Accurate Relay Selection for Improved Multipath Throughput in Internet Overlays

Distributed applications deployed over the Internet, such as in grids, clouds, or P2P networks, crucially depend on the speed of their Internet connections. A common solution is to exploit underlying network parallelism via multipath routing within an application-level overlay network. Accurately selecting appropriate relay hosts within an overlay, however, remains a challenging problem, frequently addressed by simple heuristics. We propose a modeling-based approach to solving the relay selection problem for multipath overlay routing. Our system enables formula-based, dynamic, flexible and overlay-specific model building and prediction of end-to-end metrics such as TCP throughput on multiple paths based on any number of underlying predictor metrics. For evaluation, we use Planetlab and, as prediction data, we leverage 9 months worth of available bandwidth traces collected from the S3 project, in addition to our own measurements. Comparing with existing heuristic methods, we show that relay selection using our modeling-based approach enables consistent performance gains across the Internet.