Efficient detection of communication in multi-cores
Andreas Sandberg, Stefanos Kaxiras · 2009
Several methods have been proposed to model communication in systems with coherent caches, e.g. multi-cores, however they usually incur a large overhead on the application being analyzed. In this work we describe a low-overhead statistical communication model that is driven by a sparse sample of the memory accesses in the target application. Our model allows detection of hot-spots where coherence communication occurs between different threads in an application. Preliminary results suggest that we are able to detect most of the communication hot-spots in real applications with lower overhead than previously proposed models.