MPI的全称是Message Passing
Interface即标准消息传递界面,可以用于并行计算。MPI有多种实现版本,如MPICH,
CHIMP以及OPENMPI。这里我们采用MPICH版本。
一、MPICH安装
tar -xzvf soft/mpich-3.0.4.tar.gz
cd mpich-3.0.4/
./configure --prefix=/usr/local/mpich
make && make install
安装后加入环境变量/etc/profile,并执行source
/etc/profile
PATH=$PATH:/usr/local/mpich/bin
MANPATH=$MANPATH:/usr/local/mpich/man
export PATH MANPATH
二、单节点测试
复制源代码包下的examples目录到安装目录下
cp -r examples/ /usr/local/mpich
执行
mpirun -np 10 ./examples/cpi
输出结果如下:
Process 0 of 10 is on server150
Process 9 of 10 is on server150
Process 1 of 10 is on server150
Process 4 of 10 is on server150
Process 5 of 10 is on server150
Process 7 of 10 is on server150
Process 2 of 10 is on server150
Process 3 of 10 is on server150
Process 6 of 10 is on server150
Process 8 of 10 is on server150
pi is approximately 3.1415926544231256, Error is
0.0000000008333325
wall clock time = 0.020644
如果我们现在想编译文件: 在/home/houqingdong下执行: mpicc -o
hello hello.c 这时候会提醒:-bash:mpicc command not found
这是因为我们还没有配置路径
在命令行下输入: export
PATH=/home/houqingdong/mpiexe/bin:$PATH 注意:这里仅仅是暂时的设置路径,在重启后效果会消失,如果想一劳永逸的配置,请google查询
看一下我们配置是否成功可以执行一下 echo $PATH 看一下输出结果中是否有我们的路径
三、集群配置
1、集群机器上面需要配置ssh登录权限。参考:Hadoop-0.21.0在linux分布式集群配置 中的ssh配置(密钥无密码登录)部分。
2、复制编译程序到其他机器上面
scp -r mpich server140:/usr/local/
scp -r mpich server151:/usr/local/
scp -r mpich server130:/usr/local/
scp -r mpich server143:/usr/local/
同时在每台机器上面相应加入环境变量中。
3、
在/usr/local/mpich 下新建servers文件,内容如下:
server150:2 #run 2 process
server140:2
server130:2
server143:2
server151:2
执行下面命令,并指定servers文件
mpiexec -n 10 -f servers ./examples/cpi
输出
Process 0 of 10 is on server150
Process 1 of 10 is on server150
Process 4 of 10 is on server140
Process 5 of 10 is on server140
Process 6 of 10 is on server143
Process 7 of 10 is on server143
Process 8 of 10 is on server130
Process 9 of 10 is on server130
Process 2 of 10 is on server151
Process 3 of 10 is on server151
pi is approximately 3.1415926544231256, Error is
0.0000000008333325
wall clock time = 0.018768
四、参考资料