Abstract
Large-scale data centers leverage virtualization technology to achieve excellent resource utilization, scalability, and high availability. Ideally, the performance of an application running inside a virtual machine (VM) shall be independent of co-located applications and VMs that share the physical machine. However, adverse interference effects exist and are especially severe for data-intensive applications in such virtualized environments. We design and implement Swiper, a framework which uses a carefully designed workload to incur significant delays on the target VM with minimum cost. A comprehensive set of experiments in Amazon EC2 clearly demonstrates that Swiper is capable of significantly slowing down various server applications while consuming a small amount of resources. Our following research on the interference effect leads us to construct mathematical models of resource contention and leverage the modeling results in task scheduling. We therefore present TRACON, a novel Task and Resource Allocation CONtrol framework that mitigates the interference effects from concurrent data-intensive applications and greatly improves the overall system performance. While TRACON utilizes interference models to make VM performance predictable, virtualization technology has yet achieved the vision of an ``efficient, isolated duplicate of a real machine", that is, a VM shall be able to provide the performance close to what a user would expect from a specific physical machine. Therefore, we design Matrix to enable better control of shared resources while delivering predictable VM performance in data centers. After developing models and tools to provide improved and predictable virtualization system by well handling the virtualization overhead and performance interference, we then zoom into the virtualization architecture and propose innovative prefetching method for I/O virtualization.In brief, this dissertation shows that virtualization overheads and architectures in cloud computing environments are very critical, and proposes effective novel approaches which successfully advance the state of the art. More specifically, Swiper and TRACON construct mathematical models and scheduling algorithms to mitigate the interference problem; Matrix leverages machine learning and optimization techniques to realize the ``equivalence'' property of virtualization with the best cost-efficiency; VIO-prefetching fundamentally changes the prefetching scheme in virtualization architecture and improves virtual I/O throughput. The results of this dissertation also envision numerous possibilities to thrust the virtualization and cloud computing technology.