HDD/SSD混合存储系统的数据布局和缓存管理策略研究

HDD/SSD hybrid storage systems, satisfying the requirements of mass data storage in terms of large capacity, low price, and high access performance, have become one of the research hotspots in the storage field. Existing research results do not model and analyze the performance for HDD/SSD hybrid storage systems. It lacks of theoretical support. Moreover, they do not consider the internal relation between the data migration and buffer management, so unnecessary page migration operations will be incurred during the data migration process. To solve these problems, this project first does research on modeling and analyzing the performance of HDD/SSD hybrid storage systems, which provides the theoretical basis for guiding the building and deploying the HDD/SSD hybrid storage systems for practical application environments. This project also does research on proposing a transparent hybrid storage architecture, which hides the different physical features between the HDD and SSD, determining its functional modules, putting forward the on-demand address translation table management strategy and address redirecting strategy based on the I/O classification. Thirdly, this project does research on proposing a buffer-aware data migration strategy to reduce the unnecessary page migrations. Finally, this project does research on putting forward a buffer management strategy based on three-region, which divides the buffer into three regions and introduces the finite state machine to assign probabilities to each page in the buffer. All of these research results show the important research significance and good engineering application value to promote the wide application of new storage systems in the mass data storage environments and enterprise level storage.

HDD/SSD混合存储系统满足海量数据存储的大容量、低成本以及高存取性能等需求，已成为存储领域的研究热点之一。目前的研究成果没有对HDD/SSD混合存储系统进行性能建模和分析，缺乏理论支撑；没有考虑数据迁移和缓存管理之间的内在联系，数据迁移过程会引发不必要的数据页迁移操作。鉴于此，本项目研究HDD/SSD混合存储系统的性能建模和分析，为指导构建和部署实际应用环境下的HDD/SSD混合存储系统提供理论依据；研究透明的混合存储架构，屏蔽机械硬盘和固态硬盘的异构物理特性，确定其功能模块，提出按需加载的地址映射表管理策略和基于I/O分类的地址重映射策略；研究缓存信息感知的数据迁移策略，减少不必要的数据页迁移操作；研究基于三区域的缓存管理策略，通过划分缓存、引入有限状态机等方法，最大化数据迁移的性能；对于推动新型存储系统在海量数据环境和企业级存储中的普及应用具有重要的研究意义和很好的工程应用价值。

本项目针对HDD/SSD混合存储系统中HDD和SSD的物理特性差异，重点围绕性能建模和分析、混合存储架构、数据迁移、缓存管理等四个方面，研究并探索HDD/SSD混合存储系统性能优化的理论与方法。近年来，项目组取得了一系列研究成果，共发表学术论文28篇，其中，SCI/SSCI检索论文22篇，EI检索论文3篇，北大核心期刊论文3篇；申请国家发明专利3项；获得IEEE ISPA-2020 Outstanding Paper Award；项目负责人林铭炜晋升省级人才。项目组的代表性成果包括：（1）提出了固定优先级混合关键偶发任务能耗感知算法，对混合存储系统进行性能建模和性能分析，提升存储任务的执行速度并降低能耗；（2）提出面向混合存储架构的交替编码机制，充分考虑固态硬盘的物理特性，以便减少磁盘I/O开销；（3）提出了缓存感知的垃圾回收机制，不迁移已被缓存的有效数据，减少垃圾回收的额外开销；（4）提出智能内容缓存策略，充分考虑本地缓存设备、小型基站以及边缘服务器的缓存数据，并采用强化学习算法优化不同存储设备上的缓存机制。上述研究成果对于推动固态硬盘等新型非易失性存储设备在海量数据环境和企业级存储中的普及应用具有重要的研究意义和很好的工程应用价值。