基于梯度的低功耗Cache划分算法

闫萌萌; 张盛兵

基于梯度的低功耗Cache划分算法

The Gradient-based Low-power Cache Partitioning Algorithm

摘要

摘要: 在 GPA （Gradient-based Partition Algorithm）算法的基础上提出一种对多核系统的共享 L3 cache 进行划分的方法 GLPCPA（Gradient-based Low-Power Cache Partition Algorithm），该方法通过监测 cache 行为的局部特性和全局特性的方法打开或关闭一些 cache 路对 cache 的容量进行动态调整，在保持性能的同时降低功耗。该方法使用循环 OBL（Robin One-Bit LRU）替换算法，其硬件开销小且性能高。仿真实验表明 GLPCPA 算法能将共享 L3 Cache 的总功耗减少20%以上，最多可以减少40%，但性能下降不超过10%。而且 GLPCPA 具有健壮性、可扩展性以及稳定性，并能够和现有的 cache 设计很好地结合在一起。

Abstract: The mechanism called GLPCPA that based on the GPA (Gradient-based Partition Algorithm) has introduced to partition the shared L3 cache in multi-processor systems. This mechanism resizes the cache dynamically through observing the local and global behavior of the cache to active or inactive a way.So the system could maintain the high performance while it is reducing the power consumption. Robin OBL (One-Bit LRU) replacement policy is used because of its low-cost and high performance.Simulation results show that GLPCPA reduces more than 20% power consumption of the L3 cache,and the best result could reach 40%.However,the performance degradation is less than 10%. GLPCPA has the robustness, scalability and stability, and it is integrated well with most existing cache designs.

HTML全文

参考文献(0)

施引文献

资源附件(0)