基于大规模异构众核系统的代数多重网格解法器算法研究

The Architecture of heterogeneous many-core has been an important technologic route to build the large scale parallel computer system, there have been several supercomputers based on heterogeneous accelerators around the world. Therefore, how to design scaling parallel algorithms based on large scale heterogeneous systems is an urgent task. Considering the importance of algebraic multigrid method in the field of high performance computing, the algorithms of the AMG solver based on the large scale heterogeneous many-core system would be investigated in the present project, as the followings: 1) Aiming at the sparse matrix vector multiplication kernel in AMG, the executing performance law of the matrix model and architecture features is studyed, in order to fully set free the computing power of the acceleration components.2)Considering the diversity of the architecture level parallelism in heterogeneous systems, The method of multil level parallel optimal mapping and granularity selection in AMG algorithm is investigated, which can effectively utilize the parallelism in large scale heterogeneous systems. According to the differences between heteogeneous systems and general components, high performance AMG algorithm achieved by the collaborative optimization on the general and accelerate components is investigated to improve the scalability of the parallel program runing in the full scale system. Through the research on this project, we will achieve the high scalability and high performance AMG parallel algorithm and application program on the large-scale heterogeneous many-core system, and promote the development of high performance computing

异构多核体系结构已经成为构建大规模并行计算机系统的一个重要的技术路线，国内外已经有多台基于异构加速部件的千万亿次超级计算机，如何设计基于大规模异构系统的高扩展并行算法是迫切需要研究的课题。考虑到代数多重网格（AMG）方法在高性能计算领域的重要性，本项目拟研究基于大规模异构众核系统的AMG 解法器算法：1）针对AMG中的稀疏矩阵向量乘核心，研究矩阵模式和体系结构特征在执行性能上存在的规律，以充分发挥加速部件的计算能力；2）考虑到异构系统中体系结构级并行的多样性，研究AMG 算法中多级并行的最优映射和粒度选择的方法，以有效利用大规模异构系统中的并行性；3）针对异构系统中通用和加速部件在处理能力上的差异，研究通用和加速部件协同调优的高性能AMG 算法，以提高并行程序在全系统规模的扩展性。通过本项目的研究，将在大规模异构多核系统上实现高扩展、高性能的AMG 并行算法和应用程序，促进高性能计算的发展