SLURM Queueing and Partitions
Partitions
SLURM partitions are separate queues for submitted jobs. Each partition has a different set of constraints that controls which jobs should be in that queue.
In most cases, you will not need to specify a partition when submitting a job. Given a specified (or default) account, SLURM's job submit plugin control will automatically send your job to the correct partition. For information regarding changing your account when submitting and ensuring access to buy-in nodes, please see the page on Buy-In and Account Management
The jobs in each partition are evaluated by the scheduler to determine the order in which they are run. See How jobs are scheduled for more information.
Types of Partitions
| Name | Purpose |
|---|---|
| general-short | This partition includes all nodes--buy-in and non-buy-in--and runs jobs that request a wall time of four hours or less. Jobs in this partition are considered for scheduling after jobs in buy-in partitions and the general-long partition. To prevent these jobs from being continuously bumped by general-long/buy-in jobs, they are also submitted to general-long. |
| general-long | This partition includes non-buy-in nodes and allows jobs to run for up to seven days. Jobs with a requested wall time of four hours or less are also submitted to the general-short partition. |
| general-long-bigmem | This partition includes non-buy-in nodes with more memory and CPU cores than most nodes and allows jobs to run for up to seven days. Jobs requesting more than 256GB or 40 CPUs per node are automatically submitted to this partition. This partition ensures large jobs get priority access to large nodes over jobs that can run elsewhere. Jobs with a requested wall time of four hours or less are also submitted to the general-short partition. |
| general-long-gpu | This partition contains non-buy-in nodes with GPUs and allows jobs to run for up to seven days. Jobs requesting GPUs are automatically submitted to the partition. This partition ensures jobs requesting GPUs get priority access to nodes with GPUs over jobs not requesting GPUs. Jobs with a requested wall time of four hours or less are also submitted to the general-short partition. |
| Buyin Partition (names vary) | A partition is created for each buy-in account. Each buy-in partition includes all non-buy-in nodes, allowing buy-in jobs to span buy-in and non-buy-in nodes. These jobs get equal consideration for scheduling on non-buy-in nodes as jobs in general-long. When jobs submitted to these partitions request a wall time of four hours or less, they are also submitted to the general-short, enabling them to use other available buy-in nodes and ensuring they are scheduled as fast as possible. |
How jobs are assigned to queues
Depending on whether you are using the general account or a buy-in specific account, your job will be submitted to each of the following queues depending on the wall time, memory, and GPUs requested.
| Default Account | Wall Time<=4 Hours | Wall Time<=4 Hours and >=256GB or 40CPU | Wall Time<=4 Hours and GPUs Reqested | Wall Time>4 Hours | Wall Time>4 Hours and >=256GB or 40CPU | Wall Time>4 Hours and GPUs |
|---|---|---|---|---|---|---|
| general | general-long, general-short | general-long-bigmem, general-short | general-long-gpu, general-short | general-long | general-long-bigmem | general-long-gpu |
<buyin> |
<buyin>, general-long, general-short |
<buyin>, general-long-bigmem, general-short |
<buyin>, general-long-gpu, general-short |
<buyin>, general-long |
<buyin>, general-long-bigmem |
<buyin>, general-long-gpu |
where <buyin> is the name of buyin account.