MPICH2 Plugin

By default StarCluster includes OpenMPI. However, for either performance or compatibility reasons, you may wish to use MPICH2 which provides an alternate MPI implementation. The MPICH2 plugin will install and configure MPICH2 on your clusters.

Usage

To use this plugin add a plugin section to your starcluster config file:

[plugin mpich2]
setup_class = starcluster.plugins.mpich2.MPICH2Setup

Next update the PLUGINS setting of one or more of your cluster templates to include the MPICH2 plugin:

[cluster mycluster]
plugins = mpich2

The next time you start a cluster the MPICH2 plugin will automatically be executed and MPICH2 will be installed and configured on all nodes. If you already have a cluster running that didn’t originally have MPICH2 in its plugin list you can manually run the plugin using:

$ starcluster runplugin mpich2 mycluster
StarCluster - (http://star.mit.edu/cluster)
Software Tools for Academics and Researchers (STAR)
Please submit bug reports to starcluster@mit.edu

>>> Running plugin mpich2
>>> Setting up MPICH2 hosts file on all nodes
3/3 |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||| 100%
>>> Setting MPICH2 as default MPI on all nodes
3/3 |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||| 100%
>>> MPICH2 is now ready to use
>>> Use mpicc, mpif90, mpirun, etc. to compile and run your MPI apps

Building and Running MPI Programs

To build your MPI code use the standard procedure of compiling with mpicc, mpic++, or mpif90:

$ starcluster sshmaster mycluster -u myuser
myuser@master $ cat << EOF > mpihelloworld.c
/* C Example */
#include <stdio.h>
#include <mpi.h>
#include <unistd.h>
int main (argc, argv)
     int argc;
     char *argv[];
{
  char hostname[1024];
  gethostname(hostname, 1024);
  int rank, size;
  MPI_Init (&argc, &argv);      /* starts MPI */
  MPI_Comm_rank (MPI_COMM_WORLD, &rank);        /* get current process id */
  MPI_Comm_size (MPI_COMM_WORLD, &size);        /* get number of processes */
  printf("Hello world from process %d of %d on %s\n", rank, size, hostname);
  MPI_Finalize();
  return 0;
}
EOF
myuser@master $ mpicc -o mpihw mpihelloworld.c

Note

You should put the MPI binary in your $HOME folder on the cluster in order for the binary to be NFS-shared to the rest of the nodes in the cluster.

To run the code simply use mpirun and pass in the number of processors you wish to use on the cluster via the -np option:

myuser@master $ mpirun -np 3 ./mpihw
Hello world from process 0 of 3 on master
Hello world from process 2 of 3 on node002
Hello world from process 1 of 3 on node001

This will use 3 processors to run the hello world MPI code above. You do not need to create or specify a hosts file. The hostfile is automatically created and set as the default MPICH2 hosts file by the plugin.