首页> 外文会议>IEEE/SEMI international semiconductor manufacturing science symposium : Theme: Semiconductor manufacturing >Performance modeling and measurement of parallelized code fordistributed shared memory multiprocessors
【24h】

Performance modeling and measurement of parallelized code fordistributed shared memory multiprocessors

机译:分布式共享内存多处理器并行代码的性能建模和测量

获取原文
获取原文并翻译 | 示例

摘要

This paper presents a model to evaluate the performance andnoverhead of parallelizing sequential code using compiler directives fornmultiprocessing on distributed shared memory (DSM) systems. Wenparallelized the sequential implementation of NAS benchmarks usingnnative Fortran77 compiler directives on an Origin2000, which is a DSMnsystem. We report measurement based performance of these parallelizednbenchmarks from four perspectives: efficacy of parallelization process;nscalability; parallelization overhead; and comparison withnhand-parallelized and -optimized version of the same benchmarks. Ournresults indicate that sequential programs can conveniently benparallelized for DSM systems using compiler directives but realizingnperformance gains as predicted by the performance model dependsnprimarily on minimizing architecture-specific data locality overhead
机译:本文提出了一个模型,用于评估在分布式共享内存(DSM)系统上进行多处理的编译器指令的并行化顺序代码的性能和开销。 Wen在Origin2000(DSMn系统)上使用nFortran77编译器伪指令并行执行了NAS基准的顺序实现。我们从以下四个方面报告了这些并行化基准的基于测量的性能:并行化过程的有效性;可扩展性;并行化开销;并与相同基准的手动并行化和优化版本进行比较。我们的结果表明,可以使用编译器指令为DSM系统方便地并行化顺序程序,但是要实现性能模型所预测的性能提升,主要取决于使特定于体系结构的数据局部性开销最小化

著录项

相似文献

  • 外文文献
  • 中文文献
  • 专利
获取原文

客服邮箱:kefu@zhangqiaokeyan.com

京公网安备:11010802029741号 ICP备案号:京ICP备15016152号-6 六维联合信息科技 (北京) 有限公司©版权所有
  • 客服微信

  • 服务号