This research program is motivated by problems that arise in assessing and monitoring the quality of service (QoS) characteristics of computer and communication networks. It focuses on active network tomography where end-to-end traffic measurements are obtained by actively probing the network. The project examines a number of important and interesting statistical issues dealing with the collection, modeling, and analysis of network data using active tomography. Specifically, it introduces a new, flexible class of probing experiments that has advantages over the current schemes. The investigators will study several research issues related to the design of efficient flexicast probing experiments: questions of identifiability, construction of efficient designs, optimal allocation of test probes (according to some suitable metric), and real-time monitoring strategies to detect and localize network problems. A substantial portion of the research deals with estimation and inference for a large-scale inverse problem with missing data. The focus is on fast and scalable algorithms for estimating packet loss rates and delay distributions and their asymptotic properties. Furthermore, new models for spatial and temporal dependence that capture the main features of network traffic will be introduced and the behavior of various estimation methods under these models will be analyzed. Finally, an important component of this research program is the validation of the proposed models and techniques on real data. Present day computer networks have evolved into large, complex, decentralized systems. There is a great deal of interest in estimating key network performance parameters and quality of service indicators, such as traffic intensities, link delays and dropped packet rates, and in identifying bottleneck nodes and detecting network failures in the form of connectivity, routing faults and malicious activity. However, the lack of centralized control makes the quantitative assessment of network performance a rather difficult task, since detailed queueing and traffic models do not capture their complexity and characteristics well. Network tomography is capable of assessing the performance of large-scale networks and in localizing anomalous behavior to individual components and subnetworks. The understanding and insights gained as a result of the proposed research will lead to a core of basic principles and a toolkit of key statistical methods and techniques for the analysis and design of network monitoring schemes. Efficient and scalable network monitoring could significantly benefit network control and provisioning by incorporating the necessary information into admission, flow and routing control protocols. The proposed techniques and algorithms will be integrated into an open source tool, called FLEXICAST. The research program will also be integrated into the educational activities of the Department of Statistics. Planned efforts include training of graduate and undergraduate students, the development of a new course, stimulation of collaborations with students and faculty in Engineering, and interaction with the networking community to help transfer the research results to industry.
