Stochastic control of computer networks

Although transmission control protocol (TCP) is a proven protocol in networks with congestion-dominated packet losses, it is a poor match for mobile networks with bit-error-dominated packet losses. This article considers a control theoretic approach for more effective flow control in reliable network transmissions. An extended Kalman filter is developed to estimate two key network state indicators: individual queue sizes and packet arrival rates to each queue. A source node estimates these parameters for each of the downstream queues currently servicing packet streams originating from itself. These estimates can be used by an linear quadratic Gaussian (LQG) `steady-state` linear perturbation proportional-plus-integral (PI) controller to regulate the size of downstream queues by altering the packet rate sent towards a particular downstream queue. Although this work demonstrates the effects of a single controller, it is anticipated that a network wide deployment could achieve significant throughput and stability benefits. The primary contributions of this work include: (a) development of the necessary Kalman filter theory, (b) design of a network state estimator using a transient queue behaviour model, (c) equations for the LQG synthesis of a steady-state linear perturbation PI controller and (d) demonstration of the resulting controller performance, which shows significant improvement over traditional TCP implementations.