Gadi

Revision as of 23:42, 22 January 2020 by C.carouge (talk | contribs)
Template:Needs Update This page needs updating

Gadi is NCI's primary supercomputer since January 2020. The supercomputer is composed of an assortment of processors, the majority of those are Cascade Lake processors. Different queues give access to different processors and have a different charging rate

Queue Memory Priority Charging rate per walltime-hour CPU per node Processor type
normal 192 GB Normal 2SU 48 Cascade Lake (CL), 3200 nodes
express 192 GB High 6SU 48 Cascade Lake (CL), 3200 nodes
copyq 192 GB Normal 2SU 1 cpu jobs only Cascade Lake (CL), 3200 nodes
gpuvolta 340 GB Normal 3SU 48CPUs, 4 GPU 640 Nvidia V100 GPUs, 160 nodes

More processors will be added in the coming weeks.

NCI does not yet have a User Guide for Gadi. But the notes to get prepared for Gadi will provide you with a wealth of information on the machine and its use.

Getting an account

To get a new account at NCI, you will need to get connected to a NCI project. Before you start the process, talk to your CI or supervisor to know which project code to use. You will need to apply via my.nci.org.au. NCI will send you a password via SMS once your application has been processed, this usually takes under a day to do.

Once you have an account, my.nci.org.au will allow you to ask for membership of other projects you might need. Those could be projects for additional compute time or projects to access data etc.

Connecting to Gadi

To connect to Gadi, you'll need to use a SSH connection to gadi.nci.org.au. If you're using Windows, you'll need to use something like | PuTTY, or if you're connecting from linux or mac run on the commandline (substitute abc123 with your own username)

ssh -Y abc123@gadi.nci.org.au

You can make a shortcut for this by editing (or creating) the file ~/.ssh/config and adding the lines:

Host              gadi
HostName          gadi.nci.org.au
User              abc123
ForwardX11        true
ForwardX11Trusted true

This way you just need to type 'ssh gadi' to connect.

Swapping Projects

If you use more than one project you can swap between them with the command 'switchproj', e.g.

switchproj w35

will change your current project to w35.

You can also change your default project by editing the file on Raijin ~/.rashrc, it should have a line like

setenv PROJECT w35

Resources on Raijin

To see how much compute time you have available run the command

nci_account -P $PROJECT -q 2013.q3

To see how much storage space you have available run the command

lquota -P $PROJECT

Submitting Jobs

To run a job on the supercomputer you submit it to a job queue using the 'qsub' command. Jobs are shell script files, they contain special markers to say what resources the job needs.

As an example the script "hello.sh"

#!/bin/bash
#PBS -l ncpus=2
#PBS -l walltime=10:00
#PBS -l mem=1gb
#PBS -v PROJECT

echo "Hello"

says to run with 2 cpus for a maximum time of 10 minutes. The job can use up to 1 GB of memory. Anything after the #PBS lines is what gets run on the supercomputer, in this instance it just prints "Hello" (any output goes to files in the directory you submitted the job named like "hello.sh.o123456", error messages go to files named like "hello.sh.e123456). The command '-v PROJECT' means run using the current project, you can also specify a project to use like '-v PROJECT=w35'.

If the job tries to use more resources than it's asked for it will be automatically stopped. The less resources you ask for the more likely it is that your job will run quickly however, you should try to request an amount close to what the job actually uses.

Managing Jobs

To see a list of your submitted & currently running jobs run

nqstat

This also shows how much resources each job has requested & is currently using.

Each job in the queue has a run id number associated with it (this is also printed when you submit a job with qsub). To get more information on a job run

qstat -s 123456 # Show any status information, e.g. why the job isn't currently running
qstat -f 123456 # Show full information, including resources requested & environment variables

To remove a job from the queue use qdel

qdel 123456 # Remove the job 123456 from the queue

Changes from Vayu

Vayu was NCI's previous supercomputer. There are some changes that need to be made to run jobs designed for Vayu run on Raijin.

The PBS flags

#PBS -l vmem=2gb
#PBS -wd

should be changed to

#PBS -l mem=2gb
#PBS -l wd

The environment variable $PROJECT should be set before submitting a job, or a line like

#PBS -v PROJECT=w35

should be added to scripts.

Shared ACCESS data that used to be in the path

/data/projects/access

is now available under

~access/data