RoCESim is a lossless-Ethernet benchmark harness for RoCEv2-style fabrics. It sweeps queue thresholds, ECN marking points, and oversubscription to estimate throughput collapse, PFC pressure, and tail latency risk.
Leaf-spine fabric discussions often stop at "use PFC and ECN," but do not quantify where those settings stop being safe under collective traffic bursts.
Gap statement:
Existing RoCE deployment guides describe lossless settings, but fail to show how PFC thresholds, ECN trigger points, and oversubscription interact under bursty AI training traffic.
- xAI: Ethernet fabric scaling and congestion triage
- NVIDIA: RoCEv2 and Spectrum-style lossless-fabric reasoning
make setup
make sweep
make test