Platform R: Slurm Partitions
A Slurm partition acts as a queue. It is a collection of nodes which can differ by hardware resources, the kinds of jobs suitable for those nodes, as well as job priority.
Available Partitions
Platform R has the following partitions:
cpu: the default partition. Anyone can submit batch jobs to this partition.gpu: for access nodes with a GPU. Anyone can submit batch jobs to this partition.interactive: for interactive jobs started withsrun. Anyone can use this partition, but it should not be used for batch jobs. You must also set the QoS to use this partition, with--qos=interactive_qos. The interactive partition does not currently have any GPUs.- Lab-owned: these partitions are named after the PI or lab that owns the machine. Only members of the lab can submit jobs to these partitions.
Partitions can overlap. It is possible to submit a job to the cpu partition which the Slurm scheduler could start on a lab-owned node. In that case, if a lab member submits jobs to their partition, your job might be preempted (quit and possibly restarted on another node).
Run scontrol show partition <partition-name> to get details about a particular partition.
Selecting Partitions
The -p or --partition option is used to select partitions. Most of the time you will only select one, but you can select several, separated by commas and no spaces. When you specify several partitions, the one that can offer the soonest start time is used.
When using the interactive partition you must also specify the interactive QoS:
srun --partition=interactive --ntasks=1 --mem=8G --qos interactive_qos --pty bash -i