GPU data-wait ratio · workload
Delivered bandwidth · workloadL3 → GPU total · IOPS
Zero-copy path ratioGDS + GDR bytes ÷ total
Cache/tier hit ratioLookups at L1 and below
Host CPU · node averageKPI 7 · target ≤ 5%
Data path
GPU → storage, selected workloadForecast flow · by layer
Storage layer = KETI (eBPF + Informer) · other layers = Gluesys ·Forecast bottlenecks · proactive actions
GPU wait and , last 2 min
Dashed line = policy changeBottleneck diagnosis
Recommended actions · toggle to preview impact
Agent query
Answers by querying metrics and events, with evidencePick a suggested question or type your own.