作者: Bernd Mohr , Allen Malony , Sameer Shende , Felix Wolf
DOI:
关键词: Event (computing) 、 TRACE (psycholinguistics) 、 Programming language 、 Parallel computing 、 Performance tool 、 Source code 、 Interface (Java) 、 Rewriting 、 Directive 、 Computer science 、 Code (cryptography)
摘要: Abstract. In this article we propose a “stan-dard” performance tool interface for OpenMP, sim-ilar in spirit to the MPI profiling its in-tent define clear and portable API that makesOpenMP execution events visible li-braries. When used together with profil-ing interface, it also allows tools be built hy-brid applications mix shared distributedmemory programming. We describe an instrumen-tation approach based on OpenMP directive rewrit-ing generates calls passescontext information (e.g., source code locations) ina efficient way. Our proposed OpenMPperformance further user functions andarbitrary regions marked perfor-mance measurement controlled using new pro-posed directives. The transfor-mations are implemented source-to-source translation called Opari. have usedit integrate Tau analysis frame-work [13] automatic event trace analyzerexpert [17, 18] interface. Together, these show aportable robust solution analy-sis of hybrid is possible.