Latency versus Cost Optimizations in Hierarchical Overlay Networks

Overlay networks have emerged as a fundamental technique for efficient data dissemination. Current overlay construction techniques typically face a number of limitations, including: i) limited scalability because of reliance upon global knowledge, and ii) performance characteristics targeting the requirements of a particular class of applications. In this paper, we propose dynamically organizing overlay nodes into clusters that form the basis for a hierarchical overlay network. Overall, we find that the exponential improvement in both scalability and network overhead to typically be worth the resulting reduction in overlay quality. Next, we evaluate the relative benefits of existing graph and overlay construction algorithms with respect to varying application requirements. We find that light approximate shortest-path trees (LAST) allow application developers to flexibly trade communication cost with delay along a continuous spectrum. More importantly, for our target network topologies, we find LAST settings that result in cost and delay values within 15% of optimal.