| Recent advances in overlay networks have enabled rapid development of new applications over the Internet. The performance of these applications is highly dependent on the ability to operate over a set of good quality paths (or a multipath) for load balancing and robustness. However, the logically fully connected overlay topology and the sharing of different overlay links at the underlay physical network level impose many challenges for managing multipath in overlays.; In this dissertation, we conduct a systematic study of multipath resource management in overlay networks. To reduce the complexity of managing a fully connected overlay topology, we propose a distributed overlay construction algorithm to build a mesh so that an application can work on it afterwards. After showing the NP-hardness of finding minimal correlated overlay paths, we propose a novel heuristic to select multipath. We then solve the problem of resource allocation among multiple paths using a mean-variance framework based on statistical QoS information of the paths. In addition to solving the problem for a single overlay, we design an overlay congestion control framework to fairly allocate resource among multiple overlays, thereby enabling a greater level of control that cannot be achieved by traditional congestion control schemes. Finally, we design multipath selection and allocation algorithms for robust traffic engineering in the underlay network core to deal with the uncertain traffic fluctuation generated by overlay networks. The resulting framework is the first one that integrates overlay VPN bandwidth provisioning and traffic engineering. |
