HPC Application Performance Prediction with Machine Learning on New Architectures
Dewi Yokelson, Marc Charest, Ying Wai Li
Abstract
We explore a modeling approach for scientific application performance on high-performance computer architectures using machine learning techniques. Multiple linear regression models and neural networks were evaluated for effectiveness in constructing performance models to predict the execution time of an application. Performance metrics collected during run time, together with hardware specifications, were used as input features for the performance models. Our two-step machine learning approach improved the R^2 score for performance prediction: we first performed feature selection to select a subset of metrics that are the most relevant for execution time prediction; machine learning models were then trained to predict this subset of performance metrics, which then served as the inputs for the final performance model construction in the second step. This two-step approach resulted in promising results during our case study. Regression models achieved an R^2 score up to 93% and a neural network model achieved an R^2 score of over 94% when applied to predict the execution time on an unseen computer architecture. These results are comparable to existing methods that require more upfront hardware and systems knowledge, implying that our method is more approachable for application developers without extensive performance knowledge.