An Information Theoretic Approach for Shared Bottleneck Inference Based on End-to-end Measurements

Abstract

Recent years have marked a growing interest in studying Internet path characteristics. However, most of the currently available tools to an end system to perform such measurements are slow inaccurate and generate an excessive amount of probing traffic. This paper introduces entropy as a novel and efficient metric for discovering Internet path characteristics based on data collected by an end system. In particular, the paper presents an entropy-based technique that enables an end system to cluster flows it receives according to their shared bottleneck. Our mechanism relies solely on information extracted from the packets' inter-arrivals at the receiver. It does not generate any probing traffic and can use data extracted from both TCP and UDP flows. Moreover, it requires only a small number of packets from each flow, which makes it useful for short-lived flows. We report the result of running the algorithm on simulated data and Internet traffic.