Queuing Network Topology Inference Using Passive Measurements

In this work, we revisit a classic problem of inferring a tree topology from end-to-end measurements originated by a single source, with two critical differences: (i) instead of relying on measurements with specific correlation across paths that often require active probing, we do not rely on any correlation and can thus utilize passive measurements; (ii) instead of inferring a logical topology that ignores certain nodes, we want to recover the physical topology. Our key idea is to utilize the detailed queuing dynamics inside the network to estimate the number of queues and a certain parameter (residual capacity) of each queue on each measurement path, and then use the estimated parameters as fingerprints to detect shared queues and infer the topology. To this end, we develop a Laplace-transform-based estimator to extract the parameters of a tandem of queues from end-to-end delays, and efficient algorithms to identify the parameters associated with the same queue and infer the topology accordingly. The inferred topology is guaranteed to converge to the ground truth, up to a permutation of queues traversed by the same paths, as the number of measurements increases. Our evaluations validate the proposed solutions against benchmarks and identify potential directions for further improvements.