Minimizing round-trip times for high-performance transport communication

Current transport protocols leave significant exposure to round trip times in communication, most notably in connection management. Round trip times are also an important factor in the performance of flow control, congestion control, and error control strategies. Minimizing the number of round trip times incurred is essential for providing the highest possible performance for transport protocols on emerging network technologies.;This dissertation investigates approaches to minimize the impact of round trip times on communication. Exploiting locality in communication reduces the number of round trip times in connection management. Additional feedback from the network reduces other round trips as well. By making more state information available at the endpoints of the network, senders can make informed decisions about the current state of the network, and minimize the number of round trip times incurred.;The dissertation uses empirical measurement, mathematical analysis, and simulation modeling as its tools for performance evaluation. Network measurement data collected in an existing network environment are used to assess current traffic characteristics, and to evaluate alternate approaches to connection management. Simulation and theoretic approaches are used to evaluate congestion control, flow control, and error control strategies.;The most significant contribution of this thesis is a novel feedback-based congestion control strategy for datagram computer networks. The approach is based on loss-load curves. Gateways in the network monitor local load and provide the curves as feedback to senders. The loss-load curve allows each host to choose its own tradeoff between throughput and packet loss, factoring the round trip time and error recovery cost into the choice.;Loss-load curves can be defined to encourage sender cooperation, provide protection from greedy senders, and dynamically keep total offered load close to the network capacity. Properties of the algorithm are proved formally, and verified by simulation. For cooperating senders, the network converges to a stable operating point, with each host operating at a local optimum. Furthermore, the scheme works well even in a network with uncooperative hosts.