Efficient data management on 3D stacked memory for big data applications

Big data processing has been an increasingly important field which has attracted a lot of attention from academia and industry. However, it worsens the memory wall problem for processor design, which means a large performance gap between processor computation and memory access. The 3D stacked memory structure has been put forward as a promising method to relieve this problem. As non-volatile memory(NVM) become available and common nowadays, they can be fused into the 3D memory structure to provide a fast and large memory. DRAM + NVM have been designed as a novel, faster and larger memory structure. Flash is the maturest NVM material currently so that flash takes the role of NVM in our experiment. However, as DRAM has totally different characteristics from Flash, such combined structure shows bad support for big data applications. Thus, efficient data manage strategies are greatly needed. We implemented two data manage strategies(granularity strategy (GS) and Read/Write partition strategy (RWPS)) to improve performance. Our experiment results are very positive. When using 64 channels, GS can improve performance by 11.3%, and the RWPS can improve performance by 20.4%. Combined the two strategies together, the performance can be increased by 26.1%. In addition, for RWPS, it can obviously increase the write performance by 29.8% because of its novel design.