Characterizing traffic-aware overlay topologies: a machine learning approach

Abstract

Overlay networks are application-layer networks that are constructed using the existing Internet infrastructure. Nodes in an overlay network construct logical links toward other nodes to form an overlay topology. Common routing algorithms, such as the link state and distance vector algorithms, are then used to determine how to route data in the overlay network. Previous work has demonstrated that overlay networks can be used to improve routing performance in the Internet. These quality of service improvements make overlay networks attractive for a variety of network applications. Recently, game-theoretic approaches to constructing overlay network topologies have been proposed. In these approaches, nodes establish logical links toward other nodes in a decentralized and selfish manner. Despite the selfish behavior, it has been shown that desirable global network properties emerge. These approaches, however, neglect the traffic-demand between nodes. In this thesis, a game-theoretical approach is presented to constructing overlay network topologies that considers the traffic-demand between nodes. This thesis shows that the traffic-demand between nodes has a significant effect on the topologies formed. Nodes with statistically higher traffic-demand from others become members of the graph center, while nodes that have statistically higher traffic-demand toward others establish logical links toward members of the graph center. This thesis also shows that a traffic-demand aware overlay network topology is better suited to transport the required traffic in the overlay network. Unfortunately, the game-theoretic approach is intractable. In order to construct larger overlay networks, approximate or heuristic approaches are required. In this thesis, a machine learning approach is proposed that characterizes the attributes of neighbor nodes during the construction of the overlay network topology. The approach proposed uses this knowledge and experience to learn a set of human-readable rules. This rule set is then used to decide whether to construct a logical link toward a node. This thesis shows that the machine learning approach results in similar overlay network topologies as the game-theoretic approach. Additionally, it is shown that the machine learning approach is tractable and scales to larger networks.