IT Performance: Statistical Visualization and Modeling for Network Data (Understanding Internet Performance from the User Perspective)

Introduction	The Statistical Engineering Division and the Advanced Network Technologies Division at the National Institute of Standards and Technology are collaborating on developing metric and tools to measure the performance of Internet services with the goal of understanding Internet performance from the user perspective. The project was initiated as a collaboration with an industrial consortia, which included the Cross Industry Working Team (XIWT) and the Internet Performance Metrics Working Group of the Internet Engineering Task Force. The activities of these consortia are driven by the pressing need for metrology that enables customer service level agreements with ISP's, faster methods for troubleshooting network problems and identifying network-wide inefficiencies, and resource allocation techniques to meet the Quality of Service requirements of emerging applications.
Background/ Impetus	The critical issues for network traffic modeling and simulation are characterizing the tail behavior for the "heavy-tailed" empirical distributions of network measurements and detecting changes in distribution patterns. We have proposed a statistical methodology for measuring network "slowness" and extreme-value-based metrics for network performance. We have developed graphical tools for visualizing Internet data that do not depend on a particular Internet model by focusing on the upper quantiles rather than means or medians. A spatial representation of the two different scenarios of heavy-tailed distributions in the previous plot is shown below. Working closely with customers from industry and academia, we have started on a major improvement on the statistical time series models that are used to drive the NIST Net emulation tool at http://www.antd.nist.gov/nistnet/. The new version will allow the feed of live trace data and successful adaptation of Rice's MWM software. The NIST Net emulation tool is a Linux-based general-purpose network emulation, and is widely used by the industry to test Internet communication software. This work is funded in part by DARPA.
Customers	The customers for the network visualization project are: NIST: David Su, Doug Montgomery, Mark Carlson (Advanced Network Technologies Division, ITL) DOD the Internet industry
Goals	The goals for the network visualization and modeling project are: Develop sound statistical methodology for network traffic data analysis and simulation Develop new metrics for network performance based on upper-tail distribution: graphical tools and analysis tools. Developing NIST Net to improve statistical models and integration to NMS simulation tools.
Impact	Our efforts in statistical modeling and visualization are intended to create statistical models sophisticated enough to cover a broad range of real network behavior, and yet simple and intuitive enough to be easily employed by researchers developing network simulators, emulators, and control models. The Internet is rapidly becoming the most critical information infrastructure in the United States, supporting commerce, entertainment, scientific research, and general collaboration and communication.
FY03 Milestones	Milestones for the network visualization and modeling project are: Develop and implement state-of-the art time series models for network traffic simulation. Publish graphical monitoring tools for network performance based on upper-quantile characteristics and mixture statistical modeling. Develop statistical modeling and tools for network performance, visualization, and anomaly detection.
FY 02 Accomplishments	Develop the statistical methodology for tail metrics for network performance based on GPD and mixture modeling. Develop a graphical tool for upper-quantile monitoring of network performance. Develop and improve statistical time series models used in NIST Net. Non-guassian time series models with short-term and long-term correlations are studied. Successful installation of the MWM data synthesis tool in NIST Net.
R&D Team	Hung-Kung Liu, Statistical Engineering Division, ITL John Lu, Statistical Engineering Division, ITL Nell Sedransk, Statistical Engineering Division, ITL David Su, Advanced Network Technologies Division, ITL Doug Montgomery, Advanced Network Technologies Division, ITL Mark Carlson, Advanced Network Technologies Division, ITL
Achievements	Achievements of the network visualization project include: Developed a new methodology for potential applications to Quality of Service (QoS) and computer security (to be documented in the Sedransk and Lu paper given below). Developed a graphical tool for upper-quantile monitoring of network performance. Initiated collaboration with ANTD division (Mark Carlson and Doug Montgomery) on NIST NET and worked on the improvement of statisitcal simulation models. Installed multifractal wavelet model code (Matlab code from Rudolf Riedi's group at Rice) and adapted successfully for NIST Net emulation of RTT data.
Publications	Expected publications to result from the network visualization project include: John Lu and Nell Sedransk, "Generalized Pareto and Mixture Approach to Statistical Data Network Modeling and QoS Provision", to be submitted to IEEE Transactions on Information Theory, October, 2002. Hung-kung Liu, "Statistical and Graphical Tools for Network Modeling", to be submitted to Journal of Computational and Graphical Statistics, September, 2003.
Presentations	Presentations resulting from the network visualization project include: John Lu and Nell Sedransk, "Tail Metrics for Network Performance Based on GPD and Mixture Modeling", DARPA workshop on Network Modeling and Simulation, April, 2002. Nell Sedransk and Hung-kung Liu, "Measurements, Modeling and Applications", DARPA workshop on Network Modeling and Simulation, October, 2001.

Date created: 2/8/2002
Last updated: 6/21/2002
Please email comments on this WWW page to sedwww@nist.gov.