Many Gaudi clusters host more than one team or customer workload side by side. To keep tenants from interfering with each other:
Kubernetes: each tenant gets its own namespace with resource quotas capping how many Gaudi accelerators it can claim, plus network policies blocking cross-tenant traffic.
Slurm: each tenant gets its own account with a quality-of-service (QOS) policy limiting their share of the cluster.
Network fairness: on shared fabric, we apply traffic-priority rules so one tenant's heavy job can't starve another's network traffic.
Daily/weekly checks: confirm no nodes are stuck in a drained/cordoned state unexpectedly, and that the scheduler's control plane is healthy.
← Previous | Guide Index | Next →