Genius quick start guide#

Genius is a KU Leuven/UHasselt Tier-2 cluster which entered production in 2018. Currently only its login nodes and most of its GPU nodes remain in production.

Access to the cluster#

Genius can be accessed from the Genius login nodes, or from your web browser via the Open OnDemand service.

For example, you can log in to any of the login node using SSH:

$ ssh vscXXXXX@login.hpc.kuleuven.be

Running jobs on Genius#

The Slurm Workload Manager is the scheduler, resource manager and credit accounting manager on Genius (and wICE).

In case you are not yet familiar with Slurm and/or the Genius hardware, you can find more information on the following pages:

Genius hardware
Slurm jobs (basics): Running jobs in Slurm and Example job types
Slurm jobs (advanced)
Site-specific Slurm info

Note

The examples given on this page only serve as illustrations. We expect that you adapt the number of nodes, tasks, cores, memory, walltime, … depending on what your compute task requires.

Submitting to a GPU node#

The GPU nodes are accessible via the following partitions:

Partition	Walltime	Resources	CPU model
gpu_p100	3 days	13 nodes, 4x Nvidia P100 GPUs per node	Skylake
gpu_p100_long	7 days	13 nodes, 4x Nvidia P100 GPUs per node	Skylake
gpu_v100	3 days	2 nodes, 8x Nvidia V100 GPUs per node	Cascadelake
gpu_v100_long	7 days	2 nodes, 8x Nvidia V100 GPUs per node	Cascadelake

As usual, GPU nodes can be shared by different jobs from different users. However, every user will have exclusive access to its allocated GPUs.

If you e.g. want to get one P100 GPU and two CPU cores:

$ sbatch --account=lp_myproject --clusters=genius --partition=gpu_p100 \
         --nodes=1 --ntasks=2 --gpus-per-node=1 myjobscript.slurm

For a V100 GPU, select the gpu_v100 partition instead. Note that these nodes also offer a much larger amount of CPU memory.

You are free to request more GPU devices and/or CPU cores if needed, but take note of the limits on CPU resources per allocated GPU).

For easier development and testing, also a gpu_p100_debug partition is available which accepts jobs with walltimes up to 1 hour. Note that you can only have a single debug job in the queue at any time.