基于不规则性的并行计算方法  

A parallel computing method for irregular work

作　　者：杨鑫[1] 许端清[2] 杨冰[2] 

机构地区：[1]大连理工大学计算机学院,辽宁大连116023 [2]浙江大学计算机学院,浙江杭州310027

出　　处：《浙江大学学报（工学版）》2013年第11期2057-2064,共8页Journal of Zhejiang University：Engineering Science

基　　金：国家科技支撑计划资助项目(2012BAH03F02);国家自然科学基金资助项目(61300084);中国博士后科学基金资助项目(2012M520625);大连理工大学基础科研经费(DUT12RC(3)63)

摘　　要：为了有效使用异构多核架构强大的并行计算能力,根据硬件架构的特点重新组织数据并合理调度任务的执行是非常有必要的.提出一个基于不规则性的并行计算方法,是一个融合数据并行、任务并行、管道并行的多重并行计算方法,特别适合具有动态特征执行行为和不规则数据结构的复杂算法,能够在程序运行时根据存储局部性原则和单指令多数据流(SIMD)操作机制对任务执行进行基于优先级的动态调度和数据管理,能够最大限度地有效使用CPU和GPU的硬件计算资源和存储资源.实验结果表明,该方法能够提高图形并行绘制算法关于动态执行过程和不规则数据结构构造和维护的性能.In order to effectively use the powerful computing provided by the heterogeneous multi-core ar- chitecture, re-organize the data and a reasonable schedule for the execution of tasks is very necessary, ac- cording to the characteristics of the hardware architecture. This paper presents a parallel computing meth- od for irregular work, the method is an multiple parallel integration of data parallelism, task parallelism, pipeline parallelism, is particularly suitable for the implementation of the work with dynamic behavior and complex irregular data structures algorithms, and run the program according to the storage locality and SIMD character, using priority-based dynamic scheduling and data management, to maximize the efficient use of CPU and GPU hardware computing resources and storage resources. The experiments results show that the approach can improve the parallel rendering algorithm performance for the dynamic execution and irregular data structures construction and maintenance.

关 键 词：编程模型 不规则性 并行 CUDA 

分 类 号：TP312[自动化与计算机技术—计算机软件与理论]