An adaptive approach to optimizing QoS throughputs in software-defined networks

Advisors: Nicholas T. Karonis.Committee members: Kirk Duffin; Rajkumar Kettimuthu; Michael Papka.Includes bibliographical references.Includes illustrations.Maximizing bandwidth availability to critical science data flows is a key strategy in guaranteeing that current infrastructure can keep pace with the increasing demands of distributed science workflows across multiple facilities. Much of the research in this area has gone into developing parallel file transfer protocols such as GridFTP and leveraging circuit-based reservation systems like ESnets OSCARS. Software Defined Networks (SDN) is a recent development that presents a programmable approach to networking that moves routing logic out of the network fabric and abstracts the physical network behind a unified API. While SDN has gained significant traction in the data center, its potential is just beginning to garner attention in the scientific computing space and much investigation has yet to be done. In this study we examine an adaptive, priority-driven algorithm that leverages SDN capabilities to monitor flows, enforce per-application bandwidth guarantees and reallocate unutilized bandwidth between virtual circuits in real time. Trials were conducted on a simplified, six-switch topology with a single bottleneck link to demonstrate the correct behavior of the algorithm and compare it to competing traffic management techniques. We show that by intelligently adjusting throughput limits---effectively loaning out bandwidth from flows that are under-utilizing reserved bandwidth to flows that can make better use of the extra bandwidth---more optimal throughputs can be achieved for priority applications at a minimal cost to total bandwidth usage compared to such classic policing schemes as Differentiated Services queues.M.S. (Master of Science