Power Efficiency Aware Kubernetes Scheduler
summary
PEAKS is a Kubernetes scheduler that uses Kepler (eBPF → Prometheus) metrics and an online multi‑objective learner (A.D) to pick pod replica counts and node placements. On a 6‑node heterogeneous lab cluster (256 CPUs, 128 CPUs used) running DeathStarBench Hotel Reservation, the optimizer finds configurations that deliver 5× P99 improvement at similar power, and a best low‑power run at ~109 W versus ~159 W baseline. The system is functional but configs take ~5 minutes to settle and convergence is slow. PEAKS scheduler
AI-written summary · about
chapters (10)
- 0:00Peaks project introduction
- 1:50Peaks with Kepler metrics
- 3:40Online feedback scheduling
- 5:47Cluster setup and workload
- 8:57Power and performance trade-offs
- 11:29Learning agent for pod placement
- 15:27Optimal configurations analysis
- 19:40Key findings from experiments
- 21:04Limitations and future work
- 24:18Q&A on agent inputs
related talks
35mSustainability Research the Cloud Native Way
20mKepler - Optimized power usage monitoring for Kubernetes
29mEdge DC Energy Efficiency: A K8s Workload Allocation Optimizer with Generic Server Power Model
27mKepler Watt Store: Kepler Software Watt Watcher Store
23mFrom Pods To PUE: A Digital Twin Approach To Sustainable Kubernetes Infrastructure
42mDeep Dive Into the Latest Kubernetes Scheduler Features
30mDeep Dive: Scheduling SIG
15mOptimizing Application Performance on Kubernetes