A performance model for finegrain accesses in UPC

UPC’s implicit communication and fine-grain program-ming style make application performance modeling a chal-lenging task. The correspondence between remote refer-ences and communication events depends on the internals of the compiler and runtime system. This correspondence is often hidden from application developers. Aggressive optimizations allowed by the relaxed memory consistency model further blur this correspondence by transforming code structure. A modeling approach based on UPC plat-form benchmarking and code analysis is proposed. This ap-proach abstracts a UPC platform according to its potential to apply a few common optimizations, then divides remote references in the application code into groups, based on a dependence analysis, that are amenable to each optimiza-tion. Each group is associated with a cost, obtained via benchmarking each potential optimization. The aggregated cost of these groups is the predicted cost of the applica-tion. Three simple UPC applications modeled using this ap-proach usually yielded performance predictions within 15 percent of actual running times.