Overlay Networks Formation : Models and Algorithms Doctoral Dissertation of : Jocelyne Elias

Overlay networks have recently emerged as an effective means to provide a flexible, robust, and scalable platform for distributed applications, while leaving the underlying Internet infrastructure unchanged. This work tackles the overlay network design problem considering both centralized and fully distributed approaches. On one hand, we address the centralized overlay network design problem using the Service Overlay Network (SON) paradigm, and we propose several mathematical models and heuristics for the optimal design of SONs. More specifically, we introduce two network optimization models that determine the optimal assignment of users to access overlay nodes, as well as the capacity reserved for each overlay link, while taking accurate account of traffic routing. We also propose two overlay network design models that further select the optimal number and location of the overlay nodes to be deployed, as well as the optimal coverage of network users to maximize the SON operator’s profit. Furthermore, we develop a set of efficient SON design heuristics to get near-optimal solutions for large-scale network instances in a reasonable computation time. Finally, we perform an extensive performance evaluation of the proposed centralized optimization framework in several realistic network scenarios. On the other hand, this thesis also proposes two novel socially-aware overlay network design games to deal with the fully distributed overlay network formation problem. The first game combines both individual and social concerns in a unified and flexible manner, and the second game uses a Stackelberg approach, where the overlay network administrator leads the users to a system-wide efficient equilibrium by buying an appropriate subset of the overlay network links. We evaluate the performance of the proposed games, through the determination of bounds on the Price of Anarchy and other efficiency measures, as well as by simulating several realistic network scenarios, including real ISP topologies. Numerical results demonstrate that: (1) our proposed SON design models and heuristics plan very effective overlay networks, even in the case of very large network instances, and (2) our proposed distributed network formation algorithms are able to lead client users to form stable and efficient overlay networks, obtaining in