Compute Partitions
There are several different Slurm partitions for different purposes that have different characteristics. The partitions are grouped into CPU and GPU partitions based on their purpose:
Which partition to choose?
If you do not request a partition, your job will be placed in the default partition, which is the NHR standard96 partition for all users.
This means if you are a non-NHR user, you will either receive an error like Invalid account or account/partition combination specified or your job will be pending indefinitely with the reason PartitionConfig.
Which partitions you have access to is listed on Types of User Accounts.
For example, SCC users should specify --partition scc-cpu instead for their non-GPU batch jobs.
Many partitions have variants with a suffix appended to their name (e.g. medium40:shared).
The :test partitions are, as the name suggests, intended for shorter and smaller test runs.
These have a higher priority and a few dedicated nodes, but are more limited in walltime and number of nodes you can use.
For NHR users, there are :shared partitions, where it is possible to reserve less than a full node.
This makes them suitable for pre- and postprocessing for instance.
Non-NHR partitions are shared by default (with the exception of kisski) and don’t enforce reservation of entire nodes, without the suffix in their names.
A job running on a shared node is only accounted for the number of CPU cores/GPUs/memory fraction it uses.
All non-shared nodes are exclusive to one job, which implies that the full compute cost per node are accounted, even if fewer resources were requested.
The available home/local-ssd/work/perm storages are discussed in Storage Systems.
An overview of all partitions and status of nodes can be shown by running:
sinfo -r
For detailed information on a specific node, run:
scontrol show node <nodename>