利用堆栈特征的片上末级缓存访问模式在线识别方法

很多优化处理器缓存利用效率的方法依赖于对访问请求序列的特征的探测或识别,例如,预取和绕开等。如何在线有效识别访问请求序列的特征依然是一个开放的问题。通过对典型访问模式的深入分析,发现其堆栈距离频度的分布展示出鲜明的特征。而模拟实验数据表明访问请求序列的特征具有一定的持续性和稳定性,具有检测和预测的可行性。因而提出了一种基于堆栈直方图峰值的在线识别访问模式的机制和方法,空间和时间开销都较小。对SPEC CPU2000/2006的15个程序的实验表明,所提方法均可正确识别测试程序的访问模式。

An Online Access Pattern Identification Method Based on the stack characteristic in the On-chip Last-Level-Cache

Many methods optimizing the on-chip cache utilization are dependent on the profile or identification of the access sequence characteristic, for instance, prefetcher or bypass etc. How to identify these characteristics is still an open problem. Through a detailed theoretical analysis of typical access patterns, it is shown that the frequency of stack distance has obvious features. Furthermore, according to the results of the Simics simulation, these features present persistent and stable to a certain extent, therefore, they are feasible to be identified and predicted. An online method based on the peak value of the collected stack histogram attaching to each core is provided. In addition, the storage and time overhead is small. The experimental results based on 15 benchmarks from SPEC CPU2000/2006 show that it identifies all correctly.