Introduction

A common use case for MWA ASVO command line clients (manta-ray-client or giant-squid) is as part of a workflow running on a HPC system (e.g. Setonix or Garrawarla at Pawsey). Since HPC resources are almost always a scarce and in-demand resource we need to ensure that the way that we use MWA ASVO is not going to waste those precious resources.

The Submit, Wait, Download Anti-Pattern

The below is a common workflow for MWA ASVO use:

A common anti-pattern we see is that users will create a single HPC job which implements steps 1, 2, 3 and 4. Step 2 in red (where the client is waiting for the MWA ASVO job to complete) is often set to the largest possible WALLTIME (in Pawsey’s case it is 24 hours). During that time, the HPC node which is running that job is not doing anything, except polling MWA ASVO, and worse, the resources that node is occupying are not available for other users for all of that time. This can result in a large fraction of the HPC system being tied up waiting for MWA ASVO and unavailable for other users.

The Best Practice Pattern: Submit->Check->Download->Process

It is therefore highly recommended that users split their workflows up into parts that each do a specific job and therefore can completely eliminate any wastage of HPC resources.

Job 1: Submit

• Create a job which only submits new jobs to MWA ASVO.

• This job can be run from anywhere- e.g. your laptop or a HPC login node, as it is fast and does not require HPC resources.

giant-squid example

$ giant-squid submit-vis OBSID

manta-ray-client example

$ mwa_client -s joblist.csv

Job 2: Check

• Create a job which then periodically checks if the job is ready to download.

• Polling more than once per several minutes is unnecessary and not recommended.

• This job can be run from anywhere- e.g. your laptop or a HPC login node, as it is fast and does not require HPC resources.

giant-squid example

$ giant-squid list --json

manta-ray-client example

$ mwa_client -l

Job 3: Download

• When Job 2 detects the MWA ASVO job is ready for download it can then launch a HPC job (please use a data-mover or copyq node rather than a compute node) to download the data. If you opted to deliver the data to /scratch or /astro then you can skip this step.

giant-squid example

$ giant-squid download JOBID

manta-ray-client example

$ mwa_client -w JOBID

Job 4: Process

• With the data downloaded you can then launch your processing job on a compute node in the HPC cluster.