mpi4py

mpi4py is a Python package that provides binding to functions in Message Passing Interface (MPI) library written in C language. It allows distributed memory applications to communicate between Linux processes either on the same or remote machine.

There are more than one ways to use mpi4py on KSL systems. You can either use the preinstalled package with a managed Python, or install your own in a conda or virtual environment. Either way, mpi4py links to a preinstalled MPI library to, may it be OpenMPI or MPICH to fulfill the calls for interprocessor communications.

mpi4py on Shaheen III

Shaheen III uses Cray MPICH library which is based on MPICH distribution of MPI library.

To use the mpi4py package maintained by the KSL applications team, simply load the module and run your parallel job. Below is an example jobscript demonstrating the same:

#!/bin/bash

#SBATCH -n 4
#SBATCH -N 4
#SBATCH --hint=nomultithread

module load python

srun -n ${SLURM_NTASKS} -N ${SLURM_NNODES} --hint=nomultithread python myMPIscript.py

For using mpi4py in a self-managed Python, in a conda environment, first the package needs to be installed correctly. Assuming you have created a conda environment using the instructions provided in _conda_shaheen3 mpi4py can be compiled as show below. Here the example assumes the name of the conda environment myenv and the desired version of mpi4py is 3.1.4.

conda activate myenv
env MPICC=cc pip install mpi4py==3.1.4

To use the installed package the jobscript can be modified as follows:

#!/bin/bash

#SBATCH -n 4
#SBATCH -N 4
#SBATCH --hint=nomultithread

source $MY_SW/miniconda3-amd64/bin/activate myenv

srun -n ${SLURM_NTASKS} -N ${SLURM_NNODES} --hint=nomultithread python myMPIscript.py