非一致Cache体系结构技术综述  

A Survey of the Non-Uniform Cache Architecture

作　　者：吴俊杰[1] 杨学军[1] 

机构地区：[1]并行与分布处理国防科技重点实验室,湖南长沙410073

出　　处：《计算机工程与科学》2011年第2期51-60,共10页Computer Engineering & Science

基　　金：国家自然科学基金资助项目(60621003;60873014;60633050)

摘　　要：存储墙问题使得Cache技术的研究始终非常重要。面对日益增长的片上Cache容量,线延迟逐渐成为制约Cache设计的重要因素。为了提供统一的访问延迟,传统的Cache设计方法不得不迁就离处理器最远的Cache Bank的访问时间。为此,研究人员提出了一种非一致Cache结构(NUCA),NUCA几乎成为未来处理器中大容量Cache设计的一种趋势。处理器访问NUCA时,如果在离处理器较近的Bank中发生命中,处理器的等待时间就较短;如果在离处理器较远的Bank中发生命中,处理器的等待时间就较长。本文综述了NUCA技术产生的原因、发展,以及当前最典型的NUCA系统;并且指出了对NUCA技术研究有借鉴的两种多机存储系统技术——NUMA和COMA;最后,提出了NUCA技术研究的关键问题,并给出了相应的解决思路。Because of the memory wall problem, cache memory technologies have been very important. In the future, growing wire delays will become one of the key factors restricting the design of large caches. To provide a uniform access delay, the traditional cache design methodology has to accommodate cache access time to the farthest cache bank from processors. Thus, researchers proposed a kind of NonUniform Cache Architecture （NUCA）, which is almost the inevitable outcome in intending processors. If a processor hits the data stored in the nearby banks of NUCA, the cache access time will be short. Otherwise, the access time will be long. This paper summarizes the origin, development and typical cases of NUCA. It is important that two memory technologies of multiprocessors, from which the research of NUCA may borrow ideas, are introduced. Finally, the key problems of the NUCA research as well as solutions are proposed in the authors＇ opinion.

关 键 词：非一致Cache 线延迟 局部性 多核 非一致存储访问 全Cache存储结构 

分 类 号：TP303[自动化与计算机技术—计算机系统结构]