Using linear regression to characterize data coherency traffic

Abstract

This paper proposes an algorithm to dynamically characterize the coherency traffic occurring in DSM architectures. This algorithm strongly relies on linear regressions to isolate lines among the traffic. The main features are a dynamic algorithm, robustness toward the noise and production of fine characterizations of the traffic. At the end the regularity is summarized in a set of regression lines found and some statistics are provided. The driving idea is while scientific code is widely considered as highly structured, a precise quantification may expose the underlying regularity due the code data structures. We describe the algorithm step by step and give results that show the relevance of the approach.