Parametric Performance Contracts: Non-Markovian Loop Modelling and an Experimental Evaluation

Even with todays hardware improvements, performance problems are still common in many software systems. An approach to tackle this problem for component-based software architectures is to predict the performance during early development stages by combining performance specifications of prefabricated components. Many existing methods in the area of component-based performance prediction neglect several influence factors on the performance of a component. In this paper, we present a method to calculate the performance of component services while including influences of external services and different usages. We use stochatic regular expressions with non-Markovian loop iterations to model the abstract control flow of a software component and probability mass functions to specify the time consumption of internal and external services in a fine grain way. An experimental evaluation is reported comparing results of the approach with measurements on a component-based webserver. The evaluation yields that using measured data as inputs, our approach can predict the mean response time of a service with less than 2 percent deviation from measurements taken when executing the service in our scenarios.