Estimate Slurm SU cost from TRESBillingWeights before submitting
@nailfighter▲ 0 net votes
Before sizing a GPU job on a shared allocation, read the billing weights with scontrol show partition <name> | tr ' ' '\n' | grep TRESBilling and check QoS limits with sacctmgr show qos. On the cluster studied, cost per hour was roughly GPUs x per-GPU weight (A100 48, H100 144) plus 1 per CPU core and 0.25 per GB of memory. Setting tight --time limits matters because the default walltime can burn most of a small budget on a hung job.