Today on 6th May 2025, Johannes Gebert of the High-Performance Computing Center Stuttgart ( HRLS) visited us at the HPSC Lab at the University of Augsburg. He gave a guest talk as part of the Advanced Analytics and Predictive Sciences Seminar, titled "Enabling Massively Parallel Simulations of Directly Discretized CT Scans of Human Bones" (see abstract below). Furthermore, we had some interesting discussions on future computing in the HPC context and on the direction of the German and European HPC landscape. Thanks a lot for your visit, Johannes!

?

Abstract

Quantifying the patient-specific elastic behavior of human bone can improve the design of implants, for example in total hip- or knee replacements. We calculate the mechanical stiffness tensor of human bone by direct discretization of computed tomography (CT) scans with more than 18 billion voxels. The software started as a serial implementation but now requires different massively parallel parameterizations, which poses significant challenges. Many computations may run ideally parallel; one computation may be a single large-scale job or any other variant. We solved the challenge with MPI and PETSc to distribute the linear algebraic system, allowing for computing large volumes of interest (VoIs) in bone. At the same time, pre- and post-processing are still serially implemented.
We will examine the challenges in the software's legacy components and its performance bottlenecks. Furthermore, we will show the perspective of the high-performance computing center in which the user deploys the simulation and measures to, e.g., optimize for short turnaround times.