TriBHMM: An Energy-Efficient and Latency-Aware Hybrid Main Memory

Abstract

Hybrid main memory has the advantages both of DRAM and PCM while overcome the disadvantages both of them. However, the existing hybrid main memory systems are not good at the advantages of DRAM and PCM. So we propose triple-based hybrid main memory (TriBHMM), a novel hybrid main memory system which consists of L1 DRAM and L2 DRAM and L2 PCM. How places data and migrates pages to DRAM or PCM with the page properties will significantly affect the main memory system performance. In this paper, we propose a new algorithm TriBHMM-LRU algorithm for sloving the page replacement problem in TriBHMM. TriBHM-LRU uses the locality of programm and the status of different memory types to place data to reduce PCM writes for improving the lifetime while reducing the latency of the TriBHMM. When pages migration happening, the recently updated pages are migrated from L1 DRAM to L2 DRAM. If some of these pages will be used again at soon, then the pages are replaced from L2 DRAM to L2 PCM. We evaluate the performance of TriBHM to others main memory system, and show the performance of TriBHM is more significant than others. The experiment shows that our TriBHMM is able to efficiently improve the whole performance and reduce PCM writes for improving the lifetime while reducing the latency.