Prediction of end-to-end single flow characteristics in best-effort networks

The nature of user traffic in coming years will become increasingly multimediaoriented which has much more stringent Quality of Service (QoS) requirements. The current generation of the public Internet does not provide any strict QoS guarantees. Providing Quality of Service (QoS) for multimedia application has been a difficult and challenging problem. Developing predictive models for best-effort networks, like the Internet, would be beneficial for addressing a number of technical issues, such as network bandwidth provisioning, congestion avoidance/control to name a few. The immediate motivation for creating predictive models is to improve the QoS perceived by end-users in real-time applications, such as audio and video. This research aims at developing models for single-step-ahead and multi-stepahead prediction of end-to-end single flow characteristics in best-effort networks. The performance of path-independent predictors has also been studied in this research. Empirical predictors are developed using simulated traffic data obtained from ns-2 as well as for actual traffic data collected from PlanetLab. The linear system identification models Auto-Regressive (AR), Auto-Regressive Moving Average (ARMA) and the non-linear models Feed-forward Multi-layer Perceptron (FMLP) have been used to develop predictive models. In the present research, accumulation is chosen as a signal to model the end-to-end single flow characteristics. As the raw accumulation signal is extremely noisy, the moving average of the accumulation isused for the prediction. Developed predictors have been found to perform accurate single-step-ahead predictions. However, as the multi-step-ahead prediction horizon is increased, the models do not perform as accurately as in the single-step-ahead prediction case. Acceptable multi-step-ahead predictors for up to 240 msec prediction horizon have been obtained using actual traffic data.