DOCS : JOBSCRIPTS : GPU

Support for GPU jobscripts

If your application requires an NVIDIA GPU, then you can request one by giving the option --gpu to the justin create-stage or justin simple-workflow commands as described in the justin man page.

Currently there are a limited number of sites offering GPUs to DUNE and you may need to wait significantly longer (hours?) than usual for jobs in the workflow to start running.

The CUDA libraries, drivers, /dev/nvidiaX devices, and tools like nvidia-smi are made available to your jobscript in the usual way. $CUDA_VISIBLE_DEVICES is set to the UUID of the GPU allocated to your job by the site, in the newer form GPU-uuid not as 0, 1, 2 etc. Please do not try to use any other GPUs you might be able to access: by default CUDA should respect $CUDA_VISIBLE_DEVICES as given and do what the site wants.

Once the wrapper job starts, it reports to justIN information about the GPU it has discovered, including the GPU model name, the driver version, the compute capability, the VBIOS version, and the nonreserved memory in MiB. This information is shown on the job's own page in the dashboard.