Université Paris Diderot
Efficient exploitation of modern multicore architectures relies on a hierarchical structuration of computation as well as execution concurrency. GPU's are massively multicore architectures managing several thousands of concurrent threads. This concurrence, maintained through several schedulers, is necessary to keep high performance but negatively impact predictability. In this paper, the authors first propose measures of predictability as well as CUDA tests to estimate this measure regarding warp and block scheduler for architectures from G80 to GK104. Finally, they evaluate the impact of hardware reset, underclocking and heavy synchronization on predictability.