| dc.description.abstract | Processor speed has been increasing at a higher rate than the speed of memories over the last years. Caches were designed to mitigate this gap and, ever since, several cache management techniques have been designed to further improve performance.
Most techniques have been designed and evaluated on non-inclusive caches even though many modern processors implement either inclusive or exclusive policies. Exclusive caches benefit from a larger effective capacity, so they might become more popular when the number of cores per last-level cache increases.
This thesis aims to demonstrate that the best cache management techniques for exclusive caches do not necessarily have to be the same as for non-inclusive or inclusive caches. To assess this statement we evaluated several cache management techniques with different inclusion policies, number of cores and cache sizes.
We found that the configurations for inclusive and non-inclusive policies usually performed similarly, but for exclusive caches the best configurations were indeed different. Prefetchers impacted performance more than replacement policies, and determined which configurations were the best ones. Also, exclusive caches showed a higher speedup on multi-core.
The least recently used (LRU) replacement policy is among the best policies for any prefetcher combination in exclusive caches but is the one used as a baseline in most cache replacement policy research. Therefore, we conclude that the results in this thesis motivate further research on prefetchers and replacement policies targeted
to exclusive caches. | en |
