Mixture Models of Endhost Network Traffic

In this work we focus on modeling a little studied type of traffic, namely the network traffic generated from endhosts. We introduce a parsimonious parametric model of the marginal distribution for connection arrivals. We employ mixture models based on a convex combination of component distributions with both heavy and light-tails. These models can be fitted with high accuracy using maximum likelihood techniques. Our methodology assumes that the underlying user data can be fitted to one of many modeling options, and we apply Bayesian model selection criteria as a rigorous way to choose the preferred combination of components. Our experiments show that a simple Pareto-exponential mixture model is preferred for a wide range of users, over both simpler and more complex alternatives. This model has the desirable property of modeling the entire distribution, effectively segmenting the traffic into the heavy-tailed as well as the non-heavy-tailed components. We illustrate that this technique has the flexibility to capture the wide diversity of user behaviors.