CNS Core: Small: REYAZ: Reliability-Aware Job Scheduling for HPC Systems

CNS 核心：小型：REYAZ：HPC 系统的可靠性感知作业调度

• 批准号: 1910601
• 负责人: Devesh Tiwari
• 金额: $ 49.99万
• 依托单位: Northeastern University
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2019
• 资助国家: 美国
• 起止时间: 2019-10-01 至 2024-09-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1910601&HistoricalAwards=false
• 关键词: CNS; Core; Small; REYAZ; Reliability

High-performance computing (HPC) enables breakthroughs in different science domains that lead to an improvement in national economics, health, welfare, and defense. Unfortunately, the ability of HPC systems to deliver productive science is now beginning to be significantly hampered by hardware-related errors and failures. Consequently, computational applications of national importance will need to spend a large fraction of execution time in resilience mechanisms to make forward progress in the presence of failures. Despite that, a huge amount of resources and time will be wasted on future HPC systems due to the high frequency of failure interruptions during application execution. To address these challenges, this project, called REYAZ, explores new territory in HPC job scheduling: maximizing the amount of useful work done on reliability-constrained HPC systems by jointly exploiting dynamic reliability state of the system components and resilience characteristics of applications. REYAZ will enable two novel capabilities: (1) a reliability-aware job scheduling approach that optimizes useful work done per unit time on unreliable large-scale computing systems while individual applications are guaranteed "fair" performance. (2) a family of techniques to reduce the input/output (I/O) overhead - a side-effect of widely used resilience mechanisms such as checkpoint-restart - while retaining the performance improvements obtained via reliability-aware scheduling. Maximizing the useful work per unit time on future reliability-constrained HPC systems will directly translate into more productive science - leading to faster advancements of different science fields and societal impact. Capabilities developed in this project will also help reduce the wastage of energy on large-scale systems resulting in economic benefits for the society. This project will integrate the research tasks and outcomes into educational activities to train the next generation of engineers who will face the challenges of operating unreliable large-scale systems. Undergraduate students from underrepresented groups will be engaged and trained in the field of large-scale fault-tolerant parallel computing.The project website (https://github.com/GoodwillComputingLab/REYAZ) will host all the documentation of research findings and software artifacts developed as a part of the project, including system software, runtime systems, analytical tools, modeling methodologies, experimental data, and traces. The project website will be maintained actively for at least five years beyond the project end date.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.

高性能计算（HPC）可以在不同的科学领域实现突破，从而改善国民经济、健康、福利和国防。不幸的是，HPC系统提供生产科学的能力现在开始受到硬件相关错误和故障的严重阻碍。因此，具有国家重要性的计算应用程序将需要在弹性机制中花费很大一部分执行时间，以便在出现故障的情况下取得进展。 尽管如此，由于应用程序执行期间的高频率故障中断，未来的HPC系统将浪费大量的资源和时间。为了应对这些挑战，这个项目，称为REYA Z，探索新的领域在HPC作业调度：最大限度地提高可靠性约束的HPC系统所做的有用的工作量，通过联合利用系统组件的动态可靠性状态和应用程序的弹性特性。REYAZ将实现两个新的功能：（1）可靠性感知的作业调度方法，优化不可靠的大规模计算系统上单位时间内完成的有用工作，同时保证单个应用程序的“公平”性能。(2)一系列技术，用于减少输入/输出（I/O）开销-广泛使用的弹性机制（如检查点重启）的副作用-同时保留通过可靠性感知调度获得的性能改进。在未来可靠性受限的HPC系统上，最大限度地提高单位时间的有效工作量将直接转化为更有成效的科学，从而加快不同科学领域的进步和社会影响。 该项目开发的能力还将有助于减少大型系统的能源浪费，从而为社会带来经济效益。该项目将把研究任务和成果整合到教育活动中，以培养下一代工程师，他们将面临操作不可靠的大规模系统的挑战。该项目的网站（https：//github.com/GoodwillComputingLab/REYAZ）将托管作为项目一部分开发的所有研究成果和软件工件的文档，包括系统软件、运行时系统、分析工具、建模方法、实验数据和跟踪。该项目网站将在项目结束日期后至少五年内积极维护。该奖项反映了NSF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。

项目成果

Examining Failures and Repairs on Supercomputers with Multi-GPU Compute Nodes

检查具有多 GPU 计算节点的超级计算机上的故障和修复

• DOI: 10.1109/dsn48987.2021.00043
• 发表时间: 2021
• 期刊: 2021 51st Annual IEEE/IFIP International Conference on Dependable Systems and Networks (DSN
• 作者: Taherin, Amir and

Revisiting I/O behavior in large-scale storage systems: the expected and the unexpected

重新审视大规模存储系统中的 I/O 行为：预期和意外

• DOI: 10.1145/3295500.3356183
• 发表时间: 2019
• 期刊: Storage and Analysis
• 作者: Tirthak Patel, Suren Byna

Uncovering Access, Reuse, and Sharing Characteristics of I/O-Intensive Files on Large-Scale Production HPC Systems.

揭示大规模生产 HPC 系统上 I/O 密集型文件的访问、重用和共享特征。

• 发表时间: 2020
• 作者: Tirthak Patel, Suren Byna

Qraft: reverse your Quantum circuit and know the correct program output

• DOI: 10.1145/3445814.3446743
• 发表时间: 2021-04
• 期刊: Proceedings of the 26th ACM International Conference on Architectural Support for Programming Languages and Operating Systems
• 作者: Tirthak Patel;Devesh Tiwari

Operating Liquid-Cooled Large-Scale Systems: Long-Term Monitoring, Reliability Analysis, and Efficiency Measures

• DOI: 10.1109/hpca51647.2021.00078
• 发表时间: 2021-02
• 期刊: 2021 IEEE International Symposium on High-Performance Computer Architecture (HPCA)
• 作者: Rohan Basu Roy;Tirthak Patel;R. Kettimuthu;W. Allcock;Paul M. Rich;Adam Scovel;Devesh Tiwari