CAREER: Learning-Based Hardware and Software Techniques for Quality-of-Service-Aware Cloud Microservices

数据中心支持世界数字计算能力的很大一部分，并且不断增长，包括搜索引擎、社交网络和机器学习分析。随着现代云服务越来越受欢迎，它们的设计从支持复杂的单片应用程序转变为支持专门的、松耦合的微服务集合。这种微服务通过要求快速的网络处理和低延迟的内存访问来实现其服务质量（QoS）约束，从而影响资源需求。微服务之间的依赖关系也会使计算集群管理复杂化，并可能导致级联QoS违规，损害可用性和服务可靠性。在有效使用数据中心的同时保证云服务的预期响应需要一种联合的硬件软件方法。该项目从整体角度出发，为运行在大规模数据中心上的交互式云微服务设计一个系统堆栈，该系统具有qos意识，并且资源高效。通过追求自动化、基于学习的技术，该项目强调了利用实用机器学习技术来更好地驾驭日益复杂的云计算的价值，因为越来越多的数据中心服务切换到这种新的应用程序模型。在硬件层面，本项目首先量化了微服务对服务器设计的影响，其次，探索了它们在硬件加速方面的潜力。在软件层面，这项工作是开发一个新的集群管理器，它以一种自动化的、对用户透明的方式来处理微服务之间的依赖关系，并保证端到端的性能。最后，为了消除微服务之间违反QoS的级联效应，该项目包括一个数据驱动的在线性能预测系统。该系统利用云系统收集的大量监控数据来预测即将发生的QoS违规，并在它们降低性能之前对其采取行动。通过在硬件和软件上进行创新，这项工作将获得仅靠硬件和软件无法提供的性能和效率提升。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。

Datacenters support a large and ever-increasing fraction of the world's digital computation power, including search engines, social networks, and machine learning analytics. As modern cloud services grow in popularity, their design shifts from supporting complex monolithic applications, to supporting collections of specialized, loosely-coupled microservices. Such microservices impact resource requirements by requiring fast network processing and low-latency memory accesses to achieve their quality-of-service (QoS) constraints. Dependencies among microservices also complicate compute cluster management, and can cause cascading QoS violations, hurting availability and service reliability. Guaranteeing the responsiveness expected from cloud services while using datacenters efficiently requires instead a joint hardware-software approach. This project takes a holistic view towards designing a system stack for interactive cloud microservices running on large-scale datacenters that is QoS-aware, and resource-efficient. By pursuing automated, learning-based techniques, this project highlights the value of leveraging practical machine learning techniques to better navigate the increasing complexity of the cloud, as more datacenter services switch to this new application model.At the hardware level, this project first quantifies the implications microservices have on server design, and second, explores their potential for hardware acceleration. At the software level, this work is developing a new cluster manager that accounts for the dependencies among microservices in an automated and transparent-to-the-user way, and guarantees end-to-end performance. Finally, to eliminate the cascading effects of QoS violations between microservices, this project includes a data-driven, online performance forecasting system. This system leverages the massive amount of monitoring data collected by cloud systems to anticipate upcoming QoS violations, and act on them before they degrade performance. By innovating in both hardware and software, this work will achieve performance and efficiency gains that neither hardware- nor software-only approaches can provide.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.