基于申威众核处理器的海冰模式并行加速方法

Sea ice model is an important part of the earth system model which uses a finite difference grid and time stepping to simulate the sea ice vicissitudes. Sea ice model simulations can take from hours to days to complete due to the compute-intensive nature of the model. As a result, the size and resolution of simulations are constrained by the performance limitations of modern computing hardware. Sunway TaihuLight supercomputer is the world's first system with a peak performance greater than 100 Pflops. It brings a new opportunity for high resolution earth system model simulations but exploiting the parallelism in this architecture is not trivial. Some applications have been ported and parallelized on many core architecture platforms. But when compared with other HPC application domains, the porting of the climate models onto many-core architectures has been relatively slow. Most of efforts are based on GPU. And most early-stage efforts focused on the physics modules. Since the modules are usually compute-intensive and do not involve communications, GPU-based acceleration can generally achieve a speedup of one order of magnitude. Different from the projects mentioned above, our work focuses on CICE 5.1 and uses Sunway many-core processor as our hardware platform. Since CICE sea ice model has many relationships with other climate components, optimizing communications between Sunway many-core processors becomes an important issue we need to address. Besides, tuning the original CICE algorithms to fit the new calculation elements is another problem we need to resolve. To address these issues, in this work we implement a new parallelization of sea ice model algorithm on Sunway many-core processor. Sunway many-core processor architecture is much different from the architecture of CPU which we use now. Each Sunway many-core process contains 4 CGs (Core Groups), each CG contains 1 MPE (Management Process Element) and 64 CPEs (Computing Process Element). To exploit the massive parallelism offered by Sunway many-core processor, we propose a parallel algorithm for sea ice model which contains the ways for data transmission, data dividing as well as the calculation method. The main purpose of our work is reducing the data transfer time and exploiting the parallelism offered by the platform. The methods we used to achieve this goal is easy-understanding and we believe simply methods are more efficient in this scenario. Some tricks are also used to achieve good performance. We benchmark with CICE test dataset as well as real world data collected from COREv2 dataset to test our implementation. Two sea ice model algorithms have been ported on Sunway many-core processor and our implementation yields speedups of up to 11.6x and 9.8x over the parallel implementations running on MPEs, while demonstrating comparable performance to the basic parallel implementation on Sunway many-core processor which improve the performance up to 40%.