Porting LASG/ IAP Climate System Ocean Model to Gpus Using OpenAcc

Jinrong Jiang,Linghan Zhang,Joey Wang,Wu Wang,Xuebin Chi,Pengfei Lin,Hailong Liu,Yuzhu Wang,Huiqun Hao

doi:10.1109/access.2019.2932443

Abstract

GPUs have become important solutions for accelerating scientific applications. Most of the existing work on climate models now use code rewritten using CUDA to achieve a limited speedup. This restriction also greatly limits followup development and applications. In this paper, we designed and implemented a GPU-based acceleration of the LASG/IAP climate system ocean model (LICOM) version 2, called LICOM2-GPU. Considering the extremely large codebase of the model and the occasional need to modify the code, we implemented the model completely in OpenACC. Several accelerated methods, including OpenACC data locality optimization, loop optimization, and interprocess communication optimization are presented. Developing for GPUs using OpenACC is substantially simpler than using the CUDA port. Thus, the OpenACC is a suitable GPU programming model for complex systems, such as the earth system model and its components. Our experimental results using 4 NVIDIA K80 cards achieved up to a 6.6$ \times $ speedup compared with 4 Intel(R) Xeon(R) CPU E5-2690 v2 GPUs.

Highlights

There is a growing need for ever more accurate climate and weather simulations to be delivered at higher resolution and shorter timescales
Using LASG/Institute of Atmospheric Physics (IAP) Climate system Ocean Model (LICOM) as a representative oceanic model, we demonstrate how to parallelize an oceanic model to make it run effectively on GPU architecture
We extend the buffered arrays send_buf and recv_buf to send_buf, recv_buf to allow communications to complete all at one time

Summary

Introduction

There is a growing need for ever more accurate climate and weather simulations to be delivered at higher resolution and shorter timescales. The charges and execution times associated with these applications are two main concerns of the supercomputer users; parallel implementations that can improve application execution speeds are important [18]. Systems [22]–[25], such as SUMMIT and SIERRA, which are Top-1 and Top-2 in the TOP500 supercomputer list, respectively. Hardware acceleration, such as using GPUs, can potentially result in much shorter runtimes or in simulations with higher accuracy. Researchers have spent a great deal of energy porting GPU-compatible computer code well-suited to GPUs because GPUs have become a feasible approach to accelerating high -resolution models

Objectives

Methods

Findings

Conclusion