Cache Line Sharing and Communication in ECP Proxy Applications.

Scientific computing codes rely on efficient parallelization to achieve performance. This parallel efficiency is reduced by factors such as communication, serialization, and data sharing. In this work, we examine interactions between OpenMP threads in the context of a Chip-multiprocessor (CMP). We first analyze cache line sharing to observe how often multiple threads are accessing the same data. We then look at producer-consumer and write-invalidation interactions between these threads. These interactions are implemented with cache coherence operations and demonstrate interference between threads. We find that none of the codes studied show prohibitive amounts of communication and many interactions between threads follow simple patterns. Our work discovers opportunities to increase parallel efficiency in the analyzed codes and provides motivating data for research into CMP design.