Networking Research 

  Chun You

BS, MS Shanghai Jiao Tong University, P.R. China, 1992, 1995 

MSEE, University of Massachusetts Lowell 1999

Work: Lucent Technologies, Westford, Massachusetts

Thesis Title: Time Series Models for Internet Data Traffic

Thesis (pdf)


In this thesis two types of data traffic are examined, one is traffic from Ohio State University (OSU) to the Internet and the other is Bellcore Ethernet traffic. The statistical features of these data sets are studied and traffic models proposed. The OSU traffic trace is studied at the protocol and application level. The traffic from applications with quasi-deterministic features are filtered out of the data trace prior to modeling. The Bellcore traffic trace is studied at the packet size level, and decomposed into three ranges according to the packet size. The application of linear autoregressive moving average (ARMA) models to the OSU traffic is investigated first. The random walk and exponentially weighted moving average models are used to capture the mean trend in the data. The comparison of the linear ARMA model simulated data and the OSU data shows that the linear model is unable to capture the queuing loss statistics exhibited by OSU traffic. The nonlinear threshold autoregressive (TAR) model is used to extend the linear model approximations. The TAR model is comprised of a series of linear AR models that are valid in disjoint subregions of the time series amplitude. This model is derived based on the lagged regression function estimates which show that the OSU trace amplitudes are piece-wise linear. At a given time the subregion selected will depend on the amplitudes observed over delayed time values. The AR order parameters are calculated using the minimum mean square error criterion. For selected amplitude and delay parameters, the optimal AR parameters are obtained using the minimum AIC value. The residuals in each subregion of the TAR model for OSU data are shown to be zero mean Gaussian random variables. The Bellcore data is also modeled using the TAR process and the simulated data and the Bellcore measurements show close match in statistical features. The TAR parameters are used in the design of a traffic shaper. The traffic shaper services to whiten the short term correlation in the data traffic and is parametrized by the AR coefficients of the TAR model. A voice-data multiplexer is designed based on this traffic shaper. Data from the traffic shaper shares a common buffer with voice and the delay of voice traffic is shown to decrease significantly. This process can be useful in meeting the QoS requirements for voice without the need of other feedback control schemes.