Dirty Quota-Based VM Live Migration Auto-Converge
summary
Implements a kernel-level microstun that assigns per-vCPU delta quota windows and a shared common quota so dirty rate stays ≤ network_bandwidth/x (default x=2). For a 16 GB VM over 1 Gbps the algorithm halves allowed dirty rate each iteration and reaches a final blackout of ~125 ms (<200 ms). Uses 100 ms windows in experiments, cuts total migration time about 50% in many workloads, converges cases that previously never did (examples: type‑5 ~6 minutes, type‑6 ~30 minutes) while preserving guest read throughput.
AI-written summary · about
chapters (8)
- 0:00Introduction to migration
- 1:06Pre-copy migration overview
- 2:26Limitations of current methods
- 4:26Micro-scanning algorithm design
- 10:09Quota management strategy
- 11:54Algorithm mechanics
- 16:00Performance results
- 19:10Conclusion and future work
related talks
21mKeep Your VM Dancing With Volume Migration
13mUsing CPU & Memory Hotplug for VM Vertical Scaling
23mOptimizing Live-Migration for Minimizing Packet Loss With KubeVirt and...
20mKVM-devirt: Extending KVM to a Zero-Overhead Partition Hypervisor
28mLive-Update: An Update on the Journey to Replace the Hypervisor with Limited Downtime
27mEvaluate Implementation Options of KVM-based Type1 (or 1.5) Hypervisor
24mOptimizing K8s and unlock blockers with Pod Live Migration
27mZero copy migration from VMware to Kubernetes