Redistributing and Optimizing High-Resolution Ocean Model POP2 to Million Sunway Cores

Abstract

The high-resolution CESM is widely applied in climate simulations, while a simulation speed of 5.0 simulated years per day has traditionally been considered the minimum necessary for long-term simulations. When Sunway TaihuLight supercomputer was open, the atmosphere model CAM5, one of CESM’s major component models, was already ported. But the ocean model POP2, another major component model, has not been fully done yet as known. In this paper, the high-resolution POP2 coupled in CESM is fully ported to Shenwei many-core infrastructure. Although many methods accumulated, there are still some new challenges when it comes to POP2. If just simply translated, its performance may not be well to support long-term simulations. In order to achieve high performance, three stages are adopted. Firstly, the original POP2 is ported with athread interface and fine-grained optimized to Shenwei many-core. Secondly, the grid decomposition is redesigned, and a new slave-core partition method is proposed to solve the problem that some two-dimension array related kernels after athreaded may be insignificant or even false speedup under large scale processes. Then many two-dimension array related kernels in POP2 are effectively redistributed to slave-cores. Lastly, some case-oriented skills are intensively utilized as necessary supplements. Some experiments show that the simulation speed of the finally optimized POP2 in high-resolution CESM G-compset is over 5.5 simulated years per day under 18,300 processes with 1,189,500 cores, compared with 1.43 simulated years per day of the original version, and its speed-up ratio is still over 3.8.