¡@

In this work, we present the results of our empirical study on 802:11 wireless LAN network traffic. We collect the packet trace from existing campus wireless LAN infrastructure. We analyze four different data sets: aggregate traffic, upstream traffic, downstream traffic, tcp only packet trace from aggregate traffic. We analyze the time series aspects of underlying traffic (byte count process and packet count process), marginal distribution of time series, and packet size distribution. We found that in all four data sets there exist long-range dependent properties in terms of byte count and packet count process. Inter-arrival distribution is well fitted with Pareto distribution. Upstream traffic, i.e. from the user to Internet, exhibits significant difference in packet size distribution from the rests. Average packet size of upstream traffic is 151:7byte while average packet size of the rest of the data sets are all greater than 260bytes. Packets with full data payloads constitute 3% and 10% in upstream traffic and downstream traffic, respectively. Despite the significant difference in packet size distribution, all four data sets have similar Hurst values. The Hurst alone does not properly explain the stochastic characteristics of the underlying traffic. We model the underlying traffic using fractional-ARIMA (FARIMA) and fractional Gaussian Noise (FGN). While the fractional Gaussian Noise based method is computationally more efficient, FARIMA exhibits superior performance in accurately modeling the underlying traffic.

Keywords: network traffic, modeling, analysis, self-similarity, long-range dependence, fractional-ARIMA, fractional Gaussian noise

Retrieve PDF document (200911_08.pdf)

Received December 24, 2007; revised April 22 & July 16, 2008; accepted October 2, 2008.
Communicated by Ten-Hwang Lai.
^* This work was funded by National Research Lab Grant (ROA02007-000-200114-0) by KOSEF.