机群Open MP系统的设计与实现  

Design and Implementation of OpenMP on Cluster Systems

作　　者：吴少刚[1,2] 章隆兵[3] 蔡飞[3] 顾丽红[1] 唐志敏[3] 

机构地区：[1]石油大学(华东)计算机与通信工程学院 [2]中国科学院计算技术研究所北京100080 [3]中国科学院计算技术研究所

出　　处：《计算机学报》2004年第7期904-912,共9页Chinese Journal of Computers

基　　金：国家自然科学基金 ( 60 3 0 3 0 16);国家"九七三"重点基础研究发展规划项目基金"大规模科学计算研究"(G19990 3 2 80 0 );中国科学院全国优秀博士学位论文作者专项基金;中国科学院计算技术研究所领域前沿青年基金 ( 2 0 0 2 6180 7)资助

摘　　要：OpenMP以其易用性和支持增量并行的特点成为共享存储体系结构的编程标准 .目前机群系统已成为高性能计算的主流平台 ,研究机群OpenMP系统对推进并行应用的开发和普及非常有意义 .该文作者以软件DSM系统JIAJIA作为OpenMP的运行时系统 ,结合一个前端编译器OMP2JIA ,在机群系统上实现了OpenMP/JIAJIA计算环境 ,同时在提高性能方面根据机群系统特点扩展了OpenMP制导 ,优化了后端运行时库 .通过 11个OpenMP应用 ,作者比较了该计算环境和一个支持OpenMP的硬件cc NUMA系统 (SGI 2 10 0 )的性能 .结果表明 ,作者的机群OpenMP系统的 7机平均加速比为 4 .6 2 ;SGI 2 10 0系统为 4 .5 5 ,二者性能相当 .The OpenMP Application Programming Interface(API) is an emerging standard for parallel programming on shared memory multiprocessors because of its ease of use and incremental approach to the parallelization of sequential programs. At present, clusters of workstations or PCs have been becoming the mainstream platform for high performance computing. It is increasingly attractive to develop OpenMP parallel applications on cluster systems. This paper presents the design and implementation of OpenMP computing environment on clusters, which is based on the combination of software DSM and compiler technologies. In order to improve the system performance and enlarge the application range, many methods are introduced to adapt to clusters architecture, such as OpenMP directive extension, runtime library optimization, loop scheduling algorithms etc. Furthermore, eleven OpenMP programs from some standard benchmarks have been measured on a hardware cc-NUMA machine (SGI 2100) and the OpenMP/JIAJIA system on authors' commodity cluster of PCs. The experimental results show that the mean speedup of the OpenMP/JIAJIA system on seven processors is 4.62 and that of the SGI 2100 machine is 4. 55.

关 键 词：OPENMP CC-NUMA 软件DSM系统 机群计算 JIAJIA 

分 类 号：TP302[自动化与计算机技术—计算机系统结构]