Accurate Cache Miss Ratio Prediction from Time Distance Histograms

This paper discovers an important phenomenon, the "shadowing" effect, of time distance traces which are the basis of time-distance-histogram-based cache models. The definition and basic properties of the "shadowing" effect are presented in the paper. Although several analytical cache models have been proposed recently to predict cache miss ratio from time distance histograms, none of them considers the "shadowing" effect. This paper shows that the "shadowing" effect has a significant impact on the accuracy of those time-distance-histogram-based cache models. To demonstrate how those cache models' accuracy can be improved by considering the "shadowing" effect, a modified StatStack model, the "Shadowing"-effect Aware StatStack (SASS) model, is proposed in this paper. The accuracy of the SASS model is evaluated through two kinds of address traces, synthesized ideal address traces and address traces extracted from SPEC2006 benchmarks. Experimental results show that compared with the StatStack model, the SASS model can improve the accuracy of miss ratio prediction significantly.