KVM-devirt: Extending KVM to a Zero-Overhead Partition Hypervisor
summary
KVM-devirt extends KVM to run bare-metal partitions that eliminate VM exits and remove stage-2 and IOMMU DMA-remap translations, passing APIC registers, physical ICR and PSC deadline MSRs straight to VMs and mapping GFNs↔PFNs at boot. Implementation works on Intel and AMD with QEMU and Cloud Hypervisor; microbenchmarks show IPI and timer latency equal to native and a 20–30% end-to-end latency improvement over upstream KVM; with four partitions, tail latency improves about 9%. Live migration and upstream patches are planned.
AI-written summary · about
chapters (9)
- 0:00Start
- 0:44Motivation
- 2:41KVM Virtualization Overhead
- 4:38Interrupt Passthrough
- 5:56Interrupt Remap
- 6:58IPI Passthrough
- 8:11Timer Passthrough
- 9:24Memory De-virtualization
- 10:49DMA De-virtualization
related talks
27mEvaluate Implementation Options of KVM-based Type1 (or 1.5) Hypervisor
32mNotes on Taking KVM-on-KVM Nested Virtualization for a Spin
28mLive-Update: An Update on the Journey to Replace the Hypervisor with Limited Downtime
24mHow Hard Could it be to Flip a bit? KVM PV Feature Enablement up the Virtualization Stack
17mMitigating Excessive Pause-Loop-Exiting in VM-Agnostic KVM
18mHypervisor-less Virtio: Assembling Multi-OS Systems Using Standards-based Pro...
11mBuild your own virtual machine with /dev/kvm and Rust!
10mDemo: Nested Virtualization Acceleration