Efficient Loss Inference Algorithm Using Unicast End-to-End Measurements

We address the problem of loss rates inference from end-to-end unicast measurements. Like other network tomography problems, it requires solving a system of equations that involve measurement values and the loss rates of links. However, the equations do not have a unique solution in general. One kind of method imposes unrealistic assumption on the system, e.g. the uniform prior probability of a link being congested. Other methods use multiple probe measurements to acquire more information about the system that may generate many additional overhead costs. In this paper, we demonstrate that a considerable portion (more than 95 %) of links could be uniquely identified by current measurements directly. Then we utilize the information of these determined links to acquire the global distribution of the system that can help to infer the rest loss rates. Moreover, we derive an upper bound on the accuracy of a congestion localization problem using the Bayesian network that provides a necessary condition for achieving the 0—error diagnosis. Finally, we evaluate our new method and a former representative method by both the simulation and the real implementation in the PlanetLab network. The results show that our method not only makes a great improvement on the accuracy, but also reduces the probe costs and the running time to an extremely low level. Furthermore, our method can also perform well in large and more congested networks.