On Implementation High-Scalable CFD Solvers for Hybrid Clusters with Massively-Parallel Architectures

摘要

New approach for solving of compressible fluid dynamic problems with complex geometry on Cartesian grids is proposed. It leads to algorithmic uniformity for whole domain and structured memory accesses which are essential for effective implementations on massively-parallel architectures - GPUs. Methods used axe based on implicit scheme and LU-SGS method. Novel parallel algorithm for last one is proposed. In-depth analysis of CUDA+MPI implementation (interoperability issues, libraries tuning) scalable up to hundreds GPUs is performed.