摘要
本文提出了一种直接映象式高速缓存块冲突预测方法 ,即借助于高缓块的最近替换行为动态预测冲突发生 .基于该方法 ,我们设计了一种高缓结构 冲突预测高缓 ,主体为一个直接映象式高缓和一个较小的全相联高缓 ,利用冲突预测表实行高缓块的动态分配 .应用于片上数据高缓的SPEC95仿真结果表明 ,与 16kB直接映象式高缓相比 ,(8+1)kB冲突预测高缓命中率平均可提高 12 .2 % ,与类似结构的高缓 (如NTS高缓、PCS高缓等 )相比降低了硬件开销 ,简化了控制机构 ,易于实现 。
A method to predict block conflicts in direct-mapped caches is proposed, which is based on the block replacement behavior. Accordingly, a cache scheme named Conflict Prediction(CP) Cache is presented. Its architecture contains a direct-mapped cache and a small fully-associative cache, and a conflict prediction table(CPT) which dynamically decides the allocation of fetched blocks from next memory hierarchy. SPEC95 simulation results show that the performance of CP cache is always better than the traditional direct-mapped cache with twice the size of CP cache. With the same block size of 32 bytes, the average improvement of miss ratio of (8+1)kB CP cache with 8-entry CPT is about 12.2% over the traditional 16 kB direct-mapped cache. By comparison with other similar architectures such as NTS cache and PCS cache, CP cache not only requires less hardware tradeoff and simpler control but also improves hit ratio and bus traffic.
出处
《电子学报》
EI
CAS
CSCD
北大核心
2003年第5期724-727,共4页
Acta Electronica Sinica
关键词
高缓性能
命中率
冲突预测
Computer simulation
Data flow analysis
Management
Numerical analysis