3.2. Working Remotely: NCI Gadi
Australian users can run THUNER on the National Computational Infrastructure (NCI) high performance computing system, Gadi . Scripts for working with THUNER on Gadi are available in the workflow/gridrad_severe_gadi folder on the THUNER GitHub repository.
The simplest way to install THUNER on Gadi is to first install a standalone version of
conda in your
g/data/<project>/<user> directory. Ask the computational support team if you need help
doing this. You can then clone the THUNER repository to your g/data/<project>/<user>
directory, and create the THUNER conda environment, as described in the From GitHub
section of the Installation page. After cloning you can modify
the scripts in the the workflow/gridrad_severe_gadi
folder so that paths etc are correct.
Note the GridRad Severe data has been copied to Gadi tape storage under the v46 project,
and currently exists in the directory esh563/d841006/volumes/<year>, with each GridRad
Severe “event” a compressed TAR file. Years can be copied to scratch and extracted
by navigating on the command-line to the workflow/gridrad_severe_gadi directory in
the cloned THUNER repository, and running the command
sh mdss_year_get.sh <year>
replacing <year> with the actual year you need, e.g. 2010.
THUNER can be run with the same options as the GridRad Severe demo notebook using the script
sh gridrad_year_job.sh <year>
Be wary that processing all 13 years will consume 20-30 KSU. Note that all 13 years have
already been processed with these options, and are available on tape under the v46 group
at esh563/gridrad_severe.tar.gz. You can set different options by modifying the gridrad.py file in the
workflow/gridrad_severe_gadi directory. In particular, commenting out line 102,
visualize.attribute.mcs_series(*args, **kwargs), will switch off the generation of
figures, and will reduce the KSU cost by at least half.
If required, you can then run the copy_local.sh script on your local machine to
copy data to local disk. The workflow/gridrad_severe_analysis contains scripts
for analyzing the data in a fashion analogous to
Ewan Short’s previously published work