How to measure available bandwidth on the Internet

: Available bandwidth is the most useful measurement to network adaptive applications and transports. Unfortunately, it is also very difficult to measure in a network that cannot be well approximated by a weighted fair queuing model, such as the Internet. At the same time, the Internet scalable architecture calls for end-to-end measurements. Current available bandwidth measurement techniques are based in observing packet dispersion in a packet train or pair. The available bandwidth is sampled by using a “bytes divided by dispersion” (or “bytes over time”) calculation and then filtered. In this paper, we deal with sampling the available bandwidth from packet dispersion. We argue that existing techniques, being heuristics, are unintuitive, are not based on any network model, and introduce an error in the process right from the sampling. We propose a different calculation of the available bandwidth from packet dispersion observation, we call ab-probe. It is also simple and as robust, and follows a simple network model. It is an improvement to the “bytes over time” calculation because it also captures and deducts a portion of the argued sample calculation error. We study this new model by exploring the differences between the two methods and its robustness to parameter estimation errors. We then apply our model to a well known available bandwidth measurement technique called cprobe. We find that our model explains cprobe’s behavior as recorded by its researchers’ published real experiments. We further validate the measurement using simulations in some simple wired scenarios.