Towards an Optimized Internet Backbone Network

The explosive growth of traffic demand, fueled by the expansion of the Internet in reach and capacity, imposes severe stress on backbone networks. A tier-1 Internet Service Provider (ISP) like AT&T delivers more than 7 petabytes of data per day through its backbone fabric where as little as 1% higher traffic loss will incur several terabytes of retransmissions, almost equivalent to the daily load of a medium-sized regional network, and billions of dollars are spent annually on infrastructure construction and equipment upgrade to satisfy the increasing bandwidth requirements. An optimized design for Internet backbone networks thus benefits both service providers and Internet users worldwide.;Despite the large body of research targeted at optimizing backbone networks, it remains challenging to identify the actual major factors driving the design and create a realistic and tractable model with appropriate design metrics. The development of networking technologies further complicates the problem by continuously shifting the bottleneck factors from some elements to others. Moreover, with the advent of the concept of green Internet, legacy models aimed at maximizing the throughput are tuned to a new energy-smart perspective aimed at minimizing the energy footprint. Such rapid evolution of backbone networks leaves many critical issues unsolved, inspiring more investigations and discussions in the research community.;In this dissertation, we study the WDM backbone networks from the bottom up, starting with physical topology and then extending to virtual topology and traffic routing. We make the following contributions. First, we present a physical topology model to determine the number and the choice of constituent fiber links. Our model captures the physical design principles including cost, performance, resilience and geographical constraints, and considers the problem as a tradeoff among all feasible meshes that yield best performance for a given budget. The results achieve a similarity of more than 90% with the published ISP structures. Second, existing virtual topology models are highly dependent on the network context. A model that fits one network can perform poor for others. We thus abstract the individual design objectives by iteratively identifying the bottleneck elements and setting up suitable virtual channels accordingly. Our heuristic approach is plug-and-play and averages 28% higher throughput than existing models in all tested networks with different traffic demands and technologies. Third, emerging network design models aimed at maximizing energy savings by aggregating traffic at a small set of resources is offset by legacy models aimed at maximizing the throughput by spreading the load across network resources. We solve this dilemma with a new design perspective which targets optimal power usage for common traffic demand, while accommodating traffic fluctuations. The proposed heuristic matches the optimality of both factors within 10%, while enjoying polynomial time complexity. Fourth, we make a case that existing backbone traffic routing schemes that minimize the aggregate router power usage are mislead by ignoring the impact of cooling consumption. The actual router power spectrum is polynomial in traffic demand and increases rapidly when the router is loaded. We compare the efficacy of two distinct routing philosophies, traffic aggregation vs. traffic balancing, and find that mitigating network bottlenecks, rather than creating ones, can save at least 25% energy. Our conclusions challenge the common wisdom about the merit of load concentration strategies.