XPS: EXPL: CCA: Collaborative Research: Nixing Scale Bugs in HPC Applications

XPS：EXPL：CCA：协作研究：消除 HPC 应用程序中的规模错误

• 批准号: 1439002
• 负责人: Ganesh Gopalakrishnan
• 金额: $ 15万
• 依托单位: University of Utah
• 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 2018-08-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1439002&HistoricalAwards=false
• 关键词: XPS; EXPL; CCA; Collaborative; Research

Large-scale simulation is a fundamental component of modern science and engineering. Unfortunately, programs written to perform simulations on large-scale parallel computers frequently suffer from software defects that result from the sheer scale and the variety of parallelization approaches employed. Especially egregious are software bugs that occur when large resource allocations (e.g., memory requests) are made. Formally based active-testing techniques are essential to locate such defects. However, these testing tools are themselves seldom run on parallel machines, let alone at large scale, making it difficult and very time consuming to find scale bugs with high assurance. Efforts to parallelize verification tools should reuse existing technology for easy parallelization, result collection, and fault handling. Key innovations of this project include the insight that large-scale verification runs can be described through work-flows, which makes it possible to take advantage of already available distributed computing platforms, in particular Swift/T from Argonne. The complementary backgrounds of the PIs are well matched with the need to push both formal aspects and distributed verification in the context of three widely-used concurrency models, namely MPI, OpenMP, and CUDA. This work will help create a public distributed formal active testing framework. The tools and case-study software driving this research will be maintained by the PIs and released freely under open-source licenses through websites and repositories. They will facilitate large-scale debugging of scientific simulation codes by researchers and software developers in academia, government labs, and industry. The project will also generate pedagogical material and best practices, helping educate students in the use of existing work-flow based problem solving approaches. It will help train present and future scientists, engineers, and programmers, thus assisting in maintaining our nation's leadership in computing, homeland and energy security, and STEM education.

大规模仿真是现代科学与工程的重要组成部分。不幸的是，程序编写执行大规模并行计算机上的模拟经常遭受软件缺陷，导致纯粹的规模和各种并行化方法。尤其令人震惊的是当大量资源分配时发生的软件错误（例如，存储器请求）。基于形式的主动测试技术对于定位此类缺陷至关重要。然而，这些测试工具本身很少在并行机上运行，更不用说在大规模上运行，这使得很难并且非常耗时地找到具有高保证的规模错误。并行化验证工具的努力应该重用现有技术，以便于并行化、结果收集和故障处理。该项目的主要创新包括：大规模验证运行可以通过工作流来描述，这使得利用现有的分布式计算平台成为可能，特别是来自阿贡的Swift/T。PI的互补背景与在三种广泛使用的并发模型（即MPI，OpenMP和CUDA）的背景下推动正式方面和分布式验证的需求非常匹配。这项工作将有助于创建一个公共的分布式正式的主动测试框架。推动这项研究的工具和案例研究软件将由PI维护，并通过网站和存储库在开源许可证下免费发布。它们将促进学术界、政府实验室和工业界的研究人员和软件开发人员对科学模拟代码进行大规模调试。该项目还将产生教学材料和最佳做法，帮助教育学生使用现有的基于工作流程的问题解决方法。它将帮助培养现在和未来的科学家，工程师和程序员，从而帮助保持我们国家在计算，国土和能源安全以及STEM教育方面的领导地位。