Performance modeling of scientific applications: scalability analysis of LAPW0

This paper presents a high-level approach for assessing the performance behavior of complex scientific applications running on a high-performance system through simulation. The proposed methodology relies on high-level descriptions of both application and system. The application is described in MetaPL, an XML-based description language, and the system is modeled and simulated by using HeSSE, an extensible distributed heterogeneous system (DHS) simulator. This modeling technique is applied to the performance analysis of a real-world scientific application (LAPWO), running on a cluster of SMP nodes. Extendibility features of both MetaPL and HeSSE were largely used, developing extensions for the MetaPL language and new components for the simulator. The paper closes with the validation of the performance model, obtained through the comparison of the predicted performance results with measurements on test runs of the application on the real system.