per-user figures from the .cotterell_user_data_usage.txt report Euler writes at the root of each volume every night, volume totals and quotas from lquota; sampled daily at 04:30. Per-user rows can add up to more than the volume figure (files of members in several groups are counted for each group).
GPU queues
es_cott has three SLURM accounts, one per partition family, each with its own share and queue. Click a card to select a queue; every chart and table below shows that queue only.
Showing queue
Concurrent GPUs
hourly, from job start/end intervals; black line = entitled steady state (NormShares × partition GPUs)
GPU-hours per day
Fairshare over time
sampled every 15 min from sshare since the dashboard started; Euler keeps no history of this
Running GPUs by card model
Pending GPUs by reason
Priority: waiting on fairshare. AssocGrpGRES: over the coordinator cap. QOSMaxGRESPerUser: over the QOS cap. Resources: nothing free at that size.
Queue wait, last 30 days
GPU utilization of finished jobs, weighted by GPU-hours (60 days)
gpuutil is a 30 s time average per job, divided by the job's GPU count. Jobs shorter than 10 min excluded. Bursty data-loading workloads read low legitimately.
Members
members with any use or queue activity in the window; click a row for detail. FairShare and LevelFS from sshare; caps from sacctmgr; hours from sacct intervals.
Live queue
Partition capacity
GPU counts per node state from sinfo; partition totals include drained nodes.
QOS per-user caps in force (recomputed by the site over time)