摘要
在分析分布共享存储系统中假共享及多写算法对系统性能影响的基础上，提出了一种简单有效的有限多写方法——直接收集（ＤｉｒｅｃｔＣｏｌｅｃｔｉｎｇ，ＤＣ）方法．ＤＣ方法兼收单写和多写算法的优点，实现了单写与多写的自动衔接，同时取消了通常多写方法中的Ｔｗｉｎ和Ｄｉｆ，简化了多写的实现．文中还给出了与传统的单写和多写方法的比较测试结果．初步测试表明，ＤＣ方法对一些面向科学计算的应用程序（如ＳＯＲ，ＦＦＴ和ＬＵ）比通常的多写方法有较好的（２０％到３３％）性能改善，对矩阵乘（ＭＡＴ）一类计算比通常的单写方法也有很好的（３６％）性能改进．
Single writer approach to false sharing in software DSM systems is simple but it can not avoid the ping pong effect of the date page thrashing, while multiple writer approach is effective to false sharing but the cost is high. This paper proposes a new approach called Direct Collecting (DC) to deal with false sharing in software DSMs. It distinguishes two kinds of false sharing in programs with lock based and barrier based forms, and handles them with single writer and multiple writer approaches respectively. It possibly discards the Twin and Diff in traditional multiple writer ways and simplifies the implementation of multiple writer in software DSM systems. The implementation of DC in CVM, which is a software DSM system based on a network of workstations, is introduced in this paper. Evaluating results show that for some applications such as SOR, FFT, and LU, DC provides a significant reduction in execution time (20% to 33%) compared to traditional multiple writer approach, and for matrix multiplication it can improve the performance compared to single writer approach by 36% on the platform.
出处
《计算机学报》
EI
CSCD
北大核心
1999年第8期821-827,共7页
Chinese Journal of Computers
基金
国家攀登计划项目
国家自然科学基金
关键词
多写方法
直接收集
DSM系统
存储系统
Memory consistency, false sharing, single /multiple writer, multiple protocol, direct collecting.
