Stop One Slurm Job From Starving Your Other Workloads
One data-heavy job hits shared storage, and every high-priority workload behind it slows down. Slurm schedules compute, but it can't see storage bandwidth. Noisy neighbors surface only after GPUs sit idle.
This demo shows how VAST AI Operating System (VAST AI OS) closes that gap. Slurm injects its job QoS and job ID into each job's environment. VAST maps those identifiers to a storage-side QoS policy, so limits follow job identity.
You can throttle one job to make room for another without touching the application. When a training run slows down, you can name the job, host, and mount path driving the traffic, and separate heavy reads from metadata operations.
Watch a three-phase test: a synthetic distributed training workload with heavy random reads and deep queues over NFS, across three Slurm jobs and two tenants. Just fill out the form to watch the 3-minute demo and see QoS enforced end-to-end.