JEDI-Specific Instructions
GPU and CPU Per Task
As JEDI runs with a different Slurm installation compared to current production
systems at JSC, one needs to manually specify the number of
GPUs and CPUs per task in the job script. In JUBE, this can be done by setting the
additional_job_config
parameter from the platform.xml
as follows:
<parameter name="additional_job_config">#SBATCH --gpus-per-task=1</parameter>
To ensure that you are only writing this line for the JEDI system, and not for other systems, you use a conditional configuration like shown below:
...
<parameter name="additional_job_config" mode="python">
{
"jedi": "#SBATCH --gpus-per-task=1",
}.get("${systemname}","")
</parameter>
Remember to use --cpus-per-task=72
to utilize all cores of a Grace socket.