Empirical Evaluation of Techniques for Measuring Available Bandwidth

The ability to measure end-to-end available bandwidth (AB) on a network path is useful in several domains, including overlay-routing infrastructure, network monitoring, and design of transport protocols. Several tools have, consequently, been proposed to estimate end-to-end AB. Unfortunately, existing evaluations of these tools are either not comprehensive or are biased by the current state of implementation technology. In this paper, we conduct a comprehensive empirical evaluation of algorithmic techniques for measuring AB. In order to eliminate implementation-related biases, we rely on a simulated network environment and develop a generic implementation framework for instantiating different tools. We implement our framework in NS-2 and reproduce traffic from real Internet links in order to evaluate tools under different conditions of (i) traffic load, (ii) sampling intensities, (iii) measurement timescales, (iv) number of bottleneck links, and (iv) location of bottleneck. We evaluate the tools for their accuracy, run-time, overhead, intrusiveness, and impact on responsive traffic. Our results contradict some of those in prior work that does not eliminate implementation biases.