Coherency Behavior on DSM: A Case Study

摘要

This paper summarizes a characterization effort of coherency traffic in shared memory scientific applications. In particular, based on a systematic experimentation study of the well known Splash 2 benchmarks, two properties are detailed: locality of coherency activity within data set and within application code. Characterizing properly these properties is essential for both restructuring applications to improve coherency behavior and/or design new cost effective coherency mechanisms. Consequently, as a result of our analysis, from the exposed fact that data balance between two strongly marked behaviors and that a small fraction of application code is responsible of the majority of coherency traffic, we propose various research directions for improving performance of coherency actions.