Postcopy live-migration with vhost user backend
summary
Adds a post-copy mode to the DPDK vhost-user backend and measures how it changes packet and VM behavior versus pre-copy. Describes the implementation: bitmap tracking of dirty pages, a notify/socket handshake to request missing pages, and the atomic remote-start of the destination VM so execution resumes before all pages arrive. Shows benchmarks: iteration counts and downtime drop but higher page-fault rates and extra network roundtrips, so throughput improves in some NFV workloads while latency-sensitive apps can regress.
AI-written summary · about
chapters (8)
- 0:00Postcopy live-migration intro
- 3:35Pre-copy vs post-copy concepts
- 7:05Memory page tracking
- 10:31Transfer missing pages
- 14:30Cooperative vs non-cooperative
- 17:55Benchmarking performance
- 20:52Fault detection and error handling
- 22:03Conclusion
related talks
25mDPDK performance optimization for edge computing scenario
11mSupporting Hypervisor Kexec with Modern Devices
51mAccelerate containers and VMs networking with vDPA
52mSuggestions to Enhance DPDK to Enable Migration of User Space Networking Applications to DPDK
23mOptimizing Live-Migration for Minimizing Packet Loss With KubeVirt and...
02h 53mPreview Sessions
37mEfficient and Flexible Virtual Machine Networking Through eBPF
36mBringing vDPA to Life in Kubernetes
