GPU test node (front3)

The GPU front-end node front3 is provided for

• interactive GPU access for development and testing,
• testing GPU batch jobs.

For interactive work front3 is the equivalent of the CPU front-end nodes front1 and front2.

In addition, small batch jobs can be run on front3 be employing --partition=gputest (the default time limit is 10 minutes). The idea is that functionality can be tested. Since all resources of the node are shared all test runs should use as little memory (RAM and VRAM) as possible.

Recall that

• RAM allocation can be checked with the free, top and ps -l commands,
• VRAM allocation can be checked with the nvidia-smi command.

Running batch jobs in the gputest partition

A batch job that is running in the gputest partition can use GPU but must not specify GPU resources (--gpus must not be set, see the example job below).

Differences to the gpu partition

• --gpus (or equivalent) must not be specified.
• /home is writable.
• Resources including the GPU are shared with other batch- or interactive jobs, hence time measurements are not reliable.

Example job

line no.	/sw/batch/examples/gputest-partition/gputest-job.sh
1 2 3 4 5 6 7 8 9 10 11 12	#!/bin/bash #SBATCH --account=WorkingGroupName_gpu #SBATCH --partition=gputest #SBATCH --export=NONE source /sw/batch/init.sh module load cuda/12.5.1 $CUDA_HOME/extras/demo_suite/busGrind exit