多GPU-CPU混合异构平台下的光谱计算优化

发布时间：2018-04-28 06:32

本文选题：数值积分 + 负载平衡　；参考：《天津大学》2016年硕士论文

【摘要】：太阳系外的所有宇宙天体的信息几乎都是通过光谱计算获得的,能够观察到的光谱包含了大量的重要信息,例如恒星的温度、年龄、金属丰度以及星系组成等。目前天文领域有一些经典的光谱计算工具包,例如XSPEC,ISIS,XSTART,APEC等等,虽然这些工具包可以精确地进行光谱计算的求解,但程序结构仍停留在传统的串行模式上,目前并没有任何一个基于并行架构的光谱计算工具。光谱计算的核心部分是数值积分,随着GPU通用性的不断提高,计算性能的稳步增长,许多经典的数值积分算法已经发展出来GPU加速版本。但是目前现有的GPU加速版本的数值积分算法都是针对大区间的高维积分,并不适用于光谱的计算,光谱计算中数值积分的特点是大量的、一维的、积分区间非常小的积分计算。因此要在多GPU-CPU的混合异构平台上加速光谱计算,不但需要解决光谱计算的核心算法向GPU的迁移,而且必须对GPU与CPU进行合理的动态任务调度,充分发挥GPU和CPU各自的优势。本文提出了一种多CPU-GPU混合异构并行方法来加速求解光谱计算。首先将计算密集型的积分部分放到了GPU上来计算,通过合理的任务粒度划分减少主机和设备之间频繁的数据拷贝来提高计算的性能。其次,提出了一种基于多个CPU与多个GPU之间的动态任务调度策略,该策略基于任务队列和共享内存的方法,该种方法相对于传统的客户机-服务器的体系结构可以很好地减少额外的通讯开销。最后,综合理论分析和实验验证了本文所提出的方法的有效性和准确性,实验显示,使用24个CPU核、3个GPU设备,相对于传统的串行APEC实现方法,可以将整体的计算加速300倍,相对于纯CPU的MPI并行方法,整体的计算加速也有22倍。此外,本文提出的方法也可以适用于单个任务计算量小,但总体任务量巨大的应用,使用本文提出的方法,在求解非电离平衡的常微分方程的计算中,相对于纯MPI的并行方法提速了15倍。
[Abstract]:Almost all the information of cosmic celestial bodies outside the solar system is obtained by spectral calculation. The observable spectra contain a lot of important information, such as star temperature, age, metal abundance and galaxy composition. At present, there are some classical spectral calculation toolkits in the field of astronomy, such as XSPECO ISISI XSTARTAPEC and so on. Although these toolkits can accurately solve the spectral calculation, the program structure is still in the traditional serial mode. At present, there is no spectral computing tool based on parallel architecture. The core part of spectral computation is numerical integration. With the improvement of the universality of GPU and the steady increase of computational performance, many classical numerical integration algorithms have been developed GPU accelerated version. But the current GPU accelerated version of the numerical integration algorithm is aimed at the high-dimensional integral of large interval and is not suitable for spectral calculation. The characteristics of numerical integration in spectral calculation are a large number of one-dimensional integral calculations with very small integral interval. Therefore, in order to accelerate spectral computation on mixed heterogeneous platforms with multiple GPU-CPU, it is necessary not only to solve the migration from the core algorithm of spectral computation to GPU, but also to schedule reasonably the dynamic tasks between GPU and CPU, so as to give full play to the respective advantages of GPU and CPU. In this paper, a multi CPU-GPU hybrid heterogeneous parallel method is proposed to accelerate the spectral computation. Firstly, the computation-intensive integral is put on the GPU to improve the computing performance by reducing the frequent data copy between the host and the device by reasonable task granularity partition. Secondly, a dynamic task scheduling strategy based on multiple CPU and multiple GPU is proposed, which is based on task queue and shared memory. Compared with the traditional client-server architecture, this method can reduce the extra communication cost. Finally, the effectiveness and accuracy of the proposed method are verified by comprehensive theoretical analysis and experiments. The experimental results show that using 24 CPU cores and 3 GPU devices can speed up the whole calculation by 300 times compared with the traditional serial APEC implementation method. Compared with the pure CPU MPI parallel method, the overall computational acceleration is 22 times. In addition, the method proposed in this paper can also be used in the calculation of ordinary differential equations of non-ionization equilibrium, which has a small amount of computation but a large amount of total task. Compared with the parallel method of pure MPI, the speed of the parallel method is increased by 15 times.
【学位授予单位】：天津大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：P144.1-39

【相似文献】