CRII: III: Buffer Management for Non-Volatile Memory

数据库系统五十年来的研究主要集中在典型的两层存储结构上，即易失性存储和持久存储技术。数据库系统的缓冲区管理器通过将数据迁移到易失性内存来提供对存储在存储器上的数据的有效访问。缓冲区管理器在现有系统中采用的数据迁移策略是基于这样的假设：存储比易失性内存慢几个数量级。但新兴的非易失性存储技术正准备颠覆这些设计假设。它们提供像易失性存储器那样的快速访问，但具有像存储器那样的持久写入。传统的缓冲区管理方法与这种新的硬件环境不兼容。本项目概述了研究和开发非易失性存储器的新型缓冲区管理技术的研究议程。该项目将产生用于减少大规模内存数据处理成本的算法。从这个项目中吸取的主要经验教训将被整合到佐治亚理工学院的数据库系统研究生课程中，教育学生关于非易失性存储器数据库系统的挑战和机遇。该项目的技术目标分为三个相互作用的研究重点。第一部分为易失性和非易失性存储器技术开发了数据迁移策略的分类。与以前的尝试不同的是，缓冲区管理器将动态地利用非易失性内存支持的所有数据流路径来提高数据库系统的性能。第二个重点是开发一个机器学习模型，用于自动调整任意工作负载和存储层次结构的数据迁移策略。由于非易失性内存的独特价格、容量和性能特征，数据库管理员很难确定多层存储层次结构的不同级别。第三个重点是开发存储系统推荐，以帮助管理员确定其目标工作负载的最佳层次结构。这三项研究重点将由一项综合评价计划加以补充。这项研究成果将对非易失性存储技术的采用产生深远的影响，并将实现经济高效的大规模数据处理。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Fifty years of research in database systems has focused on a canonical two-tier storage hierarchy with the volatile memory and durable storage technologies. A database system's buffer manager provides efficient access to data stored on storage by migrating it to volatile memory. The data migration policy employed by the buffer manager in existing systems is predicated on the assumptions that storage is orders of magnitude slower than volatile memory. But emerging non-volatile memory technologies are poised to upend these design assumptions. They provide fast accesses like volatile memory but with persistent writes like storage. The traditional approaches for buffer management are incompatible with this new hardware landscape. This project outlines a research agenda to study and develop novel buffer management techniques for non-volatile memory. The project will result in algorithms for shrinking the cost of large-scale in-memory data processing. The key lessons learned from this project will be integrated into the graduate-level course on database systems at Georgia Tech, educating students about the challenges and opportunities in non-volatile memory database systems.The technical aims of the project are divided into three interacting research thrusts. The first thrust develops a taxonomy of data migration policies for volatile and non-volatile memory technologies. What makes this different than earlier attempts is that the buffer manager will dynamically leverage all the data flow paths enabled by non-volatile memory to improve the database system's performance. The second thrust develops a machine-learning model for automatically tuning the data migration policy for an arbitrary workload and storage hierarchy. Due to the unique price, capacity, and performance characteristics of non-volatile memory, it will be challenging for database administrators to size the different levels of a multi-tier storage hierarchy. The third thrust develops a storage system recommender to help administrators identify the optimal hierarchy for their target workloads. These three research thrusts will be complemented by a comprehensive evaluation plan. This research effort will have a profound impact on the adoption of non-volatile memory technologies and will enable cost-effective, large-scaledata processing.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.