Don't Just Go with the Flow: Cautionary Tales of Fluid Flow Approximation

Fluid flow approximation allows efficient analysis of large scale PEPA models. Given a model, this method outputs how the mean, variance, and any other moment of the model's stochastic behaviour evolves as a function of time. We investigate whether the method's results, i.e. moments of the behaviour, are sufficient to capture system's actual dynamics. We ran a series of experiments on a client-server model. For some parametrizations of the model, the model's behaviour can accurately be characterized by the fluid flow approximations of its moments. However, the experiments show that for some other parametrizations, these moments are not sufficient to capture the model's behaviour, highlighting a pitfall of relying only on the results of fluid flow analysis. The results suggest that the sufficiency of the fluid flow method for the analysis of a model depends on the model's concrete parametrization. They also make it clear that the existing criteria for deciding on the sufficiency of the fluid flow method are not robust