SIParCS 2020 - Shay Liu

Xuechang "Shay" Liu

Xuechang "Shay" Liu, Indiana University

Analysis of FastEddy Model Data on GPUs

Recorded Talk

The data analysis, which is often embarrassingly parallel and compute intensive, is traditionally done on CPUs. This project aims to prototype an accelerated FastEddy® data analysis process on single- and multi- node GPU(s) using Python scientific libraries such as Cupy and Dask. FastEddy® is a GPU-based large eddy simulation (LES) atmospheric model that solves fundamental dynamical equations and computes turbulence at a high resolution, producing large datasets that reside on GPUs. Analysis of a single data file was performed on one GPU, and then scaled to multiple data file analysis in parallel across multiple GPU nodes. Using GPU acceleration for the data analysis provided speed up in both cases. Acceptable differences between CPU- and GPU- computed variables are observed. The difference is well-constrained for simple perturbation calculations and is increased for more complicated flux calculations.

Mentors: Supreeth Suresh, Cena Miller