A PGAS-based Implementation for the Unstructured CFD Solver TAU
Christian Simmendinger, Jens Jägersküpper, Rui Machado, Carsten Lojewski · elib (German Aerospace Center) · 2022
Whereas most applications in the realm of the partitioned global address space make use of PGAS languages we here demonstrate an implementation on top of a PGAS-API. In order to improve the scalability of the unstructured CFD solver TAU we have implemented an asynchronous communication strategy on top of the PGAS-API of GPI. We have replaced the bulk-synchronous two-sided MPI exchange with an asynchronous, RDMA-driven, one-sided communication pattern. We also have developed an asynchronous shared memory strategy for the TAU solver. We demonstrate that the corresponding implementation not only scales one order of magnitude higher than the original MPI implementation, but that it also outperforms the hybrid OpenMP/MPI programming model.