CNS Core: Small: Integrating Real-Time Learning and Control for Large and Dynamic Networked Computer Systems

大型计算机和网络系统（例如数据中心）是推动信息社会发展的主力。然而，由于其巨大的规模、快速变化的工作负载以及资源需求和可用性的巨大不确定性，它们也越来越难以控制和操作。传统的控制和优化方法依赖于精心构建的所研究系统的模型，但在快速变化的环境中，当系统模型的关键组件未知或不断变化时，它们就变得不够了。相反，该项目旨在开发新方法，可以从新鲜的实时数据中快速学习更新的模型，并将这种实时学习与实时控制相结合，以提高大规模动态网络计算机系统的效率、适应性和服务质量（QoS）。具体来说，该项目重点关注服务于大数据分析和深度学习训练工作负载的大型数据中心的运营，并开发新的实时学习和随机控制策略，这些策略不仅高效，而且可扩展、能够解释和自适应。 智能优点包括：（i）实时学习和控制策略，可以从实时反馈、计算和网络作业的服务器相关特征中学习，从而大大提高运行大型异构工作负载的数据中心的吞吐量，减少作业完成时间并满足服务期限； （ii）针对深度学习训练工作负载的独特特征量身定制的实时学习和控制策略，可以快速估计总训练时间和异构处理单元之间的依赖性，以优化吞吐量和延迟。所提出的研究有可能对知识发现和教育产生持久影响。结果可以使数据中心更快地运行作业并更快地完成它们，从而使计算行业受益，既可以提高运行多样化且快速变化的工作负载的数据中心的整体效率，也可以提高依赖数据中心进行业务决策和数据分析的用户的满意度。研究结果可能有助于在线学习和随机控制的一般理论，这对于具有不确定系统动力学和不确定代理特征的其他计算机和网络系统也很有用，例如无线网络和在线服务平台。该项目的学生将接受理论工具（包括在线学习、随机控制和数据分析）和系统构建技能（包括集群计算和数据中心网络）方面的培训，这对于未来的大数据经济至关重要。此外，与计算研究相结合的外展活动将拓宽高中生对在线学习和大数据关键原理的知识。该奖项反映了 NSF 的法定使命，并通过使用基金会的智力价值和更广泛的影响审查标准进行评估，被认为值得支持。

Large computer and network systems (such as data centers) are the workhorses driving our information society. However, they are also increasingly difficult to control and operate due to their enormous size, fast-changing workload, and significant uncertainty in resource requirement and availability. Traditional approaches to control and optimization rely on carefully-constructed models of the systems under study, but they become insufficient in such a fast-changing environment when crucial components of the system model are either unknown or constantly changing. Instead, this project aims to develop new methods that can quickly learn an updated model from fresh real-time data, and that integrate such real-time learning with real-time control to improve the efficiency, adaptability, and quality-of-service (QoS) of large-scale and dynamic networked computer systems. Specifically, the project focuses on the operation of large data centers serving big-data analytics and deep-learning training workloads, and develops new real-time learning and stochastic control policies that are not only efficient, but also scalable, able to interpret, and adaptive. The intellectual merits include: (i) real-time learning and control policies that can learn, from real-time feedback, server-dependent features of the computing and network jobs, to greatly improve the throughput of data centers running large and heterogeneous workload, reduce job completing times, and meet service deadlines; and (ii) real-time learning and control policies tailored to the unique features of deep-learning training workload, which can quickly estimate the total training time and the dependency across heterogeneous processing units, to optimize both throughput and delay.The proposed research has the potential to have a lasting impact to knowledge discovery and education. The results could enable data centers to run jobs faster and complete them sooner, and therefore benefit the computing industry, both by improving the overall efficiency of data centers running diverse and fast-changing workload, and by improving the satisfaction of users who rely on data centers for business decisions and data analytics. The research findings may contribute to the general theory of both online learning and stochastic control, which will also be useful for other computer and network systems with both uncertain system dynamics and uncertain agent features, such as wireless networks and online service platforms. Students on the project will be trained on both theoretic tools (including online learning, stochastic control, and data analytics) and system building skills (including cluster computing and data-center networking), which are essential for the future big-data economy. Further, the outreach activity integrated with the research computed will broaden the knowledge of high school students on the key principles of online learning and big-data.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.