Empirical modeling of end-to-end delay dynamics in best-effort networks

• Main Author: Doddi, Srikar
• Other Authors: Parlos, Alexander
• Format: Others
• Language: en_US
• Published: Texas A&M University 2005
• Subjects: QoS; predictors; ARX; ARMAX; Neural networks; Packet traces,
• Online Access: http://hdl.handle.net/1969.1/2244

Quality of Service (QoS) is the ability to guarantee that data sent across a network will be recieved by the desination within some constraints. For many advanced applications, such as real-time multimedia QoS is determined by four parameters--end-to-end delay, delay jitter, available bandwidth or throughput, and packet drop or loss rate. It is interesting to study and be able to predict the behavior of end-to-end packet delays in a Wide area network (WAN) because it directly a??ects the QoS of real-time distributed applications. In the current work a time-series representation of end-to-end packet delay dynamics transported over standard IP networks has been considered. As it is of interest to model the open loop delay dynamics of an IP WAN, the UDP is used for transport purposes. This research aims at developing models for single-step-ahead and multi-step-ahead prediction of moving average, one-way end-to-end delays in standard IP WAN??s. The data used in this research has been obtained from simulations performed using the widely used simulator ns-2. Simulation conditions have been tuned to enable some matching of the end-to-end delay profiles with real traffic data. This has been accomplished through the use of delay autocorrelation profiles. The linear system identification models Auto-Regressive eXogenous (AR) and Auto-Regressive Moving Average with eXtra / eXternal (ARMA) and non-linear models like the Feedforwad Multi-layer Perceptron (FMLP) have been found to perform accurate single-step-ahead predictions under varying conditions of cross-traffic flow and source send rates. However as expected, as the multi-step-ahead prediction horizon is increased, the models do not perform as accurately as the single-step-ahead prediction models. Acceptable multi-step-ahead predictions for up to 500 msec horizon have been obtained.