Black Box Approach for Selecting Optimization Options Using Budget-Limited Genetic Algorithms

Modern compilers present a large number of optimization options covering the many alternatives to achieving high performance for dierent kinds of applications and workloads. Selecting the optimal set of optimization options for a given application and workload becomes a real issue since optimization options do not necessarily improve performance when combined with other options. The ESTO framework described here searches the option set space using various types of genetic algorithms, ultimately determining the option set that maximizes the performance of the given application and workload. ESTO regards the compiler as a black box, specified by its external-visible optimization options. For the IBM XLC compiler, with some 60 optimization options, we achieved +13% gain over an aggresive base, using 60 iterations on average. We studied a number of search policies given a fixed iterations budget, and showed that exponentially decreasing the width of the search beam gives the best results.