From Pods To PUE: A Digital Twin Approach To Sustainable Kubernetes Infrastructure
summary
Claims a Kubernetes-native digital twin named Julie that uses Prometheus telemetry, Kepler/eBPF and the DRA API to compute per-node power headroom and predict PUE impact, then enforces power caps via RAPL and NVML/DCGM. Architecture: node agent, operator, scheduler extender and two CRDs. Evaluation runs in kind+quark with a workload/power simulator over 5,000-node simulations, showing non-negligible reductions in total energy and PUE with no meaningful throughput loss. Installer: Helm charts and a kubectl plugin; next step is bare-metal validation.
AI-written summary · about
chapters (9)
- 0:00Energy constraints in data centers
- 1:03Power usage effectiveness
- 3:32Digital twins for Kubernetes nodes
- 5:37Introducing Julie project
- 10:33Simulating workloads and energy
- 13:24Experimental results and PUE
- 15:13Deployment and monitoring with Julie
- 17:00Live demo of Julie
- 19:50Q&A on energy-aware scheduling
related talks
17mKepler: Powering up the efficiency of your cluster
29mPower Efficiency Aware Kubernetes Scheduler
40mSustainability the Container Native Way
20mKepler - Optimized power usage monitoring for Kubernetes
24mUsing Kubernetes with Data Processing Units to Offload Infrastructure
25mIntelligent Kubernetes Workload Optimization
42meBPF-Powered Observab...
34mKubernetes: Reducing Infrastructure cost and complexity