Multi-fidelity Surrogate Modeling for Application/Architecture Co-design

Yiming Zhang, Aravind Neelakantan, Nalini Kumar, Chanyoung Park, Raphael T. Haftka, Nam H. Kim, and Herman Lam

The HPC community has been using abstract, representative applications and architecture models to enable faster co-design cycles. While developers often qualitatively verify the correlation of the application abstractions to the parent application, it is equally important to quantify this correlation to understand how the co-design results translate to the parent application. In this paper, we propose a multi-fidelity surrogate (MFS) approach which combines data samples of low-fidelity (LF) models (representative apps and architecture simulation) with a few samples of a high-fidelity (HF) model (parent app). The application of MFS is demonstrated using a multi-physics simulation application and its proxy-app, skeleton-app, and simulation models. Our results show that RMSE between predictions of MFS and the baseline HF models was 4%, which is significantly better than using either LF or HF data alone, demonstrating that MFS is a promising approach for predicting the parent application performance while staying within a computational budget.