Virtualization · Cluster
Live Migration
State transfer and downtime bounds.
- 01 The Stop-and-Copy Boundary in Live Migration Why a migrating virtual machine must freeze its vCPUs at one exact instant, what state that instant must capture, and how the residual copy sets the floor on downtime.
- 02 Dirty Page Rate and the Convergence Condition Pre-copy migration only terminates when the guest dirties memory slower than the link can carry it. The ratio of those two rates decides the pass count, the total data moved, and whether the migration finishes at all.
- 03 Hardware Dirty-Bit Tracking for Guest Memory How second-level page tables record which guest pages changed, why the first pass needs write protection, and what the bitmap costs in exits when the CPU cannot set the bit itself.
- 04 Encoding Guest Memory for the Migration Stream Zero pages, compression, and delta encoding decide how many bytes a migration actually sends. The choice trades host CPU and per-page framing overhead against bandwidth, and each option fails on a different workload.
- 05 Serializing Device State Across a Migration Emulated devices carry configuration registers, queue indices, and pending interrupts that must be reconstructed exactly on the destination. The format is versioned, and the failure mode is a guest that resumes with a device it does not recognise.
- 06 Network Path Continuity Across a Migration Moving a guest changes the port its frames egress from. Switch forwarding tables, ARP caches, and in-flight TCP segments all have to be re-pointed, and the guest must not observe a reset connection.
- 07 Storage Handoff: Moving a Guest's Disk With It Block devices cannot be frozen and copied like RAM. Either the disk stays put and is re-attached, or its contents move while the guest writes, and each option has a different consistency failure.
- 08 Post-Copy Migration and Demand Paging from the Source Instead of shrinking the dirty set, post-copy moves the vCPUs first and faults memory back from the source on demand. It bounds downtime at the cost of fault latency and a source that cannot be released.
- 09 Abort Paths and the Blackout Window A migration can fail at any point, and what the guest experiences depends on which side still holds a complete state. The ordering of commit, acknowledgement, and source teardown defines what rollback is possible.
- 10 Migrating Guests Whose Memory the Hypervisor Cannot Read Confidential computing encrypts guest memory under keys the hypervisor does not hold, which removes compression, zero-page detection, and inspection from the migration path and adds an attestation step at the destination.