Inferring path sharing based on flow level TCP measurements

We develop methods to infer path or bottleneck sharing among TCP flow classes based on flow level measurements available from the current traffic monitoring tools. Our premise is that flows that temporally overlap on the congested resources have correlated throughputs. We propose to use factor analysis to explore the correlation structure of flow class throughputs in order to hypothesize which flow classes might share congested resources. The effectiveness of this "black box" approach is studied using the empirical data. We show that making such inferences based on flow level statistics is viable in practice, and can serve as an effective, novel tool for network design and configuration decisions. Our work on inferring bottleneck sharing differs significantly from the previous work in that we consider flow level instead of packet level statistics, and hence may potentially influence research in that area. Possible applications of this technique include network monitoring and root cause analysis of poor performance.