High-Performance, High-Level Tools for Statistical Inference and Unsupervised Learning

用于统计推断和无监督学习的高性能、高级工具

• 批准号: 1622501
• 负责人: John Owens
• 金额: $ 48万
• 依托单位: University of California-Davis
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2016
• 资助国家: 美国
• 起止时间: 2016-09-15 至 2020-12-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1622501&HistoricalAwards=false
• 关键词: Performance; Level; Tools; Statistical; Inference

Using the "Julia" language for scientific computing developed at MIT, the UC Davis, MIT, and Julia Computing, Inc. teams funded by this project will extend the Julia language and runtime to utilize massively-parallel graphics processing units (GPUs) as first-class processors for scientific computing. Julia offers the twin advantages of straightforward, high-level programmability as well as excellent performance; adding GPU capability within Julia opens the door to even greater performance. The team will use Julia and its new GPU capabilities to address emerging important problems in statistical inference and unsupervised learning, an application area that aims to draw useful conclusions from massive amounts of data. Using a high-level, high-performance language such as Julia will allow non-computer-science experts to address these important problems.The project team brings together three threads of expertise to address the challenge of delivering best-of-breed performance from a high-level language in the context of the important application domain of statistical inference and unsupervised learning: (1) application experts in this domain; (2) the designers of the programming language Julia, which allows its users to express their ideas in high-level abstractions that are natural to statisticians and mathematicians; and (3) parallel computing experts, who will develop the new support within Julia to target high-performance GPUs as first-class processors. The major outcome of this project will be a significantly enhanced Julia language and runtime that will deliver both high-level programmability, targeted at scientists who are not parallel computing experts, and best-of-breed performance.

使用由麻省理工学院开发的用于科学计算的“Julia”语言，加州大学戴维斯分校、麻省理工学院和Julia Computing，Inc.由该项目资助的团队将扩展Julia语言和运行时，以利用并行图形处理单元（GPU）作为科学计算的一流处理器。Julia提供了简单、高级可编程性和卓越性能的双重优势;在Julia中添加GPU功能为更高的性能打开了大门。该团队将使用Julia及其新的GPU功能来解决统计推断和无监督学习中新出现的重要问题，这是一个旨在从大量数据中得出有用结论的应用领域。使用像Julia这样的高级、高性能语言将允许非计算机科学专家解决这些重要问题。项目团队汇集了三个专业领域的知识，以解决在统计推断和无监督学习的重要应用领域中使用高级语言提供最佳性能的挑战：（1）该领域的应用专家;（2）Julia编程语言的设计者，它允许用户以统计学家和数学家自然的高级抽象表达他们的想法;（3）并行计算专家，他们将在Julia中开发新的支持，以将高性能GPU作为一流的处理器。该项目的主要成果将是显著增强的Julia语言和运行时，它将提供高级可编程性（针对非并行计算专家的科学家）和最佳性能。

项目成果

Distributed MCMC Inference in Dirichlet Process Mixture Models Using Julia

• DOI: 10.1109/ccgrid.2019.00066
• 发表时间: 2019-05
• 期刊: 2019 19th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing (CCGRID)
• 作者: Or Dinari;A. Yu;O. Freifeld;John W. Fisher III