#!/bin/bash -l

#$ -l h_rt=12:00:00
#$ -N mpi_test
#$ -pe mpi_16_tasks_per_node 32

module load openmpi/3.1.4


# Two examples here.  This command launches 1 MPI process per node.
# The --map-by command will round-robin through the nodes, assigning
# 1 process per node before moving on to the next.
mpirun -np 2 --map-by node ./hello


# To launch 2 processes per node use a different --map-by option to 
# make sure the processes are on separate CPU sockets. There is 1
# memory controller per socket (that's what the "numa" refers to)
# and the "span" modifier tells mpirun to use socket 0 on node 0,
# then socket 1 on node 0, then socket 0 on node 1, etc.
mpirun -np 4 --map-by numa:span ./hello