| As Emerging applications desire larger memory capacity and more bandwidth,the memory scale problem and the memory wall problem will become ever more critical and serious.To alleviate these problems,researchers in both academia and industry swarm into emerging memory techniques.Although techniques such as Process-in-Memory(PIM)and Non-Violate-Memory(NVM)can provide more capacity or bandwidth,they need a lot revise on the software stack.Sometimes,emerging techniques develop slowly from creation to commerce.But,if we look back to the traditional DDR DRAM,we will find that the refresh mechanism is one of the most factors caused memory bandwidth and memory space degradation.In the nearly future,DRAM Refresh will impact latency,bandwidth,capacity and power greatly.On the other hand,emerging applications such as data analytic applications and graph processing applications are sensitive to memory read performance.They desire more bandwidth and lesser latency.To fully unlash main memory access performance,this paper enhances conventional DRAM refresh mechanism and hardware architecture.Through analyzing memory access patterns among various applications,this paper presents a set of main memory access optimization methods.Focusing on memory refresh itself,this paper proposes DR(Device Refresh and Data Recovery)refresh mechanism and DR DRAM hardware module.Moreover,by conducting traditional refresh and DR refresh in a single memory system,we present Hybrid Refresh Main Memory(HRMM).HRMM is a software-hardware collaboration system which can designate refresh scheme on different physic memory spaces.The simulation result shows DR refresh and HRMM can improve memory access performance on most applications.For memory read reference,DR DRAM can achieve at most 11% performance on both bandwidth and latency.While HRMM can improve around 30% latency performances on average at SPEC CPU2006 benchmarks. |
