Towards a Multi-Level Cache Performance Model for 3D Stencil Computation

• DOI: 10.1016/j.procs.2011.04.235
• Publication Date: 2011-05-22T00:50:01Z
• Authors: Ràul de la Cruz; Mauricio Araya-Polo
• Keywords: Stencil; Speedup; Code (set theory)

It is crucial to optimize stencil computations since they are the core (and most computational demanding segment) of many Scientific Computing applications, therefore reducing overall execution time. This not a simple task, actually it lengthy and tedious. because large number optimizations combinations test, which might consume days computing time, process tedious due slightly different versions code implement. Alternatively, models that predict performance can be built without any actual execution, thus cumbersome optimization task. Previous works have proposed cache misses time for specific optimizations. Furthermore, them been designed 2D datasets or sizes only suit low order numerical schemes. We propose flexible accurate model wide range up high schemes, captures behavior 3D using platform parameters. The has tested in group representative hardware architectures, realistic dataset sizes. Our predicts successfully times misses. However, predictions accuracy depends on platform, instance x86 architectures prediction errors ranges between 1-20%. Therefore, reliable help speed computation process. To end, other techniques added this model, essentially providing framework covers state-of-the-art.