Resource provisioning for virtualized server applications

Data centre virtualization creates an agile environment for application deployment. Applications run within one or more virtual machines and are hosted on various servers throughout the data centre. One key mechanism provided by modern virtualization technologies is dynamic resource allocation. Using this technique virtual machines can be allocated resources as required and therefore, occupy only the necessary resources for their hosted application. In fact, two of the main challenges faced by contemporary data centres, server consolidation and power saving, can be tackled efficiently by capitalising on this mechanism. This dissertation shows how to dynamically adjust the CPU resources allocated to vir-tualized server applications in the presence of workload fluctuations. In particular it employs a reactive approach to resource provisioning based on feedback control and introduces five novel controllers. All five controllers adjust the application allocations based on past utilisation observations. A subset of the controllers integrate the Kalman filtering technique to track the utilisa-tions and based on which they predict the allocations for the next interval. This approach is particularly attractive for the resource management problem since the Kalman filter uses the evolution of past utilisations to adjust the allocations. In addition, the ad-aptive Kalman controller which adjusts its parameters online and dynamically estimates the utilisation dynamics, is able to differentiate substantial workload changes from small fluctuations for unknown workloads. In addition, this dissertation captures, models, and builds controllers based on the CPU resource coupling of application components. In the case of multi-tier applications, these controllers collectively allocate resources to all application tiers detecting saturation points across components. This results in them acting faster to workload variations than their single-tier counterparts. All controllers are evaluated against the Rubis benchmark application deployed on a prototype virtualized cluster built for this purpose. 3 Acknowledgements I would like to thank my two supervisors, Steven Hand and Ian Pratt. I am very grateful to Steven Hand for his invaluable suggestions on this dissertation and helping me reach the finish line. As his student I have learnt so many things in such a short time. I am also grateful to Ian Pratt for his advice and help for most of my studies. My gratitude goes to Jon Crowcroft for his comments and enthusiasm on my research over the years. Similarly , to Rebecca Isaacs for an enjoyable summer during my internship at MSR Cambridge and for her advice from then on. I am grateful to …