MPI集群搭建指南

由于有时需要使用到MPI集群来进行迭代运算，就尝试着自己去搭建着MPI集群。不过这也是到去年工作的事情了。我现在凭我的记忆去把它回忆出来。

 

从零开始，搭建MPI集群是一件很费周折的事情。MPI是一种编程模型，业界有很多种实现，其中开源版本就有好几个版本。对比过openfabrics公司提供的OFED和俄亥俄州立大学的mvapich2，我最终在选型时选取了mvapich2。

 

服务器系统有 centos 5.8/6.3，安装的版本mvapich2-1.9a。成功编译及安装上mvapich2-1.9a前，我安装了很多依赖包。安装过程如下：

 

yum install  gcc-c++.x86_64 redhat-rpm-config kernel-devel  bison flex zlib-devel tcl-devel gcc-gfortran elfutils-libs-x86_64 rpm-build  pciutils tcsh tk libtool libibverbs libibverbs-devel mpi-selector.noarch libibumad-devel libibverbs-devel librdmacm-devel mvapich2-psm-devel infinipath-psm-devel

cd /opt/mvapich2-1.9a && ./configure --enable-shared --enable-g=all --enable-error-messages=all --prefix=/usr --libdir=/usr/lib64 --with-device=ch3:nemesis && make && make install

 

 Helloword 测试代码：

 

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

int main(argc, argv)
int argc;
char *argv[];
{
        int myid, np;
        int namelen;
        char proc_name[MPI_MAX_PROCESSOR_NAME];
        MPI_Init(&argc, &argv);
        MPI_Comm_rank(MPI_COMM_WORLD, &myid);
        MPI_Comm_size(MPI_COMM_WORLD, &np);
        MPI_Get_processor_name(proc_name, &namelen);
        fprintf(stderr, "Hello, I am proc, %d of %d on %s\n", myid, np, proc_name);
        MPI_Finalize();
}

 

 Send_Recv 测试代码：

 

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

int main( int argc, char *argv[])
{
        int myid, np;
        char greeting[199];
        int i, t;
        int partner;
        int namelen;
        char proc_name[MPI_MAX_PROCESSOR_NAME];
        MPI_Status stat;
        MPI_Init(&argc, &argv);
        MPI_Comm_rank(MPI_COMM_WORLD, &myid);
        MPI_Comm_size(MPI_COMM_WORLD, &np);
        MPI_Get_processor_name(proc_name, &namelen);
        fprintf(stdout, "[fprintf]Hello, I am proc, %d of %d on %s\n", myid, np, proc_name);
        sprintf(greeting, "[sprintf]Hello, I am proc, %d of %d on %s\n", myid, np, proc_name);
        if (myid == 1) for (i = 0; i < 10000; i++) t = i;
        if(myid == 0){

                fputs(greeting , stdout);
                for (partner = 1; partner < np; partner++){
                        MPI_Recv(greeting, sizeof(greeting), MPI_BYTE, partner, 1, MPI_COMM_WORLD, &stat);
                        fputs(greeting, stdout);
                }
        }else{
                MPI_Send(greeting, sizeof(greeting) , MPI_BYTE, 0, 1, MPI_COMM_WORLD);
        }
        if (myid== 0) printf("That is all for now!\n");
        MPI_Finalize();
}