基于众核架构的并行SPH算法的研究与实现

发布时间：2018-05-05 13:47

本文选题：光滑粒子动力学 + 流体运动　；参考：《上海交通大学》2013年硕士论文

【摘要】：流体运动现象是周围环境的一个重要组成部分，对该现象在计算机上快速的模拟仿真,是航空航天、海洋船舶等许多行业领域所迫切需要的。目前对混合介质的流体运动模拟较为理想的方法是光滑粒子动力学(SPH)方法，它是一种无网格的纯Lagrange方法，以粒子理论为基础，通过计算整合各个粒子的状态来得到整个流场运动的规律。但是该方法在实际应用时需要用到大量的计算，以至于模拟速度缓慢。至今为止，已有一些在SPH并行方法上的研究成果，但大部分实现模型过于简化或者仅将部分算法移植到GPU上实现，未能充分利用GPU的计算能力。本文首先对SPH方法的串行程序代码做了详细分析，了解该算法的具体流程。同时对算法中的两个关键步骤（相邻粒子对构建以及粒子属性值计算）做了深入研究，归纳整合出了具体的执行过程。然后通过对串行代码的热点分析，找到程序性能瓶颈所在,寻找方法加以改善。为充分利用GPU众核架构强大的计算能力，作者对SPH的并行算法的实现主要做了如下几点贡献：将SPH算法的整体过程，包括相邻粒子对构建以及粒子物理属性值计算，移植到GPU众核架构上实现在相邻粒子对存储过程中，利用对存储空间下标的原子操作，代替使用临界区锁定整个存储空间，实现存储器的写并行，加快模拟速度。在粒子属性值计算过程中，利用共享存储器作为中间变量减少对全局存储器的访问次数，，加速计算过程。利用不同线程维度，以及对GPU资源使用的分析找到最合适的并行粒度，取得最佳性能。在本文实验中，我们使用CUDA作为GPU的编程语言。在性能测试对比实验中，分别使用NVIDIA Tesla C2050以及NVIDIA Tesla K20分别作为GPU硬件设备，进行并行算法性能测试，使用Intel至强CPU W3520进行串行算法的性能测试。将并行程序的性能与串行程序作对比得到加，使用C2050GPU单卡加速可以取得相对于串行程序8倍的加速比，而使用K20GPU显卡可以取得20倍的加速比。
[Abstract]:The phenomenon of fluid motion is an important part of the surrounding environment. The rapid simulation of the phenomenon on the computer is urgently needed in many fields, such as aerospace, ocean and ship. At present, the ideal method for fluid motion simulation in mixed media is smooth particle dynamics (SPH) method, which is a meshless pure Lagrange method based on particle theory. The motion of the whole flow field is obtained by calculating the state of the integrated particles. However, this method needs a lot of calculation in practical application, so the simulation speed is slow. Up to now, there have been some research results on SPH parallel methods, but most of the implementation models are too simplified or only some algorithms are transplanted to GPU, so they can not make full use of the computing power of GPU. In this paper, the serial program code of SPH method is analyzed in detail, and the flow of the algorithm is understood. At the same time, the two key steps in the algorithm (the construction of adjacent particle pairs and the calculation of particle attribute values) are deeply studied, and the concrete execution process is summarized and integrated. Then, by analyzing the hot spot of serial code, we find the bottleneck of program performance and find ways to improve it. In order to make full use of the powerful computing power of the GPU multi-core architecture, the author has made the following contributions to the implementation of the parallel algorithm of SPH: The whole process of SPH algorithm, including the construction of adjacent particle pairs and the calculation of particle physical attribute values, is transplanted to the GPU multi-kernel architecture. In the storage procedure of adjacent particle pairs, the atomic operation of subscript to storage space is used instead of locking the entire storage space in the critical region to realize the memory write parallelism and speed up the simulation. In the process of calculating the particle attribute value, the shared memory is used as the intermediate variable to reduce the number of visits to the global memory, and the calculation process is accelerated. Using different thread dimensions and analyzing the usage of GPU resources, the most suitable parallel granularity is found and the best performance is obtained. In our experiment, we use CUDA as the programming language of GPU. In the performance test contrast experiment, NVIDIA Tesla C2050 and NVIDIA Tesla K20 are used as GPU hardware devices to test the performance of parallel algorithm, and Intel to strong CPU W3520 are used to test the performance of serial algorithm. By comparing the performance of parallel program with serial program, the speedup ratio of 8 times that of serial program can be obtained by using C2050GPU single card acceleration, and 20 times speed ratio can be obtained by using K20GPU graphics card.
【学位授予单位】：上海交通大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP338.6

【相似文献】