Design and use of managed overlay networks

As the role of the Internet has been steadily gaining in importance, overlays are increasingly being used to provide new services and to deploy older ones. Some of the services for which overlays have been proposed include multicast, quality of service (QoS), search, and resilient networks. The use of overlays, in turn, has led to more interest in improving their performance. The performance of an overlay network depends significantly on how the network is structured, i.e., the placement of the nodes in the underlying network topology, the links between the overlay nodes and the access links of these nodes. This thesis focuses on algorithms for improving the performance of managed overlays, in which the capabilities and locations of the participating nodes are known. Managed overlay networks allow for greater optimization and design than unmanaged overlays due to this increased knowledge of the network. Our first technique applies to managed overlays that are used for streaming real-time content. Many such networks are composed of end-systems with a single access link to the rest of the network. We propose Time Division Streaming (TDS) as a technique to schedule the serial access to these links. We show that overlays constructed using TDS reduce the average delay experienced by the participating nodes as compared to previous designs. Our second technique focuses on the placement of nodes in service overlay networks. We propose RouteSeer, an algorithm to place nodes in service overlays to create resilient paths between overlay nodes. RouteSeer works by examining local routing information at the overlay nodes to place a few intermediate nodes such that the overlay links are protected by a disjoint path through an intermediate node. We extend RouteSeer to topologies with incomplete and dynamic routing information and demonstrate its ability to protect overlay links in such cases as well. Finally, we propose and evaluate a managed overlay network for a highly reliable, multi-attribute query service. As opposed to previous designs, our architecture ensures bounded query response times and can handle dynamically updated data.