How UCloud Optimizes KVM Live Migration for Faster, Safer Cloud Operations

Introduction

In the previous article, the principles of KVM cross‑host live migration were analyzed. To address the complexity of cloud environments and diverse user needs, UCloud optimizes several aspects of the migration process.

Preparation Stage

During preparation, the host must match the business type and have sufficient free memory, disk, and appropriate CPU specifications. Bandwidth limitations are considered, and parallel migrations between the same two hosts are generally prohibited unless the network capacity is ample.

Migration Stage

Live migration transfers all VM disks and memory without downtime, requiring efficient handling of incremental data. Reducing transferred data shortens migration time and improves dynamic resource adjustment.

1. Zero‑Data Optimization

UCloud stores VM disks as sparse files; only used blocks occupy space. Native QEMU discards sparsity during migration, causing full‑size transfers. UCloud patches QEMU to skip sending all‑zero blocks, preserving sparsity and dramatically reducing transferred data and storage usage.

2. UDisk Network Disk Optimization

Since UDisk is network‑attached, its migration is unnecessary. UCloud filters out UDisk during migration and supports multi‑point mounting, allowing transparent migration without moving UDisk data.

3. Migration End Optimization

High memory‑write workloads generate continuous dirty pages, extending incremental transfer beyond the allowed downtime. UCloud improves QEMU's auto‑converge by increasing throttling speed and adjusting trigger thresholds, reducing vCPU execution time and limiting dirty page generation.

4. Compression Optimization

UCloud leverages QEMU's XBZRLE compression, which caches previously sent memory pages and transmits only XOR‑encoded differences, significantly lowering memory transfer volume and speeding up migration.

Switch Stage

Source‑side paused optimization: Migration control moves from QEMU to Libvirt, and UCloud’s management platform handles source host recovery, preventing VM downtime.

OVS switch optimization: By tuning Open vSwitch flow rule propagation, network interruption during the final switch is reduced to a few hundred milliseconds, making the migration virtually transparent to users.

Typical Application Scenarios

Fast Migration

For VMs with large data disks or high I/O workloads, UCloud skips disk transfer by sharing storage between source and destination, migrating memory and CPU first, then pulling the disk data, achieving rapid VM startup on the target host.

Cross‑Machine‑Type Migration

UCloud enables migration between different host types (e.g., standard, SSD, Ark) by converting disk formats (qcow2 → raw) via Libvirt configuration, allowing seamless upgrades without VM downtime.

Local Hot Upgrade

A local hot‑migration method upgrades QEMU without moving VM disks, transferring only memory. A new VM with the updated QEMU runs paused, receives memory from the old VM, and then switches over, completing the upgrade in seconds.

Conclusion

UCloud has implemented full‑stack optimizations for live migration—including host selection, disk and memory handling, network switching, fast migration, cross‑type migration, and local hot upgrades—ensuring uninterrupted, high‑performance VM operation and supporting efficient, elastic cloud services.