| Organizations deploying mission-critical applications are increasingly turning to computing utility providers for highly reliable and scalable computing and network infrastructures. Utility providers attain economies of scale by hosting applications on shared infrastructures, but are continually seeking new ways to improve resource utilization, and thereby reduce costs. One way of improving utilization is through cooperatives, in which providers leverage the network and computing infrastructure of partners. Cooperatives allow providers to handle surges with limited over-provisioning, reduce the cost of dedicated infrastructure, and leverage the specialization and pricing of partners. Despite these advantages, there are significant challenges in designing a system that effectively distributes workload among separately-administered infrastructures.; This thesis presents analyses, models, and algorithms for resource management in distributed computing cooperatives. Central to our study is a detailed investigation of the operating environment of computing utilities. Our analysis includes settlement models widely used in the industry, server assignment techniques employed by commercial providers to minimize network latency, and Internet measurement techniques for effective server assignment. We begin with an analysis of the potential cost reductions providers can expect, either by serving offloaded traffic or by offloading workload to a partner, as their needs dictate. We also illustrate challenges in making sound capacity planning decisions when multi-provider resource pools are available.; We go on to develop models and algorithms for cooperatives offering large-scale network services. We expose issues with current network measurement techniques and quantify their impact.{09}We develop a client clustering technique so the assignment of clients (potentially numbering in the millions) to servers can be managed in a scalable manner. We create models of commercially deployed computing utilities to enable prediction of the expected level of service from given partners. Finally, we tie all of our analyses, models, and algorithms together in an innovative peering system featuring scalable and effective workload assignment for large-scale content delivery cooperatives. We show our system is significantly more efficient than greedy alternatives, in terms of minimizing cost and respecting network and server constraints, over a broad range of real-world scenarios. |
