Multicast inference of temporal loss characteristics

Multicast-based inference has been proposed as a method of estimating average loss rates of internal network links, using end-to-end loss measurements of probes sent over a multicast tree. We show that, in addition to loss rates, temporal characteristics of losses can also be estimated. Knowledge of temporal loss characteristics has applications for services such as voip which are sensitive to loss bursts, as well as for bottleneck detection. Under the assumption of mutually independent, but otherwise general, link loss processes, we show that probabilities of arbitrary loss patterns, mean loss-run length, and even the loss-run distribution, can be recovered for each link. Alternative estimators are presented which trade-off efficiency of data use against implementation complexity. A second contribution is a novel method of reducing the computational complexity of estimation, which can also be used by existing minc estimators. We analyse estimator performance using a combination of theory and simulation.