Regression-Based Multi-Model Prediction of Data Reuse Signature
Xipeng Shen, Yutao Zhong, and Chen Ding
As a locality metric, the distance of data reuses has been used in
designing compiler, architecture, and file systems. Recently, Ding
and Zhong described a method that predicts reuse distance histograms
across all inputs of a program. In this paper we extend their method
in two ways. First, we consider more than two training inputs using
regression analysis. Second, we use a method called multi-model
prediction to overcome the limitation due to small training inputs or
coarse-grain data collection. Compared to Ding and Zhong's method,
the new locality prediction can reduce about half of the prediction
error, remove 95% of space cost, and use much smaller inputs and
faster data collection in training.
Download the paper in pdf format.