Rank-order statistics for the campus local area network traffic

Abstract

We analyze the statistical properties of the local area network (LAN) traffic data collected from a part of the campus of St. Petersburg Electrotechnical University. The entire LAN is subdivided into several subnets, thus we focus on the analysis of the downlink traffic fluctuations at several levels, including individual end users' devices (end-IPs), individual subnets as well as the entire network. Following a recently suggested approach, we also split the daily traffic logs into short fragments where the traffic could be treated as stationarity to a first approximation. Next traffic anomalies were excluded from further analysis based on their Cook's distance exceeding a certain threshold. We find that at all studied levels the distribution of the traffic intensity to different end devices as well as separate subnets obeys Zipf's law. We show that, while the asymptotic behavior is characterized by a power law, in practical setting for a small number of IPs and subnets in the LAN leading to inevitable finite size effects this law to a reasonable approximation can be described by a gamma distribution. Particular parameters of the fitting distribution were obtained by regression analysis leading to a phenomenological scalable model. Finally, we show that the parameters of this model depend for a given network are determined by the number of active users.