EXPLOITING TEMPORAL LOCALITY FOR ENERGY EFFICIENT MEMORY MANAGEMENT

Abstract

Memory is becoming one of the major power consumers in computing systems. Therefore, energy efficient memory management is essential. Modern memory systems employ sleep states for energy saving. To utilize this feature, existing research activities have concentrated on increasing spatial locality to deactivate as many blocks as possible. However, they did not count the unexpected activation of memory blocks due to cache eviction of deactivated tasks. In this paper, we suggest a software-based power state management scheme for memory, which exploits temporal locality to relieve the energy loss from the unexpected activation of memory blocks from cache eviction. The suggested scheme SW-NAP makes a memory block remain deactivated during a certain tick, which has no cache miss over the block. The evaluation shows that SW-NAP is 50% better than PAVM, which is an existing software scheme, and worse than PMU, which is another approach based on the specialized hardware by 20%. We also suggest task scheduling policies that increase the effectiveness of SW-NAP and they saved up to 7% additional energy.