Extreme Value Analysis of Network Traffic Time Series and Tail Performance Evaluation

• Activity Number: 367
• Type: Contributed
• Date/Time: Wednesday, August 6, 2003 : 10:30 AM to 12:20 PM
• Sponsor: Section on Physical and Engineering Sciences
• Author(s): Z. Q. John Lu*+
• Companies: National Institute of Standards and Technology
• Keywords: tail performance evaluation ; network metrics ; generalized Pareto distribution ; non-Gaussian time series ; distributional comparison ; mixture models

Abstract

Internet network modeling and simulation is characterized by the efforts in "searching for invariant" characteristics of data, in the sense of finding the most parsimonious statistical models for large and multiscale time series data. Two most commonly observed features of data network traffic are the heavy-tailed marginal distribution and long range (or self-similarity) dependence. There is, however, lack of robust and stable parametric models that can work across different timescales and on a variety of datasets. By focusing on the most important upper tail property, we introduce the generalized Pareto model as a general model for normally behaving network traffic. To account for anomalies and network perturbation in "jamming" or "slowdown" sessions, we advocate the mixture model framework. Both ideas have led to a fruitful approach to non-Gaussian network time series modeling and simulation. Applications to the Round Trip Time database collected at NIST reveal interesting comparison on the tail performance behavior of NIST's network change in 1998. (Supported in part by DARPA's Network Modeling and Simulation Program.)