Diplopodus - How to run mpi

---

All jobs on the cluster should be started through the queue system PBS. This only applies to jobs you want to run on the nodes, not compilation etc. which are done on the server.

mpirun

#!/bin/bash
# Which queue
#PBS -q tpv
# This many nodes
#PBS -l nodes=4
# The output goes to
###PBS -o pbs.out
# Std. out and Std. err into same file.
###PBS -j oe
#PBS 
# cd to working directory
cd $PBS_O_WORKDIR
# Add /usr/local/bin to the PATH
PATH=$PATH:/usr/local/bin

# run the command.
pmpirun ./rd2D 100 100 2000 2

Explanation

• #PBS -q tpv indicates that this job should be run in tpv-queue. This is the default queue, and for now the only queue, hence this setting is not required.
• #PBS -l nodes=4 request four nodes, one cpu on each node. If you want more than one cpu on each node, you can use #PBS -l nodes=4:ppn=2 which gives you eight nodes in total.
• #PBS -o name set the output file to name
• #PBS -j oe put standard out and standard error into the same file.
• If you put more than one hash infront of PBS, the line turns into a standard comment
• $PBS_O_WORKDIR is the current directory when qsub is called. Change into this directoryif you don't want to give the full pathname for your executable.
• The command mpirun should not be used directly, since this command not recognize which cpus PBS has allocated for you. Instead, you should use /usr/local/bin/pmpirun, which automagically computes the right number of nodes from the #PBS nodes setting.