Towards a More Efficiently Synchronization in KVM
summary
Identifies three concrete KVM issues: busy‑waiting wastes host CPU when a vCPU is preempted and causes long waits for TLB shootdowns, so the patch boost user-mode vCPU for IPI/TLB work and measures a ~10% pbz2 speedup in a 96‑vCPU guest; scheduler packing causes lock‑holder starvation so disabling weaker fine placement increases a benchmark by ~17% on large-scale VMs; and RCU reader preemption in guests leads to unbounded callback memory growth. Patches are implemented for boosting and placement; RCU remains a tougher corner case.
AI-written summary · about
chapters (10)
- 0:00Agenda
- 3:48Kernel Mode
- 4:45Vistapo Mode
- 5:09Visible Sled
- 5:50Scheduler
- 7:00Scheduler Stacked
- 7:33Disable VCPU
- 8:09Issue
- 8:32Critical Section
- 9:02Conclusion
related talks
25mThe Challenges of Asynchronous Teardown of KVM Guests
33mPreserving IOMMU States During Kexec Reboot
15mKVM Live Upgrade with Properly Handling of Passthrough Devices
39mKVM Modifications for Clustered Operation
29mPostcopy Preemption
29mQEMU & KVM Automated Performance Benchmarking @ SUSE
22mQEMU/KVM Upgrade Test (Stable Guest ABI and in Place Upgrade)
28mHyperscale Kernel Live Patching at Meta