Defense Event
Women in Computing

Resource-Efficient Replication and Migration of Virtual Machines

Kai-Yuan Hou

Monday, November 24, 2014
08:15am - 10:15am
3725 Beyster Bldg.

Add to Google Calendar

About the Event

In a virtualized environment, applications run in Virtual Machines (VMs), and multiple VMs may be consolidated in a physical host. VM replication and migration are two prevalent ways to relocate VMs between physical hosts for maintaining high-availability and enhancing system performance. However, current replication and migration approaches are resource-expensive. This thesis explores ways to replicate and migrate VMs using resources efficiently. First, we investigate the tradeoffs in using different checkpoint compression methods to reduce the network bandwidth required by continuous check- point replication. We conduct a detailed evaluation of three checkpoint compression methods, including gzip, delta compression, and similarity compression. Based on this evaluation, we suggest guidelines for selecting and using them for workload types and resource constraints. Next, we propose HydraVM, a storage-based high-availability system, to eliminate the unproductive memory reservations made in backup hosts as passive receptacles of redundant VM state. Finally, we propose application-assisted VM live migration, which skips transfer of VM memory that need not be migrated for the execution of running applications in the destination host. This not only reduces the migration traffic sent over the network, but also reduces the time required to finish migration of a VM and the impact of VM migration on running applications’ performance. We develop a generic framework for application-assisted live migration, and then use the framework to build JAVMM, a system that migrates VMs running various types of Java applications skip- ping transfer of garbage in Java memory. Our experimental results show that compared to Xen live migration, which is agnostic of the applications running in the migrating VM, JAVMM can reduce the completion time, network traffic and application downtime caused by JavaVM migration, all by up to over 90%.

Additional Information

Sponsor(s): CSE

Faculty Sponsor: Professor Kang G. Shin

Open to: Public