XPS: FULL: Integrating Programming Model, Runtime, Algorithmic, and Architectural Support To Use Inexact and Heterogeneous Hardware for Scientific Computations

XPS：完整：集成编程模型、运行时、算法和架构支持，以使用不精确和异构硬件进行科学计算

• 批准号: 1629392
• 负责人: Radu Teodorescu
• 金额: $ 87.5万
• 依托单位: Ohio State University
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2016
• 资助国家: 美国
• 起止时间: 2016-08-01 至 2021-07-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1629392&HistoricalAwards=false
• 关键词: XPS; FULL; Integrating; Programming; Model

As the High Performance Computing (HPC) field moves towards even more powerful (Exascale) systems, it faces two key challenges: resilience and power efficiency. Increasing computing power while not significantly increasing the power budget potentially involves new architectural designs, such as the ones with low power and/or low margins. Though future systems are expected to experience an increase in the number of faults (because of increasing number of cores and decreasing feature sizes), energy efficient designs will likely suffer even more errors due to tighter margins and other design compromises. Of particular concern are the errors that escape hardware detection: such errors are said to cause Silent Data Corruption (SDC). Maintaining correctness of a numerical simulation in the presence of SDCs is a very challenging problem. The intellectual merit of this project is in combining ideas from numerical methods, programming model design and architecture design to address the correctness of numerical simulations. The broader significance and importance of the project includes impact on scientific and high performance computing, system software for parallel computing, and architectural designs and research. This project will also make several contributions towards education, human resource development, and increasing diversity, with activities like teaching parallel computing (and programming) to diverse audience, mentoring of doctoral students, including those from underrepresented groups, and an interdisciplinary training program in Mathematical Biology for undergraduates. Technically, the project addresses the challenge of developing and executing scientific applications with energy efficient low-power/margin architectures that experience occasional faults, while maintaining programmer productivity and accuracy of results. This project develops a synergistic research program combining advances in HPC programming models, runtime systems, Near Threshold Voltage (NTV) architectures, and numerical methods (algorithms). Specifically, the project involves close collaboration between researchers from three areas: (a) parallel programming models, applications, and runtime systems, (b) architecture, and (c) finite difference and finite volume numerical models.

随着高性能计算（HPC）字段朝着更强大（Exascale）系统发展，它面临两个关键挑战：弹性和功率效率。增加计算能力的同时没有显着提高功率预算可能会涉及新的建筑设计，例如具有低功率和/或低利润率的建筑设计。尽管预计未来的系统会遇到故障数量的增加（由于核心数量的增加和功能尺寸减少），但由于更紧密的利润和其他设计妥协，节能设计可能会遭受更多的错误。特别令人担忧的是逃避硬件检测的错误：据说这种错误会导致无声数据腐败（SDC）。 在存在SDC的情况下，保持数值模拟的正确性是一个非常具有挑战性的问题。 该项目的智力优点在于结合了数值方法，编程模型设计和架构设计的思想，以解决数值模拟的正确性。该项目的更广泛的意义和重要性包括对科学和高性能计算的影响，并行计算的系统软件以及建筑设计和研究。该项目还将为教育，人力资源发展和不断增长的多样性做出一些贡献，诸如向多样化受众进行平行计算（和编程）的活动，包括博士生的指导，包括来自支出不足的群体的学生以及一项在本科生数学生物学领域的跨学科培训计划。 从技术上讲，该项目解决了开发和执行科学应用程序的挑战，该应用具有节能的低功率/利润体系结构，这些架构偶尔会出现故障，同时保持程序员的生产力和结果的准确性。该项目开发了一项协同研究计划，结合了HPC编程模型，运行时系统，接近阈值电压（NTV）架构和数值方法（算法）的进步。具体而言，该项目涉及来自三个领域的研究人员之间的密切合作：（a）并行编程模型，应用程序和运行时系统，（b）体系结构，以及（c）有限差和有限的卷数值模型。

项目成果

Using Undervolting as an on-Device Defense Against Adversarial Machine Learning Attacks

• DOI: 10.1109/host49136.2021.9702287
• 发表时间: 2021-07
• 期刊: 2021 IEEE International Symposium on Hardware Oriented Security and Trust (HOST)
• 作者: Saikat Majumdar;Mohammad Hossein Samavatian;Kristin Barber;R. Teodorescu