Running GPU Jobs

Running GPU Batch Jobs

Requesting GPU resources

MonARCH is equipped with the following GPU cards:

  • P100 cards

  • K80 cards

When requesting a Tesla K80 GPU, you need to specify --gres=gpu:K80:<no of cards>

#SBATCH --gres=gpu:K80:1

When requesting a P100 GPU, you need to specify --gres=gpu:P100:<no of cards>

#SBATCH --gres=gpu:P100:1"

Sample GPU Slurm scripts

To submit a job, if you need 1 node with 3 cores and 1 GPU, then the slurm submission script should look like:

#!/bin/bash
#SBATCH --job-name=MyJob
#SBATCH --time=01:00:00
#SBATCH --ntasks=3
#SBATCH --cpus-per-task=1
#SBATCH --gres=gpu:K80:1

If you need 6 nodes with 4 cpu cores and 2 GPUs on each node, then the slurm submission script should look like:

#!/bin/bash
#SBATCH --job-name=MyJob
#SBATCH --time=01:00:00
#SBATCH --ntasks=24
#SBATCH --ntasks-per-node=4
#SBATCH --cpus-per-task=1
#SBATCH --gres=gpu:K80:2

Compiling your own CUDA or OpenCL codes for use on M

MonARCH has been configured to allow CUDA (or OpenCL) applications to be compiled (device independent code ONLY) on the Login node (no GPUs installed) for execution on a Compute node (with GPU).

../../_images/massive-gpu-structure.png

Login node: can compile some of CUDA (or OpenCL) source code (device independent code ONLY) but cannot run it

Compute node: can compile all CUDA (or OpenCL) source code as well as execute it.

We strongly suggest you compile your code on a compute node. To do that, you need to use an smux session to gain access to a compute node

smux new-session --gres=gpu:K80:1

Once your interactive session has begun, load the cuda module

module load cuda

To check the GPU device information

nvidia-smi
deviceQuery

Then you should be able to compile the GPU code. Once compilation has run to completion, without error, you can execute your GPU code.

Attention

If you attempt to run any CUDA (or OpenCL) application (compiled executable) on the Login node, ‘no CUDA device found’ error may be reported. This is because no CUDA-enabled GPUs are installed on the Login node. You must run GPU code on a compute node.