首页 | 本学科首页   官方微博 | 高级检索  
     

多核机群下基于神经网络的MPI运行时参数优化
引用本文:王洁,曾宇,张建林.多核机群下基于神经网络的MPI运行时参数优化[J].计算机科学,2010,37(6):229-232.
作者姓名:王洁  曾宇  张建林
作者单位:1. 首都师范大学信息工程学院,北京,100048;中国科学院计算技术研究所,北京,100090;奥地利因斯布鲁克大学分布式与并行计算研究实验室,因斯布鲁克,6020
2. 北京计算中心,北京,100005
3. 首都师范大学信息工程学院,北京,100048
基金项目:奥地利蒂罗尔州未来基金会基金 
摘    要:多核处理器的新特性给MPI应用带来了新的优化空间,其中调优MPI运行时参数被证明是优化MPI应用的有效方法.然而最优的运行时参数不仅与多核机群的体系结构有关,也决定于MPI应用的程序特征.提出并分析了一种在给定多核机群下基于人工神经网络的优化模型,用于自动为未知的MPI程序预测接近最优的运行时参数.两个不同基准的实验证明了本方法的有效性.实验证明,基于本方法得到的运行时参数所产生的加速比平均达到了实际最大加速比的95%以上.

关 键 词:多核机群  运行时参数优化  神经网络
收稿时间:2009/7/14 0:00:00
修稿时间:2009/10/23 0:00:00

MPI Runtime Parameters Tuning Based on Neural Network on Multi-core Clusters
WANG Jie,ZENG Yu,ZHANG Jian-lin.MPI Runtime Parameters Tuning Based on Neural Network on Multi-core Clusters[J].Computer Science,2010,37(6):229-232.
Authors:WANG Jie  ZENG Yu  ZHANG Jian-lin
Affiliation:(College of Information Engineering,Capital formal University,Beijing 100048,China),(Institute of Computing Technology,Chinese Academy of Sciences, Beijing 100090,China), (Beijing Computing Center, Beijing 100005,China),(University of Innsbruck Distributed and Parallel Systems Croap,Innsbruck 6020,Aastria)
Abstract:The new features of multi core add the optimization space for MPI applications, and besides tuning MPI runtime parameters is a common practice perceived to optimize the MPI application performance. However, the best configuration of the runtime parameters not only depends on the underlying architecture of a specific multi-core cluster but also on the features of MPI application. We constructed and analyzed an effective tuning model bases on artificial neural network to automatically predict the near-optimal configuration of runtime parameters for any unseen input programs under the current multi-core cluster. Experimental results from two different benchmarks were presented to show effectiveness of our approach. We observed that the speedup gained by the predicted runtime parameters can averagely achieve 95% of the speedup gained by the best parameters configuration.
Keywords:MPI
本文献已被 万方数据 等数据库收录!
点击此处可从《计算机科学》下载全文
设为首页 | 免责声明 | 关于勤云 | 加入收藏

Copyright©北京勤云科技发展有限公司  京ICP备09084417号