On-line Cache Resizing for Low-Power Microprocessors

We propose a novel scheme, called on-line cache resizing (OCR), to dynamically resize the cache and meet the size requirement of each application. At each periodic interval, the scheme gathers the cache hit-miss statistics at runtime using an extra tag array. These executing statistics serve as inputs to an analytical model of cache energy. The scheme uses energy as a primary metric to dynamically increase/decrease the number of active cache ways for the next interval. The scheme minimizes the active cache size to save energy with minimal performance loss. The simulation with SPEC 2000 benchmarks shows that OCR results in an average of 38.4% energy saving compared with fixed-size caches, with only 2.0% performance loss.