Improving data cache performance by value correlation

The most important processor performance bottleneck is the ever-increasing gap between the memory and the processor speeds. To close this gap, current microprocessors employ cache hierarchies. Caches hide the memory latency by keeping the most frequently and recently used data closer to the processor. However, caches are limited in size and must be managed intelligently. In this research, we introduce a notion of Reference Value Caches (RVC). We explored the potential by tracking the blocks that contains reference values more than given threshold. Our findings show that frequent values embrace at least %50 of the loads. We also show that with only using four reference values for our implementation, we achieved speedups up to 10%