Array Data Layout for the Reduction of Cache Conflicts
Naraig Manjikian, Tarek S. Abdelrahman · 2001
The performance of applications on large-scale shared-memory multiprocessors depends to a large extent on cache behavior. Cache conflicts among array elements in loop nests degrade performance and reduce the effectiveness of locality-enhancing optimizations. In this paper, we describe a new technique for reducing cache conflict misses. The technique, called cache partitioning, logically divides cache capacity into equal parts, and allocates arrays in memory such that each array maps into a separate partition in the cache. We present experimental results from KSR and SGI machines to demonstrate the effectiveness of cache partitioning in eliminating cache conflicts and in realizing the full benefit of locality-enhancing techniques for both sequential and parallel execution.