Exploring an Architecture-neutral KVM MMU
summary
KVM’s TDP MMU gets refactored so about 90% of the tdpmu becomes architecture-neutral and portable to ARM, keeping features like eager page splitting, lockless write-unprotection for dirty logging and parallel fault handling. Google Cloud scale motivates this: VMs with up to 400 vCPUs and 12 TB RAM, and recent ARM t2a VMs (up to 48 vCPUs) expose the same bottlenecks. Work is feasible but needs ARM fixes (weak memory barriers, break‑before‑make and cache maintenance) and won’t initially support pkvm.
AI-written summary · about
chapters (8)
- 0:00Introduction to KVM MMU
- 1:12Scalability and development
- 4:21Arm implementation challenges
- 7:24Architecture neutral design
- 8:53Bridging x86 and Arm
- 17:18Integrating pKVM
- 19:23Conclusion and Q&A
- 22:14Technical discussion
related talks
41mBuilding the Future Trends in Modern Application Architecture
40mVirtual Topology for Virtual Machines: Friend or Foe?
32mOptimizing Resource Usage in Kubernetes
27mInsights on the Re-Architecting to a Single Codebase, Running on Different Hardware...
46mAsymmetric Multiprocessing and Embedded Linux
39mOptimizing CephFS with Combining MDS QoS Scheduling and Static-Dynamic Subtree Partitioning
15mvBOIDs: Taming Chaos via Coarse-Grained Scheduling Abstraction for Containers
11mBuilding a Test for Diverse Fleet of Platforms, How Hard That Could Be?