A 311 model farm’s leaf shows discard signatures during peak inference. Which Spectrum service should operators use for real-time event analysis?

When a 311 model farm’s leaf shows discard signatures during peak inference, operators should use NVIDIA What Just Happened (WJH) for real-time discard and congestion event analysis on Spectrum. WJH is the Spectrum service aimed at real-time visibility into why packets were dropped or congested, which is exactly what a 311 peak-inference discard signature demands. For municipal model farms, that real-time event analysis shortens mean time to identify whether the leaf is buffering, pausing, or discarding under bursty inference loads.

NVIDIA NetQ fabric monitoring for periodic inventory validation and configuration compliance on Spectrum fails by concept: it is periodic and focused on inventory and configuration compliance, not real-time discard and congestion analysis during peak 311 inference. Spectrum SPAN mirroring to dedicated collectors for offline packet capture and forensic review fails because offline packet capture and forensic review is unrelated to real-time Spectrum buffer and discard behavior during peak inference. Scheduled SNMP counter polling for long-term bandwidth trending and capacity planning on Spectrum fails both as engineering and as observability—it is scheduled and long-term, so it cannot capture real-time discard reasons or congestion events as they occur.

Exam caveat: WJH identifies discard/congestion events; fixing them still requires acting on PFC/ECN, QoS, ECMP balance, or capacity—not merely watching the events. Operational check: enable and query WJH on the implicated 311 Spectrum leaf during peak inference, capture discard reasons and interfaces, correlate with model-farm latency or error spikes, apply remediation, and verify WJH discard signatures subside while 311 SLOs recover.