Capacity Management Approaches for Compute Clouds

Cloud computing provides the illusion of a seamless, infinite resource pool with flexible on-demand accessibility. However, behind this illusion there are thousands of servers and peta-bytes of storage, running tens of thousands of applications accessed by millions of users. The management of such systems is non-trivial because they face elastic demand, have heterogeneous resources, must fulfill diverse management objectives, and are vast in scale. Autonomic computing techniques can be used to tackle the complex problem of resource management in cloud data centers by introducing self-managing elements known as autonomic managers. Each autonomic manager should be capable of managing itself while simultaneously contributing to the fulfillment of high level systemwide objectives. A wide range of approaches and mechanisms can be used to define and design these autonomic managers as well as to organize them and coordinate their actions in order to achieve specific goals. This thesis investigates autonomic approaches for cloud resource management that aim to optimize the cloud infrastructure layer with respect to various high level objectives. The resource management problem is formulated as a problem of optimization with respect to one or more management objectives such as cost, profitability, or data center utilization, as well as performance concerns such as response time, quality of service, and rejection rates. The aim of the reported investigations is to address the problems of cost-efficient elastic resource provisioning, unified management of cloud resources, and scalability in cloud resource management. This is achieved by introducing three new concepts in capacity management: the Repacking, Holistic, and Peer to Peer approaches.