Abstract: Based on the analysis of benchmark program’s memory access behavior, this paper proposes stride and pointer based prefetching scheme which can capture regular data access pattern and pointer access pattern, and implements a prefetcher between L2 cache and memory using a global history buffer. Full system simulation shows that this prefetching scheme can improve the commercial benchmark’s performance by 14%, and improve the scientific benchmark’s performance by 34.5%.

Key words: chip multiprocessor, stride prefetching, pointer prefetching

摘要： 在对大量程序访存行为进行分析的基础上，提出基于步长和指针的预取方法。能捕获规整的数据访问模式和指针访问模式。在L2 cache和内存之间采用全局历史缓存实现该预取方法。全系统模拟结果表明，该预取方法对商业应用测试程序的性能平均提高14%，对科学计算测试程序的性能平均提高34.5%。

关键词: 片上多处理器, 步长预取, 指针预取

CLC Number: 

• TP311