Flow level IP traffic characterization

In this paper we perform a detailed statistical analysis of IP traffic at flow level, based on data traces collected on the France Telecom wide area IP network. We particularly focus on the assessment of assumptions commonly used in flow level performance models, i. e. Poisson flow arrival process, both for UDP and TCP traffic. We first discuss the concept of flow and present a method to estimate an adequate Time Out value to distinguish different flows within a traffic stream. To model the flow arrival process, we find that Gamma and Weibull distributions provide excellent fits to the empirical flow inter-arrival time distributions. In addition, a fairly low level of correlation is detected between successive flow arrivals. Flow size distributions, though more difficult to model due to their composite aspect, show a very well marked heavy-tailed character of the Pareto type. We found that these results were very consistent among all the data traces we analyzed.