TAO: A Topology-Adaptive Overlay Framework

Abstract

Large-scale distributed systems rely on constructing overlay networks in which nodes communicate with each other through intermediate overlay neighbors. Organizing nodes in the overlay while preserving its congruence with the underlying IP topology (the underlay) is important to reduce the communication cost between nodes. In this thesis, we study the state-of-the-art approaches to match the overlay and underlay topologies and pinpoint their limitations in Internet-like setups. We also introduce a new Topology-Adaptive Overlay organization framework, TAO, which is scalable, accurate and lightweight. As opposed to earlier approaches, TAO compiles information resulting from traceroute packets to a small number of landmarks, and clusters nodes based on (1) the number of shared hops on their path towards the landmarks, and (2) their proximity to the landmarks. TAO is also highly flexible and can complement all existing structured and unstructured distributed systems. Our experimental results, based on actual Internet data, reveal that with only five landmarks, TAO identifies the closest node to any node with 85% - 90% accuracy and returns nodes that on average are within 1 millisecond from the closest node if the latter is missed. As a result, TAO overlays enjoy very low stretch (between 1.15 and 1.25). Our results also indicate that shortest-path routing on TAO overlays result in shorter end-to-end delays than direct underlay delays in 8-10% of the overlay paths.