Parallel Princeton Ocean Model based on OpenACC

Abstract

With the development of the ocean economy, accurate forecasting using ocean models has become increasingly important. Existing parallel versions of the Princeton Ocean Model (POM) often feature complex code and limited portability. To address these issues and meet the computational demands of high-resolution ocean models while reducing program runtime, we developed an OpenACC-based parallel version of POM. Our approach migrates all computational components to the GPU using OpenACC, providing better maintainability and portability. We identified parallelizable sections and used Nsight Systems to analyze bottlenecks, reducing the transfer time efficiently between CPU and GPU. We tested the model's accuracy and performance under various simulation durations and resolutions. The results show a slight reduction in accuracy, while the speedup improved significantly, ranging from 11.75 to 45.04 with increased simulation duration and resolution. This work enhances the usability and efficiency of POM, making it more suitable for ocean forecasting and advanced research applications.