Characterizing Non-stationarity in the Presence of Long-range Dependence

Based on ours and other studies, it is becoming evident that Internet traffic is often nonstationary. Since such traffic is generally found to be long-range dependent as well, and the two properties can get confused, it is important to have a method for separately characterizing them. This paper proposes such a methodology aimed specifically for traffic scenarios where nonstationarity comes into play at a relatively large time-scale. This methodology allows us to accurately characterize nonstationarity and and longrange dependence and thereby synthetically generate traffic exhibiting it. Such synthetic traffic can be used for performance engineering, dynamic resource allocation and overload control. We illustrate the methodology by applying it to traffic traces from three very different types of commercial web-service environments.