Optimization Parameter Selection by Means of Limited Execution and Genetic Algorithms

Abstract

In this article, we present an optimization parameter selection framework based on limited execution and genetic algorithm. In this framework, the parameter selection problem is transformed into a combinatorial minimization problem. We first perform reduction transformation to reduce the program’s runtime while maintaining its relative performance as regard to different parameter vectors. Then we search for the near-optimal optimization parameter vector based on the reduced program’s real execution time. The search engine is guided by genetic algorithm, which converges to near-optimal solution quickly. The reduction transformation reduces the time to evaluate the quality of each parameter vector. And the genetic algorithm reduces the number of candidate parameter vectors evaluated. This makes execution-driven optimization parameter search practical. Our experiments for 5 scientific applications on 3 different platforms suggest that our approach can find excellent optimization parameters in reasonable time. It obtains highly architecture specific optimizations in an architecture independent manner and can solve nearly all combined optimization parameter selection problems.