무선랜 트래픽의 분석과 모델링

Abstract

In this thesis, we present the result of our empirical study on IEEE 802.11b wireless LAN network traffic, to effectively exploit the underlying network bandwidth while maximizing user perceivable QoS, mandatory to make proper estimation on packet loss and queuing delay of the underlying network. This issue is further emphasized in wireless network environment where network bandwidth is scarce resource. We focus our effort on developing performance model for wireless network. We collect the packet trace from existing campus wireless LAN infra-structure. We analyze four different data sets: aggregate traffic, upstream traffic, downstream traffic, tcp only packet trace from aggregate. 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 term 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 151.7 byte while average packet sizes of the rest of the data sets are all greater than 260 bytes. 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 extract key performance parameters of the underlying network traffic. We model the underlying traffic using fractional-ARIMA (FARIMA) fractional Gaussian Noise (FGN), Markov Modeling Poisson Process (MMPP) and develop an analytical model for Token bucket with Fractional Brown Motion (FBM), buffer overflow probability, channel capacity with dependent of H parameter, and waiting time. We obtain the tail probability of the queuing system using FBM. We present average queuing delay from queue length model. While the fractional Gaussian Noise based method is computationally more efficient, FARIMA exhibits superior performance in accurately modeling the underlying traffic. Through our study based upon empirical data, it is found that our performance model well represent the physical characteristics of the IEEE 802.11b network traffic.