Prediction-Based Power Oversubscription in Cloud Platforms

Explore a 16-minute conference talk from USENIX ATC '21 that delves into prediction-based power oversubscription in cloud platforms. Learn how Microsoft Research and Microsoft Azure tackle the challenges of increasing oversubscription in public clouds while minimizing performance impact on critical workloads. Discover innovative approaches to identifying performance-critical workloads from opaque VMs, implementing criticality-aware power management, and addressing hardware and software challenges in Microsoft Azure. Gain insights into the results of doubling oversubscription with minimal impact on critical workloads and understand the lessons learned from deploying this technology in production environments.

Intro
Motivation
Prior Work: Power Capping and Oversubscription
Oversubscription Challenges for Cloud Providers
Fine-grained Power Capping for Oversubscription
Per-VM Power Capping Overview
Inferring criticality of opaque VMs (challenge #1)
Full-server throttling (challenge #3)
Oversubscription Strategy with Per-VM Capping
Production Impact and Lessons
Conclusions