Authors: Bogdan Florin Cornea Julien Bourgeois
Publish Date: 2012/09/14
Volume: 62, Issue: 3, Pages: 1609-1634
Abstract
Predicting distributed application performance is a constant challenge to researchers with an increased difficulty when heterogeneous systems are involved Research conducted so far is limited by application type programming language or targeted system The employed models become too complex and prediction cost increases significantly We propose dPerf a new performance prediction tool In dPerf we extended existing methods from the frameworks Rose and SimGrid New methods have also been proposed and implemented such that dPerf would perform i static code analysis and ii tracebased simulation Based on these two phases dPerf predicts the performance of C C++ and Fortran applications communicating using MPI or P2PSAP Neither one of the used frameworks was developed explicitly for performance prediction making dPerf a novel tool dPerf accuracy is validated by a sequential Laplace code and a parallel NAS benchmark For a low prediction cost and a high gain dPerf yields accurate results
Keywords: