A Data Driven Biomechanical Workflow on NEC Vulcan

Johannes Gebert (HLRS)

Abstract

HLRS research on numerical evaluation in biomechanics, introduced by Ralf Schneider, is characterized by precisely denoted goals, but evolving questions. Transparent algorithms, a modular software stack, availability of compute resources and reproducibility are mandatory. Computed tomography data sets undergo manipulation, e.g., with filters, spatial registration, finite element analysis and they require visualization. Heterogeneous clusters provide the desired range of CPUs, GPUs, accelerators and memory configurations. While the demand on hardware varies with progress and interdependent programs, data sets are a well known and static element of the research. This leads to a format describing the raw data for all intermediate results with keywords in a meta-file. Auxiliary files with the same base name, but different suffixes, are considered a unique data set. It may contain information (i.e., results) in various forms, for example binary or PureDat. All steps are reproducible by the preceding state of the raw data and the suffix files, even if they already contain follow-up information. The talk highlights our software stack used on NEC Vulcan and shows our data driven workflow. Reproducibility is ensured by the data sets from Ruf et al. and the publication bound releases of our source codes, as proposed by Karthik Ram.

<= Go back