UC Santa Cruz

Computer networks are complex constructs consisting of many entities that simultaneously interact with one another. Understanding various aspects of such complex systems and their scaling properties is therefore a challenging task. This thesis studies the performance of computer networks through several layers of abstraction—namely communication, social and information—and investigates the interplay between these layers at a large scale and from an analytical point of view. The results derived from this analysis are crucial in early identifying of design issues and potential weaknesses of large-scale networks such as the Internet for which performing simulations is prohibitive.

The first part of this dissertation studies how the spatial diversity of social contacts affect the scalability of communication networks and identifies classes of social models that let computer networks properly scale. From this analysis, it is established that scalability is achieved under social models in which social contacts are statistically concentrated within a confined geographical region around each node. We shall recognize that the true distribution of social contacts in real networks does not generally meet this requirement, imposing a scalability gap on today's networks.

The second part of the dissertation studies information-centric networking (ICN); a framework that utilizes distributed content caching that can be used to bridge the foregoing scalability gap. Three dominant methods of distributed content caching—namely uniform-, optimal- and edge-caching—are compared. We shall see that caching only at the edge of the network outperforms uniform-caching (the de facto standard of ICN) in terms of end-to-end latency while offering slightly inferior results compared to the more complex optimal-caching strategy. This result is further augmented by the observation that higher degrees of reference locality in space and/or time improve the performance of edge-caching, making it a viable alternative to optimal solution.

Finally, an analysis of the forwarding plane of ICN is presented. In the absence of host addressing, ICN routers are required to keep track of all requests (Interests) passing through them in Pending Interest Tables (PITs) that are needed to deliver data back to the requesters as well as to enable optimization mechanisms such as Interest aggregation. Through careful analysis of the PIT size distribution and the probability of Interest aggregation at PIT, we shall see not only are the true benefits from the stateful forwarding plane of ICN much smaller than anticipated, but also they come at the high expense of maintaining very large PITs. These results reveal that the forwarding plane of ICN must be rethought; an important finding that provokes the investigation of a stateless forwarding plane for future ICNs.