An Experimental Study of Global and Local Search Algorithms in Empirical Performance Tuning

The increasing complexity, heterogeneity, and rapid evolution of modern computer architectures present obstacles for achieving high performance of scientific codes on different machines. Empirical performance tuning is a viable approach to obtain highperforming code variants based on their measured performance on the target machine. In previous work, we formulated the search for the best code variant as a numerical optimization problem. From a mathematical optimization standpoint, two classes of algorithms are available to tackle this problem: global and local algorithms. In this paper, we investigate the effectiveness of some global and local search algorithms for empirical performance tuning. We present an experimental study of these algorithms on a number of problems from the recently introduced SPAPT test suite. We show that local search algorithms are particularly attractive for empirical performance tuning, where finding high-preforming code variants in a short computation time is crucial.