Building and maintaining overlay networks for bandwidth-demanding applications

The demands of Internet applications have grown significantly in terms of required resources and types of services. Overlay networks have emerged to accommodate such applications by implementing more services on top of IP (Internet Protocol). However; while overlay networks are successful in circumventing limitations of IP, the task of building and maintaining an overlay network is still challenging. In an overlay network, participating hosts are virtually fully-connected through the underlying Internet. However, since the quality of overlay connections varies, the performance of the overlay network is dependent on which connections are chosen to be utilized. Therefore, maintaining a good overlay network topology is crucial in achieving high performance. To demonstrate how much performance gain can be achieved through topology changes, a distributed algorithm to build an overlay multicast tree is proposed for streaming media distribution. The algorithm finds an optimal tree such that the average bandwidth of receivers is maximized under an abstract network model. However, increasing bandwidth does not necessarily lead to a better overlay topology; in overlay networks, interference between overlay connections should be taken into account. Since such interference occurs when different overlay connections pass through a congested link simultaneously, detecting congestion shared by multiple overlay connections is necessary to avoid bottlenecks. For shared congestion detection, a novel technique called DCW (Delay Correlation with Wavelet denoising) is proposed. Previous techniques to detect shared congestion have limitations in applying to overlay networks; they assume a common source or destination node, drop-tail queueing, or a single point of congestion. However, DCW is applicable to any pair of paths on the Internet without such limitations. It employs a signal processing method, wavelet denoising, to separate queueing delay caused by network congestion from various other delay variations. The proposed technique is evaluated through both simulations and Internet experiments. They show that for paths with a common synchronization point, DCW provides faster convergence and higher accuracy while using fewer packets than previous techniques. Furthermore, DCW is robust and accurate without a synchronization point; more specifically, it can tolerate a synchronization offset of up to one second between two packet flows. (Abstract shortened by UMI.)