A 3D radiative transfer framework: XII. Many-core, vector and GPU methods

Abstract 3D detailed radiative transfer is computationally taxing, since the solution of the radiative transfer equation involves traversing the six dimensional phase space of the 3D domain. With modern supercomputers the hardware available for wallclock speedup is rapidly changing, mostly in response to requirements to minimize the cost of electrical power. Given the variety of modern computing architectures, we aim to develop and adapt algorithms for different computing architectures to improve performance on a wide variety of platforms. We implemented the main time consuming kernels for solving 3D radiative transfer problems for vastly different computing architectures using MPI, OpenMP, OpenACC and vector algorithms. Adapted algorithms lead to massively improved speed for all architectures, making extremely large model calculations easily feasible. These calculations would have previously been considered impossible or prohibitively expensive. Efficient use of modern computing devices is entirely feasible, but unfortunately requires the implementation of specialized algorithms for them.