Experimental Evaluation of Statistical Model Checking Methods for Probabilistic Timing Analysis of Multiprocessor Systems

International audienceTiming prediction of complex parallel data flowapplications on multiprocessor systems represents a difficult taskdue to complex interferences observed by running software onplatforms shared resources. In this domain, classical analyticalor simulation-based approaches demonstrate scalability issues todeliver fast yet accurate predictions. In this work, we presentan experimental evaluation of new simulation-based statisticalmethods for timing analysis of multiprocessor systems. We adopta measurement-based approach for the creation of probabilisticsystem-level models of the studied systems. Efficiency of statisticalmethods is evaluated for platforms with different levels ofcomplexity from the point of view of shared resources. Wecompare our approach against measurement and traditionalsimulation methods on two case-studies from the computer visiondomain: a Sobel filter and a JPEG decoder. We show that theaccuracy and execution time of our simulation approach has goodpotential for fast yet accurate design space exploration