Collecting and exploiting cache-reuse metrics

The increasing gap of processor and main memory performance underlines the need for cache-optimizations, especially on memory-intensive applications. Tools which are able to localize code regions with high cache miss ratio seem to be appropriate for access optimizations. However, a programmer often does not know what to do with the collected information. We try to improve this situation by providing cache reuse metrics which are supposed to give more precise hints on how to optimize memory access behavior. We enhanced the cache simulator Callgrind to give metrics on temporal and spatial cache utilization for a given memory block, relating this information to the code line where the block was loaded into cache. We show what is needed for hardware-supported measurement for such metrics, and give example code where the collected information directly points to optimization possibilities.