RoCE traffic suffers under congestion on an Ethernet fabric. Which features are key to making it perform?
Why C is the answer
PFC and ECN for lossless, congestion-aware delivery
RDMA over Ethernet is very sensitive to loss. PFC prevents drops and ECN signals congestion early.
In detail
- PFC and ECN give lossless, congestion-aware delivery.
- RDMA over Ethernet is very sensitive to loss.
Use case
An illustrative scenario, not a real incident.
Training slows under congestion, because RoCE traffic is losing packets. PFC stops the drops, and ECN warns senders early.
See it in different ways
flowchart LR A["RoCE traffic"] --> B["QoS class"] B --> C["ECN signals congestion"] C --> D["PFC prevents drops"] D --> E["Adaptive routing spreads load"]
sequenceDiagram participant S as Sender participant W as Switch participant R as Receiver S->>W: RDMA traffic W->>R: Marks packets with ECN R-->>S: Congestion notification S->>S: Slows down W-->>S: PFC pause, if a queue fills
mindmap
root(("RoCE tuning"))
n1["PFC"]
n2["Prevents drops"]
n3["ECN"]
n4["Signals congestion early"]Key point: RoCE needs PFC and ECN.
Why the other options are wrong
- A. DHCP and DNS tuning on the leaf switches
- Those are about addresses and names.
- B. NAT and PAT at the fabric edge
- Those are about address translation.
- D. STP and LLDP for a loop-free, discovered topology
- Those are about loops and discovery.