Saturation aware TCP throughput prediction

With the ever growing network traffic, capacity of paths in the networks nowadays has become increasingly easy to saturate. This brings new challenge to TCP throughput prediction. The main problem is loss rate and RTT during the TCP flow are often expected as inputs for the throughput models used in the prediction; however, only loss rate and RTT before the flow are available and are used instead. If the flow itself causes significant changes in loss rate and RTT, e.g., when the TCP flow attempts to saturate the underlying available bandwidth, the prediction error can be unacceptably large. Though new prediction approaches are being proposed, they basic require record of previous TCP transfers and are applicable only when TCP transfers are performed repeatedly, which limits their application. In this work, by properly using a measurement of the underlying available bandwidth, we develop an analytical TCP throughput model which can explicitly capture changes in loss rate and RTT caused by the target TCP flow and hence, can largely improve the prediction accuracy by making the prediction aware of capacity saturation of paths while at the same does not require any history record of previous TCP transfers as current newly proposed works do. Results show that when the changes in loss rate and RTT are large, the errors by traditional models can be as large as over 200%, whereas the error by the proposed model is usually very small, e.g., with the average error below 10% and the maximum error below 20% for general settings, and is bounded by the measurement error of available bandwidth in the worse case; when the changes in loss rate and RTT are small, even a very rough estimation of available bandwidth, e.g., with an error of around 50%, can lead to very accurate prediction by the proposed model.