Shared nodes

Shared queues allow multiple users to run jobs on the same node simultaneously. Unlike exclusive nodes, they make efficient use of hardware when your job doesn’t need a whole node, and queue waits are often shorter when you request only what you need. All shared queues are reached through the milan1 / milan2 login nodes.

CPU shared queues

Intel Skylake: 40 cores, 192 GB per node

• short-40core-shared: 1 h default / 4 h max
• long-40core-shared: 8 h default / 24 h max
• extended-40core-shared: 8 h default / 3.5 days max

AMD EPYC Milan: 96 cores, 256 GB per node

• short-96core-shared: 1 h default / 4 h max
• long-96core-shared: 8 h default / 24 h max
• extended-96core-shared: 8 h default / 3.5 days max

GPU shared queues

Intel Ice Lake nodes: 64 cores and 4× NVIDIA A100 per node:

• a100: 1 h default / 8 h max · up to 2 nodes, 2 simultaneous jobs per user
• a100-long: 8 h default / 2 days max · 1 node, 2 simultaneous jobs
• a100-large: 1 h default / 8 h max · up to 4 nodes, 1 simultaneous job

Request memory explicitly

The critical difference from exclusive queues: you share the node’s memory with other jobs, so you must say how much you need:

#SBATCH --mem=<amount>

Best practices

• Check the queue’s per-user limits before submitting.
• Monitor your jobs with squeue -u $USER.
• Request cores and memory precisely: smaller, well-sized requests schedule faster.