面向智能终端的非易失SLC/MLC混合内存能耗与耐久性优化研究

With the development of edge computing, a lot of data-intensive applications have been provided on edge smart devices. Compared with the traditional DRAM, the multiple-level cell (MLC) in emerging non-volatile memory (NVM) offers even higher density and larger capacity. However, this comes with the price of longer access energy and weaker endurance compared to its single-level cell (SLC) counterpart. Therefore, considering the limited resources in smart devices, how to effectively take advantage of the high-density MLC and the low-energy SLC in NVM becomes an important issue. Therefore, in order to meet the requirements of low energy and reliability on smart devices, this project focuses on exploring the system software optimizing techniques of SLC/MLC hybrid memory. At first, combining with the static and dynamic memory access features on the application-specific smart devices, energy-aware data allocation and memory management optimizing techniques are studied. Then, this project fully explores software compiler and lifetime management of memory page to improve the endurance of NVM. Finally, we will implement a prototype system based on real development board for evaluating our proposed techniques. The proposed techniques in this project can effectively lead to the promotion of emerging NVM in smart devices, and provide the technical support for the development of edge computing.

随着智能边缘计算的快速发展，网络边缘的智能终端将支持更多的数据密集型应用。新型非易失性存储器多层存储单元（MLC）在存储密度上的研究进展，为突破终端存储瓶颈带来了机遇。然而，相比其低密度的单层存储单元（SLC），MLC存在读写能耗高与耐久性低的缺陷。基于智能终端诸多特性，如何优化系统软件以充分发挥SLC/MLC在访问能耗与存储密度的互补优势，并降低耐久性低的负面影响仍是亟待解决的关键问题。为此，本项目围绕智能终端低能耗与耐久的关键目标，着力于非易失SLC/MLC混合内存的系统软件优化研究，具体包括：探索智能终端面向特定应用的静态与动态访存行为，分别在软件编译层与系统管理层，提出能耗感知的数据分布与页面分配优化方案；考虑SLC/MLC在耐久方面的差异化，设计多粒度的耐久性优化方案；最后构建原型验证平台，实现关键技术验证。本项目研究对智能终端和智能边缘计算的发展具有一定的理论意义和应用价值。

本项目围绕智能终端低能耗与耐久的关键目标，着力于非易失SLC/MLC混合内存的系统软件优化研究，具体包括：在软件编译层的数据布局优化方面，提出状态可变的非易失性内存的数据分配能耗优化方案、基于可变SPM的实时任务调度与数据分配方案、基于多核可变SPM的数据分配方案、基于保留状态的高速缓存管理方案；在操作系统层的页面管理优化方面，提出基于混合SLC/MLC内存架构的页面转换策略、基于3D TLC存储单元的数据分配与重编程技术、虚拟超级页机制；在耐久性优化方面，提出状态可变的非易失性内存的耐久性优化方案与MLC页面的磨损均衡优化方案。然后，通过构建原型验证平台，实现关键技术验证，实验结果显示我们的方案在能效、性能、寿命等方面取得较大提升。研究成果在国际重要学术期刊上发表6 篇学术论文，其中3篇CCF A类期刊，1篇CCF B类期刊，2篇CCF C类期刊；此外，录用未刊出学术论文2篇。并在关键技术点，申请 11项国家发明专利。培养研究10名（5人已毕业），就业去向均为领域内重要企事业单位。本项目研究对智能终端和智能边缘计算的发展具有一定的理论意义和应用价值。