Kepler - Optimized power usage monitoring for Kubernetes
summary
Kepler reports real‑time power at process, container and pod granularity and exports metrics to Prometheus (scraped from port 9102) so dashboards show values like hundreds of joules per sample; it uses eBPF for per‑process usage and reads CPU/GPU sensors via RAPL and NVML, then multiplies utilization ratios by dynamic power. Kepler runs as a Helm chart on GKE in the demo, trains a regression power model for VMs because direct sensors aren’t available, and currently does not convert wattage to carbon emissions.
AI-written summary · about
chapters (9)
- 0:00Introduction and Speaker Background
- 1:02Agenda Overview
- 1:34Understanding Sustainability
- 3:27Challenges in Sustainability
- 6:00Solutions from CNCF
- 7:52Introduction to Kepler
- 9:48Kepler's Architecture and Functionality
- 13:31Kepler Demo
- 15:54Q&A Session
related talks
05mProject Lightning Talk: Kepler Project Updates
20mMonitoring Energy Consumption and Building Energy Efficient Systems the Cloud Native Way
27mKepler Watt Store: Kepler Software Watt Watcher Store
20mEnergy Efficient Placement of Edge Workloads
28mExploring Kepler's Next Chapter: Achieving Cloud Native Sustainability With MCP Integration
35mSustainability Research the Cloud Native Way
27mGreening the AI Cloud: Validating Power Models for Kubernetes Containers
42meBPF-Powered Observab...