Collaborative Research: CISE: Large: Cross-Layer Resilience to Silent Data Corruption

超缩放器(即大型云服务提供商)报告其数据中心基础设施内频繁出现静默数据损坏(或SDC)。SDC是软件错误，其唯一症状是结果不正确。值得注意的是，规模上的SDC的错误发生率约为每一百万台设备发生1000个故障。与此同时，硬件制造商努力在商业和汽车领域分别实现100%和接近于零的瑕疵部件。制造商的目标和超标度器的观察之间的这种差异表明，SDC是对所有现代计算系统的可靠性的真正威胁，进而威胁到它们的安全性和可持续性。该项目探索是否有可能合作地为SDC设计测试、检测和缓解方法，以最大限度地减少对软件应用程序的性能影响，以及与制造和运行计算系统相关的额外碳足迹支出。该项目的主要创新包括：(1)利用软件(如流行的机器学习应用中的矩阵乘法)和现代专用硬件(如人工智能处理器)中重复出现的计算基元来设计特定领域的SDC解决方案；(2)利用SDC测试可以在数据中心的整个设备生命周期中执行这一事实，而不是几秒钟到几分钟-这在制造测试平台上是一个严格的限制；(3)考虑将可持续性和碳足迹作为核心设计指标。该项目的核心影响将是我们今天委托计算系统使用的无数应用程序在可靠性和安全性方面的关键改进。第二个核心影响是计算设备寿命的提高，这对可持续计算具有重要的积极影响。研究团队还将培训学生，并与行业合作伙伴合作。为了应对SDC的挑战，研究团队进行了四项跨越不同领域的协同研究：硅设备、计算机体系结构、软件和算法。在每一次推进中，团队将通过以下几个方面来研究SDC挑战：测试、检测、缓解和安全影响。推力1通过新颖的测试模式指标和连续扫描测试部署来探索设备级测试。推力2研究系统级测试(改进错误检测延迟和测试覆盖率，并调整测试以更好地代表数据中心工作负载)、特定于核心的测试、缺陷表征、用于测试和缓解的硬件支持，以及系统安全影响。推力3通过(部分)冗余、适当的扫描和系统级测试调度、测试-应用程序融合(应用程序自我测试)以及针对缺陷导致的漏洞的软件安全强化来调查软件检测和缓解。Struts 4致力于算法检测和缓解，特别强调为重要的数据中心工作负载(如神经网络)启用稳健的非线性计算。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Hyperscalers (i.e., large cloud service providers) are reporting frequent silent data corruptions (or SDCs) within their datacenter infrastructures. SDCs are software errors for which the only symptom is an incorrect result. Remarkably, SDCs at-scale exhibit error occurrence rates on the order of one thousand faults per one million devices. Meanwhile, hardware manufacturers strive to achieve one hundred and close to zero defective parts per million for the commercial and automotive domains, respectively. This discrepancy between manufacturers’ goals and hyperscalers’ observations suggests that SDCs are a real threat to the reliability of all modern computing systems, and by extension their security and sustainability. This project explores whether it is possible to cooperatively design testing, detection, and mitigation approaches for SDCs that minimize performance impact on software applications, as well as additional carbon footprint expenditures associated with manufacturing and running computing systems. The project’s key novelties include: (1) leveraging reoccurring computational primitives in software (e.g., matrix multiplication in popular machine learning applications) and modern special-purpose hardware (e.g., Artificial Intelligence processors) to design domain-specific SDC solutions; (2) exploiting the fact that SDC testing can be performed throughout a device’s lifetime in the datacenter rather than for a few seconds to minutes — a strict limitation on the manufacturing test floor; (3) considering sustainability and carbon footprint as a core design metric. This project’s core impact will be a critical improvement in reliability and security for the countless applications to which we entrust computing systems today. A secondary core impact is an improvement in the longevity of computing devices, which has significant positive implications for sustainable computing. The research team will also train students and work with industry partners. To address the SDC challenge, the research team pursues four synergistic research thrusts that cut across diverse domains: Silicon Devices, Computer Architecture, Software, and Algorithms. Within each thrust, the team will study the SDC challenge through the lenses of: Testing, Detection, Mitigation, and Security implications. Thrust 1 explores device-level testing through novel test pattern metrics and continuous scan test deployment. Thrust 2 studies system-level testing (improving error detection latency and test coverage and adapting tests to be more representative of datacenter workloads), core-specific testing, defect characterization, hardware support for testing and mitigation, and system security implications. Thrust 3 investigates software detection and mitigation through (partial) redundancy, appropriate scan and system-level test scheduling, test-application fusion (where applications test themselves), and software security hardening against defect-induced vulnerabilities. Thrust 4 pursues algorithmic detection and mitigations with a particular emphasis on enabling robust non-linear computation for important datacenter workloads, like neural networks.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.