Explicit rate congestion management for packet switched networks

This dissertation develops two distributed explicit rate algorithms that implement network layer congestion management in packet switched networks. These algorithms are alternates to bandwidth probing and loss-based congestion signaling, and enable sources to rapidly adjust to changing network conditions. Network layer implementation provides congestion management for all transport layer protocols including unresponsive UDP flows. The controllers maximize utilization, fairly allocate link bandwidth resources, and minimize delay, packet loss, and rate fluctuations. Network nodes convey to sources the minimum explicit rate along a flow's forward path. Each node determines a feasible explicit rate that provides a max-min fair allocation of link bandwidth while maintaining a positive queue length at the output buffer to assures full utilization of the link bandwidth.; The communication channel between two hosts, consisting of the bottleneck link with an infinite FIFO buffer, and fixed round trip delay and control period, is modeled as a discrete-time transfer function. Rate compensation is a function of the round trip delay and packet scheduling is performed by hosts. Link bandwidth and queue resources are partitioned among the virtual active flows. The model provides tractability while capturing the essence of the delayed feedback problem. Controller performance is evaluated with course-grained simulations with a large number of flows and random traffic statistics.; The proportional controller explicit rate is a function of the round trip delay maintained by sources, deviation of the queue length from the queue setpoint, and the number of active flows. Values of delay-dependent gain for stable operation are determined. The controller provides better queue length management, but the link is underutilized when the aggregate contains constrained flows. The optimal controller is a discrete LQR that calculates the explicit rate using rate matching history and queue length deviation from setpoint. A feedforward path estimates the number of virtual flows used to manage the queue length. Parameters are determined that satisfy performance objectives. The optimal controller provides better link utilization at the cost of greater buffer resources.