A new open source framework for multiscale modeling of fibrous materials on heterogeneous supercomputers

TL;DR

MuMFiM is an open source multiscale modeling framework for fibrous materials that leverages heterogeneous supercomputers and GPU acceleration to achieve significant speedups and scalable performance.

Contribution

The paper introduces MuMFiM, a novel open source framework enabling efficient multiscale modeling of fibrous materials on supercomputers with GPU acceleration.

Findings

GPU-accelerated microscale problems achieve 1000x speedup over single RVE.

MuMFiM demonstrates nearly optimal strong and weak scaling on 128 nodes.

Application to human spine ligament modeling illustrates practical utility.

Abstract

This article presents MuMFiM, an open source application for multiscale modeling of fibrous materials on massively parallel computers. MuMFiM uses two scales to represent fibrous materials such as biological network materials (extracellular matrix, connective tissue, etc.). It is designed to make use of multiple levels of parallelism, including distributed parallelism of the macro and microscales as well as GPU accelerated data-parallelism of the microscale. Scaling results of the GPU accelerated microscale show that solving microscale problems concurrently on the GPU can lead to a 1000x speedup over the solution of a single RVE on the GPU. In addition, we show nearly optimal strong and weak scaling results of MuMFiM on up to 128 nodes of AiMOS (Rensselaer Polytechnic Institute) which is composed of IBM AC922 nodes with 6 Volta V100 GPU and 2 20 core Power 9 CPUs each. We also show how MuMFiM can be used to solve problems of interest to the broader engineering community, in particular providing an example of the facet capsule ligament (FCL) of the human spine undergoing uniaxial extension.