Sustainability the Container Native Way
summary
A Kubernetes-native exporter called kepler exporter collects per-node and per-pod CPU, GPU and DRAM counters via eBPF and vendor APIs, sends snapshots to Prometheus, and uses a remote model server running linear regression models to convert perf counters (instructions, cycles, average frequency) into energy estimates (reported as energy per 3s). Demo runs on microshift with NVIDIA GPU, shows CPU vs GPU workloads, and reports very low agent overhead (~0.02 CPU cores, ~150 MB) while requiring a bootstrap: node-level stats train and validate models (target ~95% validation).
AI-written summary · about
chapters (7)
- 0:00Start
- 1:33Power Measurement Theory
- 2:51Energy Measurement Methodology: Ideal vs Reality
- 5:34Energy Consumption Attribution Methodology
- 6:58Kepler: Kubernetes-based Efficient Power Level Exporter
- 9:14Kepler Architecture
- 28:11Community
related talks
35mSustainability Research the Cloud Native Way
17mKepler: Powering up the efficiency of your cluster
42meBPF-Powered Observab...
05mLightning Talk: Running Energy-Aware AI Inference on Edge Kubernetes With Kepler
23mFrom Pods To PUE: A Digital Twin Approach To Sustainable Kubernetes Infrastructure
40mMeasuring Energy Consumption in Clusters with Kepler
26mAccelerating Science, Decelerating Carbon: Sustainable Computing...
28mHow Green Is My OpenTelemetry Collector?