D. Zheng∗ and G.Y. Lazarou∗∗


TCP, performance evaluation, stochastic model


In this paper, we first develop a new model for the slow-start phase based on the discrete evolutions of the congestion window. By examining the evolution of the congestion window size under the effects of the delayed ACK mechanism, we show that the early rounds of the congestion window evolution in the slow-start phase can be well approximated with a Fibonacci sequence. This greatly simplifies the derivation of the relationship between the number of transmitted packets and the congestion window size. Using this new slow-start phase model, we then construct a complete and more accurate TCP steady-state model. Major improvement in modeling the steady-state is further achieved by relaxing key assumptions and enhancing critical approximations that have been made in existing popular models. Finally, based on our slow-start phase and improved steady-state models, we develop a stochastic model which can more accurately predict the throughput and latency of short-lived TCP connections as a function of loss rate, round-trip time, and file size. We validate our models with simulations and compare them against existing models. The results show that our extended steady-state model is up to 75% more accurate than the model proposed in [3]. In addition, our model for the short-lived flows yields more accurate performance predictions (up to 20%) than the ones developed in [4] and [5].

Important Links:

Go Back