Research On Performance And Availability Evaluation Methods Of Virtualization Systems

In recent years, virtualization has become a hot reaserch topic as a technology for computing resources consolidation. Alghough it holds numerous merits such as flexibility, security, encapsulation, isolation, reliability, high resource utilization, ease management and configuration, saved power, reduced TCO and so forth, virtualization has also some unexpected problems. For example, virtualization depresses the performance of some specific operations, the introduction of virtual machine monitor brings several unsteady factors, and the availability of virtualization systems displays the phenoment of Single Point of Failure (SPOF). Therefore, it's urgent for improving the design of virtualization systems to evaluate their performance and availability, characterize their features of performance and availability, and identify their bottlenecks of performance and availability. This thesis has carried out a lot of research works on the performance and availability evaluation methods of virtualization systems. Specifically, the main contributions of this thesis are as follows:First, this thesis proposes a multi-granularity performance combinative evaluation method by analyzing the architecture and princinple of various virtualization systems, and provides a new approach to resolve the drawback of traditional performance evaluation methods on identifying effectively the performance bottlenecks of virtualization systems. At the same time, a tool named vTestKit is developed to implement the multi-granularity performance combinative evaluation. vTestKit can configure flexiblely testing environments and deploy automatically testing tasks.Second, this thesis proposes two workload-related analytic models of Xen system by investigating its execution mechanism of network I/O operations and CPU scheduling strategies:an analytic model to limit the network I/O request number of guest domains and an analytic model to analyze the average performance and scalability of guest domains executing compute-intensive workload, which offer some analyzing ways to resolve the overloading of Xen system incurred by intensive network I/O operations from guest domains and the combinatorial optimization between the average performance and corresponding number of guest domains executing compute-intensive workload.Third, this thesis establishs the typical failure models (including fault type, fault location, occurrence mechasim and duration) and their implementation by statistically analyzing common fault symptoms of virtualization systems, and designs a fault injection tool for Xen system-SFIVE. These works build the foundation for the availability evaluation of virtualization systems.Finally, this thesis builds the state transition diagram and Markov chain-based availability analytic model of three basic virtual cluster nodes by exeploring their architecture characteristics and life cycle states, and settles the elementary problem for analyzing the availability of virtual cluster systems.In general, researches on performance and availability evaluation of virtualization systems mainly focus on performance testing, performance modeling, availability testing and availability modeling. This thesis has obtained several archievements in these four fields and presented some reference value in theory and application.