How to submit an MPI parallel job

This article explains the submission of an MPI parallel job to Slurm/ LOTUS.

What is an MPI parallel job?  

An MPI parallel job runs on more than one core and more than one host using the Message Passing Interface (MPI) library for communication between all cores. A simple script, such as the one given below “my_script_name.sbatch "

#!/bin/bash
#SBATCH -p par-multi
#SBATCH -n 36
#SBATCH -t 30
#SBATCH -o %j.log
#SBATCH -e %j.err

# Load a module for the gcc OpenMPI library  (needed for mpi_myname.exe)
module load eb/OpenMPI/gcc/4.0.0

# Start the job running using OpenMPI's "mpirun" job launcher
mpirun ./mpi_myname.exe

-n refers to the number of processors or cores you wish to run on. The rest of the #SBATCH input options, and many more besides, can be found in the sbatch manual page or in the related articles. You must only use the par- multi queue for parallel jobs using MPI.

For those familiar with LOTUS running Platform MPI and Platform LSF, please note that the job is started using the OpenMPI native “mpirun” command not the “mpirun.lotus” wrapper script that was previously required. We have provided an mpirun.lotus script for backward compatiblity but it just runs the native mpirun

To submit the job, do not run the script, but rather use it as the standard input to sbatch, like so:

sbatch --exclusive my_script_name.sbatch

The --exclusive flag is used to group the parallel jobs onto hosts which are allocated only to run this job. This ensures the best MPI communication consistency and bandwidth/latency for the job and ensures no interference from other users’ jobs that might otherwise be running on those hosts.

MPI implementation and Slurm  

The OpenMPI library is the only supported MPI library on the cluster. OpenMPI v3.1.1 and v4.0.0 are provided which are fully MPI3 compliant. MPI I/O features are fully supported only on the LOTUS /work/scratch-pw* volumes as these use a Panasas fully parallel file system. The MPI implementation on CentOS7 LOTUS/Slurm is available via the module environment for each compiler as listed below:

eb/OpenMPI/gcc/3.1.1 
eb/OpenMPI/gcc/4.0.0       
eb/OpenMPI/intel/3.1.1

Note: OPenMPI Intel compiler is due shortly as is 4.0.3

Parallel MPI compiler with OpenMPI  

Compile and link to OpenMPI libraries using the mpif90, mpif77, mpicc, mpiCC wrapper scripts which are in the default unix path. The scripts will detect which compiler you are using (Gnu, Intel) by the compiler environment loaded and add the relevant compiler library switches. For example:

module load intel/20.0.0
module load eb/OpenMPI/intel/3.1.1
mpif90

will use the Intel Fortran compiler ifort and OpenMPI/3.1.1. Whereas

module load eb/OpenMPI/gcc/3.1.1
mpicc

will call the GNU C compiler gcc and OpenMPI/3.1.1.

The OpenMPI User Guides can be found at https://www.open-mpi.org/doc/  .