A dynamic rate-based flow control scheme for high-speed networks based on arrival rate measurements
A. Atai · 1994
This dissertation presents a flow control scheme that adjusts rates by taking into account a non-negligible propagation delay between multiple homogeneous bursty sources and target. Traditionally in existing low to medium speed networks, instantaneous queue occupancy in relation to some predefined thresholds has been used to take network wide congestion control actions. While the classical techniques work reasonably well for non-bursty traffic, they perform poorly in a bursty traffic environment where the queues are near empty most of the time, and when the number of simultaneous bursts exceeds capacity, queues will rapidly grow and reach their physical limits. During such events many cells will be discarded. In this research, a congestion control framework that takes into account the number of cells in transit and the burstiness of traffic is developed. A measurement scheme is proposed which attempts to quickly and reliably estimate the number of simultaneous bursts in progress at any time, as well as to provide some smoothing function in order to reduce the number of false alarms. The output of this measurement device is then used by a rate controller module that assigns rates to sources such that the probability of cell loss is met. The main approach used for designing the measurement module is based on modeling the problem in the context of control theory and then building an optimal state estimator (Kalman Filtering). The rate controller module is designed based on estimating the trajectory of the arrival rate signal in the near future (the round trip delay). The probabilistic behavior of the trajectory of the arrival rate signal in relation to predefined Onset and Abatement thresholds is used as a basis to extend the length of bursts that are generated by traffic sources such that the amount of aggregate traffic arriving in the near future in relation to system capacity results in meeting the probability of cell loss objective. Results show that a rate-based control with feedback can maintain a higher throughput than an optimum open-loop policy for propagation delay values of tens of milliseconds. It is also found that the rate-based scheme requires less buffer space to achieve a certain probability of loss compared to the open-loop policy. Estimation of the state of traffic sources using Kalman filtering provides superior performance when short-term measurements are needed for taking control actions in a timely manner.