中数据同步和分配的重要性。

1. 函数 MPI_Scatter

MPI_Scatter (&sendbuf,sendcnt,sendtype,&recvbuf,recvcnt,recvtype,root,comm)

Distributes distinct messages from a single source task to each task in the group.

2. 举例

#include<stdio.h>
#include"mpi.h"
#define SIZE 4

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

        float sendBuf[SIZE][SIZE] = {
                {1.0,   2.0,    3.0,    4.0},
                {5.0,   6.0,    7.0,    8.0},
                {9.0,   10.0,   11.0,   12.0},
                {13.0,  14.0,   15.0,   16.0}
        };

        MPI_Init(&argc, &argv);
        MPI_Comm_rank(MPI_COMM_WORLD, &rankID);
        MPI_Comm_size(MPI_COMM_WORLD, &totalNumTasks);

        if(totalNumTasks == SIZE){
                int source = 0;
                int sendCount = SIZE;
                int recvCount = SIZE;
                float recvBuf[SIZE];
                //scatter data from source process to all processes in MPI_COMM_WORLD
                MPI_Scatter(sendBuf, sendCount, MPI_FLOAT, 
                            recvBuf, recvCount, MPI_FLOAT, source, MPI_COMM_WORLD);

                printf("my rankID = %d, receive Results: %f %f %f %f, total = %f\n", 
                        rankID, recvBuf[0], recvBuf[1], recvBuf[2], recvBuf[3],
                        recvBuf[0] + recvBuf[1] + recvBuf[2] + recvBuf[3]);
        }else if(totalNumTasks == 8){
                int source = 0;
                int sendCount = 2;
                int recvCount = 2;
                float recvBuf[2];

                MPI_Scatter(sendBuf, sendCount, MPI_FLOAT, 
                            recvBuf, recvCount, MPI_FLOAT, source, MPI_COMM_WORLD);

                printf("my rankID = %d, receive result: %f %f, total = %f\n",
                        rankID, recvBuf[0], recvBuf[1], recvBuf[0] + recvBuf[1]);
        }else{
                printf("Please specify -n %d or -n %d\n", SIZE, 2*SIZE);
        }

        MPI_Finalize();
        return 0;
}

3. 编译执行

[amao@amao991 mpi-study]$ mpicc collectiveScatter.c 
[amao@amao991 mpi-study]$ mpiexec -n 4 -f machinefile ./a.out 
my rankID = 0, receive Results: 1.000000 2.000000 3.000000 4.000000, total = 10.000000
my rankID = 1, receive Results: 5.000000 6.000000 7.000000 8.000000, total = 26.000000
my rankID = 3, receive Results: 13.000000 14.000000 15.000000 16.000000, total = 58.000000
my rankID = 2, receive Results: 9.000000 10.000000 11.000000 12.000000, total = 42.000000
[amao@amao991 mpi-study]$ mpiexec -n 8 -f machinefile ./a.out 
my rankID = 6, receive result: 13.000000 14.000000, total = 27.000000
my rankID = 7, receive result: 15.000000 16.000000, total = 31.000000
my rankID = 4, receive result: 9.000000 10.000000, total = 19.000000
my rankID = 0, receive result: 1.000000 2.000000, total = 3.000000
my rankID = 1, receive result: 3.000000 4.000000, total = 7.000000
my rankID = 5, receive result: 11.000000 12.000000, total = 23.000000
my rankID = 3, receive result: 7.000000 8.000000, total = 15.000000
my rankID = 2, receive result: 5.000000 6.000000, total = 11.000000
[amao@amao991 mpi-study]$ mpiexec -n 2 -f machinefile ./a.out 
Please specify -n 4 or -n 8
Please specify -n 4 or -n 8

4. 总结

(1) 由于集合通信是阻塞的，因此，每个进程自动等待，完成接收后，执行printf语句

(2) 为了能够平均划分数据，本例执行时最好-n 4 或者-n 8

(3) 在每个进程接受到数据后，可以对其求和。若要把每个进程的求和结果汇总到某个进程，要用n->1即就是MPI_Gather或者MPI_Reduce函数。

-----------------------------------------------------

1. 函数

MPI_Bcast (&buffer,count,datatype,root,comm)

Broadcasts (sends) a message from the process with rank "root" to all other processes in the group.

2. 图示

3. 举例

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

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

        MPI_Init(&argc, &argv);
        MPI_Barrier(MPI_COMM_WORLD);
        double elapsed_time = -MPI_Wtime();

        MPI_Comm_rank(MPI_COMM_WORLD, &rankID);
        MPI_Comm_size(MPI_COMM_WORLD, &totalNumTasks);

        int sendRecvBuf[3] = {0, 0, 0};

        if(!rankID){
                sendRecvBuf[0] = 3;
                sendRecvBuf[1] = 6;
                sendRecvBuf[2] = 9;
        }

        int count = 3;
        int root = 0;
        MPI_Bcast(sendRecvBuf, count, MPI_INT, root, MPI_COMM_WORLD); //MPI_Bcast can be seen from all processes

        printf("my rankID = %d, sendRecvBuf = {%d, %d, %d}\n", rankID, sendRecvBuf[0], sendRecvBuf[1], sendRecvBuf[2]);

        elapsed_time += MPI_Wtime();
        if(!rankID){
                printf("total elapsed time = %10.6f\n", elapsed_time);
        }

        MPI_Finalize();
        return 0;
}

4. 编译执行

[amao@amao991 mpi-study]$ mpicc broadcast.c

[amao@amao991 mpi-study]$ mpiexec -f machinefile -n 5 ./a.out
my rankID = 0, sendRecvBuf = {3, 6, 9}
total elapsed time =   0.000045
my rankID = 2, sendRecvBuf = {3, 6, 9}
my rankID = 4, sendRecvBuf = {3, 6, 9}
my rankID = 1, sendRecvBuf = {3, 6, 9}
my rankID = 3, sendRecvBuf = {3, 6, 9}

5. 总结

(1) 在记录起始时间之前，用MPI_Barrier(MPI_COMM_WORLD);对所有进程做同步

(2) MPI_Bcast(sendRecvBuf, count, MPI_INT, root, MPI_COMM_WORLD);

要对所有进程可见，若只对process 0可见，是错误的。其他进程也需要知道它而接收啊！