Prediction-Based Power Oversubscription in Cloud Platforms
summary
Claims a 2x increase in safe power oversubscription by combining an ML model in Resource Central (99% accuracy at identifying user‑facing VMs) with enhanced VM placement and hardware per‑VM power capping (per‑core DVFS). Measured gains: ~30 W reclaimed per server versus full‑server throttle, 12.1% chassis power harvest (vs 6.2% prior) and roughly $155M DC savings estimate. System is implemented in Azure (thousands of servers), reduces throttling of critical VMs, and uses feedback to correct mispredictions.
AI-written summary · about
chapters (11)
- 0:00Intro
- 0:19Motivation
- 1:11Prior Work: Power Capping and Oversubscription
- 2:25Oversubscription Challenges for Cloud Providers
- 3:27Fine-grained Power Capping for Oversubscription
- 4:22Per-VM Power Capping Overview
- 6:37Inferring criticality of opaque VMs (challenge #1)
- 8:54Full-server throttling (challenge #3)
- 10:26Oversubscription Strategy with Per-VM Capping
- 12:09Production Impact and Lessons
- 14:45Conclusions
related talks
19mResource Management Innovation for AI Workloads in Cloud Computing
12mAI-Powered Predictive Analytics for Multi-Cloud Management
29mEdge DC Energy Efficiency: A K8s Workload Allocation Optimizer with Generic Server Power Model
11mUnlocking the Power of Machine Learning for Cloud Workload Opt.
47mScaling FinOps with a holistic approach to resource optimization
27mGreening the AI Cloud: Validating Power Models for Kubernetes Containers
20mPredictive Analytics for 99.98% Oracle Cloud Availability
35mSustainability Research the Cloud Native Way