MARTE Models Based System Performance Assessment

With the wide use of computer systems, system performance, or non-functional requirements of the systems, gets more and more attentions. Traditionally, the performance of a system is evaluated after it is implemented. Some problems found at this stage may need great effort to resolve, because they are caused by improper system architecture designed in an early development stage. In this paper, we present model-based methods to evaluate the performance of systems. Our methods can find the potential performance bottlenecks during modeling stage and provide advices for more efficient designs.We use MARTE models to describe the behavior and relative performance parameters of systems. The performance criteria we consider are the system reliability and the resource utilization.UML(Unified Modeling Language) can be used to describe the structure and behavior characteristics of the systems from different perspectives and is widely used. It is a de facto standard of modeling language. UML Profile for MARTE is an extension of UML in the domain of real-time and embedded systems. MARTE profile is used to model the non-functional properties. The MARTE model we consider includes a use case diagram, a deployment diagram and a set of activity diagrams.The system reliability is the probability of failure-free operation in the given condition and period. A MARTE model is transformed into a network of Markov Decision Process, which is then analyzed by the model checking tool PRISM. We obtain the estimation of system reliability by analyzing the resulting model. We use stress testing to find out the peak value of the resource utilization under different environment configurations. We extract all execution paths from a MARTE model and then obtain the maximum amount of the resource used and then generate the schedules as test cases. We find maximum values through two methods: constraint programming and genetic algorithms. The former is an exact method and the latter is a heuristic.We have implemented our methods and have carried out them on several models. Our case studies show that our methods are feasible for some kinds of applications.