Network Monitoring: It Depends on your Points of View

End-to-end active network monitoring infers network characteristics by sending and collecting probe packets from the network edge, while probes traverse the network through multicast trees or a mesh of unicast paths. Most reported methods consider given source and receiver locations and study the path selection and the associated estimation algorithms. In this paper, we show that appropriately choosing the number of sources and receivers, as well as their location, may have a significant effect on the accuracy of the estimation; we also give guidelines on how to choose the best “points of view” of a network for link loss monitoring purposes. Though this observation applies across all monitoring methods, we consider, in particular, networks where nodes are equipped with network coding capabilities; our framework includes as special cases the scenarios of pure multicast and network coding. We show that, in network-coding enabled networks, multiple source active monitoring can exploit these capabilities to estimate link loss rates more efficiently than purely tomographic methods. To address the complexity of the estimation problem for large networks, we also propose efficient algorithms, including the decomposition into smaller multicast inference problems, belief-propagation, and a MINClike algorithm.