A school-district GPU lab enables PFC on Spectrum-X without ECN and then sees head-of-line blocking across leafs. What lesson should operators take?
Select an answer to reveal the explanation.
Short Explanation
PFC without ECN is like slamming every school driveway gate closed whenever one hallway is busy—pretty soon the whole campus stalls. Pair PFC with verified ECN and QoS so congestion is signaled early instead of pausing everything into head-of-line blocking.
Full Explanation
A school-district GPU lab that enables PFC on Spectrum-X without ECN and then sees head-of-line blocking across leafs should learn that PFC without verified ECN/QoS can cause pause-driven HOL blocking; the fix is to enable and verify ECN with PFC. Pause frames stop whole priority groups; without early ECN marking, congestion spreads stalls across leafs instead of slowing senders smoothly. Pairing PFC with working ECN and sensible QoS restores controlled lossless behavior for district AI jobs.
Extending PFC to multiple traffic classes on leaf uplinks does not by itself provide safe lossless delivery for collective flows—the lab’s HOL symptom shows that expanding pauses without ECN just spreads the stall. Raising pause thresholds on the lossless queue does not safely absorb incast bursts with default ECN thresholds left in place; higher pause thresholds without ECN just delay and widen the pause-driven stall across leafs. Tuning DCQCN reaction timers on GPU hosts while continuing PFC-only operation is not enough to manage congestion across leafs; without ECN enabled and verified alongside PFC, timer changes cannot stop pause-driven HOL blocking.
Exam caveat: PFC-without-ECN is a classic cause of pause storms and HOL blocking on AI Ethernet. Operational check: enable ECN alongside PFC on the school-district Spectrum-X RoCE class, verify marks and reduced pause propagation in telemetry, and keep telemetry on while retesting collectives across leafs. Retest multi-leaf collectives after ECN is verified so the school-district lab can show that pause propagation and head-of-line blocking dropped under the same load profile that previously stalled Spectrum-X; keep the before/after counter evidence in the lab runbook.