On Improving The Performance And Resource Utilization of Consolidated Virtual Machines: Measurement, Modeling, Analysis, and Prediction

Abstract

This dissertation addresses the performance related issues of consolidated \emph{Virtual Machines} (VMs). \emph{Virtualization} is an important technology for the \emph{Cloud} and data centers. Essential features of a data center like the fault tolerance, high-availability, and \emph{pay-as-you-go} model of services are implemented with the help of VMs. Cloud had become one of the significant innovations over the past decade. Research has been going on the deployment of newer and diverse set of applications like the \emph{High-Performance Computing} (HPC), and parallel applications on the Cloud. The primary method to increase the server resource utilization is VM consolidation, running as many VMs as possible on a server is the key to improving the resource utilization. On the other hand, consolidating too many VMs on a server can degrade the performance of all VMs. Therefore, it is necessary to measure, analyze and find ways to predict the performance variation of consolidated VMs. This dissertation investigates the causes of performance variation of consolidated VMs; the relationship between the resource contention and consolidation performance, and ways to predict the performance variation. Experiments have been conducted with real virtualized servers without using any simulation. All the results presented here are real system data. In this dissertation, a methodology is introduced to do the experiments with a large number of tasks and VMs; it is called the \emph{Incremental Consolidation Benchmarking Method} (ICBM). The experiments have been done with different types of resource-intensive tasks, parallel workflow, and VMs. Furthermore, to experiment with a large number of VMs and collect the data; a scheduling framework is also designed and implemented. Experimental results are presented to demonstrate the efficiency of the ICBM and framework.