Network Tomography

The Internet has evolved from a small tightly controlled network serving only a few users in the late 1970's to the immense decentralized multi-layered collection of heterogeneous terminals, routers and other platforms that we encounter today when surfing the web. The lack of centralized control has allowed Internet service providers (ISP)'s to develop a rich variety of user-services at different quality-of-service (QoS^?) levels. However, in such a decentralized environment quantitative assessment of network performance is difficult. One cannot depend on the cooperation of individual servers and routers to freely transmit vital network statistics such as traffic rates, link delays, and dropped packet rates. Indeed, an ISP may regard such information as highly confidential. On the other hand, sophisticated methods of active probing and/or passive monitoring can be used to extract useful statistical quantities that can reveal hidden network structure and detect and isolate congestion, routing faults, and anomalous traffic. The problem of extracting such hidden information from active or passive traffic measurements falls in the realm of statistical inverse problems; an area which has long been of interest to signal.

A fundamental ingredient in the successful design, control and management of coming networks will be the accurate measurement and characterization of its dynamics. For the task of network management and monitoring, the computation cost is high, and currently Yu's group has developped a pseudo likelihood approach to cut down the computational cost. Also we view the data network project for a constrained independent component analysis (CICA) point of view. Our proposed research focuses on the following goals:

• Development of rigorous statistical estimation methodology for multicast internal link distribution through end-to-end measurements.
• Develop and evaluate parametric and non-parametric estimators for link-level performance metrics such as loss rate, delay statistics.
• Multicast network topology identification.
• Network origin-destination (OD) matrix inference.

People Involved

• Gang Liang
• Bin Yu

Papers

• R. Castro, M. Coates, G. Liang, R. Nowak, and B. Yu. (2003) Internet Tomography (Recent Developments), Statistical Science (invited)
• G. Liang and B. Yu. (2003) Maximum Pseudo Likelihood Estimation in Network Tomography, IEEE Transaction on Signal Processing (special issue on data network).
• M. Coates, A. Hero, R. Nowak and B. Yu. (2002) Large scale inference and tomography for network monitoring and diagnosis, Signal Processing Magazine.