EAGER: A Demonstration of the IMP Programming Model

EAGER：IMP 编程模型的演示

• 批准号: 1451204
• 负责人: Victor Eijkhout
• 金额: $ 9万
• 依托单位: University of Texas at Austin
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2014
• 资助国家: 美国
• 起止时间: 2014-09-01 至 2015-08-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1451204&HistoricalAwards=false
• 关键词: EAGER; Demonstration; IMP; Programming; Model

The machine architectures used for scientific computing are ever growing in scale and complexity. Software has not been keeping up with this development, and application codes often use multiple parallelism modes and parallel programming systems (typically MPI and a co-processor or threading library) to extract all possible performance. This makes current architectures unnecessarily hard to use for any but the most experienced programmers.This project addresses parallel programming on contemporary supercomputer architectures. Based on a new theoretical model, called the Integrative Model for Parallelism (IMP), this project develops a framework for such hybrid designs. Ultimately, the goal is to develop a production software system that allows domain scientists to write applications independently of the underlying supercomputer hardware. This project creates a demonstration prototype, a `mini-application' for N-body dynamics, a simplified code containing the essential structure of the parallelism in an N-body application, to be used for the demonstration of the principles of the IMP model. This mini-app can be run, unchanged, using different types of parallelism, tailored for the underlying hardware. Having such a software system implies for application scientists both increased productivity over current approaches, and a guarantee of comparable efficiency of execution.

用于科学计算的机器架构的规模和复杂性不断增长。软件并没有跟上这种发展，应用程序代码通常使用多种并行模式和并行编程系统（通常是 MPI 和协处理器或线程库）来提取所有可能的性能。这使得当前的架构对于除了最有经验的程序员之外的任何人来说都不必要地难以使用。该项目解决了当代超级计算机架构上的并行编程问题。该项目基于一种称为并行集成模型 (IMP) 的新理论模型，为此类混合设计开发了一个框架。最终的目标是开发一个生产软件系统，允许领域科学家独立于底层超级计算机硬件编写应用程序。该项目创建了一个演示原型，一个用于 N 体动力学的“迷你应用程序”，一个包含 N 体应用程序中并行性基本结构的简化代码，用于演示 IMP 模型的原理。该迷你应用程序可以使用针对底层硬件量身定制的不同类型的并行性来运行，无需更改。拥有这样的软件系统对于应用科学家来说意味着比当前方法提高生产力，并保证可比较的执行效率。