Large-scale acceleration algorithms for a deep convective physical parameterization scheme on GPU

Early warning of geological hazards requires monitoring extreme weather conditions, such as heavy rainfall. Atmospheric circulation models are used for weather forecasting and climate simulation. As a critical physical process in atmospheric circulation models, the Zhang-McFarlane (ZM) deep convective physical parameterization scheme involves computationally intensive calculations that significantly impact the overall operational efficiency of the model. However, many of these calculations are independent and can be computed in parallel. Therefore, this paper proposes a GPU-based acceleration algorithm for the ZM scheme. Based on the computation characteristics of the ZM scheme, we propose its one-demensional and two-demensional acceleration algorithms based on GPU. These algorithms are implemented using CUDA C and compared against a single Kunpeng-920 (Dual Socket) CPU core and the OpenMP version on multi-core CPUs. In the absence of I/O transmission, the proposed algorithm achieves a speedup of 413.6×. Experimental results demonstrate the significant acceleration effect of the proposed algorithms and methods. It is of great significance for the development of deep convective parameterization schemes and their further generalization in climate models. Additionally, we propose a performance optimization method utilizing the CUDA streaming technology to improve data transmission efficiency between CPU and GPU. In the presence of I/O transmission, the proposed algorithm achieves a speedup of 350.1× on A100 GPU.