卫星轨道递推的GPU集成式并行加速方法

GPU monolithic parallel acceleration method for satellite orbit prediction with SGP4/SDP4 model

为克服传统卫星轨道模型预报方法的速度瓶颈,为实现卫星在轨自主规划变轨奠定基础,利用图形处理器(GPU)并行计算方法对多卫星轨道解算进行加速,构建了轨道预报并行计算模块,成功实现了卫星轨道预报的大幅加速.为提高低计算量时解算速度,提出了集成式GPU加速方法,将简化常规摄动模型(SGP4)解算模型整体代入核函数,计算机内存仅需与GPU进行一次调用及数据交互,大大缩短调用核函数时间,较模块化GPU加速方法在中低规模计算量时速度有明显提高.本研究于两种设备上基于统一计算设备架构(CUDA)实现了集成式加速方法并进行了加速试验,在小型嵌入式开发板NIVIDA TX2设备上可实现在5 s内进行500颗星一天时间86400步的轨道预报,笔记本设备上GPU加速比也可达到中央处理器(CPU)的4.6倍,且加速后精度损失极低.实验结果表明:集成式加速方法适用于中低规模星数(总步数小于400万步)的并行解算任务,模块化加速方法适用于大规模星数(总步数大于400万步)的并行解算任务.

To overcome the speed limit of the traditional satellite orbit prediction method and lay foundation for independent orbital transfer planning of on-orbit satellites,the graphics processing unit(GPU)parallel computing method was utilized to accelerate the multi-satellite orbit calculation,and the parallel prediction module of orbit prediction was constructed,which realized the acceleration of satellite orbit prediction.In order to improve the calculation speed when the calculation amount is low,a monolithic GPU acceleration method was proposed,which substituted the simplified general perturbation version 4(SGP4)calculation model into the kernel function.The computer memory only needed to interact with the GPU once,which greatly shortened the data transmission time between the memory and the GPU.Compared with the modular GPU acceleration method,the speed for medium or low scale calculations was increased greatly.The proposed monolithic acceleration method was implemented on two devices based on compute unified device architecture(CUDA)library.On NIVIDA TX2,a small embedded development board,it could realize the orbit prediction of 500 satellites for one day in 5 s(86400 steps for each satellite),while the GPU acceleration ratio on the laptop was 4.6 times more than that of the central processing unit(CPU),and the precision loss after the acceleration was low.The experiment showed that the monolithic acceleration method was suitable for the parallel calculation of low and medium scale calculations(the number of steps is less than four million),and the modular acceleration method was suitable for the parallel calculation of large scale calculations(the number of steps is more than four million).