Characterizing L2 Cache Behavior of Programs on Multi-core Processors: Regression Models and Their Transferability

Abstract

Contention for shared resources on multi-core processors has been a performance bottleneck. A solution to manage contention would be to apply knowledge about the shared resource utilization behavior of programs running on multi-core processors. In our previous work we used machine learning techniques to predict solo-run-L2-cachestress, which can be utilized as a metric to characterize such behavior of programs. In this study we investigate the transferability of trained regression models that estimate solo-run L2 cache stress of programs running on multi-core processors. Machine learning techniques were used to generate the trained regression models. Transferability of a regression model is the utility of a regression model trained on one architecture to predict the solorun L2 cache stress on another architecture. The statistical methodology to assess model transferability is discussed. We observed that regression models trained on a given L2 cache architecture are reasonably transferable to another L2 cache architecture and vice versa.