Using Machine Learning to Explore Huge Parameter Spaces for High End Computing Applications: Tools and Examples

With constantly increasing software and architectural complexities and machine scales, creating accurate performance models for applications with large parameter spaces becomes extremely challenging. Approaches using analytic models are difficult and time consuming to construct, limited in scope, and can fail to capture full system and application complexity. To retain these details and at the same time reduce the user’s effort in constructing models, we automatically build models using execution samples. We use multilayer neural networks due to their ability to represent arbitrary functions. Our approach predicts many aspects of performance, and can capture full system complexity. In this tutorial, we focus on SMG 2000, a semicoarsening multigrid solver. We have gathered sample data for this code across large, multi-dimensional parameter spaces, and our models can predict performance with high accuracy. On a Xeon-based cluster, 50% and 75% of predictions achieve error rates of 3.4% and 5.8%, respectively, when training on 1K points. Furthermore, the tools are not limited to HEC applications, but can be used for general predictions (e.g., they can be used for power/performance/temperature predictions in exploring intractably large architectural design spaces). Here we describe our approach, demonstrating how to use our tools to generate application performance predictions.