The GPU resources on Shamu have been configured as a "consumable resource". This means you can request individual GPU cards, nodes, etc... We have two types of GPU resources on Shamu, Tesla K80s and V100s. Below are some sample scripts showing how you can request these resources.

For the Tesla K80 cards:

#SBATCH --partition="gpu"
#SBATCH --nodes=1
#SBATCH --gres=gpu:k80:1

. /etc/profile.d/modules.sh

module load my_app

my_app < job arguments >

For the V100 cards:

#SBATCH --partition="gpu-v100"
#SBATCH --nodes=1
#SBATCH --gres=gpu:v100:1
 
. /etc/profile.d/modules.sh

module load my_app

my_app < job arguments >

We have also introduced one V100 GPU node as a shared resource. It will accept up to 4 jobs at one time. If their are more than 4 jobs the remaining jobs will be held in PENDING status and will be executed on a first come first served basis. To access this new share you have to use the following SBATCH commands or the job will be rejected . The "--memory=XXXX" SBATCH command has to be included. Please change "XXXX" to the amount of memory your job will require. The format can be in MB (megabytes):

#!/bin/bash
#SBATCH --partition=gpu-v100-shared
#SBATCH --memory=XXXX

. /etc/profile.d/modules.sh
module load my_app
my_app < job arguments >

-- AdminUser - 21 Apr 2020