In this thesis we design, analyze, and develop accurate and scalable algorithms and techniques for performance analysis and inference of large communication networks.

In the network plan phase, stochastic network models such as loss networks and queueing networks are often used to analyze and predict the network performance. We propose a new approach to analyze stochastic networks based on probabilistic graphical models. We use factor graphs to model the stationary distribution of a network and apply the sum-product algorithm to compute the performance measures of interest. We demonstrate via both analysis and numerical experiments that the algorithm has excellent performance for networks with a variety of topologies. In addition, the algorithm scales gracefully to large networks and can be implemented in a distributed manner.

In the network operation phase, inference of the network structure and dynamics is an essential component of many network design and management tasks to achieve high network performance. We first address the network inference problem of estimating the routing topology and link performance. We propose a new, general framework for designing network inference algorithms based on additive metrics using tools from phylogenetic inference. The framework can flexibly fuse information from different measurements to improve estimation accuracy. Based on the framework we develop several computationally efficient distance-based inference algorithms with provable performance. We also design topology inference procedures which can handle dynamic node joining and leaving efficiently. Using the procedures we develop a novel sequential topology inference algorithm which greatly reduces the probing scalability problem under unicast probing.

For networks with general, known topology, we develop explicit link performance parameter estimators that are easy to implement and compute. The derived asymptotic properties of the explicit estimators provide a basis for us to analyze and bound the performance of more sophisticated estimators.

We then address the network inference problem of estimating the origin-destination traffic matrices. Using convex optimization theory, we propose a dual approach to convert the constrained primal optimization problem of the gravity model approach into an unconstrained dual optimization problem. We develop a distributed traffic matrix estimation algorithm based on the dual problem.