MPI消息传递MPI_Sendrecv的用法

sweli

　　利用mpi求解微分方程时，经常会遇到不同进程的通讯，特别是如下形式的通讯：
　　进程0->进程1->进程2->进程3...->进程n->进程0
　　这时，若单纯的利用MPI_Send, MPI_Recv函数进行通讯的话，容易造成死锁，下面介绍MPI_Sendrecv的来解决这个问题。顾名思义，MPI_Sendrecv表示的作用是将本进程的信息发送出去，并接收其他进程的信息，其调用方式如下：



MPI_Sendrecv( void *sendbuf //initial address of send buffer
int sendcount //number of entries to send
MPI_Datatype sendtype //type of entries in send buffer
int dest //rank of destination
int sendtag //send tag
void *recvbuf //initial address of receive buffer
int recvcount //max number of entries to receive
MPI_Datatype recvtype //type of entries in receive buffer
int source //rank of source
int recvtag //receive tag
MPI_Comm comm //group communicator
MPI_Status status //return status;        

　　下面给出一个实例：



#include<stdio.h>
#include "mpi.h"
#include <math.h>
#define n 4
int
main(int argc, char* argv[]){
int nProcs, Rank, i;
double A0[n],A1[n];
MPI_Status status;
MPI_Init(&argc, &argv);
MPI_Comm_size(MPI_COMM_WORLD, &nProcs);
MPI_Comm_rank(MPI_COMM_WORLD, &Rank);
for(int i=0; i<n; i++){
A0 = Rank;
A1 = Rank;
}
printf("\nBefore exchange A0 A1:\n");
for(i=0;i<n;i++){
printf("rank:%d\t%f\t%f\n",Rank, A0, A1);
}   
int rightrank = (Rank + 1) % nProcs;
int leftrank = (Rank + nProcs-1)%nProcs;
MPI_Barrier(MPI_COMM_WORLD);
MPI_Sendrecv(A0, n, MPI_DOUBLE, rightrank,990,
A1, n, MPI_DOUBLE, leftrank,990,  MPI_COMM_WORLD,&status);
MPI_Finalize();
printf("After exchange A0 A1\n");
for(i=0;i<n;i++){
printf("rank:%d %f\t%f\n",Rank,  A0, A1);
}
}

　　下面这条语句表示：将进程为Rank的A0发送到rightrank进程，并接收来自leftrank的A1。



  MPI_Sendrecv(A0, n, MPI_DOUBLE, rightrank,990,
A1, n, MPI_DOUBLE, leftrank,990,  MPI_COMM_WORLD,&status);

　　得到的数值结果如下：



Before exchange A0 A1
rank:00.0000000.000000
rank:00.0000000.000000
rank:00.0000000.000000
rank:00.0000000.000000
Before exchange A0 A1
rank:11.0000001.000000
rank:11.0000001.000000
rank:11.0000001.000000
rank:11.0000001.000000
Before exchange A0 A1
rank:22.0000002.000000
rank:22.0000002.000000
rank:22.0000002.000000
rank:22.0000002.000000
Before exchange A0 A1
rank:33.0000003.000000
rank:33.0000003.000000
rank:33.0000003.000000
rank:33.0000003.000000
After exchange A0 A1
rank:1 1.0000000.000000
rank:1 1.0000000.000000
rank:1 1.0000000.000000
rank:1 1.0000000.000000
After exchange A0 A1
rank:2 2.0000001.000000
rank:2 2.0000001.000000
rank:2 2.0000001.000000
rank:2 2.0000001.000000
After exchange A0 A1
rank:3 3.0000002.000000
rank:3 3.0000002.000000
rank:3 3.0000002.000000
rank:3 3.0000002.000000
After exchange A0 A1
rank:0 0.0000003.000000
rank:0 0.0000003.000000
rank:0 0.0000003.000000
rank:0 0.0000003.000000