持久性内存（persist memory，PM）具有非易失、字节寻址、低时延和大容量等特性，打破了传统内外存之间的界限，对现有软件体系结构带来颠覆性影响.但是，当前PM硬件还存在着磨损不均衡、读写不对称等问题，特别是当跨NUMA （non uniform memory access）节点访问PM时，存在着严重的I/O性能衰减问题.提出了一种NUMA感知的PM存储引擎优化设计，并应用到中兴新一代数据库系统GoldenX中，显著降低了数据库系统跨NUMA节点访问持久内存的开销.主要创新点包括：提出了一种DRAM+PM混合内存架构下跨NUMA节点的数据空间分布策略和分布式存取模型，实现了PM数据空间的高效使用；针对跨NUMA访问PM的高开销问题，提出了I/O代理例程访问方法，将跨NUMA访问PM开销转化为一次远程DRAM内存拷贝和本地访问PM的开销，设计了Cache Line Area （CLA）缓存页机制，缓解了I/O写放大问题，提升了本地访问PM的效率；扩展了传统表空间概念，让每个表空间既拥有独立的表数据存储，也拥有专门的WAL （write-ahead logging）日志存储，针对该分布式WAL存储架构提出了一种基于全局顺序号的事务处理机制，解决了单点WAL性能瓶颈问题，并实现了NUMA感知的事务处理、检查点和灾难恢复等优化机制及算法.实验结果表明，所提出的方法能够有效提升NUMA架构下PM存储引擎的性能，在YCSB多种测试场景下分别提升了105%-317%，在TPC-C场景下提升了90%-134%.关键词：数据库；存储引擎；持久性内存；NUMA （non uniform memory access）；WAL （write-ahead logging）

Persistent memory (PM) has the characteristics of non-volatility, byte addressable, low latency, and large capacity, which breaks the boundary between traditional internal and external memory and has a has a disruptive impact on the existing software architecture. However, the current PM hardware still has problems such as uneven wear and asymmetric read and write. Especially serious I/O performance degradation problem will occur when the CPU accesses the PM across NUMA (non uniform memory access) nodes. An NUMA-aware PM storage engine optimization design is proposed and applied to Zhongxing’s new generation database system GoldenX, which significantly reduces the overhead of database system accessing persistent memory across NUMA nodes. The main innovations include: a data space distribution strategy and distributed access model across NUMA nodes are proposed under a DRAM+PM hybrid memory architecture, which realizes the efficient use of PM data space; aiming at the high latency problem of accessing PM across NUMA nodes, an I/O proxy routines access method is proposed, which converts the overhead of accessing PM across NUMA into the overhead of a remote DRAM memory copy and local access to PM. The Cache Line Area cache page mechanism is designed to alleviate the problem of I/O write amplification and improve the efficiency of local access to PM. The concept of traditional table space is extended, so that each table space has both independent table data storage and dedicated WAL (write-ahead logging) storage. For the distributed WAL storage architecture, a transaction processing mechanism based on global sequence numbers is proposed, which addresses the problem of single-point the WAL performance bottleneck, and implement NUMA-aware transaction processing, checkpoint and disaster recovery optimization mechanisms and algorithms. Experimental results show that the method proposed in this study can effectively improve the performance of the PM storage engine under the NUMA architecture, by 105%-317% in various test scenarios of YCSB and 90%-134% in TPC-C.

国家重点研发计划（2019YFB2102002）；江苏省重点研发计划（BE2019012）