提升稀疏目录缓存一致性系统性能的方法

发布时间：2018-05-19 23:26

本文选题：ccNUMA系统 + 稀疏目录　；参考：《计算机研究与发展》2014年09期

【摘要】：稀疏目录技术在缓存一致性非一致存储访问(cache coherent non-uniform memory access,ccNUMA)系统中有广泛应用.但是,稀疏目录技术的一个主要缺陷在于目录项替换在目录热点存在的情况下会严重降低系统的性能.针对此问题,研究如何提升稀疏目录性能.首先,从定性和定量的角度对主流的稀疏目录替换算法进行了分析,并且发现最为原始的最久未使用(least-recent-used,LRU)算法的性能实际上要优于一些近期提出的目录替换算法,如最少共享者(least-sharer-count,LSC)算法.其次,将victim cache的思想应用到稀疏目录上,提出了victim目录(victim directory),该技术在主稀疏目录模块上加入了一个小的全相联二级目录存储.最后,提出了选择性victim目录(selective victim directory)技术,使得victim目录选择性地仅存储有用的目录项,从而进一步减少了目录替换的数量.选择性victim目录通过向所有节点的cache发出探查消息(probe messages),从而选择性地存储较为有用的目录项.实验证明,选择性victim目录取得了比简单victim目录更好的性能;而且在仅增加了1KB左右的硬件开销的代价下,选择性victim目录节省了35.7%的程序运行时间.
[Abstract]:Sparse directory technology is widely used in cache consistent storage access coherent non-uniform memory access nu MAs. However, one of the main defects of sparse directory technology is that the substitution of directory items in the presence of directory hotspots will seriously reduce the performance of the system. To solve this problem, we study how to improve the performance of sparse directories. First of all, the main sparse directory replacement algorithms are analyzed qualitatively and quantitatively, and it is found that the performance of the most original algorithm, which has not been used for a long time, is actually better than that of some recently proposed directory replacement algorithms. For example, least-share-count LSC algorithm. Secondly, the idea of victim cache is applied to the sparse directory, and the victim directory is proposed, which adds a small fully associative secondary directory storage to the main sparse directory module. Finally, the selective victim directory selective victim directorytechnique is proposed, which makes the victim directory selectively store only useful directory entries, thus further reducing the number of directory replacements. The selective victim directory selectively stores useful directory entries by sending a probe message to the cache of all nodes. Experiments show that the selective victim directory achieves better performance than the simple victim directory, and that the selective victim directory saves 35.7% of the program running time at the cost of only increasing the hardware overhead about 1KB.
【作者单位】：中国科学院计算技术研究所高性能计算机研究中心;中国科学院大学;
【基金】：国家“九七三”重点基础研究发展计划基金项目(2011CB302501) 国家自然科学基金项目(61173007,61100013,61100015,61202059)
【分类号】：TP333

【相似文献】