Abstract: Cloud computing has become a hot technology. The virtualization technology, one of its core support technologies, is also developing quickly. In recent years, a lot of research and practice have been done in the virtualization field. Live virtual machine migration has drawn much attention as a representative technology of the virtualization field. This paper summarizes the research progress of live virtual machine migration technology including pre-copy, post-copy and CR/TR-Motion analyzes the future research trend.
Introduction
Live virtual machine (VM) migration is often described as the "magic trick" of modern cloud computing. It allows system administrators to move active, running servers between physical hardware hosts without users ever noticing a service interruption. This capability is fundamental for load balancing, hardware maintenance, and energy management in data centers. However, the technical execution of moving a VM—specifically its memory—while it remains operational presents significant challenges.
The Memory Migration Challenge
The core difficulty in live migration lies in memory management. While CPU states and I/O device information are relatively small, a VM’s RAM can span several gigabytes. Furthermore, this memory is constantly changing or "dirtying" as the VM continues to process requests.
Administrators must balance two competing metrics:
- Downtime: The period during which the VM is suspended and service is unavailable.
- Total Migration Time: The entire duration from the start of the migration command until the VM is fully operational on the destination host.
Primary Migration Strategies
Pre-copy: The Iterative Standard
Pre-copy is currently the most widely used strategy in the industry. In this approach, the source host copies all memory pages to the destination while the VM is still running. Because the VM continues to write to memory during this process, the system performs several subsequent rounds of copying only the "dirty pages"—those that changed during the previous iteration.
As shown in the figure above, the challenge with standard pre-copy is that frequently updated pages are transmitted repeatedly, consuming significant network bandwidth. To mitigate this, researchers have proposed optimizations that identify "high-dirty" pages and defer their transmission until the final stop-and-copy phase.
Post-copy and CR/TR-Motion
Alternative strategies attempt to solve the convergence problem differently:
- Post-copy: The VM is suspended briefly to move the minimal execution state and then resumed immediately at the destination. Memory is pulled from the source only when the VM attempts to access it (demand paging). While this ensures a deterministic total migration time, it risks VM failure if the network drops mid-migration.
- CR/TR-Motion (Trace and Replay): This method logs the execution of the source VM and "replays" that log on the destination. By using the destination's CPU to reconstruct the state, it drastically reduces the volume of data sent over the network, making it ideal for low-bandwidth environments.
Optimizing Migration with Delta Compression
To further reduce the amount of data transferred, many modern frameworks employ Delta Compression. Instead of sending an entire 4KB memory page that has only changed by a few bytes, the system creates a "delta page" containing only the differences.
As illustrated in the flowchart, the source host checks if a page is already cached; if so, it generates a delta, compresses it using Run-Length Encoding (RLE), and transfers the smaller RLE page. The destination host then decodes the RLE data to re-create the full page, significantly lowering network overhead.
Conclusion and Future Outlook
While live migration has matured significantly, several hurdles remain. Future research is trending toward "Green Data Centers," where migration policies are integrated with energy-saving goals to consolidate hardware. Additionally, the rise of mobile cloud computing introduces the need for migration between mobile devices and the cloud, requiring even more bandwidth-efficient algorithms. As cloud applications become more write-intensive, hybrid approaches that combine the reliability of pre-copy with the efficiency of delta compression and log-based replay will become essential.
Summary updated