CAREER: Application-centric, Reliable and Efficient High Performance Computing

职业：以应用为中心、可靠且高效的高性能计算

• 批准号: 1553645
• 负责人: Dong Li
• 金额: $ 50万
• 依托单位: University of California - Merced
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2016
• 资助国家: 美国
• 起止时间: 2016-02-01 至 2023-01-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1553645&HistoricalAwards=false
• 关键词: CAREER; Application; centric; Reliable; Efficient

Mission-critical scientific simulations (e.g., climate simulation and fluid dynamics simulation) and enterprise workloads (e.g., search and encryption) running on large-scale computing systems are jeopardized by the increase of faults and errors in hardware and software. Understanding the vulnerability of these large-scale applications is important to minimize performance and power. Lack of the knowledge of application vulnerability forms a major bottleneck of execution efficiency, and jeopardizes HPC simulation capabilities. Previous works rely on random fault injection or detailed architecture analysis to evaluate application vulnerability. They can be slow and inaccurate. There is a big gap between the needs of reliable and efficient HPC and what the current methodologies can provide. This research explores a new methodology to understand application vulnerability. It investigates new analytical and statistical models to quantify and characterize application vulnerability based on a novel metric and application semantics (including algorithm semantics and data semantics). The PI integrates modeling techniques into a broader context for vulnerability analysis to improve the modeling accuracy and explore reliable and efficient protection for applications while examine the interplay between reliability, power, and performance.The outcome from this research will provide support for execution correctness and efficiency of large-scale applications running on future computing systems that demand high data integrity. The proposed research will affect design of reliable applications and algorithms. Built upon the collaboration with industry, the research outcome is expected to be tangible and have direct impact on realistic scientific problems. Furthermore, the tight coupling between research components and education components creates a HPC learning culture to engage students in HPC, addressing HPC workforce shortage in the nation.

关键任务科学模拟（例如，气候模拟和流体动力学模拟）和企业工作负载（例如，在大规模计算系统上运行的企业工作负载（例如，搜索和加密））因硬件和软件中的故障和错误的增加而危害了大规模计算系统。了解这些大规模应用的脆弱性对于最大程度地降低性能和力量很重要。缺乏应用程序脆弱性的知识构成了执行效率的主要瓶颈，并危害了HPC模拟功能。以前的工作依靠随机的故障注入或详细的体系结构分析来评估应用程序漏洞。它们可能会很慢且不准确。可靠和有效的HPC的需求与当前方法可以提供的需求之间存在很大的差距。这项研究探讨了一种了解应用程序漏洞的新方法。它研究了新的分析和统计模型，以量化和表征基于新型指标和应用语义（包括算法语义和数据语义）的应用程序漏洞。 PI将建模技术集成到更广泛的环境中，以提高建模准确性，并探索对应用程序的可靠和有效保护，同时研究可靠性，权力和性能之间的相互作用。这项研究的结果将为实现高度数据完整性的未来计算系统的执行性和效率提供支持。拟议的研究将影响可靠的应用和算法的设计。基于与行业的合作，预计研究结果将是有形的，并直接影响现实的科学问题。此外，研究组件和教育组成部分之间的紧密结合创造了HPC学习文化，使学生参与HPC，以解决全国HPC劳动力短缺。

项目成果

Lobster: Load Balance-Aware I/O for Distributed DNN Training

• DOI: 10.1145/3545008.3545090
• 发表时间: 2022-08
• 期刊: Proceedings of the 51st International Conference on Parallel Processing
• 作者: Jie Liu-;Bogdan Nicolae;Dong Li