Optimizing Loop Tiling in Computing Systems through Ensemble Machine Learning Techniques

Abstract

This work investigates the use of ensemble machine-learning algorithms to optimize loop-tiling in computing systems, with the goal of improving performance by predicting optimal tile sizes. It compares two approaches: independent training and averaging (soft voting) and an ensemble technique (hard voting) that employs models such as linear regression, ridge regression, and random forests. Experiments on an Intel Core i7-8565U CPU with several benchmark programs revealed that the hard voting Ensemble Approach beat the soft voting technique, providing more dependable and accurate predictions across a range of computing environments. The hard voting technique reduced execution time by around 87.5% for dynamic features and 89.89% for static features, whereas the soft voting approach showed an average drop of 75.45% for dynamic features and 78.13% for static characteristics. This work demonstrates the effectiveness of hard voting ensemble machine learning approaches in improving cache efficiency and total execution time, opening the way for future advances in high-performance computing settings.