UC San Diego

Dense Wavelength Division Multiplexing (DWDM), dedicated optical paths, high-speed switches and routers are giving rise to networks with plentiful bandwidth in the core. In such networks, bottlenecks and congestion are concentrated at the edge and at end nodes. Collaborative data-centric e -science applications and new multipoint-to-point and multipoint-to-multipoint communication patters raise the challenge of how to allocate bandwidth resources efficiently and fairly among active sessions. In this dissertation, we consider how end nodes should efficiently and fairly manage capacity across multiple sessions with finite and unknown demands in such networks to support long-lived bulk data transfers. We propose a novel distributed end-node bandwidth sharing algorithm that controls each source and sink independently with local information, adaptively allocating its capacity to active sessions. This algorithm is given no knowledge of the desired session rates, but rather discovers them. We prove analytically that our distributed algorithm converges to the unique global max-min fair rate allocation from any initial or transitional states. Simulations with different algorithm parameters, network topology and traffic patterns validate the convergence and fairness properties of our distributed algorithm. Simulations further show that the system convergence is in practice fast in networks of 32 to 1024 nodes, and our proposed approach achieves better efficiency and fairness than other high- speed transport alternatives. We design and implement a prototype of our distributed bandwidth sharing algorithm. Experimental results on the OptIPuter networks and in emulated environments conform closely to analytical studies and simulation results, showing that the proposed approach is practically feasible