Efficient Auto-Tuning of Parallel Programs with Interdependent Tuning Parameters via Auto-Tuning Framework (ATF)

Auto-tuning is a popular approach to program optimization: it automatically finds good configurations of program’s so-called tuning parameters whose values are crucial for achieving high performance particular parallel architecture and characteristics input/output data. We present three new contributions the Auto-Tuning Framework (ATF), which enable key advantage in general-purpose auto-tuning : efficiently optimizing programs have interdependencies among them. make following main phases auto-tuning: (1) ATF generates search space interdependent with by exploiting parameter constraints; (2) stores such spaces memory, based on novel chain-of-trees structure; (3) explores these faster, employing multi-dimensional strategy its representation. Our experiments demonstrate that, compared state-of-the-art, frameworks, substantially improves generating, storing, exploring parameters, thereby enabling an efficient overall process important applications from domains, including stencil computations, linear algebra routines, quantum chemistry data mining algorithms.