IP traffic characterization for planning and control

IP traffic modeling and engineering is a challenging area that has attracted an extensive research effort in recent years. Many studies show that the packet-level traffic from data networks, such as LANs, WANs, ATM, Frame Relay, Internet, etc., exhibits slowly decaying autocorrelation or long range dependence. In this paper, we discuss the key differences between traditional voice traffic and IP traffic, and the challenging issues in IP traffic engineering and modeling. By using packet data collected on a corporate Intranet, we make several observations and recommendations that are new and have significant implications for the future of IP traffic characterization. We find that the slowly decaying autocorrelation phenomenon is mostly caused by flows of correlated packets and the existence of some periodic network management traffic. On the control time scale, we propose to identify the sources of the peaks that cause most of the performance problems, partition the traffic into classes of similar applications, and do characterization by class. On the planning time scale, we characterize traffic variations by a mixture of distributions that closely fit the empirical histogram formed from round-the clock measurements during a fixed calendar period.