On the efficiency of fluid simulation of networks

Performance evaluation of computer networks through traditional packet-level simulation is becoming increasingly difficult as networks grow in size along different dimensions. Due to its higher level of abstraction, fluid simulation is a promising approach for evaluating large-scale network models. In this study we focus on evaluating and comparing the computational effort required for fluidand packet-level simulation. To measure the computational effort required by a simulation approach, we introduce the concept of “simulation event rate,” a measure that is both analytically tractable and adequate. We identify the fundamental factors that contribute to the simulation event rate in fluidand packet-level simulations and provide an analytical characterization of the simulation event rate for specific network models. Among such factors, we identify the “ripple effect” as a significant contributor to the computational effort required by fluid simulation. We also show that the parameter space of a given network model can be divided into different regions where one simulation technique is more efficient than the other. In particular, we consider a realistic large-scale network and demonstrate how the computational effort depends on simulation parameters. Finally, we show that flow aggregation and WFQ scheduling can effectively reduce the impact of the “ripple effect.”