Abstract

At a node of a broadband network, such as an ATM network, the prediction of quality of service plays an important role in formulating traffic control functions. The historical data from 1995 to 2001 has shown that data traffic has doubled each year and this trend is likely to continue. Assuming that traffic is generated from a number of data sources, we propose a new approach in predicting packet (or cell) loss rate, which is considered to be the quality of service of interest. The proposed approach not only does not rely on an assumption of a statistical model for the traffic patterns, but also closely approximates the cell loss rate in an output queue of a node. To do this, first, we identify a set of traffic parameters, as traffic indicator , that can describe the behavior of short-term, long-term or self-similar traffic. Then, we approximate the cell loss rate in terms of the traffic indicator using function approximation capability of a neural network system consisting of a linear combination of a number of sigmoid functions. The proposed traffic indicator and cell loss approximator can be used for traffic engineering of broadband networks, e.g., ATM networks, to maximize the utilization of an output link. As an illustrative example, we propose a new connection admission control that predicts packet cell loss rate from the aggregate of two traffic indicators: one for the existing connections and the other for the new connection. If the predicted cell loss rate for the aggregate traffic indicator is less than a pre-set threshold the new connection is admitted. Under the assumption that the users do not tight bound on the cell loss rate, we showed that the proposed admission control is twice as efficient as the Equivalent Capacity