Exploiting parallelism for CPU and GPU linear solvers on chemistry for atmospheric models

Atmospheric models are a representation of dynamical, physical, chemical, dynamical, and radiative processes in the atmosphere [1]. Most of the algorithms used in these models have been developed decades ago. With the increasing demand on supercomputing performance, these models are gradually being updated to new performance techniques and hardware options. These improvements typically focus on high time consuming components. One of these components corresponds to the resolution of chemical processes, which can take up to 90% of the total time execution [2]. Recent studies reported a relevant speedup by translating a chemical module to GPUs. For example, the Kinetic Preprocessor library (KPP) from the EMAC Earth system model achieves up to 20x speedup compared to the single-thread CPU version [3]. Other studies developed algorithms specifically for GPU computing, like the Runge-Kutta-Chebyshev for moderately stiff kinetics [4]. This algorithm achieves up to 59x speedup, which is notably greater in comparison to the KPP GPU. However, the KPP is able to solve higher levels of stiffness and its solving method has been proven multiple times on the CPU. Also, the KPP GPU strategy uses the same solving algorithm on both versions, which requires less programming effort