intro

Device query on Puhti

1) Start an interactive session with 1 GPU:

salloc --partition=gpu --time=00:15:00 --ntasks=1 --cpus-per-task=10 \
  --mem-per-cpu=8000 --gres=gpu:v100:1 --job-name=d_query \
  --account=<training_project> --reservation=<reservation name>

Log in on the running node ssh rxxgxx.

1. Load the hip module, then run the command nvidia-smi. Investigate the output
2. Load the hipmodule with module load hip and then run the command $CUDA_INSTALL_ROOT/extras/demo_suite/deviceQuery. Comment on the output
3. Check more examples from $CUDA_INSTALL_ROOT/extras/demo_suite/.

1B) Alternatively, use a non-interactive job

Instead of using an interactive session, run all the commands above in a non-interactive way by adding the commands to the submission script below.

#!/bin/bash
#SBATCH --job-name=d_query
#SBATCH --account=<training_project>
#SBATCH --reservation=<reservation name>
#SBATCH --partition=gpu
#SBATCH --time=00:05:00
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=10
#SBATCH --mem-per-cpu=8000
#SBATCH --gres=gpu:v100:1

srun nvidia-smi

2) Repeat the tasks above with 2 or more GPUs.

3) Compare Puhti GPUs with Mahti GPUs.