Control of transport dynamics in overlay networks.

摘要

Transport control is an important factor in the performance of Internet protocols, particularly in the next generation network applications involving computational steering, interactive visualization, instrument control, and transfer of large data sets. The widely deployed Transport Control Protocol is inadequate for these tasks due to its performance drawbacks. The purpose of this dissertation is to conduct a rigorous analytical study on the design and performance of transport protocols, and systematically develop a new class of protocols to overcome the limitations of current methods.; Various sources of randomness exist in network performance measurements due to the stochastic nature of network traffic. We propose a new class of transport protocols that explicitly accounts for the randomness based on dynamic stochastic approximation methods. These protocols use congestion window and idle time to dynamically control the source rate to achieve transport objectives. We conduct statistical analyses to determine the main effects of these two control parameters and their interaction effects. The application of stochastic approximation methods enables us to show the analytical stability of the transport protocols and avoid pre-selecting the flow and congestion control parameters. These new protocols are successfully applied to transport control for both goodput stabilization and maximization. The experimental results show the superior performance compared to current methods particularly for Internet applications.; To effectively deploy these protocols over the Internet, we develop an overlay network, which resides at the application level to provide data transmission service using User Datagram Protocol. The overlay network, together with the new protocols based on User Datagram Protocol, provides an effective environment for implementing transport control using application-level modules. We also study problems in overlay networks such as path bandwidth estimation and multiple quickest path computation.; In wireless networks, most packet losses are caused by physical signal losses and do not necessarily indicate network congestion. Furthermore, the physical link connectivity in ad-hoc networks deployed in unstructured areas is unpredictable. We develop the Connectivity-Through-Time protocols that exploit the node movements to deliver data under dynamic connectivity. We integrate this protocol into overlay networks and present experimental results using network to support a team of mobile robots.