Resource Limits

Understanding and working within cluster resource limits.

Account Limits

Wall Time Limits

Different QoS’s have different maximum wall times:

Check limits: sacctmgr show qos format=name,maxwall

Memory Limits

Each partition has memory limits per node:

• Standard nodes: 192 GB - 512 GB

• Big Memory nodes: 384 GB - 2 TB

• GPU nodes: 1 TB - 2 TB

Request memory: #SBATCH --mem=64G or #SBATCH --mem-per-cpu=4G

Node/Task Limits

• Maximum nodes per job: Varies by partition

• Maximum tasks per node: Depends on cores available

Queue Limits

Fair Share

The cluster uses fair-share scheduling. Jobs from users who haven’t run recently get priority.

Check your fair share: sshare -u $USER

Quality of Service (QoS)

Different QoS levels have different priorities and limits:

• normal - Standard priority

• high - Higher priority (requires approval)

• low - Lower priority, good for testing

Set QoS: #SBATCH --qos=normal

Best Practices

1. Request only what you need - Don’t ask for 100G if you need 10G

2. Use appropriate partitions - Don’t use GPU partition for CPU jobs

3. Clean up after yourself - Remove old job files and outputs

4. Monitor your usage - Check sacct regularly