Dynamic Resource Management in Resource-overbooked Cloud Data Centers

Abstract

The rapid growth of social media, mobile data traffic, and sensors that surround us are giving rise to very large volumes of data, which must then be processed in a timely and scalable manner to make informed decisions. The elastic properties of the cloud makes it suitable to address these data processing challenges. Despite this promise, however, numerous challenges remain unresolved, which pertain to operating a cloud data center in a way that lends itself to energy conservation, and provides effective resource management which improves resource utilization while satisfying application performance requirements, and security. This doctoral research makes the following four contributions to address a subset of these challenges. First, it presents a dynamic and adaptive algorithm to reconfigure the parameters of the hypervisor scheduler that effectively schedules the virtual machines (VMs) on a host in response to anticipated workload changes. Second, it provides a model-predictive algorithm that balances the need to utilize resources effectively by promoting maximal overbooking while still honoring the soft real-time requirements of applications. Third, it provides novel solutions for VM placement that accounts for VM performance interference. Fourth, it presents an effective runtime virtual machine placement technique that identifies an aptly suited host machine to host a VM that is to be migrated by considering both power and performance. The doctoral research has utilized real-world traces of cloud data centers to develop and validate the solutions. The long lasting impact of this dissertation stems from that fact that each solution provides a systematic and scientific approach that a cloud service provider can implement in their data centers to address energy consumption and resource utilization challenges.