Collaborative Research: CNS Core: Small: Scalable ACID Transactions for Persistent Memory Databases

该项目解决了当前数据库系统无法跟上从机器生成的数据集（如物联网传感器和机器学习系统）中分析和提取信息的应用程序不断增长的需求的问题。直观地说，将系统的计算资源增加一倍应该会使系统在单位时间内可以处理的负载增加一倍，但对于今天的数据库来说并非如此：在相当适度的系统大小之外，添加更多的计算资源并不能按比例增加性能增益。关键原因是数据库要正确执行，必须限制对某些关键数据结构的并发访问：添加更多资源会增加对这些数据结构访问的竞争，从而为系统性能造成瓶颈。这个项目为可扩展数据库引入了一个关键的创新。它使数据库不必在修改数据库记录时立即更新范围索引——这是数据库中一种常见的数据组织形式，当数据库试图扩展其计算资源时，它往往会成为热点。为了消除这个瓶颈，这个项目开发了一个新的可扩展接口：每处理器核心队列在后台定期吸收索引更新并将它们合并到共享范围索引数据结构中。消除对范围索引的同步更新不会削弱数据库的保证：序列化性的标准正确性标准是通过使用来自系统范围时钟的多部分时间戳对事务更新进行全局排序来实现的；数据持久性是通过将每核队列存储在非易失性内存中实现的；新的数据结构确保对单个记录执行的读取返回它们最近提交的值。数据库是现代行星级应用程序的关键组成部分。通过消除可伸缩性瓶颈和利用新兴的非易失性内存技术，该项目将显著降低提供数据库的成本。特别是，在价值数十亿美元的数据中心中，运营成本的很大一部分都花在了为越来越多的服务器供电上。提高多处理器核心的可伸缩性将增加数据库部署的密度，大幅减少提供数据库所需的服务器数量：节省的资源可以推迟对新数据中心和存储设备的需求，因为现有服务器可以完成更多有用的工作，或者减少现有工作负载的能耗。这项工作还将影响下一代数据库工程师的教育。拟议的讲座和项目材料将帮助学生在响应硬件和应用程序工作负载的未来变化时识别可扩展的数据库设计。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。

This project addresses the inability of current database systems to keep up with the ever growing demands of applications that analyze and extract information from machine-generated data sets, such as Internet-of-Things sensors and machine-learning systems. Intuitively, doubling a system's computing resources should double the load that the system can process per unit of time, but that is not true of today's databases: beyond a fairly modest system size, adding more computing resources does not scale to proportionate gains in performance. The key reason is that databases, to perform correctly, must limit concurrent access to some critical data structures: adding more resources increases competition for access to these data structures, creating a bottleneck for the system's performance. This project introduces a key innovation towards scalable databases. It frees the database from the need, whenever a databases record is modified, to immediately update range indexes---a common form of data organization in databases that tend to become a hotspot when databases try to scale up their computing resources. To remove this bottleneck, this project develops a new scalable interface: per-processor-core queues absorb index updates and merge them in the shared range index data structures periodically, in the background. Eliminating synchronous updates to range indices does not weaken the database guarantees: the standard correctness criterion of serializability is achieved by globally ordering transactional updates using multi-part timestamps derived from a system-wide clock; data durability is achieved by storing per-core queues in non-volatile memory; and a new data structure ensures that reads performed on individual records return their most-recently committed value.Databases are a critical component of modern planet-scale applications. By eliminating scalability bottlenecks and leveraging emerging non volatile memory technology, this project will dramatically reduce the cost to provision databases. In particular, a large fraction of operational cost in multi-billion-dollar data centers is spent on powering a growing number of servers. Improving the scalability of multiple processor cores will increase the density of database deployments, reducing drastically the number of servers required to provision a database: the savings can defer the need for new data centers and storage devices, as more useful work is achieved with existing servers, or reduce energy consumption for existing workloads. The work will also influence the education of the next generation of database engineers. Proposed lecture and project materials will prepare students to identify scalable database designs when responding to future changes in hardware and application workloads.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.