AI Fabric Observability — Single Pane of Glass rockynet · pod-A · 4,096×H100 · Arista 7060 + SONiC

– healthy – degraded – failed --:--:--
Job, GPU and RoCE/RDMA telemetry on one screen, joined by rank ↔ GPU ↔ NIC ↔ port. Pick a scenario to jump the fleet to a canonical incident, or click any job row to drill down to its host/network/switch verdict — watch the fabric panels stay clean for host faults and light up for fabric faults.
Scenario
Job Completion Time — ETA
MFU
Step p99
AllReduce busbw
of 400
Job avail (SLO 99.5)
RoCE loss
no-drop contract
PFC pause (TC3)
tx_prio3_pause
▚ Job health — click any row to drill down– jobs
Job / podPhaseGPU%°CbusbwQ3 bufPortSM_ACTIVE
◈ Root-cause verdict — the selected job
⌘ Remediation — copy-ready querieshost/switch CLI + PromQL
▦ Switch port × metric — leaf-14 (SONiC)gNMI stream
✧ GPU_UTIL trap — selected nodeDCGM :9400
GPU_UTIL
vs
SM_ACTIVE
◉ RoCEv2 congestion loop — TC3node_exporter · gNMI · WJH
ECN marked /s
CNP /s
packet_seq_err /s
FEC pre-BER
ECN marks vs CNP (60s) — ECN-first, PFC-rarely
Watermark → xoff · et33
⚑ Drops with reason — WJHevent-based
⑆ The Walk-Down — rank → portidle
🔔 Alerts — anomaly, not absolutes