二维磁场重联的OpenCL研究

发布时间：2018-01-09 15:32

本文关键词：二维磁场重联的OpenCL研究　出处：《中国地质大学（北京）》2013年硕士论文　论文类型：学位论文

【摘要】：太阳大气层的爆炸，最大的破坏源自日冕物质抛射，他们高速射离太阳外层大气，并将爆炸产生的巨量等离子体向地球碰撞的轨道中射入，这就是太阳耀斑爆发。耀斑爆发的主要原因是太阳磁场中磁力线的重新排布，即磁重联。本课题是以均匀网格下二维磁场重联为依据，运用磁流体力学（MHD）数值模拟对磁场重联、太阳风暴等一系列等离子体物理现象和运动过程进行研究，使用近几年发展的高效率、高分辨率、高精度的新型数据格式时间空间守恒（CESE）格式，又称时空守恒（CESE）方法进行数值求解。随着物理研究过程的日益复杂，空间分辨率的不断提高，数据规模的不断加大，计算量大、速度慢的问题日益显著，运用传统的串行CESE方法已经（遇到了无法避免的性能瓶颈）很难满足研究者的实时需求，如何高效加速数据运算成为一个非常值得关注的问题。通用GPU的出现，使得一些计算密度高、逻辑结构简单的大规模数据并行运算得到了一种新的解决方案。GPU的发展不仅是在图形渲染、图像处理等方面，在通用计算特别是浮点并行运算方面已取得很大的突破和成功，它能在有限的面积上实现很强的处理能力和很高的存储器带宽。随着Shader Mode、CUDA、OpenCL等架构体系和相应平台技术的发展运用，GPU在性能上更是有了进一步的提升，同时发展速度也远远高于CPU，基本上每年都会产生新一代的GPU。 OpenCL（Open Computing Languange，开放式计算机语言），是一种异构资源解决方案，它提供了一个高效的、相互开发的、可自由移植的平台。本文的主要研究内容是利用GPU的强大数据并行计算能力，使用OpenCL架构和MPI消息传递模型，将CESE核心算法重构优化，提出一种多线程的基于OpenCL的并行CESE算法，使之与GPU体系完好结合，得到更好的加速比。利用OpenCL的实现步骤如下：研究CESE数值模型和主体计算步骤，分析耗时的关键程序；使用MPI和OpenCL并行编程模型，编写为GPU和CPU的高效混合并行计算；研究进程间，主机和设备的通信问题，进行调优测试。经试验得出，使用OpenCL的并行CESE算法，比传统串行算法提速了5倍左右。文章还对并行解决方案的优势和不足进行了分析，从而加深了对GPU体系架构和计算模型的理解。
[Abstract]:The biggest damage to the sun's atmosphere comes from coronal mass ejections, which shoot at high speed from the sun's outer atmosphere and fire huge amounts of plasma from the explosion into the orbit of the Earth's collision. This is the solar flare burst, the main reason for the flare burst is the rearrangement of the magnetic field in the solar magnetic field, that is, magnetic reconnection. Based on the two-dimensional magnetic reconnection under uniform grid, a series of plasma physical phenomena and motion processes, such as magnetic field reconnection and solar storm, are studied by MHD numerical simulation. Use the new data format of high efficiency, high resolution and high precision developed in recent years. With the increasing complexity of the physical research process, the spatial resolution continues to improve, the scale of the data continues to increase, and the amount of calculation is large. The problem of slow speed is becoming more and more obvious, and it is difficult to meet the real-time requirements of researchers by using the traditional serial CESE method (encountered an unavoidable performance bottleneck). How to speed up data operation efficiently becomes a very important problem. With the emergence of general GPU, some large scale data parallel operations with high computing density and simple logic structure have been developed not only in graphics rendering, but also in graphics rendering. In the field of image processing, great breakthrough and success have been made in general computing, especially in floating-point parallel computing. It can achieve very strong processing power and high memory bandwidth in a limited area. OpenCL and other architecture systems and the development of the corresponding platform technology has been further improved in performance, at the same time, the speed of development is far higher than CPU. Basically, a new generation of GPUs is produced every year. OpenCL(Open Computing language, an open computer language, is a heterogeneous resource solution that provides an efficient solution. The main research content of this paper is to utilize the powerful data parallel computing ability of GPU, using OpenCL architecture and MPI message passing model. This paper proposes a multi-thread parallel CESE algorithm based on OpenCL to optimize the reconstruction of the CESE core algorithm, which is integrated with the GPU system. The implementation steps of using OpenCL are as follows: the numerical model of CESE and the main calculation steps are studied, and the key procedures of time consuming are analyzed. Using the parallel programming model of MPI and OpenCL, the high efficient hybrid parallel computing for GPU and CPU is written. The communication between process, host and equipment is studied, and the tuning test is carried out. The parallel CESE algorithm using OpenCL is obtained by experiment. The paper also analyzes the advantages and disadvantages of parallel solutions, thus deepening the understanding of GPU architecture and computing model.
【学位授予单位】：中国地质大学（北京）
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：P182.5;TP338.6

【参考文献】