闪存数据库管理系统关键技术研究

发布时间：2018-03-23 07:06

本文选题：闪存　切入点：数据库　出处：《北京大学》2013年博士论文　论文类型：学位论文

【摘要】：闪存是一种电可擦除可编程只读存储器。由于其具有体积小，重量轻，访问速度快，适用环境广等优点，闪存已经大量地被应用在移动设备上，如手机、MP3、传感器等。近年来，随着闪存容量的不断增加，价格的不断下降，基于闪存的存储设备被越来越多地用于企业级大型数据存储系统中，如数据库系统。由于闪存的访问速度，，特别是其随机访问速度大大快于磁盘，所以闪存的使用可以大大地缓解数据库外存的瓶颈问题。与传统的磁盘相比，闪存有其自身的特性。首先，作为一种电子设备，闪存并不需要机械寻道延时，所以其随机访问速度与顺序访问速度相差不大。此外，由于闪存在覆盖写前要进行擦除操作，所以闪存有读写速度不对称的特点。当前的数据库管理系统均按照磁盘特性优化系统性能，所以直接将数据库系统移植到闪存上并不能发挥出闪存的最佳性能。这就要根据闪存的特点对数据库进行调整。为了扬长避短，对于闪存的优化需要减少闪存的写入操作特别是随机写操作数量。对于随机更新较多的索引结构中，通常使用日志的方式，将随机更新操作转化为顺序操作，而如何设计日志的格式和组织方式是一个重要的问题。缓冲区管理中同样存在类似的问题，如何合理地将一些脏页面保留在缓冲区中，以减少整体I/O的性能也是闪存优化要考虑的重点问题。此外，当前闪存磁盘混合存储系统设计同样受到关注。本文以此为背景进行闪存数据库系统的优化研究，具体来说包括以下几个方面。 (1)本文提出了操作可感知的缓冲区管理策略FOR。在代价分析的基础上，FOR同时考虑页面状态与发生在页面上的操作，充分地利用页面访问信息来决定页面的替换。所以FOR方法可以更好地实现基于闪存的缓冲区管理功能。而且本文进一步针对FOR方法运行速度较慢的缺点进行了优化，提出了FOR+算法，进一步地提升了缓冲区管理的整体性能。 (2)本文提出了一种基于闪存的高维索引结构LCR树。 LCR树在传统的高维索引上加入日志区，将随机更新操作转化为顺序更新，提高了索引的更新效率。此外，LCR树将日志区组织为压缩日志方式，即多次更新的日志会保存在同一个页面内。与现有方法相比压缩日志能够减少索引的读取操作数量，达到更好的I/O性能。 (3)本文针对闪存磁盘混合存储系统进行优化，提出了基于概率和基于热度的混合存储管理策略。基于概率的混合存储方法使用两个概率参数控制数据之间的迁移操作，以减少系统各存储层次之间不必要的数据移动，从而提升整体的性能。基于热度的存储方法可以识别出数据的热度，所以能够将访问频繁的数据放在较高的存储层次中，与已有方法相比，可以以较小的代价更好地完成混合存储管理。
[Abstract]:Flash memory is a kind of electrically erasable programmable read-only memory. Because of its small size, light weight, fast access speed and wide application environment, flash memory has been widely used in mobile devices, such as mobile phone MP3, sensor, etc. In recent years, flash memory has been widely used in mobile devices. With the increasing of flash storage capacity and the decreasing of price, flash based storage devices are used more and more in large enterprise data storage systems, such as database system. Especially its random access speed is much faster than disk, so the use of flash memory can greatly alleviate the bottleneck of database storage. Compared with traditional disks, flash memory has its own characteristics. First, as an electronic device, flash memory does not require mechanical seek delay, so its random access speed is not different from sequential access speed. Since flash memory has to be erased before overwriting, flash memory is characterized by asymmetric read and write speed. Current database management systems optimize system performance according to disk characteristics. Therefore, transplanting database system directly to flash memory can not bring into play the best performance of flash memory, which should be adjusted according to the characteristics of flash memory. In order to optimize flash memory, it is necessary to reduce the number of write operations, especially the number of random write operations. For index structures with more random updates, log operations are usually used to convert random update operations into sequential operations. How to design the format and organization of the log is an important issue. There are similar problems in buffer management, how to reasonably keep some dirty pages in the buffer, To reduce the performance of the whole I / O is also a key problem to be considered in flash memory optimization. In addition, the design of flash disk hybrid storage system is also concerned. This paper studies the optimization of flash memory database system based on this background. Specifically, it includes the following aspects. In this paper, we propose an operational perceptible buffer management strategy FOR.On the basis of cost analysis, we consider both the state of the page and the operations that occur on the page. Therefore, the FOR method can better realize the function of buffer management based on flash memory. Furthermore, this paper further optimizes the FOR method because of its slow running speed. FOR algorithm is proposed to further improve the overall performance of buffer management. In this paper, a high-dimensional index structure LCR tree based on flash memory is proposed. The LCR tree adds log area to the traditional high-dimensional index, which transforms random update operation into sequential update. In addition, the LCR tree organizes the log area into a compressed log mode, that is, the log that has been updated many times will be saved in the same page. Compared with the existing method, the compressed log can reduce the number of read operations of the index. Achieve better I / O performance. In this paper, a hybrid storage management strategy based on probability and heat is proposed for flash disk hybrid storage system. The hybrid storage method based on probability uses two probabilistic parameters to control the migration between data. In order to reduce unnecessary data movement between different storage levels of the system, thereby improving the overall performance. The heat based storage method can identify the heat of the data, so the frequently accessed data can be placed in a higher storage level. Compared with the existing methods, hybrid storage management can be accomplished at a lower cost.
【学位授予单位】：北京大学
【学位级别】：博士
【学位授予年份】：2013
【分类号】：TP333

【相似文献】