Explore a 17-minute conference talk that delves into the challenges of excessive Pause-Loop-Exiting (PLE) in VM-agnostic KVM environments. Learn about the impact of oversubscribing virtual CPUs and critical sections on physical CPUs. Discover innovative approaches to improve guest performance by addressing excessive vCPU spinning, including KVM's hardware-based strategy and methods to suppress PLE events. Examine the causes and consequences of continuous PLE events, and understand the process of selecting candidate vCPUs for mitigation. Investigate two key mitigation techniques: Strict Boost and Debooster, along with their implementation and potential scheduler mismatches. Analyze experimental results showcasing the reduction of PLE occurrences and improvements in system fairness. Gain valuable insights into optimizing virtualization performance and enhancing overall system efficiency in KVM environments.
Mitigating Excessive Pause-Loop-Exiting in VM-Agnostic KVM
Overview
Syllabus
Intro
Oversubscribing virtual CPUs
Critical section on PCPU
Improving guest performance by solving excessive vCPU spinning
KVM approach with hardware feature
KVM's strategy to suppress PLE events
What happens in the worst case!
Continuous PLE events are NOT rare
Why PLE events occur continuously?
How to select candidate vCPU
Mitigation: Strict Boost
Problem: Scheduler Mismatch
Case Study: Scheduler Mismatch
Mitigation: Debooster
Implementation
Experimental Setup
Evaluation: Reduction of PLE Occurrences
Evaluation System Fairness
Conclusion
Taught by
Linux Foundation
