面向集群式内存的容错机制和数据组织策略研究

The in-depth analysis reveals that the overall performance of clustered storage is still restricted by secondary storage such as hard disks. Clustered memory constituted from available DRAM via high-speed network can accomplish high-IOPS, low-latency data accesses, thereby potentially boosting the access performance of clustered storage. Focusing on both fault-tolerance and data organization technical challenges occurred in clustered memory, this project proposes a log-structured erasure-coded memory cluster scheme, where a large-capacity storage pool is created by data segmentation, encoding and distribution, using pre-allocated memory of nodes in storage clusters. Furthermore, in order to achieve data durability, data and its replica are respectively placed on memory and disk by employing replication-declustering mechanism. Theoretically, a clustered memory framework is constructed, both fault-tolerance mechanism and data organization strategy is investigated for clustered memory, and its reliability model is built. Technically, a set of key techniques and schemes are designed for clustered in-memory storage, including declustering-based data layout, pipelined high-performance in-memory data recovery, log-structured highly-efficient in-memory data organization, and log-segment-based memory defragmentation schemes.. This project is aimed to investigate a in-memory cluster featured with high-efficiency and high-reliability, and its prospective research results may not only deepen the memory-level fault-tolerance mechanism, but also provide valuable technical cases for the performance optimization of clustered storage.

分析发现，现有集群式存储的整体性能仍受限于磁盘等外存设备。利用高速网络和节点空闲内存构建的集群式内存能够提供高IOPS、低延迟的数据访问，有助于提升集群式存储的访问性能。针对集群式内存的内存数据容错和内存数据组织挑战，本项目提出一种基于纠删码、采用日志结构的集群式内存方案，通过数据分段、编码、布局等方式，将集群中多个节点的空闲内存构建成一个大容量内存级空间，并按副本散布机制来放置内存数据和外存副本，获得数据持久性支持。理论上，构建集群式内存存储框架，研究其数据容错机制和数据组织策略，建立集群式内存可靠性模型；方案上，研究一组关键支撑技术，包括基于副本散布的内外存数据布局、基于流水线的内存数据恢复、基于日志的内存数据组织和基于日志分段的内存碎片回收方案等。. 本项目旨在研究一种高效、可靠的集群式内存方案，其预取成果将丰富内存层数据容错研究，并为优化集群式存储提供方案借鉴和技术参考。

鉴于大数据处理的多样性，目前出现的多种计算模式(如，大数据查询分析计算、图计算、内存计算等)对数据存储提出一个共性要求——让数据更加接近计算。本项目结合当前高性能计算系统架构，建立了一种分布式纠删码内存系统框架，即，Proxy节点充当内存系统的应用服务器，为前台用户提供访问接口；Proxy节点、Coordinator节点和内存节点组成三方架构，支持带外通信，构建出一个大容量集群式内存系统，为前提用户提供透明的、可靠的内存访问空间。其中，Coordinator节点负责管理所有元数据信息（包括分块的分布和组织、纠删码条带的组织、分块访问信息的管理等），Proxy节点和内存节点之间提供内存级数据访问。在纠删码内存系统框架的基础上，结合考虑了数据容错性、访问倾斜性、访问局部性、负载均衡性、访问异构性等多方面因素，本项目研究了多种内存数据组织策略和管理方案，包括一种异构感知的纠删码归档方法LEAP、一种基于预取机制和平衡策略的归档优化方案PBArch、一种纠删码内存系统中智能数据放置策略WPS、一种针对通用写的混合式更新方案Hybrid-U、一种访问热度感知的冗余管理方案PaRS和一种流量感知的内存数据副本放置策略ERP/TEA。上述研究以数据编码和数据访问作为研究切入点，同时充分考虑网络、内存、外存等组件的特性，结合‘大容量内存级空间’这一需求来研究集群式内存的容错机制和数据组织策略，旨在设计一种高空间效率、高访问性能的内存级数据存储方案，为‘大容量的热点数据集’提供高可靠内存级空间，进而为大数据分析和处理提供有力的支撑。