RoCE congestion despite an operational routed fabric
An authorized test network carries an RDMA workload over RoCEv2. Every routed link is up, and small probes between endpoints succeed. During a representative incast workload, application latency rises and some queues show drops. The intended class has PFC enabled on selected links, but one host marks traffic differently from the downstream switch's expected mapping. Pause counters rise on several links, while the endpoint congestion-feedback behavior has not been examined. A five-minute utilization chart remains moderate and is cited as proof that there cannot be a network bottleneck. The engineer must connect classification, buffering, congestion control, and measurement resolution rather than simply raise every threshold or declare all routed reachability sufficient.
1. What should be investigated about traffic treatment?
Read the worked answer
Trace actual host marking, trust and remarking, queue mapping, PFC behavior, ECN marking, and applicable endpoint rate adaptation. The configured intended class does not help packets classified elsewhere. Assess congestion spreading and relevant platform guidance before changing thresholds.
2. Why are the current health observations incomplete?
Read the worked answer
Link-up and small probes do not validate workload loss or congestion behavior. A long average can hide microbursts. Use suitable queue, pause, drop, and higher-resolution observations with the representative workload and verify the approved design objectives.