Towards realistic HPC models of the neuromuscular system

TL;DR

This paper enhances a biophysics-based neuromuscular simulation framework to achieve better parallel scalability and visualization, enabling large-scale, high-resolution simulations on supercomputers with significant runtime improvements.

Contribution

It introduces a scalable, parallelized muscle simulation framework with an integrated advanced visualization environment, improving performance and usability for large-scale biomedical simulations.

Findings

Runtime improved by up to 2.6 times

Scales efficiently on up to 768 cores

Supports large, detailed neuromuscular models

Abstract

Realistic simulations of detailed, biophysics-based, multi-scale models require very high resolution and, thus, large-scale compute facilities. Existing simulation environments, especially for biomedical applications, are designed to allow for a high flexibility and generality in model development. Flexibility and model development, however, are often a limiting factor for large-scale simulations. Therefore, new models are typically tested and run on small-scale compute facilities. By using a detailed biophysics-based, chemo-electromechanical skeletal muscle model and the international open-source software library OpenCMISS as an example, we present an approach to upgrade an existing muscle simulation framework from a moderately parallel version towards a massively parallel one that scales both in terms of problem size and in terms of the number of parallel processes. For this purpose, we investigate different modeling, algorithmic and implementational aspects. We present improvements addressing both numerical and parallel scalability. In addition, our approach includes a novel visualization environment, which is based on the MegaMol environment capable of handling large amounts of simulated data. It offers a platform for fast visualization prototyping, distributed rendering, and advanced visualization techniques. We present results of a variety of scaling studies at the Tier-1 supercomputer HazelHen at the High Performance Computing Center Stuttgart (HLRS). We improve the overall runtime by a factor of up to 2.6 and achieved good scalability on up to 768 cores, where the previous implementation used only 4 cores.