MP1: Running MPI Tasks on the Grid Appliance


This tutorial provides step-by-step instructions on how to deploy a pool of Grid Appliances configured with MPICH-2, an open-source implementation of the MPI (Message Passing Interface) framework. The Grid Appliance/MPICH2 virtual clusters can be used as a basis for running a variety of MPI tutorials.


Set Up and Submit MPI Jobs

  1. First you will build a local installation of MPI on your appliance from source. Use the following commands (note that it will take a few minutes to build MPI). The “-m32” option allows a 64-bit machine to submit MPI jobs to 32-bit machines.
    su griduser
    cd ~/examples/mpi
    ./ -m32
  2. Compile the “Hello World” example using the MPI cc compiler you just built in the previous step. (Note: This example builds a 32-bit binary with the “-m32” option):
    /mnt/local/mpich2/bin/mpicc -m32 -o HelloWorld HelloWorld.c
  3. Before proceeding, check that your machine is connected to the public pool by using ``condor_status`` command. You may need to wait for a few minutes before it becomes fully operational.

  4. Submit the “HelloWorld” program to run on 2 nodes:
    ./ -n 2 HelloWorld
  5. If all goes well, you will see an output similar to the output below; if an error occurs and the MPI ring fails to start, you can retry the command above.

    griduser@C111197176:~/examples/mpi$ ./ -n 2 HelloWorld
    serv ipop ip =
    submit condor with file tmpo21uIeMh/submit_mpi_vanilla_o21uIeMh
    Submitting job(s).
    Logging submit event(s).
    1 job(s) submitted to cluster 5.
    Waiting for 1 workers to response .... finished
    host list:
    ['C079184183', '1', 'cndrusr1', '45556', '/var/lib/condor/execute/dir_2824']
    MPD trace:
    C111197176_60625 (
    C079184183_43364 (
    Processor 0 of 2: Hello World!
    Processor 1 of 2: Hello World!

Building Your Own MPI Cluster Within or Outside FutureGrid:

If you would like to re-create your own MPI virtual cluster outside FutureGrid (e.g. in a local cluster, or on student desktops), the overall steps to accomplish this are:

  1. Create a GroupVPN and GroupAppliance for your class. Please refer to FutureGrid Tutorial GA2, which explains how to create grid appliance clusters.
  2. Download virtual appliances for deployment on your own resources, and configure them to connect to your GroupVPN/GroupAppliance with one Condor server and one or more Condor clients.
  3. On the machine that will be the submitter of MPI jobs, install gcc, make, MPI, and the scripts used to submit MPI jobs through Condor. A tutorial detailing this process is maintained at the Grid Appliance portal. See MPI over Condor with Grid Appliance (opens in new window).


Panoat Chuchaisri, David Wolinsky, Renato Figueiredo, (University of Florida)