Message Passing Interface¶
Many parallel programs take advantage of Message Passing Interface (MPI) to communicate across compute nodes and efficiently scale up computational analysis. While we can’t give instructions on how to run every program using MPI, all the programs using MPI need to be told which hosts and cpus to run. Generally, most programs use the mpirun executable that comes pre-packaged with MPI libraries to accomplish this task. The general syntax for running parallel programs with mpirun is:
mpirun -np <processor number> -machinefile <procs file> command [..args]
With -np option specifying the number of processors the job requires, and the -machinefile option with a file that contains the host address and number of CPUs to use on each host. On a distributed computing cluster with a scheduler in place (both RC clusters), the number of processors will be the same number of processors requested with the qsub command. Further, it is almost impossible to know which hosts your job will run on, since that is determined by a number of factors with the current state of the system. Therefore Moab/Torque provides a machinefile for each job that is executed through the scheduler, and it’s location is stored in the $PBS_NODEFILE variable. Therefore, on our clusters, the general syntax of running a program parallel with MPI is:
module load <mpi/modulefile>
mpirun -np <processor number> -machinefile $PBS_NODEFILE command [..args]
with being the appropriate module file for the MPI package the program was compiled with. Of course, you should consult the programs documentation for details, as some programs come with their own mpirun substitute and/or work better with certain MPI libraries than others. Also, each MPI libraries have differences with their mpirun commands, and therefore you should take a look at the MPI’s own documentation (links below).
MPI Standards¶
Message Passing Interface standards are available free for download at MPI Forum.
Useful Books¶
Parallel Programming with MPI. Peter Pecheco. ISBN-13:978-1558603394
Parallel Programming in C with MPI and OpenMP. Michael Quinn. ISBN-13:978-0072822564
MPI: The Complete Reference. Snir et al. ISBN-13:978-0262692168
Libraries¶
Current information about location and version of library can be found in their respective modulefiles.
Library |
Documentation |
Intel |
https://software.intel.com/en-us/mpi-library/documentation/get-started |
Mpich2 |
|
Mvapich2 |
|
OpenMPI |