基于CUDA的并行雷达拼图算法研究  

Study of Algorithms for Radar Networking Based on CUDA

作　　者：韩丰 高嵩 薛峰 李月安 HAN Feng;GAO Song;XUE Feng;LI Yue’an(National Meteorological Centre,Beijing 100081)

机构地区：[1]国家气象中心,北京100081

出　　处：《气象》2023年第10期1246-1253,共8页Meteorological Monthly

基　　金：国家重点研发计划(2022YFC3002904);国家气象中心气象现代化专项(QXXDH202231);中国气象局重点创新团队(CMA2022ZD07)共同资助。

摘　　要：雷达组网拼图算法是强对流天气短时临近预报系统(Severe Weather Automatic Nowcasting,SWAN)的重要基础方法之一。提高拼图算法的效率,不仅可以提升现有SWAN临近算法序列的时效性,也能更好地应用高分辨率雷达数据,具有重要的实际意义。采用中央处理器(central processing unit,CPU)和图形处理器(graphics processing unit,GPU)混合架构设计并行雷达拼图算法,其中CPU负责雷达数据的解析和调度GPU并行模块,GPU负责大规模数据的并行计算。通过分析计算统一设备架构(compute unified device architecture,CUDA)算法的并行开销和拼图算法的特点,提出并实现了GPU内存管理优化和数据交换流程优化方案,提高了组网拼图算法的效率。对比试验结果表明,基于CUDA的GPU并行拼图算法和SWAN中30线程并行的CPU算法相比,在全国1 km和500 m分辨率的拼图任务上,加速比分别达到3.52和6.82。综上,基于CUDA的并行拼图算法不仅可以提高SWAN短时临近算法序列的时效性,也为更高分辨率雷达资料的拼图提供了技术支持。Radar networking method is one of the most important methods in Severe Weather Automatic Nowcasting(SWAN)radar applications.Improving the efficiency of the radar networking method can improve the timeliness of the short-term algorithm sequence,and also can make good use of high-resolution radar data,which has important practical significance.With the development of GPU(graphics processing unit)general computing interface,GPU has also been regarded as a powerful computing resource rather than display device for rendering and images.Therefore,this study proposes a new GPU parallel radar networking method based on CUDA(compute unified device architecture).The method is designed in a hybrid architecture of CPU(central processing unit)and GPU,in which the CPU is for the decoding of radar data and scheduling the GPU parallel modules,and the GPU is for the parallel computing of large-scale data.By analyzing the parallel overhead of the CUDA and the characteristics of the radar networking method,a scheme of GPU memory management optimization and data exchange process simplification is proposed and implemented,which effectively improves the efficiency of the method.The comparative test results show that,compared with the CPU parallel algorithm in SWAN,the GPU parallel networking method based on CUDA achieves a speed-up ratio of 3.52 and 6.82,respectively,on the national puzzle tasks of 1 km and 500 m resolution.To sum up,the parallel networking method based on CUDA can not only improve the timeliness of the short-term nowcasting methods sequence,but also provide technical support for the puzzle of higher resolution radar data.

关 键 词：组网拼图 GPU并行 计算统一设备架构(CUDA) 强对流天气短时临近预报系统(SWAN) 

分 类 号：TP319[自动化与计算机技术—计算机软件与理论]