The most fruitful use of a performance model is to study deep properties of the system, and hypothetical situations that might lead to improved configurations or designs. This requires executing experiments on the model, which evaluate systematic changes. Parameter estimation methods also exploit search in a parameter space to fit a model to performance data. Estimation, sensitivity and optimization experiments can require hundreds of evaluations, and the efficiency of the analytic model solver may become an issue. Analytic models usually provide fast solutions (compared to simulations) but repetitive solutions for near-neighbour models offer opportunities for further reducing the effort. This work describes an experiment driver for a layered queueing solver which provides a factor of two improvement. It also raises the issue of domain-specific languages for model experiments, versus general languages with suitable libraries. Copyright

Additional Metadata
Keywords Efficient solution, Experiment control, Modeling languages, Performance analysis, Sensitivity
Persistent URL dx.doi.org/10.4108/ICST.VALUETOOLS2009.7807
Conference 4th International Conference on Performance Evaluation Methodologies and Tools, VALUETOOLS 2009
Citation
Mroz, M. (Martin), & Franks, G. (2009). A performance experiment system supporting fast mapping of system issues. Presented at the 4th International Conference on Performance Evaluation Methodologies and Tools, VALUETOOLS 2009. doi:10.4108/ICST.VALUETOOLS2009.7807