EAGER: HAWKEYE: A Cross-Layer Resilient Architecture to Tradeoff Program Accuracy and Resilience Overheads

随着半导体技术扩展到亚纳米级，光刻图案和制造工艺中的微小扰动使计算硬件容易受到不可预测的功能偏差的影响。此外，随着功率限制的日益严格，硬件设计现在依赖于标称区域和接近阈值区域之间的动态电压调节，这加剧了可靠性问题。这些趋势推动了在不增加功耗的情况下实现系统弹性的需求。该项目提出了一种新的方法来实现高效的弹性，其中硬件的覆盖率在一定的保证下是有界的，而计算效率的代价是影响程序的准确性。该策略专注于数据分析应用程序，因为由于输入的非结构化性质，它们自然会在程序精度方面提供折衷，以及用于解决这些本来难以解决的问题的近似算法。在软件层面，该项目正在开发方法来对程序中的关键和非关键代码区域进行分类：关键代码影响程序的正确性和结果，其功能不能以任何方式改变；非关键代码影响程序的准确性，因此它可以容忍错误。硬件随后仅为关键代码实现更强的弹性方法(例如，冗余执行)。对于非关键代码，轻量级弹性方案确保了程序的正确性。如果成功，这个项目将是朝着新的严格框架迈出的重要一步，以推理和保证系统的弹性，以及程序的性能、能力和计算的准确性。拟议的研究通过将模块整合到现有的计算机体系结构和系统设计课程中，并通过已建立的REU站点和丰富方案进行推广，对课程开发和学生培训产生了重大的更广泛的影响。该项目承诺提供透明的弹性保证，这将产生重大的社会影响：企业和任务机构将能够以可量化的方式就其软件和硬件基础设施中的弹性与效率之间的权衡进行推理。鉴于医疗保健、国防、金融、交通和汽车等许多应用领域对系统可靠性的认识不断提高，这是向前迈出的关键一步。

With semiconductor technology scaling to sub-nanometer scales, minute perturbations in lithography patterns and manufacturing processes make the computing hardware vulnerable to unpredictable deviations in functionality. Further, with increasingly strict power constraints, hardware designs are now reliant on dynamic voltage scaling between nominal and near-threshold regions, which exacerbates the reliability problem. These trends motivate the need for system resiliency without increasing power consumption. This project proposes a new method to achieve efficient resiliency, where hardware's coverage is bounded with certain guarantees, while computational efficiency is traded off at the cost of affecting program accuracy. The strategy focuses on data analytic applications since they naturally offer tradeoffs in program accuracy due to the unstructured nature of inputs, and the approximate algorithms used to solve these otherwise intractable problems. At the software level, the project is developing methods to classify crucial versus non-crucial code regions in a program: the crucial code affects program correctness and outcome, its functionality cannot be altered in any way; non-crucial code affects program accuracy, therefore it could tolerate errors. The hardware subsequently implements stronger resiliency methods (e.g., redundant execution) only for crucial code. For the non-crucial code, lightweight resiliency schemes ensure program correctness. If successful, this project will be a major step towards a new rigorous framework to reason and guarantee system resiliency alongside a program's performance, power, and accuracy of computation. The proposed research provides for significant broader impacts related to curriculum development and student training through integration of modules in existing computer architecture and system design courses, and outreach through established REU sites and enrichment programs. The project promises transparent resiliency guarantees and this will have a major societal impact: enterprises and mission agencies will be able to reason in a quantifiable way about the resiliency versus efficiency tradeoffs in their software and hardware infrastructure. This is a crucial step forward given the increased awareness of system reliability in many application domains, such as healthcare, defense, finance, transportation, and automotive.