Using non-random associations for predicting latency in WANs

In this paper, we propose a scalable performance management tool for Wide Area Applications. Our objective is to scalably identify non-random associations between pairs of individual Latency Profiles (iLPs) (i.e., latency distributions experienced by clients when connecting to a server) and exploit them in latency prediction. Our approach utilizes Relevance Networks (RNs) to manage tens of thousands of iLPs. Non-random associations between iLPs can be identified by topology-independent measures such as correlation and mutual information. We demonstrate that these non-random associations do indeed have a significant impact in improving the error of latency prediction.