SHF: Small: Indy: Toward Safe and Fast Compiler Flags

计算机体系结构的快速发展是对计算速度和能源效率日益增长的需求所必需的。不幸的是，这些进步需要对现有软件进行重大调整并重新编译，当应用于科学模拟和工程领域的软件时，这些软件的数字输出可能会发生变化。这是一个值得关注的问题，因为整个科学和工程企业都依赖于建立在先前贡献之上的可重复结果。该研究项目开发了检测和隔离那些负责更改答案的软件组件的方法，并根据需要重写的软件例程产生可操作的证据。这确保了科学软件可以真正成为服务于几代研究的资产，而不是每隔三年左右就因硬件变化而被淘汰。该项目开发了使用不同的编译器编译用户的软件应用程序或库并在不同平台上执行的方法。给定一组验收测试，在本工作中开发的工具（统称为“Indy”，意味着“快速和安全的编译器标志”）有助于检测那些产生不可接受的偏差答案的组件。然后，使用代码平分方法，Indy工具将可变性的根源根源定位到单个文件甚至单个函数符号的级别。考虑到Indy使用的方法可能是计算密集型的，该项目还探索了机器学习方法和其他静态分析方法，以避免大规模的重新编译和重新执行。Indy工具在公共存储库中维护，调查团队与用户组保持密切联系，以确保工具的采用、反馈和改进。该项目将以开源的形式发布Indy，并围绕Indy建立一个用户社区，确保感兴趣的研究人员能够为代码库做出贡献。这将使该项目得到更广泛的发展。NSF高级网络基础设施办公室的软件集群对这方面特别感兴趣，该办公室是该奖项的共同资助方。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。

Rapid advances in computer architecture are necessitated by the growing demand for computing speed and also energy efficiency. Unfortunately, these advances require significant adaptations of existing software and their recompilation, and when applied to software in scientific simulations and engineering domains, the numeric outputs of these software can change. This issue is of significant concern, as the whole scientific and engineering enterprise rests on reproducible results building on top of prior contributions. This research project develops methods to detect and isolate those software components responsible for changing answers, and produces actionable evidence in terms of the software routines that need to be rewritten. This ensures that scientific software can truly become an asset that serves generations of research, and not be obviated every three or so years when hardware changes.The project develops methods to compile a users' software application or library using different compilers and executing on different platforms. Given a collection of acceptance tests, the tools developed in this work (collectively called "Indy", connoting "Fast and Safe Compiler Flags") help detect those components that produce unacceptably deviant answers. Then, employing code bisection methods, the Indy tools root-cause the source of variability to the level of single file or even to a single function symbol. Given that the methods used by Indy can be computationally intensive, the project also explores machine-learning methods and other static analysis methods to avoid wholesale recompilation and re-execution. The Indy tools are maintained on public repositories, and the investigative team maintains strong connections with user-groups to ensure tool adoption, feedback, and refinement. The project will release Indy as open source and build a user community around Indy by ensuring that interested researchers are able to contribute to the code-base. This will allow a wider growth of the project. This aspect is of special interest to the software cluster in the NSF Office of Advanced Cyberinfrastructure, which is co-funding this award.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.