Accelerating HPC Applications Using Machine Learning-based Approximation

Speaker Bio

Wenqian Dong is an assistant professor of KFSICS department of Florida International University. She studied her Ph.D. in Computer Science and Engineering at the University of California, Merced. Her research focuses on high performance computing system (HPC). She is particularly interested in using machine learning techniques to improve performance of scientific applications. During her Ph.D., she has published a set of papers in various top HPC conferences, including two published in the prestigious Supercomputing Conference (SC), the flagship conference in the HPC field. She also received the UCM Bobcat Fellowship in 2018 and 2020. Wenqian was a research intern in the HP labs and is a research intern in the Pacific Northwest National Lab. Her work on power grid simulation was highlighted at the DOE science news and PNNL media. She is looking for self-motivated PhD starting from 2023 Spring/Fall, or research interns starting from anytime.

Description

Neural network-based approximation, aiming to use neural networks to approximate some computation in applications or generate meaningful initial states of applications to shorten application execution time, is promising in high performance computing (HPC). However, how to apply this method to an HPC application to accommodate various input problems of the application to ensure high simulation quality remains to be studied. In this talk, we will present our work on applying neural network-based approximation to HPC applications, namely the Eulerian fluid simulation and AC-OPF power grid simulation. In the Eulerian fluid simulation, we generate multiple neural networks before the simulation and introduce a runtime system that dynamically switches the neural networks to make the best efforts to reach the user’s requirement on simulation quality. We show that our method achieves 1.46x and 590x speedup, compared with a state-of-the-art neural network model and the numerical fluid simulation respectively, while providing better simulation quality than the state-of-the-art model. In the AC-OPF power grid simulation, we generate multitask-learning (MTL) neural network (NN) models to predict the initial values of variables critical to the convergence of the power grid problem. We also incorporating physical constraints into the MTL model to improve model accuracy interpretability. These techniques bring 2.60× speedup on average (up to 3.28×) computed over 10,000 problems, without losing solution optimality.

This event will be webcast live. Join via Mediasite live streaming.

The seminar will also have a Zoom session that will be monitored for Q&A.