Experiences using the UML profile for MARTE to stochastically model post-production interactive applications

We describe a practical approach applying the UML 2.0 standard MARTE profile to model stochastic interactive application workflows, using the PapyrusUML editor. We use the PaStep, PaCommStep, PaLogicalResource and GaCommHost MARTE stereotypes and find them sufficient for stochastic modelling with the exception of being unable to define non-standard probability distributions. We have investigated both Markovian stochastic models and discrete event simulation models, serializing UML deployment and state machine diagrams to automate model creation. The choice between using a stochastic model (e.g. PRISM Markov models) or discrete event simulation model (e.g. Monte Carlo simulations) depends on the complexity of the model, accuracy required and compute time needed. We find that PRISM models are fast to execute if the complexity is small and produce numerically accurate results. Discrete event simulation models are slower to execute but scale much better and are probably the default solution to a model of unknown complexity.