Classifying daily patterns in long duration network traces

Models of network traffic for use in simulation should be representative of the traffic observed on the type of day they are trying to replicate. Building a model from a single day or small number of days makes it prone to overfitting or being unduly influenced by unusual events. With very long duration traces such as the multiple-year spanning Waikato datasets captured by the WAND Network Research Group it is possible to more accurately characterise behaviour and define appropriate boundaries for when traffic is similar enough and when it is different. We present here an approach to identifying and describing discrete ldquotypesrdquo of days within these traces and what differences are important to distinguish between them. By applying machine learning techniques to the long duration traces it is possible to describe and simulate a generic day of a specific type without it being explicitly based on a particular day. The resulting parameters are used to configure a number of popular traffic generators which are then evaluated using the same criteria with which the model was built.