Kepler: Powering up the efficiency of your cluster
summary
Kepler uses eBPF probes to read per-container system counters and export Prometheus metrics so you get actual energy numbers rather than guesses. Where node power meters are unavailable Kepler’s model server runs CPU and DRAM ML models (TensorFlow/Keras, served over Flask) to estimate pod energy, and an RPM lets it work on non‑Kubernetes Linux processes too. A carbon intensity forecaster maps energy to emissions and Kubernetes labels/taints enable carbon-aware scheduling and VPA/KEDA integrations for placement and autoscaling decisions. The project targets measurable reductions against ICT’s 1.8–3.9% global emissions.
AI-written summary · about
chapters (7)
- 0:00Kepler and sustainability
- 1:45Energy measurement without hardware
- 3:30Power measurement approach
- 5:15Energy estimation models
- 8:04Kepler beyond Kubernetes
- 10:11Carbon-aware scheduling
- 13:45Q&A
related talks
40mMeasuring Energy Consumption in Clusters with Kepler
42meBPF-Powered Observab...
40mSustainability the Container Native Way
33mIs Kepler Accurate on Specific Platforms?
26mAccelerating Science, Decelerating Carbon: Sustainable Computing...
23mFrom Pods To PUE: A Digital Twin Approach To Sustainable Kubernetes Infrastructure
28mHow Green Is My OpenTelemetry Collector?
33mWhen Observability Meets Sustainability: A Real World Experience