Machine-learnt versus analytical models of TCP throughput

We first study the accuracy of two well-known analytical models of the average throughput of long-term TCP flows, namely the so-called SQRT and PFTK models, and show that these models are far from being accurate in general. Our simulations, based on a large set of long-term TCP sessions, show that 70% of their predictions exceed the boundaries of TCP-Friendliness, thus questioning their use in the design of new TCP-Friendly transport protocols. We then investigate the reasons of this inaccuracy, and show that it is largely due to the lack of discrimination between the two packet loss detection methods used by TCP, namely by triple duplicate acknowledgements or by timeout expirations. We then apply various machine learning techniques to infer new models of the average TCP throughput. We show that they are more accurate than the SQRT and PFTK models, even without the above discrimination, and are further improved when we allow the machine-learnt models to distinguish the two loss detection techniques. Although our models are not analytical formulas, they can be plugged in transport protocols to make them TCP-Friendly. Our results also suggest that analytical models of the TCP throughput should certainly benefit from the incorporation of the timeout loss rate.