Perofmrance Evaluation of HPCN Applications

摘要

While designing a parallel application, programmers have to identify a paralleliation strategy. In many cases, depending on the programming language adopted, on the characteristics of the numerical algorithms to be exploited, and on the topology of the arget architecture, several possibilities srise for distributing data and tasks onto the processors. Effective a priori predictions of the performance attained by means of different streategies are very difficult (see, for example, (1) (2)). A posteriorievaluations are based on measurements collected at run-time on the real system. These studies results in accurate performance figures and can be used to diagnose efficiency losses (see, for example, (3), (5)). One of the major issue in evaluting, on a posteriori basis, the performance of parallel applications is based on trace file analysis. This approach reveals particulary useful in that it addresses the correspondences between the structure of the application and the characteristics of the target architecture.