Zooming Statistics: A Multiscale Look at Internet Traffic Data

Monday, August 6, 2001 - 4:00pm - 5:00pm
Keller 3-180
J. Stephen Marron (University of North Carolina, Chapel Hill)
Joint work with: Jan Hannig, Colorado State University and Rolf Riedi, Rice University.

Scale of statistical analysis is an important topic for internet traffic data. A seminal body of work has revealed that apparent heavy tail TCP connection distributions result in a host of phenomena related to long range dependence and self similarity, indicating that the exponential/Poisson models at the heart of classical queueing theory are quite inappropriate. More recent purely empirical work suggests that at fine time scales, classical models are appropriate near the center of the internet. These issues are empirically illustrated, and the apparent contradictions are resolved, using zooming autocovariance and zooming SiZer, analyses. If time permits, some novel views of stationarity and heavy tails will also be presented.