January 9, 2020 –
Title: QoS in SDN-Based Large-Scale Networks
Ph.D. Candidate: Haitham Ghalwash
Major Advisor: Dr. Chun-Hsi Huang
Associate Advisors: Dr. Reda A. Ammar and Dr. Sanguthevar Rajasekaran
Day/Time: Thursday, January 9, 2020 10:00am
Location: ITE 336
Abstract: Recent research has shown that the Software-Defined Networking (SDN) technology is a promising architecture providing abstraction and programmability of modern networks and enables a more efficient solution to many of the security, performance, management, and QoS issues. The SDN is also proven to meet the growing demands of new applications for an application-oriented network. This dissertation researches an enriched SDN with an added level of QoS for network applications. Starting by investigating the potential of an SDN-based large-scale networked system, two topologies widely used in modern data centers, namely, Fat-tree and BCube, are considered. Their behavior and performance under different network scales, traffic loads, and traffic patterns are studied. Experimental results indicate the superiority of a Fat-tree network as it scales up. The potential of SDN in supporting Big-Data applications is subsequently investigated, using a Hadoop cluster with the Fat-tree interconnection. Experimental results in terms of throughput and execution time for the read/write and sorting operations demonstrate the superiority of the SDN controller over the normal forwarding mechanism. In addition, a framework adopting externally developed modules to enrich SDN capabilities for forwarding, metric retrievals, and congestion control is proposed. A QoS level is guaranteed for traffic classification, metric-based route selection, or congestion detection and control. Noticeable features of the proposed framework include: (1) facilitating dispatching applications over different paths based on the application type; (2) metric-based monitoring and rerouting, and (3) congestion detection and control to ensure balanced traffic flow. The behavior and the performance of different traffic types, namely, UDP, TCP, VOIP, and a Big-Data application, are investigated. Experimental results via such metrics as delay, jitter, and packet loss substantiate the advantage of having the developed modules on top of the controller for all traffic types. The proposed framework reduces the overall average delay, jitter, and packet loss by 54%, 32%, and 51%, respectively. Moreover, the average utilization of a monitored port is reduced by 22%.