Improving performance predictions by accounting for the accuracy of composed performance models

Performance predictions on the software architecture level support the design and evaluation of component-based systems. Composable and parameterized models are used in current prediction approaches for reasoning. The actual model for an influencing factor or component of the system is the result of a trade-off between the required accuracy, prediction speed, and validation effort. Different models can have different accuracies and the overall effect on the prediction depends on their composition and used parameter values. Existing prediction approaches neglect to take into account this potentially difference in the accuracies of models. In this paper, we present an approach in which accuracy statements attached to composable performance models allow analyzing their influence on predictions without adding restrictions on the compositionality of each model. The resulting support for risk mitigation in decision making, prediction quality evaluation, as well as inaccuracy effect propagation from parts of the analyzed system on the overall prediction results are evaluated on a case study.