Accelerating HPC Applications Using Machine Learning-based Approximation

Historically, numerical analysis has formed the backbone of supercomputing for decades by applying mathematical models of first-principle physics to simulate the behavior of systems from subatomic to a galactic scale. Recently, scientists have begun experimenting with a new approach to understanding complex systems using machine learning (ML) predictive models, primarily Deep Neural Networks (DNN), trained by the virtually unlimited data sets produced from traditional analysis and direct observation. Early results indicate that these “synthesis models” combining ML and traditional simulation, can improve accuracy, accelerate time to solution and significantly reduce costs.In this thesis, we study how to enhance the usability of machine learning models to accelerate HPC applications. We first study an application, the Eulerian fluid simulation. The Eulerian fluid simulation is an important HPC application. The current methods that accelerate the fluid simulation with Neural Networks (NNs) lack flexibility and generalization. In this application, we tackle the above limitation and aim to enhance the applicability of NNs in the Eulerian fluid simulation. We introduce Smart-fluidnet, a framework that automates model generation and application. Given an existing NN as input, Smart-fluidnet generates multiple NNs before the simulation to meet the execution time and simulation quality requirement. During the simulation, Smart-fluidnet dynamically switches the NNs to make best efforts to reach the user’s requirement on simulation quality. Evaluating with 20,480 input problems, we show that Smart-fluidnet achieves 1.46x and 590x speedup comparing with a state-of-the-art NN model and the original fluid simulation respectively on an NVIDIA Titan X Pascal GPU, while providing better simulation quality than the state-of-the-art model