High-Speed AI Networking and Fabric Design
This domain tests the lossless Ethernet design that carries RDMA: RoCEv2 classification, PFC, ECN, MTU consistency, topology choice, and the operational telemetry that keeps a congested fabric from collapsing.
Core rules
- RoCEv2 encapsulates RDMA in UDP destination port 4791. NICs tag it with a DSCP value (commonly 26 or 48); the switch maps DSCP to CoS and to a dedicated QoS group.
- PFC (802.1Qbb) provides per-priority pause so that a congested egress queue signals the sender instead of dropping. Enable it only on the RoCE class - turning it on everywhere creates head-of-line blocking.
- ECN plus DCQCN is the congestion-control loop. The switch marks CE bits in the egress queue; the NIC reacts and slows the flow. PFC is the safety net, ECN is the steering wheel.
- MTU must be consistent end to end, typically 9216. One 1500-byte hop silently drops jumbo frames.
- Topology: rail-optimized or rail-aligned leaf-spine for training, standard spine-leaf for inference and general workloads. Use per-packet dynamic load balancing where available to spread flows across equal-cost uplinks.
Worked example: lossless policy
The requirement is lossless RoCEv2 for GPU servers on a Nexus 9000 leaf.
- Classify the traffic: match DSCP 26, set CoS 3, assign QoS group 3.
- In the network QoS policy, set MTU 9216 and enable pause for CoS 3 only, then attach the policy under system QoS.
- Enable PFC watchdog with detection and recovery timers so a stuck or faulty NIC cannot pause the fabric indefinitely.
- Set ECN marking thresholds in the egress queues facing GPU servers, and confirm DCQCN is enabled on the NIC. Reasoning: PFC alone guarantees zero drops but propagates backpressure; ECN gives the NIC an early signal to reduce rate before pause frames are needed. You need both, plus the watchdog as the failure containment.
Traps that cost marks
- Enabling PFC without a watchdog, so one bad NIC can stall a whole leaf.
- ECN thresholds set so high nothing is ever marked, or so low every packet is marked.
- Mixing 1500 and 9216 MTU across the path.
- Quoting non-blocking while uplinks are oversubscribed.
- No telemetry: without per-queue and per-flow counters, congestion is invisible until jobs slow down.