Getting started with GPUs at FSL

Currently FSL has 32 nodes with GPUs, each with 2 NVIDIA K80's. Note that the devices report as 2 GPUs each, however. This means if you query the system it will report that there are 4 K80 GPUs available.

Each node has 64 GB of memory; the CPUs match the m8 cluster exactly. The NVIDIA K80's are a substantial increase in computing power compared to our previous GPUs. For more detailed information about the GPU hardware, see NVIDIA's website.

Using GPUs in Jobs

You can request GPUs in Slurm using the --gres feature. As an example you can request a whole node and its 4 GPUs by including the following in your sbatch options: --nodes=1 --mem=64G --exclusive --gres=gpu:4. Please be aware of how many GPUs your program can use and request accordingly. If you can only use one GPU and 6 processors use: --nodes=1 --ntasks=6 --mem=12G --gres=gpu:1. The environment variable CUDA_VISIBLE_DEVICES will contain a comma separated list of CUDA devices that your job has access to.

To compile or run CUDA code you'll need the CUDA libraries and runtime; get them in your path by doing: module load cuda.

For interactive development or for compiling CUDA programs, you will need to request an interactive job using salloc. The salloc program accepts the same flags as sbatch but you must provide them on the command-line since salloc isn't given a file to run (it gives you a shell instead).

Job Restrictions

Here are our current GPU restrictions:

  • Walltime of 3 days
  • No more than 26 GPU nodes total
  • No more than 26 GPU jobs at any time
  • Your jobs can escape all limits (up to 7 days walltime) by becoming preemptable: --qos=standby

These are subject to change, and administrators may impose additional restrictions as they see fit based on demand.