Available bandwidth estimation from passive TCP measurements using the probe gap model

The Internet relies on congestion control protocols and adaptive applications that adjust their data rate to achieve good performance while avoiding network congestion. An essential prerequisite is the estimation of available network resources: implicitly like prevailing TCP versions that adapt their data rate iteratively; or explicitly by available bandwidth estimation techniques, as recently also adopted by TCP HyStart. Using observations of TCP throughput, applications like MPEG- DASH adapt the video quality and data rate. We discover, however, relevant conditions where TCP throughput is not a good bandwidth estimator and observe that it is outperformed by known UDP-based active probing methods. We investigate how the theory of active probing can possibly be used to extract relevant information also from passive TCP measurements. In case of TCP, the additional difficulty is found to be due to its chaotic traffic characteristics. We define a criterion to select relevant traffic samples and apply a regression technique to estimate the available bandwidth. Noteworthy, using the feedback provided by TCP acknowledgements, we can perform the estimation from sender-side measurements only. We verify the fidelity of the approach in a variety of experiments, including different types of cross-traffic, delays, and loss of data packets as well as acknowledgements.