When you first access Katana and look at the available resources it is very tempting to use it for everything that you can. But this can cause problems as the head node is responsible for controlling the whole cluster and problems start to occur when it becomes overloaded. You also miss out on the opportunity to access even more powerful resources as the head node is one of the least powerful nodes in the cluster,
If you start a process that places a load on the head node the process may be terminated without warning. Instead of running something on the Katana head node then you should use one of the following alternatives.
Running Interactive Jobs on a Compute Node
Instead of running an interactive job on the head node, follow the instructions for running an interactive job on a compute node.
A good starting point for running an interactive job is to type
qsub -I -l nodes=1:ppn=1,vmem=12gb,walltime=10:00:00
which will give you access to 1 CPU core an a compute node for 10 hours and allow you to use up to 12Gb of memory. (You should adjust the resource requirements to match your needs.)
Running Big Jobs or More than One Job at a Time
Instead of running an interactive job which requires you to be logged on to Katana you can submit one or more batch jobs. A batch job is run in a hands off manner with no input from you once the job is submitted. It just sits in the queue until it starts to run on a compute node, runs until it finishes (or runs out of time or memory) and then emails you telling you that your results are ready.
Batch jobs are the best way to use Katana. More information on batch jobs is found on the following pages.
Copying Data using the Katana Data Mover (KDM)
If you have a large amount of data that you would like to copy to or within the cluster you should use the Katana Data Mover or KDM server rather than using the head node. This is especially the case when using the UNSW Long Term Research Data Storage,
There are specific instructions about using the KDM server on the following pages: