多核系统中共享cache的动态划分  

Dynamic Partition of Shared Cache in Multi-Core System

作　　者：杨磊[1,2] 时磊[1,2] 张铁军[1] 王东辉[1] 

机构地区：[1]中国科学院声学研究所数字系统集成实验室,北京100190 [2]中国科学院研究生院,北京100190

出　　处：《微电子学与计算机》2009年第5期56-59,共4页Microelectronics & Computer

摘　　要：引入建立单独的监控电路的方法对cache进行最优的动态划分,有效解决了线程间cache访问互相污染的问题.同时基于cache访问局部性的差异,提出了Group划分的方法,在增加很少硬件开销的条件下进一步提高了性能.通过运行SPEC CPU2000测试程序,动态划分将整体IPC提高8.7%,而Group划分方法可以提高11.8%.With the CMP/SMT technology development,a number of parallel threads are executed in processor.These threads contest shared cache,which causes great loss in system performance.This paper put forward a way of establishing a separate circuit monitoring to partition cache dynamically.The cache pollution of multi-threads can be resolved.Based on differences in local cache access,GroupPart method is proposed to get further improvement which takes little hardware increase.By running SPEC CPU2000 benchmarks,dynamic partition of cache could improve IPC by 8.7%,and GroupPart improve 11.8%.

关 键 词：高速缓存 栈距离剖析 动态划分 

分 类 号：TN4[电子电信—微电子学与固体电子学]