Resource-Efficient Replication and Migration of Virtual Machines.

Continuous replication and live migration of Virtual Machines (VMs) are two vital tools in a virtualized environment, but they are resource-expensive. Continuously replicating a VM's checkpointed state to a backup host maintains high-availability (HA) of the VM despite host failures, but checkpoint replication can generate significant network traffic. Each replicated VM also incurs a 100% memory overhead, since the backup unproductively reserves the same amount of memory to hold the redundant VM state. Live migration, though being widely used for load-balancing, power-saving, etc., can also generate excessive network traffic, by transferring VM state iteratively. In addition, it can incur a long completion time and degrade application performance. This thesis explores ways to replicate VMs for HA using resources efficiently, and to migrate VMs fast, with minimal execution disruption and using resources efficiently. First, we investigate the tradeoffs in using different compression methods to reduce the network traffic of checkpoint replication in a HA system. We evaluate gzip, delta and similarity compressions based on metrics that are specifically important in a HA system, and then suggest guidelines for their selection. Next, we propose HydraVM, a storage-based HA approach that eliminates the unproductive memory reservation made in backup hosts. HydraVM maintains a recent image of a protected VM in a shared storage by taking and consolidating incremental VM checkpoints. When a failure occurs, HydraVM quickly resumes the execution of a failed VM by loading a small amount of essential VM state from the storage. As the VM executes, the VM state not yet loaded is supplied on-demand. Finally, we propose application-assisted live migration, which skips transfer of VM memory that need not be migrated to execute running applications at the destination. We develop a generic framework for the proposed approach, and then use the framework to build JAVMM, a system that migrates VMs running Java applications skipping transfer of garbage in Java memory. Our evaluation results show that compared to Xen live migration, which is agnostic of running applications, JAVMM can reduce the completion time, network traffic and application downtime caused by Java VM migration, all by up to over 90%.PhDComputer Science and EngineeringUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttp://deepblue.lib.umich.edu/bitstream/2027.42/111575/1/karenhou_1.pd