Selective inline expansion for improvement of multi grain parallelism
Selective inline expansion for improvement of multi grain parallelism
复制标题
用于改进多粒并行性的选择性内联扩展
DOI：
--
复制
发表时间：
2004
期刊：
影响因子：
--
通讯作者：
H. Kasahara
中科院分区：
文献类型：
--
作者：
J. Shirako;Kouhei Nagasawa;K. Ishizaka;M. Obata;H. Kasahara
本文提出了一种提高多粒度并行性的选择性过程内联方案，该方案除了利用循环并行性外，还分层利用循环、子程序和基本块之间的粗粒度任务并行性和基本块内部语句之间的近细粒度并行性。使用所提出的方案，不同层（嵌套级别）之间的并行性可以被利用。在16路IBM pSeries 690 SMP服务器上使用SPEC 95 FP基准测试程序中的103.su2cor、107.mgrid和125.turb3d进行测试，结果表明，该方案的多粒度并行处理性能比IBM XL Fortran编译器提高了3.65 ~ 5.34倍，比传统的多粒度并行处理提高了1.03 ~ 1.47倍。
This paper proposes a selective procedure inlining scheme to improve a multi-grain parallelism, which hierarchically exploits the coarse grain task parallelism among loops, subroutines and basic blocks and near fine grain parallelism among statements inside a basic block in addition to the loop parallelism. Using the proposed scheme, the parallelism among different layers(nested levels) can be exploited. In the evaluation using 103.su2cor, 107.mgrid and 125.turb3d in SPEC95FP benchmarks on 16 way IBM pSeries690 SMP server, the multi-grain parallel processing with the proposed scheme gave us 3.65 to 5.34 times speedups against IBM XL Fortran compiler and 1.03 to 1.47 times speedups against conventional multi-grain parallelization.