多核处理器下事务型数据库性能优化技术综述

A Survey of Optimization Methods for Transactional Database in Multi-Core Era

传统数据库的设计假设磁盘为主要存储设备,其性能取决于基于I/O代价模型的优化.然而,当前数据库运行的平台已逐渐转移到由多核处理器、大内存和以闪存为代表的低延迟存储所构成的新型硬件平台上.在大多数情况下,工作数据集能够全部加载到内存或者闪存等高速存储器中.这样,数据库的性能瓶颈由传统的I/O转移到CPU上.而传统数据库的加锁操作、闩锁竞争、日志管理以及缓冲区管理在设计时均未考虑到多核处理器的使用,因而成为了限制CPU利用率的明显瓶颈.改变传统数据库的优化重点以适应硬件的发展对应用而言是十分必要的.该文针对当前新的应用背景,主要围绕数据库系统中锁管理、日志管理、缓冲区管理以及B树索引等核心模块在多核平台下已有的优化技术进行详细介绍和归纳总结.同时介绍了中国人民大学在数据库系统的多核处理器优化方面所做的一些工作.

The design of traditional DBMS assumes magnetic disk as the storage device. Its optimization techniques are focused on the reduction of I/O cost. However, the database platforms in the future will be dominated by multi-core processors, large main memory and low-latency semiconductor storage, such as SSD. On such platforms, the entire data set can normally fit into main memory or high-speed storage. Thus, the performance bottleneck of query execution has shifted from disk I/O to CPU. The components for locking, latching, logging and buffer management of traditional DBMS were not originally designed for multi-core processors. These components severely prohibit the scalability of DBMS in multi-core architectures. Adaptation of traditional DBMS to new hard- ware is a common and necessary practice. This paper provides a survey of the recent optimization techniques proposed for DBMS on multi-core platforms. Meanwhile, the efforts by Renmin University of China on DBMS optimization on multi-core platform are introduced.