Vendor-Neutral GPU Observability for Kubernetes
summary
GPU observability unifies vendor metrics by instrumenting scheduling rather than scraping drivers: HAMi records scheduling-layer instrumentation and reports per-workload allocation efficiency, utilization and memory pressure across NVIDIA, AMD, Ascend and any device-plugin accelerator. It injects a device config via a mutating webhook + runtime shim, normalizes vendor counters with adapters and exposes scheduler metrics (reserved vs used memory, sharing, noisy neighbour signals). Project is CNCF incubation with six+ case studies and active vendor adapters and vGPU monitors.
AI-written summary · about
chapters (11)
- 0:00Introduction to GPU observability
- 1:25Vendor-specific GPU metrics challenges
- 3:09GPU vs CPU scheduling differences
- 4:36Karmada scheduler role explained
- 6:42Harmoni metrics enrichment process
- 8:29Key workload metrics to monitor
- 10:55Standardizing metrics with Prometheus
- 12:20Vendor contributions and limitations
- 14:29Normalizing metrics across devices
- 15:52Dynamic Resource Allocation status
- 16:56Project status and community
