CAREER: Achieving Ultra-Low Latency under Heterogeneity and Uncertainty in Edge Computing

边缘计算被设想为云计算之外的一种范例，支持自动驾驶、增强现实和自动移动机器人等新兴应用。然而，为了实现边缘计算的预期延迟突破并将这种新范式投入运行，仍然缺少的关键部分是协调数据和计算以保证超低延迟的算法。本CAREER提案的总体目标是通过开发(i)动态协调边缘计算系统的数据和计算的编排算法来填补这一空白，以满足严格的延迟目标，以及（ii）描述边缘计算系统的基本资源需求和最佳操作点的理论基础。该提案中实现超低延迟的算法创新和配置见解将指导边缘计算系统的大规模部署，极大地受益于延迟敏感的边缘应用，这些应用具有强大的社会影响，如老年人和残疾人的认知辅助以及自动驾驶。本提案下的理论进展将为随机系统的研究做出基础性贡献，为电气工程、计算机科学和运筹学交叉的跨学科研究社区创造新的研究焦点。这项建议将对教育和社区产生重大影响。理论方法和实验平台都将被纳入卡内基梅隆大学研究生和本科生的课程和课程项目中。还将利用在线平台传播与该项目有关的教育和研究材料，以扩大影响范围。继续和扩大的努力将用于对K-12学生的STEM外展活动，指导来自代表性不足群体的学生进行研究，提高来自代表性不足群体的研究人员的知名度，并发起在线研讨会以向公众外展。该项目的目标是开发(i)动态协调边缘计算系统的数据和计算的编排算法，以满足严格的延迟目标，以及（ii）描述边缘计算系统的基本资源需求和最佳操作点的理论基础。具体而言，这一目标将分别在两种具有代表性的边缘系统运行模式（thrust I和thrust II）中实现，在此基础上，边缘节点被授权处理客户端生成的数据，并利用其计算能力。然后，通信和计算环境中的不确定性将在一个正交推力（推力III）基于学习的编排中得到解决。拟议的研究将导致目前缺失的算法创新和提供保证边缘计算系统超低延迟所需的见解。具体而言，将开发编排算法，共同动态地利用新兴的5G及以上无线技术下的通信资源以及边缘服务器和边缘客户端的分散计算能力。本提案中的横切方法的动机是观察到未来的边缘系统将是大规模的，并且该方法建立在最近大规模随机系统的重要结果之上。这些结果表明，使用正确的编排算法，可以在大型系统中同时实现超低延迟和高系统利用率。这一建议将进一步推进大规模随机系统的理论，以解决异质性、不确定性、不同类型资源之间的相互作用以及基于动态性能的工作执行。这些都是边缘系统和现代应用中出现的新的独特挑战，在传统方法中尚未得到充分探索。该奖项反映了美国国家科学基金会的法定使命，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。

Edge computing has been envisioned to be a paradigm beyond cloud computing that supports emerging applications such as autonomous driving, augmented reality, and automated mobile robots. However, to realize the envisioned latency breakthrough of edge computing and put this new paradigm into operation, a critical piece that is still missing is algorithms that orchestrate the data and the computation to guarantee ultra-low latency. The overall objective of this CAREER proposal is to fill this gap by developing (i) orchestration algorithms that dynamically coordinate data and computation for edge computing systems to meet stringent latency goals, and (ii) theoretical foundations to characterize the fundamental resource requirements and optimal operating points of edge computing systems. The algorithmic innovation and provisioning insights for achieving ultra-low latency in this proposal will guide the deployment of edge computing systems in large scale, greatly benefiting latency sensitive edge applications with strong societal impacts such as cognitive assistance for the elderly and disabled and autonomous driving. The theoretical advances under this proposal will make fundamental contributions to research in stochastic systems, creating new research focuses for interdisciplinary research communities at the intersection of electrical engineering, computer science, and operations research. This proposal will have significant educational and community impact. Both the theoretical approaches and the experiment platforms will be incorporated into the curriculum and course projects at graduate and undergraduate levels at Carnegie Mellon University. Online platforms will also be leveraged to disseminate educational and research materials related to this project for a greater reach. Continuing and expanded efforts will be spent on STEM outreach activities to K-12 students, mentoring students from underrepresented groups for research, promoting the visibility of researchers from underrepresented groups, and initiating online seminars to outreach to the general public.The goal of this project is to develop (i) orchestration algorithms that dynamically coordinate data and computation for edge computing systems to meet stringent latency goals, and (ii) theoretical foundations to characterize the fundamental resource requirements and optimal operating points of edge computing systems. In particular, this goal will be achieved in two representative operating modes of edge systems (Thrusts I and II), respectively, based on which edge nodes are authorized to process the data generated by clients and whose computing power is being exploited. Then the uncertainty in communication and computation environments will be addressed in an orthogonal thrust (Thrust III) learning-based orchestration. The proposed research will result in the currently missing algorithmic innovation and provisioning insights needed for guaranteeing ultra-low latency in edge computing systems. Specifically, orchestration algorithms will be developed to jointly and dynamically utilize the communication resources under the emerging 5G and beyond wireless technologies and the dispersed computing power of edge servers and edge clients. The cross-cutting approach in this proposal is motivated by the observation that future edge systems will be of large scale, and the approach builds upon significant recent results on large-scale stochastic systems. These results demonstrate that with the right orchestration algorithms, it is possible to achieve ultra-low latency and high system utilization simultaneously in large systems. This proposal will further advance the theory for large-scale stochastic systems to a much greater generality to address heterogeneity, uncertainty, interactions among different types of resources, and dynamic performance-based job execution. These are new unique challenges arising in edge systems and modern applications in general that are highly underexplored in traditional approaches.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.