A city NOC configures ETS on the municipal AI fabric. Which use of strict priority is appropriate?

An AI data center's most critical concern: a fabric that handles multi-hop RoCE without drops. The RoCE traffic must be isolated in an ETS class with dedicated bandwidth and protected by PFC so frames are paused (not dropped) when that class's queue fills. That isolation is what keeps RoCE collectives from being starved or broken by bulk background traffic. Enabling ETS strict priority only for a tiny telemetry class, and assigning bulk RDMA to bandwidth-managed ETS groups misplaces strict priority on telemetry instead of control; strict priority should be reserved for the tiny control class, not telemetry. Enabling ETS strict priority only for a tiny control class, and placing bulk RDMA in a best-effort ETS group leaves bulk RDMA without bandwidth guarantees; bulk RDMA should be in a bandwidth-managed ETS group, not best-effort. Enabling ETS strict priority only for a tiny firmware-update class, and assigning bulk RDMA to bandwidth-managed ETS groups misplaces strict priority on firmware updates instead of control; strict priority should be reserved for the tiny control class, not firmware updates. Exam caveat: PFC alone is not enough; a separate ETS class for RoCE ensures fabric resources are reserved for it, not just protected from drops after the fact. Operational check: configure ETS with a dedicated RoCE class and a separate class for bulk traffic, enable PFC on the RoCE Cos, and verify that RoCE frames never see drops even under sustained background load.