Skip to content

Latest commit

 

History

History
81 lines (51 loc) · 3.21 KB

HPCResources.md

File metadata and controls

81 lines (51 loc) · 3.21 KB
Raw

HPC Carpentry lessons

These are a great set of lessons to hand to people if they can't make a HPC tutorial. I have a modified version in my repo but this is where ours originated from : https://github.com/hpc-carpentry/hpc-intro

Looping with HPC inside Node using PBSpro

So it looks like inside a pbs script you can run a loop :

#PBS -J 1-30

is a loop that is 1 to 30 in 1 steps. You can change the steps etc. The index number its up to is represented by the env var $PBS_ARRAY_INDEX

You can add steps too. http://www.pbsworks.com/pdfs/PBSUserGuide13.0.pdf Section: 9.4.3 File Staging for Job Arrays

So for example

#!/bin/bash
#
#PBS -N simp
#                HH:MM:SS
#PBS -l walltime=02:00:00
#PBS -l nodes=3:ppn=1,mem=2g
#PBS -J 1-3

echo 'job running' $PBS_ARRAY_INDEX  $PBS_ARRAY_ID >> $HOME/output

Returns:
job running 1 4814333[].pbsserver
job running 2 4814333[].pbsserver
job running 3 4814333[].pbsserver

You can use that to loop 1-30 or 1-300 and span up to 300 at once as they become available. Each node can talk to $HOME or $SCRATCH. Do note that all jobs will run serially on the same node.

(Quick and Dirty) Looping with HPC across nodes

You should be doing proper neat parallel code. But if you've decided you need a dirty hack, you can use these two scripts to run up a number of pbs jobs- each with its own node, running the same script (but perhaps different parameters).

You can change the numbers it loops through from 1 - 5 to anything in Parser.test.sh.

Looping script: parser.test.sh

Example PBS script: test.pbs

Writing Parallel code

Long Term- Run segments of code parallel via OpenMP with c++

Pawsey HPC do a great tutorial of this here : https://support.pawsey.org.au/documentation/download/attachments/2162899/Parallelize%20your%20Code%20with%20OpenMP.pdf?api=v2

Basically segments of your code are parked with a pragma segment and that segment is looped and treated as parallel, fired off to numerous nodes.

Doing some googling on OpenMP with c++ should give you a bucketful of tutorials as well.

Moving Matlab scripts from GUI to HPC

When moving Matlab scripts to running via command line (generally for HPC), it will expect the functions to be in a separate file from the commands calling it. For a fast workaround, instead of using

matlab < script.m

You can use

matlab -r 'run("script")'

which will run matlab then call the script.m file and run the code from there. From reading in Stack Exchange, apparently Matlab 2017 doesn't do this anymore but from testing it was the same case.

Topic Name Description Free for AU/NZ Free for All URL
HPC NCI Slides Slides around HPC training with Parallel/MPI/OpenMP Yes Yes http://nci.org.au/user-support/training/
HPC Pawsey training material Slides around HPC training with Parallel/MPI/OpenMP Yes Yes https://support.pawsey.org.au/documentation/display/US/Training+Material
HPC HPC in a day lessons Lessons around HPC and how to use Yes Yes https://psteinb.github.io/hpc-in-a-day/