应用驱动的并行程序性能优化研究

Research on Application-driven Parallel Program Performance Tuning

从应用角度出发,分析、归纳各种应用中的核心计算过程,利用符合多核处理器芯片架构的并行计算模型对这些核心计算过程进行优化,得出可以被重复利用的高性能可扩展的软件库,它既可以支持新应用的高效开发,也可以保证程序性能的可扩展性。以分层并行计算模型思想为指导,从应用驱动的并行程序性能优化的角度出发,首先提出了面向多核处理器芯片体系结构的并行算法设计模型,在此基础上对并行扫描算法进行分析优化,得出新的具有良好扩展性、高性能的g-scan算法。之后深入研究13种核心计算实体之一的稀疏线性代数计算实体,应用g-scan算法设计实现了新的稀疏矩阵-向量运算算法,并将其应用于结构工程领域中广泛使用的有限元分析,大大提升了其执行效率。

Multi-core processor provides multiple threads parallel execution capability,and makes applications to have huge potential for performance improvement,but makes it enormously challenge to efficiently develop high-performance program.Meanwhile,through the old process of performance optimization,the scalability is difficult to be guarantee.From application point of view,attributing core calculation to patterns and motifs,and optimizing these motifs can produce reusable library that can support efficient develop new application,and also can guarantee the scalability of the application performance.A layered parallel computing model was used for guidance in this article.From the perspective of application-driven parallel program performance optimization,this article designed a new parallel multi-core processor computing model,which can be used in the architecture of multi-core processor chip.Based on this model,g-scan algorithm was designed which has good extendibility together with high performance after analyzing and optimizing some fundamental parallel algorithms.At the last,newly designed parallel algorithm was applied to OpenSeesSP which is a finite element software widely used in structure engineering.