CCR allows the use of NVIDIA GPU virtual machine containers on our clusters using Singularity. Users can run pre-packaged applications that make use of the GPUs without the need for installation, setup or configuration of the applications.
NVIDIA offers a container registry of many popular applications that can utilize GPU resources. Users interested in using this should register for an NVIDIA account and follow the instructions on their website.
Once you've downloaded and converted your Docker image into a Singularity container, you will be able to run this on the CCR cluster. For an example of a SLURM batch script, please see the files in this directory:
This demonstrates how to run the NAMD application on a single node on the GPU partition of the ub-hpc cluster:
Slurm script: namd_single_slurm.sh
Runs Singularity script: singularity_single.sh
NOTE: The NVIDIA scripts all have a line to download the image files. However, we recommend users download these first, store them in their home or project directories and then point to them in this script. Downloading the files each time you run your job is inefficient. Our script shows the NVIDIA download line commented out.
While CCR staff can provide some assistance with running Singularity containers on our clusters, it is recommended you exhaust all troubleshooting options from NVIDIA before contacting CCR Help, unless you're certain the problem is CCR-related.