Automatic Software Performance Optimization using Genetic Algorithms

Abstract

Software programs (SW) are who ultimately drives the underlying hardware (HW). Therefore SW should make an efficient use of all available resources in order to achieve high performance levels when executing. Compilers typically perform a number of code transformations, also known as optimizations, before generating the executable file. The purpose of applying such optimizations is to modify the use of the underlying HW by the considered SW. These are generic sequences of optimizations that are known to generally perform well. However, it is clear that distinct SW programs and HW architectures offer different characteristics, and the same sequence of code transformations cannot fully exploit them for optimal performance. Therefore, there is a need of finding ad hoc optimization sequences to accurately optimize the performance of every SW/HW pair. However, this is a titanic task, taking into account the immense variety of existing HW architectures, for which every SW must be optimized. In addition, the process of optimizing SW performance for a given HW architecture is a difficult, error prone, and expensive task, requiring the hands of expert programmers with high knowledge on the considered architecture. Automatizing this process would allow the generation of a new generation of smart compilers, able to perform ad hoc transformations for the considered SW and HW for optimal performance. We propose in this work a novel approach to automatize the process of optimizing the performance of a particular SW program on a specific HW. For that, we define a novel combinatorial optimization problem, that we call SCOP (for Software Code Optimization Problem), consisting in finding a sequence of code transformations that minimizes the runtime of the resulting SW on the considered HW architecture. This extended abstract presents our work [2], recently published in IEEE Transactions on Sustainable Computing.