Multi-tenancy patterns for shared GPU clusters

Isolate tenants at three levels: namespaces + RBAC + quotas for access, network policies for traffic, and MIG or dedicated node pools for the GPUs themselves. GPU memory remanence means shared GPUs should be scrubbed or hard-partitioned for sensitive tenants.

Use a fair-share scheduler (Kueue/YuniKorn) so no team monopolizes the pool, with priority and preemption for mixed batch/interactive workloads.

Attribute cost with per-namespace utilization and chargeback — it's the pressure that keeps utilization high and idle capacity reclaimable.