III: Medium: Learning-based Synthesis of Data Processing Engines

III：媒介：基于学习的数据处理引擎综合

• 批准号: 1900933
• 负责人: Tim Kraska
• 金额: $ 120万
• 依托单位: Massachusetts Institute of Technology
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2019
• 资助国家: 美国
• 起止时间: 2019-09-01 至 2023-08-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1900933&HistoricalAwards=false
• 关键词: III; Medium; Learning; based; Synthesis

Modern data-processing systems are designed to be general-purpose systems, in that they can handle a wide variety of applications and data. Unfortunately, this general-purpose nature causes these systems to achieve below-optimal performance for every single application and user. Rather technical compromises have to be made to support a wide range of use cases, often leading to orders-of-magnitude worse performance than what a highly customized system would be able to achieve. At the same time, developing a database system from scratch for each individual application and user is neither economical nor practical. The goal of this project is to explore how machine learning can be used to automatically customize a database system for a specific application or user to achieve so called 'instance-optimality'. If successful, this project will transform the way that modern database systems that underpin the Internet and many enterprise computing systems are built, resulting in systems with much better performance or systems that are able to process large datasets using much less hardware than current systems. Concretely, the project investigates to what extent learned models can automatically instance-optimize the various components of a large-scale data processing system: 1) data indexing, where a model can predict the location of a key in a database; 2) algorithms, including sorting and joins, where a model can predict where in a sorted list a record should go, or where joining tuples are in another relation; 3) optimizers, where a model can predict the optimal plan to use for processing queries on data, and 4) storage layout, where a model can predict the optimal layout of data for a particular query workload. This raises a number of intellectually deep questions, including what types of models work best, what theoretical guarantees we can give about the performance of these models, how such generated systems will compare to hand-tuned systems, how such systems can exploit new hardware such as TPUs/GPUs and how program synthesis will work with such modelled data, advancing the fields of databases, machine learning, and program modeling and synthesis.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.

现代数据处理系统被设计成通用系统，因为它们可以处理各种各样的应用程序和数据。不幸的是，这种通用性导致这些系统为每个应用程序和用户实现了低于最佳性能的性能。相反，必须在技术上做出妥协，以支持广泛的用例，这往往会导致性能比高度定制的系统所能实现的性能差几个数量级。同时，从零开始为每个单独的应用程序和用户开发一个数据库系统既不经济也不实用。这个项目的目标是探索如何使用机器学习来为特定的应用程序或用户自动定制数据库系统，以实现所谓的“实例最优”。如果成功，该项目将改变支撑互联网和许多企业计算系统的现代数据库系统的构建方式，导致系统具有更好的性能，或者系统能够使用比当前系统少得多的硬件来处理大型数据集。具体地说，该项目调查学习的模型可以在多大程度上自动实例优化大规模数据处理系统的各种组件：1)数据索引，其中模型可以预测关键字在数据库中的位置；2)算法，包括排序和联接，其中模型可以预测记录在排序列表中的位置，或者连接元组在另一种关系中的位置；3)优化器，其中模型可以预测用于处理对数据的查询的最优计划；以及4)存储布局，其中模型可以预测针对特定查询工作负载的数据的最优布局。这引发了许多智力上的深层问题，包括哪些类型的模型工作得最好，我们可以为这些模型的性能提供什么理论上的保证，这样的生成系统将如何与手动调整的系统进行比较，这样的系统如何利用TPU/GPU等新硬件，以及程序合成将如何与这些建模数据一起工作，推动数据库、机器学习以及程序建模和合成领域的发展。这一奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

项目成果

SNARF: A Learning-Enhanced Range Filter

• DOI: 10.14778/3529337.3529347
• 发表时间: 2022-04
• 期刊: Proc. VLDB Endow.
• 作者: Kapil Vaidya;Tim Kraska;Subarna Chatterjee;Eric R. Knorr;M. Mitzenmacher;Stratos Idreos

Optimization of Graph Neural Networks: Implicit Acceleration by Skip Connections and More Depth

• 发表时间: 2021-05
• 作者: Keyulu Xu;Mozhi Zhang;S. Jegelka;Kenji Kawaguchi

TreeLine: An Update-In-Place Key-Value Store for Modern Storage

• DOI: 10.14778/3561261.3561270
• 发表时间: 2022-09
• 期刊: Proc. VLDB Endow.
• 作者: Geoffrey X. Yu;Markos Markakis;Andreas Kipf;P. Larson;U. F. Minhas;Tim Kraska

How Neural Networks Extrapolate: From Feedforward to Graph Neural Networks

• 发表时间: 2020-09
• 期刊: ArXiv
• 作者: Keyulu Xu;Jingling Li;Mozhi Zhang;S. Du;K. Kawarabayashi;S. Jegelka

Towards instance-optimized data systems

• DOI: 10.14778/3476311.3476392
• 发表时间: 2021-07
• 期刊: Proc. VLDB Endow.
• 作者: Tim Kraska