Skip to content

SLURM partition and QOS


UBELIX provides different CPU and GPU architectures. Furthermore, we provide job different queues for different job priorities and limitations.

restructuring 04.05.2021

With the maintainace 04.05.2021 we restructured the Slurm partition to:

  • more effcient resource usage: All users get access to all resources. Investors have priviledeged access
  • general access to GPUs without preempting
  • resource sharing (fair share) based on research groups using Workspaces, instead of institute based sharing
  • simpler partition design

There are three main control mechanisms to specify queues and resources:


There are the current 3 partitions, epyc2 is default:

Partition job type CPU / GPU node / GPU memory local Scratch
epyc2 single and multi-core AMD Epyc2 2x64 cores 1TB 1TB
bdw full nodes only (x*20cores) Intel Broadwell 2x10 cores 156GB 1TB
gpu GPU
(8 GPUs per node,
varying CPUs)
Nvidia GTX 1080 Ti
Nvidia RTX 2080 Ti
Nvidia RTX 3090
Nvidia Tesla P100

The current usage can be listed on the UBELIX status page


Within these partitions, QoS are used to distinguish different job limits. In each partition there is a default QoS (bold below). Each QoS has specific limits:

Partition QoS time limit cores/node/GPU limit max jobs
epyc2 job_epyc2 4 days 512 cores array jobs up to 10000 tasks
job_epyc2_debug 20 min 20 cores 1
job_epyc2_long 15 days 64 cores 50
job_epyc2_short 6h 10 nodes 50
bdw job_bdw 24 h 50 nodes 300
bdw_debug 20 min 2 nodes 1
bdw_short 6 h 2 nodes 10
gpu job_gpu 24 h 6x GTX 1080 Ti
2x RTX 2080 Ti
1x RTX 3090
1x Tesla P100
job_gpu_debug 20 min 1 GPU 1
job_gpu_preempt 24 h 12x GTX 1080 Ti
4x RTX 2080 Ti
4x RTX 3090
4x Tesla P100

The QoS job_epyc2_short and job_gpu_preempt have access extended resources. In case of job_gpu_preempt, these jobs will be pre-empted if not enough resources for the investors.

Thus a job can be submitted to the gpu partition using a RTX3090 and allow pre-emption:

sbatch --partition=gpu --qos=job_gpu_preempt --gres=gpu:rtx3090

Please see for more details: GPUs

Investor QoS

Investors get elvated priviledges on certain resources. The membership to these investor priviledges will/are managed on basis of HPC Workspaces, see Workspace Management.

To utilize the invested resources, members need to specify one of the following investor partitions:

  • epyc2-invest,
  • bdw-invest,
  • gpu-invest

As an example, the members of an GPU investor, submit jobs with:

module load Workspace         # use the Workspace account
sbatch --partition=gpu-invest

previous empi investors now use bdw-invest partition and GPU investors use gpu-invest

Technical Details:

Within the investor partition, investor QoS are specified. These have the format job_<partitionBase>_<investorID>, e.g. “job_bdw_aschauer”. These QoS will be default for the members in the investor partition. And therewith do not need to be specified. You can check your investor QoS and the used partitions using:

sacctmgr show assoc where user=$USER format=user%20,account%20,partition%16,qos%40,defaultqos%20