A Profiling-Based Approach to Cache Partitioning of Program Data.

Cache efficiency is important to avoid unnecessary data transfers and to keep processors active. Cache partitioning, a technique to virtually divide a cache into multiple partitions, has become available in recent hardware. Cache partitioning can improve efficiency by isolating data with high temporal locality to avoid its early eviction before reuse. However, deciding on the partitioning is challenging, because it depends on the locality of reference. To facilitate the decision-making, we propose a profiling-based approach that measures locality, providing knowledge for cache partitioning without requiring manual code analysis. We present a profiling tool and confirm its benefits through experiments on Fujitsu’s A64FX processor, which supports the cache partitioning mechanism called sector cache. Our results show ways to optimize program codes to improve cache efficiency.