SHF: Medium: Quantifying and Designing Around Architectural Risk

计算机应用和技术正在以越来越快的速度发生变化。这种变化随之而来的是不确定性，这种不确定性使得人们很难就如何建立未来最有用的系统做出好的决定。虽然开发新的计算机系统总是涉及风险，但这些不确定性的新程度现在可能会导致一端的设计实践过于保守，或者另一端的设计具有“脆弱”的性能。虽然业务和投资都需要进行风险评估和管理，但这些方面通常被视为独立于计算机设计中的性能和效率问题，而实际上并非如此。随着硬件和软件特征变得不确定(即来自分布的样本)，由此产生的性能分布迅速增长，超出了我们仅凭直觉对其进行推理的能力。通过计算机系统设计者和专家在技术不确定性影响方面的合作，该项目正在开发新的基本技术，用于量化风险和以有风险意识的方式优化设计。该项目致力于改变从微型到数据中心规模的架构和系统的设计和分析方式。调查人员正在创造新的技术，但也使这些技术可通过开放知识库获得和使用，让各级本科生参与他的研究，并通过“我爱STEM”和其他努力，将这些统计设计方法的基本概念纳入外联活动。这个跨学科项目的目标是推进计算机体系结构、电子设计自动化和不确定性量化的前沿。所开发的方法还可以在原始调查领域之外找到应用，例如，在许多其他工程系统中进行影响风险分析和管理，包括可再生能源、机器人系统和自动驾驶。这项工作首次证明，定义、建模、量化和缓解计算机体系结构风险是可能的。通过将经济学中的风险和风险管理思想与不确定性量化中的快速随机算法相结合，正在创建一个新的高级别风险感知计算机体系结构分析框架。使用少于50个数据点的高效技术可以有效地估计体系结构的不确定性，从而能够在计算机体系结构设计期间对风险进行严格的量化和管理。这一框架体现在一个符号/统计分析系统中，该系统简化了对这些令人惊讶的复杂设计空间的探索。声明性语言抽象了新的概率约束优化和智能抽样方法的复杂性，而具有风险意识的微观和宏观架构改进在实践中展示了这些方法的价值。这些方法改变了人们从每个制造商可用的有限数据点中提取有用的不确定性模型的方式，通过系统的复杂组成和资源的交互来正确地传播这些不确定性，有效地量化这些传播的不确定性对系统最终品质因数的影响，将这些方法封装在语言和求解器框架中，并开发如何更积极地减轻和控制此类不确定性的范例。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。

Computer applications and technologies are changing at an ever-increasing rate. This change comes with uncertainty and that uncertainty makes it difficult to make good decisions about how to build systems that are maximally useful in the future. While developing new computer systems has always involved risk, the new magnitude of these uncertainties may now lead to either overly conservative design practices at one end, or designs that have 'fragile' performance at the other end. While risk assessment and management are expected in both business and investment, these aspects are typically treated as independent to questions of performance and efficiency in computer design when in fact they are not. As hardware and software characteristics become uncertain (i.e. samples from a distribution), the resulting performance distributions quickly grow beyond our ability to reason about them with intuition alone. Through the collaboration of computer system designers and experts in the impacts of technology uncertainty, this project is developing new and fundamental techniques for both quantifying risk and optimizing designs in risk-aware ways. This project is working to transform the way in which architectures and systems, from micro to data-center scale, are designed and analyzed. The investigators are creating new technologies, but also making those technologies available and accessible through open repositories, involving undergraduates at all levels in his research, and integrating basic concepts from these statistical design methods into outreach through "I love STEM" and other efforts. This interdisciplinary project is targeted at advancing the frontiers of computer architecture, electronic design automation, and uncertainty quantification. The developed methodologies can also find application far outside the original area of inquiry e.g., impact risk analysis and management across many other engineering systems including renewable energy, robotic systems, and autonomous driving.The work is demonstrating, for the first time, that it is possible to define, model, quantify, and mitigate computer architectural risk. By bridging ideas of risk and risk-management from economics and fast stochastic algorithms from uncertainty quantification, a new framework for high-level risk-aware computer architecture analysis is being created. Efficient techniques, using fewer than 50 data points, can effectively estimate architectural uncertainty to enable the rigorous quantification and management of risk during computer architecture design. This framework is embodied in a symbolic / statistical analysis system that eases the exploration of these surprisingly complex design spaces. A declarative language abstracts away the complexity of new probability-constrained optimization and intelligent sampling methods, while risk-aware micro and macro architectural improvements demonstrate the value of these methods in practice. The methods transform the way one can extract useful models of uncertainty from a limited selections of data points available to each manufacturer, propagate those uncertainties correctly through complex compositions of systems and the interactions of resources, efficiently quantify the impact of those propagated uncertainties on the end figures of merit for the system, encapsulate those methods in a language and solver framework, and develop exemplars of how such uncertainty can be more actively mitigated and controlled.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.