Reducing search space of auto-tuners using parallel patterns

Auto-tuning is indispensable to achieve best performance of parallel applications, as manual tuning is extremely labor intensive and error-prone. Search-based auto-tuners offer a systematic way to find performance optimums, and existing approaches provide promising results. However, they suffer from large search spaces. In this paper we propose the idea to reduce the search space using parameterized parallel patterns. We introduce an approach to exploit context information from Master/Worker and Pipeline patterns before applying common search algorithms. The approach enables a more efficient search and is suitable for parallel applications in general. In addition, we present an implementation concept and a corresponding prototype for pattern-based tuning. The approach and the prototype have been successfully evaluated in two large case studies. Due to the significantly reduced search space a common hill climbing algorithm and a random sampling strategy require on average 54% less tuning iterations, while even achieving a better accuracy in most cases.