基于显卡加速的氧化铝团簇的结构预测

发布时间：2018-06-12 04:44

本文选题：显卡计算 + 势能函数　；参考：《安徽大学》2015年硕士论文

【摘要】：显卡全称为显示接口卡(GPU),自电子计算机被发明以来,一直是计算机的基本配置之一。在计算机发展的初期阶段,显卡的性能较弱,仅在主机中承担显示图形的作用；随着近半个世纪的发展,显卡的性能突飞猛进,并凭借其特殊的架构,使得当代显卡的浮点、并行计算能力数十倍乃至于上百倍于中央处理器(CPU)。在传统的科学计算中,由于单个CPU的计算能力较弱,导致计算速度缓慢,而多个CPU并行计算虽然能提高计算速度,大型计算机集群又过于耗费资源。显卡正好可以补足传统计算的短板,因此,显卡在未来的科学计算中将处于十分重要的地位。本文中,我们采用显卡加速的策略,计算氧化铝团簇的能量和梯度,统计了显卡相比传统中央处理器的加速比,并将显卡加速运用到氧化铝团簇结构的全局优化当中。主要内容如下：1、基于显卡加速的能量和梯度的计算对于氧化铝团簇,我们把它当作一个刚性模型,它的势能函数共有四项,将势能函数项放入显卡中进行计算,最终得到了很高的加速比。得出结论如下：设计了三种不同的加速策略,一维操作、块操作和二维操作,以及两种精度,单精度和双精度。在单精度条件下,一维操作、块操作和二维操作峰值加速比分别为220、240和77；双精度条件下,一维操作、块操作和二维操作峰值加速比分别为103、107和35。对于小尺寸团簇,二维操作的加速比具有绝对优势,一维操作和块操作的加速比都非常小；对于中等尺寸团簇,二维操作无法计算,块操作加速比开始显著增加,远远高于一维操作,并更早的达到了峰值加速比；对于大尺寸团簇,一维操作与块操作加速比达到饱和,块操作峰值加速比略微大于一维操作。2、显卡加速在氧化铝团簇电子结构优化中的应用采用显卡加速结合遗传算法预测了(Al2O3)n(n=1-15)团簇的结构并分析了它们的结构特点。由于团簇尺寸较小以及对精度的高要求,我们采用的是二维的双精度操作,峰值加速比为35倍左右。得出结论如下：在n=1-3时,全局最优结构分别为笼状,笼状和茶壶形状；当n=4和5时,都是高度对称的笼状结构；当n=6时,结构倾向于无序；当n=7-9时,也是高度对称的大型笼状结构；当n=10时,结构又倾向于无序。先前文献报道过的(Al2O3)n (n=1-10)的结构在我们的方法中均得到了重现。此外我们还预测出n=11-15的结构,这些团簇结构先前从未报道过。当n=11时,最优结构和n=7结构较为相似；当n=12时,结构是无序的,但在特定角度某些原子可以重叠；当n=13和14时,最优结构对称性分别为Cs和D2;当n=15时,结构是无序的。
[Abstract]:Graphics card is called display interface card (GPU). Since the invention of electronic computer, it has been one of the basic configuration of computer. In the early stage of computer development, the performance of graphics card is relatively weak, it only plays the role of displaying graphics in the host computer. With the development of nearly half a century, the performance of graphics card develops by leaps and bounds, and by virtue of its special structure, it makes the floating point of contemporary graphics card. Parallel computing power is tens or even hundreds of times higher than the CPU. In the traditional scientific computation, the computing speed is slow because of the weak computing power of a single CPU, while the parallel computing of multiple CPUs can improve the computing speed, and the large computer cluster consumes too much resources. Graphics card can complement the traditional computing board, so graphics card will play an important role in the future scientific calculation. In this paper, we calculate the energy and gradient of the alumina cluster by using the strategy of graphics card acceleration, and calculate the speedup ratio of the display card compared with the traditional central processor, and apply the graphics card acceleration to the global optimization of the alumina cluster structure. The main contents are as follows: 1. For the alumina cluster, we regard it as a rigid model based on the calculation of the accelerating energy and gradient of the graphics card. Its potential energy function has four terms, and the potential energy function is put into the graphics card to calculate. A high speedup was finally achieved. The conclusions are as follows: three different acceleration strategies, one dimensional operation, one block operation and two dimensional operation, as well as two kinds of precision, single precision and double precision, are designed. The peak speedup ratios of one-dimensional operation, block operation and two-dimensional operation are 220240 and 77, respectively, and the peak speedup ratios of one-dimensional operation, block operation and two-dimensional operation are 103107 and 35 respectively under the condition of single precision. For small clusters, the speedup ratio of two-dimensional operation is absolutely superior, the speedup ratio of one-dimensional operation and block operation is very small, for medium-size cluster, two-dimensional operation can not be calculated, and block operation speedup begins to increase significantly. It is much higher than one-dimensional operation and reaches peak speedup earlier. For large clusters, the ratio of one-dimensional operation to block operation reaches saturation. The peak speedup ratio of block operation is slightly higher than that of one-dimensional operation .2.The application of graphics card acceleration in the optimization of electronic structure of alumina cluster is discussed. The structure of Al _ 2O _ 3N _ (nn) _ (1-15) cluster is predicted and their structural characteristics are analyzed by means of video card acceleration combined with genetic algorithm (GA). Due to the small size of clusters and the high requirement of precision, we use two dimensional double precision operation, and the peak speedup ratio is about 35 times. The conclusions are as follows: the global optimum structure is cage, cage and teapot shape when nn 1-3, highly symmetrical cage structure when nu 4 and 5, disorderly when n = 6, and n 7-9 when n = 7-9, n = 3, the global optimum structure is cage shape, cage shape and teapot shape, respectively; It is also a large cage structure with high symmetry; when n = 10:00, the structure tends to be disordered. The structure of Al _ 2O _ 3N ~ (1-10), reported in previous literature, has been reproduced in our method. In addition, we have predicted the structure of nc-15, which has never been reported before. When n = 11:00, the optimal structure is more similar to that of nni7; when n = 12:00, the structure is disordered, but some atoms can overlap at a particular angle; when n = 13 and 14:00, the symmetry of the optimal structure is Cs and D _ 2, respectively; when n = 14:00, the structure is disordered.
【学位授予单位】：安徽大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TQ133.1

【相似文献】