| The majority of current Internet traffic is based on TCP. With the emergence of new applications, especially new multimedia applications, however, UDP-based traffic is expected to increase. Furthermore, multimedia applications have sparked the development of protocols responding to congestion while behaving differently from TCP. As a result, network traffic is expected to become more and more diverse. The increasing link capacity further stimulates new applications utilizing higher bandwidths of future. Besides the traffic diversity, the network is also evolving around new technologies. These trends in the Internet motivate our research work.; In this dissertation, modeling and estimation techniques of heterogeneous traffic at a router are presented. The idea of the presented techniques is that if the observed queue length and packet drop probability do not match the predictions from a model of responsive (TCP) traffic, then the error must come from non-responsive traffic; it can then be used for estimating the proportion of non-responsive traffic. The proposed scheme is based on the queue length history, packet drop history, expected TCP and queue dynamics. The effectiveness of the proposed techniques over a wide range of traffic scenarios is corroborated using NS-2 based simulations. Possible applications based on the estimation technique are discussed.; The implementation of the estimation technique in the Linux kernel is presented in order to validate our estimation technique in a realistic network environment. Adapted from the NS-2 implementation, the Linux implementation is tuned to produce accurate estimates under a realistic testbed. The correctness of our Linux implementation is corroborated through tests.; In order to investigate the feasibility of an approach to regulate the volume of incoming non-responsive traffic on an aggregate level by utilizing the estimation information, the performance evaluation of heterogeneous traffic under different queue management schemes and network configurations is studied. Our evaluation considers the aggregate bandwidth of different classes of traffic and the delays observed at the router. Our NS-2 based evaluation shows that queue management schemes (without per-flow or per-class state) do not provide significant control over the traffic mix consisting of both long-term and short-term traffic. |
