It is well-known that the use of path diversity, i.e., the use of multiple end-to-end paths can improve the performance of applications such as multimedia streaming and Voice over IP. Leveraging path diversity requires applications to choose independent paths, i.e., paths that do not share Points of Congestion (PoCs). Although independent paths may not be available, performance can often be improved by selecting paths with the least amount of shared congestion. In this paper, we propose Shared CONgestion Estimator (SCONE), a tool that provides an estimate of shared congestion between two paths. SCONE sends probe flows along each of these paths and calculates the fraction of drops appearing in correlated bursts as the estimate of shared congestion. The assumption here is that correlated bursts typically occur at shared PoCs. Previously proposed path selection techniques only determine if two paths share a PoC or not. Also, previous techniques require that the two paths form one of only 2 topologies, while SCONE can work with two additional topologies relevant to applications such as media streaming for a single source/destination pair over multiple paths. We used both ns-2 simulations and wide area experiments on PlanetLab to evaluate SCONE. We measured SCONE's ability to correctly estimate the fraction of drops that occurred at shared PoCs. We found that the absolute estimation error of SCONE is no more than 0.2 in 95% of simulation experiments. In 80% of wide area experiments, SCONE has an absolute estimation error less than 0.3. We also show that the overhead of probe traffic can be avoided by passively observing application traffic such as multimedia streams which send adequate amounts of traffic.




Download Full History