III: Small: Database Algorithms for Modern CPU Memory Hierarchies

现代数据库服务器通常在具有大量RAM的机器集群上运行，以确保在经常访问的数据(甚至整个数据库)可以放入主内存时的快速查询响应。数据可以驻留在远程或本地内存中，也可以驻留在多个级别的高速缓存中；每种类型的内存都有其自己的特征大小和性能属性。该项目将开发查询处理算法和为这种存储器层次结构量身定做的查询处理系统。分析数据库系统已成为从从医学到科学再到商业的“大数据”集合中提取可操作信息的一项关键技术。该项目实现的核心数据库操作性能的改进将对许多应用领域产生影响。该项目还将通过提供用于讲授数据库系统实现技术的技术来促进教育。该项目将开发一个以高效分区为核心操作的数据库体系结构，在内存层次的每一级使用不同的分区技术。适当时将使用就地分区，以避免分配额外的内存，分区速度只会有很小的下降。分区将用作其他运算符的构建块，如排序、联接和聚合。将实现各种数据库算法，利用分区的效率在规模上实现良好的性能。网络传输量将使用专门的连接算法来减少，因为此类传输很可能成为涉及大型分布式连接的查询的瓶颈。这些算法将构成将在项目过程中开发的数据库系统原型的基础。该项目将提供利用现代机器进行高效分析查询处理的新技术。拟议的系统将大大提高数据密集型查询的吞吐量。有关更多信息，请参阅项目网站：http://www.cs.columbia.edu/~kar/cpumem.html

A modern database server typically runs on a cluster of machines with a large amount of RAM, to ensure fast query responses when the commonly accessed data (or even the whole database) can fit in main memory. Data may reside in remote or local memory, or in one of several levels of cache; each type of memory has its own characteristic size and performance properties. This project will develop query processing algorithms and a query processing system tailored to such memory hierarchies. Analytic database systems have emerged as a key technology for extracting actionable information from "big data" collections ranging from medicine to science to business. Improvements in the performance of core database operations achieved by the project will have impact on many application domains. The project will also contribute to education by contributing technology for use in teaching database system implementation techniques.The project will develop a database architecture with efficient partitioning as the core operation, using different partitioning techniques at each level of the memory hierarchy. In-place partitioning will be employed where appropriate to avoid allocating extra memory, with only a small decrease in partitioning speed. Partitioning will be used as a building block for other operators such as sorting, joins and aggregation. A variety of database algorithms will be implemented, taking advantage of the efficiency of partitioning to perform well at scale. Network transfer volumes will be reduced using specialized join algorithms, since such transfers are likely to be the bottleneck for queries involving large distributed joins. These algorithms will form the basis of a database system prototype to be developed during the course of the project. The project will provide new techniques to exploit modern machines for efficient analytic query processing. The proposed system will significantly improve the throughput of data-intensive queries.For further information see the project web site at: http://www.cs.columbia.edu/~kar/cpumem.html