Computational Hydrodynamics: How Portable and Scalable Are Heterogeneous Programming Paradigms?

摘要

New many-core era applications at the interface of mathematics and computer science adopt modern parallel programming paradigms and expose parallelism through proper algorithms. We present new performance results for a novel massively parallel free surface wave model suitable for advanced simulations in arbitrary size Numerical Wave Tanks.The application has already been studied in a series of works (see References) and is demonstrated to exhibit excellent performance portability and scalability using hybrid MPI-OpenCL/CUDA. Furthermore, it can be executed on arbitrary heterogeneous multi-device system sizes from desktops to large HPC systems such as superclusters and in the cloud utilizing heterogeneous devices like multi-core CPUs, GPUs, and Xeon Phi coprocessors.The numerical efficiency is evaluated on heterogeneous devices like multi-core CPUs, GPUs and Xeon Phi coprocessors to test the performance with respect to both portability and scalability. This study contributes to investigating the potential of code acceleration for reducing turn-around times of industrial CFD applications on heterogeneous hardware.