A Hybrid Machine Learning Method for Cross-Platform Performance Prediction of Parallel Applications

Accurately predicting parallel application performance across diverse architectures is crucial for cost-effective platform selection and optimization. The existing analytic predictive approaches pose challenges in building accurate, scalable, and comprehensive models with limited applicability by missing fine-grain interdependencies between system architecture and application. In this paper, we propose a hybrid machine learning methodology to map performance across heterogeneous computing platforms by their mutual performance ratios. The methodology allows users to predict the relative performance of a parallel application without fully executing it on systems by using a reference platform. We demonstrate that it is sufficient to observe brief partial executions of an application on the reference platform. Then, our trained models can predict the application's performance on several targeted platforms. We present our novel Ensemble Cluster Classify Regress method as a predictive kernel to maximize the models' accuracy, efficiency, scalability, and interpretability. We propose an automatic mechanism to map accordant CPU bursts in parallel applications to label data by computing the ratios. The models are automatically generated from the training dataset, bypassing the challenging and possibly error-prone procedure needed for creating analytic models. Consequently, our novel data-driven approach is handier for developers with limited performance knowledge, outperforming existing methods that require advanced hardware and analytics expertise. Our experiments across various platforms and applications demonstrate a predictive model cross-validation accuracy exceeding 98%, along with the capability to forecast execution times for unseen applications with an accuracy exceeding 94%. Integrating our innovative macrobenchmark kernels lead to a significant improvement in prediction accuracy.