On Modeling and Control of Network Queue Dynamics

Abstract: Internet flow control algorithms used today are separately designed and tuned for outdated setups and traffic mixes. Differnet algorithms exist in both end nodes an core routers of a network. This work presents improved techniques for network flow control. The solutions are adaptive and scalable to account for changing network structures and shifts in the traffic mix. The flow control is performed at the network routers by adaptive queue management. The problem is identified to be a control problem, which requires dynamic models to base control design on. System identification and automatic control are used to design adaptive queue management techniques. It is shown that biased autoregressive models of the queue length dynamics capture the main dynamics. The throughput is increased and the oscillations damped by using knowledge from the model, in the control.The techniques for adaptive queue management are triggered by packet arrivals, which indicate nonuniform sampling rather than uniform. Techniques for modeling dynamics based on nonuniform samples are discussed. Accurate approximations of the signal spectra can be used to adopt continuous time models. Approximations of the Fourier transform based on nonuniform samples are developed and evaluated.Continuous time models can be identified using standard identification techniques, from nonuniformly sampled data. Some of the considerations that has to made are presented. Filtering at nonuniform sample times is also discussed.There exist many control problems on different levels in a network. An overview of network control is given, based on available specification. The control problems are also presented using block diagrams and connections between the problems are highlighted. It is shown that, is important to consider the structure of underlying network control when design is done.