Edge DC Energy Efficiency: A K8s Workload Allocation Optimizer with Generic Server Power Model
Talkat Kubernetes on Edge Day North America 2024 · on Tuesday, 2024-11-12· at 15:55· Salt Palace | Level 2 | 250 A·
summary
A Kubernetes-compatible Workload Allocation Optimizer uses a generic server power model (ML models managed in MLflow, XGBoost performs best) plus custom front/back static pressure difference sensors to predict per-node power delta and score nodes for placement every minute. Evaluation on a real testbed of 200+ servers shows RMS errors ~1–2% and peak energy reductions up to 51% (varies by CPU family; AMD ~25% max), with the scheduler implemented as K8s filter/score/bind extensions and the code available on GitHub.
AI-written summary · about
chapters (8)
- 0:00Edge DC energy efficiency intro
- 1:29Global data center power trends
- 2:56K8s workload allocator concept
- 4:50Server power modeling approach
- 8:05W algorithm in K8s
- 14:47K8s scheduler power extension
- 18:48Power reduction testbed results
- 22:37Summary & Q&A
related talks
29mPower Efficiency Aware Kubernetes Scheduler
35mSustainability Research the Cloud Native Way
27mGreening the AI Cloud: Validating Power Models for Kubernetes Containers
27mEnvironmental Sustainability: An Energy-Aware Datacenter Operating System- 27mGreening the AI Cloud: Validating Power Models for Kubernetes Containers
16mPrediction-Based Power Oversubscription in Cloud Platforms
27mLeverage ML to automatically tune K8s apps
11mK8s Resources Ensuring Performance and Optimization