SC17 Denver, CO

P06: Large Scale FFT-Based Stress-Strain Simulations with Irregular Domain Decomposition

Authors: Anuva Kulkarni (Carnegie Mellon University), Franz Franchetti (Carnegie Mellon University), Jelena Kovacevic (Carnegie Mellon University)

Abstract: Large-scale stress-strain simulations involving parallel Fast Fourier Transforms (FFTs) suffer from high memory requirements and high communication overhead. We propose an irregular domain decomposition method to reduce the memory requirement of an FFT-based stress-strain simulation algorithm for composite materials, the Moulinec-Suquet Composite (MSC) - Basic Scheme. This algorithm uses Green’s functions to solve a partial differential equation. FFTs are used for convolution of large 3-D tensor fields with the Green’s function.

In this preliminary work, we propose a modified algorithm, the MSC-Alternate Scheme, to show that processing the composite with smaller, local FFTs on irregular domains (grains in the material’s microstructure) can reduce memory usage without adversely impacting accuracy of the result. Additionally, data models can reduce communication by compressing the data in the domains before the communication step. Our poster presents our proof-of-concept results and charts out the path towards a GPU implementation.

Award: Best Poster Finalist (BP): no

Poster: pdf
Two-page extended abstract: pdf

Poster Index