一种降低末级高速缓存污染的软件控制插入策略  

A Software-Controlled Insertion Policy for Reducing Last Level Cache Pollution

作　　者：黄涛[1,2] 王晶[1] 管雪涛[1] 钟祺[1,2] 王克义[1] 

机构地区：[1]北京大学信息科学技术学院微处理器及系统教育部工程研究中心,北京100871 [2]北京大学深圳研究生院,广东深圳518055

出　　处：《电子学报》2012年第12期2433-2438,共6页Acta Electronica Sinica

基　　金：国家科技重大专项(No.2009ZX01029-001-002;No.2009ZX01036-001-003)

摘　　要：现有高速缓存替换算法大多无法有效识别数据的局部性特征,导致高速缓存内即将被访问到的数据可能被未来不会被访问到的数据所替换,造成高速缓存污染问题.末级高速缓存污染引发的性能损失随着处理器和存储器之间性能差距的扩大而不断增大,成为制约系统性能提升的重要瓶颈之一.本文针对末级高速缓存污染问题,在剖视分析访存密集型程序的页一级访存行为基础上,提出一种软件控制末级高速缓存插入策略.本方法通过控制和指导页一级数据插入位置,限制局部性差的数据页在末级高速缓存中的访问空间,达到降低末级高速缓存污染的目的.实验结果表明,相对于LRU和DIP策略,本文方法能够有效降低末级高速缓存失效率,提高程序性能.Most of current cache replacement algorithms could not effectively identify the data locality and may cause cache pollution problem, which means that the reusable data cache line is replaced by non-reusable one. The performance penalty caused by the last level cache pollution is becoming larger as the widening gap between processor and memory performance. It has been one of the major system performance bottlenecks. For addressing the last level cache pollution problem, this paper proposes a software- controlled last level cache insertion policy based on the page-level memory behavior profiling for the memory intensive programs. This approach can restrict the last level cache space of the pages with poor locality and reduce the last level cache misses through controlling and guiding the last level cache insertion position for the data of each page. Our approach can achieve lower last level cache miss rate and better oerformance over LRU and DIP oolicies.

关 键 词：末级高速缓存 剖视 插入策略 

分 类 号：TP302.1[自动化与计算机技术—计算机系统结构]