一种渐变的多核共享Cache划分算法  

Proposing an Effective "Gradual Runtime Shared Cache Partitioning Algorithm" for Multi-Core

作　　者：李兵哲[1,2] 朱怡安[1] 刘琪[3] 黄姝娟[1] 

机构地区：[1]西北工业大学计算机学院,陕西西安710072 [2]解放军西安政治学院,陕西西安710068 [3]中国人民解放军空军装备部科研订货部,北京100843

出　　处：《西北工业大学学报》2012年第3期461-465,共5页Journal of Northwestern Polytechnical University

基　　金：航空科学基金(20100753022);西北工业大学基础研究基金(JC20110283);航天科技创新基金(2011XR160001)资助

摘　　要：针对多核共享Cache动态划分技术存在硬件存储开销较大、划分颠簸等问题,提出了一种渐变的多核共享Cache动态划分算法。渐变算法旨在追求最优划分性能的同时,减小划分机制实现的硬件开销。渐变算法采用基于组相联Cache路的划分机制,每次划分时最多只允许1路的Cache转移。具有硬件存储开销小、划分稳定、易于实现等特点。例如,对于1MB的16路组相联L2Cache,渐变的Cache动态划分机制仅增加0.01%的硬件存储开销,与UCP方法比较,增加的硬件存储开销可减小至6.25%。The runtime shared cache partitioning is an effective way to reduce conflicting accesses for simultaneous multiple execution applications. However, the existing cache partitioning monitor always has big hardware overhead and is prone to thrash. Aiming to reduce the hardware overhead, while keeping the IPC （instruction per clock） performance optimal, we propose a gradual runtime shared cache partitioning algorithm. Section ! of the full paper explains the algorithm mentioned in the title, which we believe is effective; section 1 consists of four subsections. Its core consists of： our partitioning algorithm uses the partitioning mechanism based on group-association cache ways, each partitioning only permitting at most one way of cache transfer. Section 2 evaluates the performance of our partitioning algorithm, explaining that, for a typical L2cache configuration, our partitioning algorithm can re- duce hardware overhead to only 6. 25% ; The simulation results, given in Fig. 2, and their analysis show prelimi- narily that our partitioning algorithm can effectively eliminate partitioning thrash, while keeping the IPC perform- ance optimal.

关 键 词：多核 CACHE划分 算法 

分 类 号：TP332[自动化与计算机技术—计算机系统结构]