Performance Optimization for Low-Leakage Caches based on Sleep-Line Access Density

As the transistor feature sizes and threshold voltages reduce, leakage energy consumption has become an inevitable issue for high-performance microprocessor designs. In order to solve the energy issue, a number of techniques to reduce the cache leakage energy have so far been proposed. However, the low-leakage caches affect negatively the processor performance due to the accesses to non-state-preserving sleep-mode lines. In this paper, we analyze the access behavior on a low-leakage cache and show a remarkable observation for the density of sleep-line accesses. Based on this observation, we propose a new cache management technique to alleviate the performance degradation caused by low-leakage caches. In our approach, a small number of cache lines which are frequently accessed in the sleep mode are forced to stay in always-active mode. Although this mode is high leakage, it saves the state. Thus, the performance overhead caused by the leakage optimization can be eliminated