Beyond fluid models: Modelling TCP mice in IP networks under non-stationary random traffic

Fluid models of IP networks are based on a set of ordinary differential equations, that provide an abstract deterministic description of the average network dynamics. When IP networks operate close to saturation, fluid models were proved to provide reliable performance estimates. Instead, when the network load is well below saturation, standard fluid models lead to wrong performance predictions, since all buffers are forecasted to be always empty, so that the packet discard probability is predicted to be zero. These incorrect predictions are due to the fact that fluid models, being deterministic in nature, do not account for the random traffic variations that may induce temporary congestion of some network elements. In this paper we discuss three different approaches to describe random traffic variations in fluid models, considering randomness at both the flow and packet levels. With these approaches, fluid models allow reliable results to be obtained also in the case of IP networks that operate well below their saturation load. Numerical results are presented to prove the accuracy and the versatility of the proposed approaches, considering both stationary and non-stationary traffic regimes.