SHF: Small: Whole-application Coordiated Parallelization Through The Optimization Of Abstraction Hierarchies

大型计算机软件基于编程抽象，例如函数和面向对象类型，来组织计算和数据，这些计算和数据反过来通过调用各种低级库支持来并行化。一个正确而高效的软件实现需要对编码细节进行细致的关注，并且常常容易出错。然而，作为软件发展和维护过程的一部分，这些实现经常需要调整以在不同的硬件平台上运行。本项目开发新的尖端编译器技术，自动提高大型计算机软件的并行化效率，从而同时提高它们的生产率、正确性保证、性能和硬件可移植性。该项目与现代软件工程中的面向对象设计原则保持一致，因此期望通过自动探索可选的并行化和数据结构设计来影响现代软件实践，以获得更好的性能、可维护性和硬件可移植性。在这个项目中开发的编译器工具集将以开源的方式传播，并将作为编程语言、编译器、高性能计算和并行编程课程的教学工具包使用。这个项目的研究方法在用户定义的抽象层次上探索了一个新的优化范围，通过将相关操作分组和封装到抽象层次中，并通过采用一组传统上在软件工程实践中使用的概念，特别是抽象数据类型、抽象适配器和应用程序抽象层次。自动发现用户定义抽象的高级语义，并将算法级结构和逻辑与其低级实现隔离开来。然后开发模式驱动的分析技术，以自动发现整个应用程序任务图，并跨抽象边界应用协调优化，以最大限度地提高整体应用程序效率。统一接口的目的是简洁地记录抽象的高级语义。经验性性能调优用于支持应用程序在各种现代计算平台上的性能可移植性。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Large computer software is based on programming abstractions, e.g., functions and object-oriented types, to organize computation and data, which are in turn parallelized by invoking a variety of lower-level library support. A correct and efficient software implementation requires meticulous attention to coding details and is often error-prone. However, these implementations frequently need to be adapted to run on different hardware platforms as part of the software evolution and maintenance process. This project develops new cutting-edge compiler technology to automatically enhance the parallelization efficiency of large computer software, thereby simultaneously enhancing their productivity, correctness assurance, performance, and hardware portability. This project is aligned with object-oriented design principles in modern software engineering and therefore is expected to impact a modern software practices by enabling automatic exploration of alternative parallelization and data structure designs for better performance, maintainability, and hardware portability. The compiler toolsets developed in this project will be disseminated open source and will be used as an educational toolkit in the teaching of programming languages, compilers, high performance computing, and parallel programming courses.The research methodology of this project explores a new scope of optimization at the user-defined abstraction level, by grouping and encapsulating related operations into hierarchies of abstractions, and by adapting a set of concepts traditionally used in the practice of software engineering, specifically abstract data types, abstraction adapters, and application abstraction hierarchy, to automatically discover higher-level semantics of user-defined abstractions and to isolate algorithm-level structures and logistics from their lower-level implementations. Pattern-driven analysis techniques are then developed to automatically discover a whole-application task graph, and coordinated optimizations are applied across abstraction boundaries to maximize overall application efficiency. A uniform interface is designed to concisely document the higher-level semantics of abstractions. Empirical performance tuning is used to support the performance portability of applications on a variety of modern computing platforms.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.