PFI AIR-TT: Improving Data Base Management System Performance Through Micro-Specialization

这个PFI: AIR技术翻译项目的重点是翻译微专业化的新技术，以满足提高数据库管理系统（DBMSes）和企业应用软件（EAS）速度的需要。商业化的下一个必要步骤是演示该过程如何自动化。该项目将产生一组概念验证软件工具，这些工具将获取一组已识别的不变量，并自动对DBMS源代码进行微专门化。这些工具的增强版本将具有以下独特的特性：它将是(i)可扩展的，即可以处理实际的DBMS代码库，多达数百万行源代码，具有可接受的性能；（ii）健壮，即能够处理对编译器来说可能很棘手的语言特性；（iii）有效，即能够生成高质量的专门代码，实现期望的性能改进。这些特性提供了以下两个主要优点。首先，它不需要对DBMS源代码进行昂贵和耗时的重写，但可以应用于现有的DBMS代码。其次，它独立于DBMS性能改进措施的其他方法，并且与之正交。这意味着DBMS供应商在开发其他性能改进技术方面所做的任何投资都不会因我们的方法而无效；相反，我们的方法通过产生额外的性能改进进一步增强了这些技术的效果。该项目解决了以下技术差距，因为它从研究发现转化为商业应用，即自动化微专业化。对于这些供应商来说，在最终用户DBMS上手动执行微专门化是不实际的，因为微专门化会根据安装的具体情况定制DBMS代码，因此对于不同的安装是不同的。相反，供应商想要的，以及我们将提供的产品，是一个软件开发环境，它允许开发人员快速、高效、可靠地对客户的DBMS执行微专门化，然后他们将其授权给客户：DBMS最终用户。因此，为了使我们的想法商业化，我们下一步必须证明这个过程可以在很大程度上自动化。该项目聘请Dataware Ventures负责项目的开发方面，专注于亚利桑那大学研究人员开发的代码专业化算法的功能和功效的大规模测试和改进。具体来说，Dataware Ventures将执行软件开发，进行大规模测试和评估，并向UA研究人员提供相关反馈，帮助将这项技术从研究发现转化为商业现实。DBMS的微专门化很重要，因为大多数具有重要数据处理需求的组织，包括几乎所有的中型和大型企业和公司，都依赖于企业应用系统进行数据处理和分析。这些应用程序通常运行在DBMS之上，DBMS处理数据的实际存储和检索。因此，底层DBMS的性能，即在可接受的时间内处理大量数据的能力，对于向业务决策者提供及时和准确的信息至关重要。此外，潜在的经济影响预计将在未来三年内增长到1570亿美元，这将有助于美国在企业应用系统和数据库管理系统基础技术方面的竞争力。

This PFI: AIR Technology Translation project focuses on translating the novel technology of micro-specialization to fill the need for improving the speed of database management systems (DBMSes) and enterprise application software (EAS). The next step necessary for commercialization is to demonstrate how the process can be automated. The project will result in a proof-of-concept set of software tools that will take a collection of identified invariants and micro-specialize the DBMS source code automatically.The enhanced version of those tools will have the following unique features: it will be (i) scalable, i.e., can handle realistic DBMS code bases, up to millions of lines of source code, with acceptable performance; (ii) robust, i.e., able to handle language features that can be tricky for compilers; and (iii) effective, i.e., able to generate high-quality specialized code that realizes the desired performance improvements. These features provide the following two main advantages. First, it does not require an expensive and time-consuming rewrite of the DBMS source code, but can be applied to existing DBMS code. Second, it is independent of, and orthogonal to, other approaches to DBMS performance improvement measures. This means that any investments a DBMS vendor may have made in developing other performance improvement technologies are not nullified by our approach; rather, our approach further enhances the effects of those technologies by producing additional improvements in performance.This project addresses the following technology gap as it translates from research discovery toward commercial application, that of automating micro-specialization. It is not practical for these vendors to manually perform micro-specialization on the end-user DBMSes since micro-specialization tailors the DBMS code to the specifics of an installation and so is different for different installations. Rather, what the vendors will want, and the product we will provide, is a software development environment that allows developers to quickly, efficiently, and reliably perform micro-specialization on their customers' DBMSes, that they then license to their customers: the DBMS end users. In order to commercialize our idea, therefore, we next have to show that the process can be largely automated.The project engages Dataware Ventures to be responsible for the development side of the project, focusing on large-scale testing and refinement of the functionality and efficacy of the code specialization algorithms developed by researchers at the University of Arizona. Specifically, Dataware Ventures will perform software development, carry out large-scale testing and evaluation, and provide relevant feedback to UA researchers, aiding in this technology translation effort from research discovery toward commercial reality.DBMS micro-specialization is important because most organizations with non-trivial data processing needs, including virtually all medium and large businesses and corporations, rely on enterprise application systems for their data processing and analysis. These applications generally run on top of a DBMS, which handles the actual storage and retrieval of data. The performance of the underlying DBMS, i.e., its ability to process large amounts of data within an acceptable amount of time, is therefore of crucial importance for providing timely and accurate information to business decision makers. In addition, the potential economic impact is expected to grow to $157B in the next three years, which will contribute to the U.S. competitiveness in enterprise application systems and in the underlying technology of database management systems.