Slurm allocation policies#

Introduction#

At SCITAS, we aim to allocate all computing resources in the clusters as efficiently as possible. Our allocation policies are driven by Quality of Service (QoS), ensuring that resources are distributed based on user needs.

Partitions are organized by node type, allowing users to select the most suitable hardware for their tasks. We do not impose per-user limits or per-account restrictions, enabling us to adapt continuously to user demands without impacting running jobs.

Both the "free" and "debug" QoS are free of charge, offering accessible options for experimentation and short-term resource needs.

Allocation policies on Jed#

The allocation policies on Jed follows a hierarchical structure, with partitions organized by node type and various job QoS levels controlling resource distribution.

Bigmen and Hugemem nodes have their own dedicated QoS, accessible exclusively to pay-per-use accounts.

The operations team allocates resources between parallel and serial jobs based on user demands.

The following diagram illustrates the current policy implemented on the Jed cluster.