Prediction of Bandwidth and Additive Metrics for Large Scale Network Tomography

For real time communication services over the Internet, it is important to be able to predict in advance the quality of a call before relaying it over a particular path. In this paper we show how to predict the distribution of the end-to-end bandwidth, latency, jitter, and loss of a call from an arbitrary user X to an arbitrary user Y through particular components of the Internet, given a dataset of millions of calls among other users. This work is the first to infer component bandwidth distributions for an arbitrary network topology, and the first to infer component bandwidth and additive metric distributions for an arbitrary network topology at large scale. On a dataset of over seven million global Skype calls, we demonstrate significant performance improvement compared to a baseline approach used in today’s commercial systems.