The Chair of Fluid Mechanics has developed a two-layer parallelization approach capable of dealing with clusters of heterogeneous compute nodes, accommodating multicores as well as GPUs. Hardware-specific adaptations employ OpenACC or OpenMP and are confined to the lower, sub-CPU level, while the upper layer is virtually device independent. Using the MPI message passing interface it bridges arbitrarily heterogeneous nodes and fits seamlessly into the computational fluid dynamics framework (paper).

This work is accompanied by the development of new algorithms which attain superior computational complexity combined with adjustable granularity, which yields better mapping and higher performance of solvers on different accelerator types (paper).