首页 | 本学科首页   官方微博 | 高级检索  
     

分布共享存储系统中的数据预送技术
引用本文:谢向辉,韩承德,唐志敏.分布共享存储系统中的数据预送技术[J].计算机学报,1999,22(3):241-248.
作者姓名:谢向辉  韩承德  唐志敏
作者单位:中国科学院计算技术研究所,北京,100080
摘    要:远程数据访问的延迟已成有分布共享存储系统发展的最大障碍。它直接影响到DSM系统的效率,尤其是对用软件实现的DSM系统。为理解和分析DSM系统中的数据行为,论文提出了一种新的分布共享存储结构模型,并在此基础上提出了一种叫做“数据预送”技术,旨在从缩小数据在系统不同层次间的语义差别入手,减少DSM中的通信次数,提高对远程访问延迟的容忍力。文中对数据预送技术的原理和实现进行了描述。经对对原形系统的测试,

关 键 词:分布共享存储  数据预送  结构模型  存储一致性  延迟
修稿时间:1997年11月7日

DATA PRE-SENDING TECHNIQUE IN DISTRIBUTED SHARED MEMORY SYSTEMS
XIE Xiang-Hui,HAN Cheng-De,TANG Zhi-Min.DATA PRE-SENDING TECHNIQUE IN DISTRIBUTED SHARED MEMORY SYSTEMS[J].Chinese Journal of Computers,1999,22(3):241-248.
Authors:XIE Xiang-Hui  HAN Cheng-De  TANG Zhi-Min
Abstract:The high latency of remote data access is critical to the performance of distributed shared memory systems. It directly influences the efficiency of the DSM system, especially, software DSM system. In this paper, a DSM model is proposed for understanding and analyzing the data behavior in DSM system. Based on the model, a method called pre sending is introduced. This technique is to reduce the amount of communication and tolerate the latency of every remote access from reducing the difference of semantic on data in different system levels. The large and small data shall be processed in different way. The owner of shared data can become the master of data transmission and automatically determine how, where, and when to send more data to the destinations. The results of this method in a prototype test environment show that for some applications such as SOR, LU, FFT, and Marix multiplication the amount of communication is decreased from 49 to 20 percent, the number of access miss drops from 49 to 28 percent, and the execution time of the programs is cut down from 22 to 5 percent. The preliminary results indicate that if the characteristics of shared variable in high level can be passed into the lower level, the utilization ratio of data in shared memory space could be improved immediately.
Keywords:Distributed shared memory data  pre  sending  architecture model  memory concictency  latency    
本文献已被 CNKI 维普 万方数据 等数据库收录!
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