Beyond Software Estimates: Measured Per-Pod...
summary
A Kubernetes NRI plugin plus a Go resctrl library reads Linux 7.0 resctrl group counters and Intel AET core energy and activity metrics, binds each container PID at the runtime start-container hook so child forks never escape, and exports per-pod joules and work estimates via Prometheus or OTLP to Grafana. Live demo shows two 100% CPU pods using ~1.5 W versus ~3.3 W, requires CAP_SYS_ADMIN and host PID, open source in containers/nri-plugins.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 1:13Challenges in energy tracking
- 2:47Technological enablers
- 4:20Resctrl background
- 5:31Intel energy telemetry
- 6:53Overcoming the fork race
- 8:41NRI plugin implementation
- 11:18Pod lifecycle management
- 12:59Software architecture
- 14:56Observability integration
- 16:08Human-readable metrics
- 17:01Production readiness
- 18:43Live demonstration
- 23:32Deployment instructions
- 23:58Future work and conclusion
