Up to now, numerical predictions of fuel preparation in the context of aircraft engines were not feasible due to the enormous computational costs. Thanks to the steadily improving availability of High Performance Computing (HPC) resources, such simulations have come into reach. In this paper we present the transition to such large scale simulations and the associated implications. An exemplary 3D simulation of a planar prefilming airblast atomizer, which is modeled by 1.2 billion particles, helps to illustrate the workflow, necessary code modifications and possible ways to assess the simulation data. Several aspects, which are prerequisites for successfully conducting the simulation and a beneficial usage of the results, are discussed. Most importantly, a massive improvement of the serial code performance has been achieved by changing the data structure and by ensuring a cache efficient order of the particle interactions. A reduction of the memory requirement during the pre- and post-processing steps has been realized using disassembled datasets. This allows the handling and analysis of large computational domains on standard desktop computers. By applying the alpha-shape algorithm, a descriptive and memory-saving visualization of the liquid surface is possible.
HPC Predictions of Primary Atomization with SPH: Challenges and Lessons Learned
11th International SPHERIC Workshop
13.06.2016 - 16.06.2016