The code is based on a 4th order accurate discontinuous Galerkin discretisation with implicit time-integration.
Tests on fundamental
benchmarks have show that this discretisation offers similar precision
as academic finite-difference and spectral codes, but in contrast is
applicable to generic geometries.
Furthermore the method offers excellent serial and parallel efficiency and excellent stability, and provides a visual means to assess grid resolution. Therefore it is expected that for similar computational cost this method will enable much more accurate industrial LES computations when compared to state of the art methods. This project is part of the roadmap towards proving the potential of the method for prediction of ao. noise generation and flow instabilities in turbomachinery.
The simulation has been run on 4000 cores of the BlueGene/P machine at Juelich Supercomputing Centre for a total cumulated time of 16 days, and a total of about 10 flow passages to allow evacuation of transients and statistical convergence of the turbulence. The first results look very promising, but as the project has only just finished, detailed postprocessing and comparison to experiments is still in progress.
Finally, during the
course of this project, the hybrid parallel (weak) scaling of the
time-implicit code has been enhanced and tested up to 16k cores,
resulting in an overall minimal efficiency of 94%.
This article has been published in Teratec's Newsletters of November 2011.