On optimizations of Virtual Machine live storage migration for the Cloud.

摘要

Virtual Machine (VM) live storage migration is widely performed in the data centers of the Cloud, for the purposes of load balance, reliability, availability, hardware maintenance and system upgrade. It entails moving all the state information of the VM being migrated, including memory state, network state and storage state, from one physical server to another within the same data center or across different data centers. To minimize its performance impact, this migration process is required to be transparent to applications running within the migrating VM, meaning that applications will keep running inside the VM as if there were no migration operations at all.;In this dissertation, a thorough literature review is conducted to provide a big picture of the VM live storage migration process, its problems and existing solutions. After an in-depth examination, we observe that a severe IO interference between the VM IO threads and migration IO threads exists and causes both types of the IO threads to suffer from performance degradation. This interference stems from the fact that both types of IO threads share the same critical IO path by reading from and writing to the same shared storage system. Owing to IO resource contention and requests interference between the two different types of IO requests, not only will the IO request queue lengthens in the storage system, but the time-consuming disk seek operations will also become more frequent. Based on this fundamental observation, this dissertation research presents three related but orthogonal solutions that tackle the IO interference problem in order to improve the VM live storage migration performance.;First, we introduce the Workload-Aware IO Outsourcing scheme, called WAIO, to improve the VM live storage migration efficiency. Second, we address this problem by proposing a novel scheme, called SnapMig, to improve the VM live storage migration efficiency and eliminate its performance impact on user applications at the source server by effectively leveraging the existing VM snapshots in the backup servers. Third, we propose the IOFollow scheme to improve both the VM performance and migration performance simultaneously. Finally, we outline the direction for the future research work.