GPU Usage

Introduction

Since we got 12 dedicated GPU nodes (mpcg[001-009]) containing one NVIDIA Tesla P100 each and four additional nodes (mpc[001-004]) containing two GTX 1080 each, its possible to run your jobs with one or multiple associated GPUs. The usage might not be self-explanatory, we created this guide to help you get everything set up and working properly.

How to request a GPU

In order to use a GPUs for your Job you will have to request a Generic resource (GRES). You can do that by adding the following line to your jobscript:

#SBATCH --gres=gpu:1

This will request one GPU per requested node. Suitable nodes will be automatically chosen by SLURM.

Of couse its possible to request more than one GPU per node, however currently GPU nodes have only one or two GPUs each. With the following line of code you will request 2 GPUs in a single node:

#SBATCH --gres=gpu:2

You will also have to add the following line to your jobscript:

#SBATCH --partition=mpcg.p

This will allow you to use one (or more) of the Tesla P100 cards. You could also add the line

#SBATCH --partition=mpcb.p

which has some nodes with GTX 1080 cards.