CAREER: Green Functions as a Service: Towards Sustainable and Efficient Distributed Computing Infrastructure

云计算不断增长的碳排放足迹（占全球排放量的2%）威胁着可持续性，尤其是随着高能耗的人工智能（AI）和物联网（IoT）应用的兴起。虽然可再生能源的采用正在增加，但由于可变性和位置依赖性，将其有效整合到网络基础设施中仍然是一个障碍。拟议的工作将为分布式计算基础设施奠定科学和技术基础，该基础设施既高效又可持续，将碳作为全系统的一级目标。它将使网络服务、人工智能、物联网和数据分析等低碳应用程序的无缝部署和使用成为可能。通过研究产生的模型、软件和数据集将是开源的，并集成到本科课程和研究中。利用新的可扩展教学软件，如“政策健身房”，我们将为计算机工程、人工智能和可持续发展领域的本科生提供跨学科的实践培训。该项目将开发“绿色功能即服务”，这是一种新的抽象，用于在边缘云连续体上脱碳对延迟敏感的应用程序。我们的方法将以需求响应、碳定价和生态反馈等可持续发展的基本原则为基础，并使用现代人工智能技术，如替代模型进行碳建模和优化。我们将扩展无服务器计算的新功能，如多态功能，在异构CPU和GPU架构上实现碳高效执行。我们的分布式资源管理算法将使用时空碳和工作负载建模和优化。地理负载平衡将结合机器学习和碳信用额，为分布式网络基础设施提供碳和性能管理。我们多方面的研究和教育计划将在系统设计和教学中引入关键的可持续性原则，并为可持续的数字世界做出贡献。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。

Cloud computing's growing carbon gas emissions footprint (2% of global emissions) threatens sustainability, especially with the rise of energy-hungry artificial intelligence (AI) and internet of things (IoT) applications. While renewable energy adoption is increasing, integrating it effectively into cyberinfrastructure remains a hurdle due to variability and location dependence. The proposed work will lay the scientific and technical foundations of a distributed computing infrastructure which is both efficient and sustainable, using carbon as the first-order system-wide objective. It will enable the seamless deployment and use of low-carbon applications such as web services, AI, IoT, and data analytics. The models, software, and datasets produced through the research will be open-sourced and integrated into undergraduate curriculum and research. Using new scalable pedagogical software such as "policy gyms", we will provide cross-disciplinary hands-on training to undergraduate students in the fields of computer engineering, AI, and sustainability.The project will develop "Green Functions as a Service", a new abstraction for decarboninzing latency-sensitive applications on the edge-cloud continuum. Our approach will be grounded in fundamental principles of sustainability such as demand response, carbon pricing, and eco-feedback, and use modern AI techniques such as surrogate models for carbon modeling and optimization. We will extend serverless computing with new capabilities such as polymorphic functions for carbon-efficient execution on heterogeneous CPU and GPU architectures. Our distributed resource management algorithms will use spatio-temporal carbon and workload modeling and optimization. The geographical load balancing will combine machine learning and carbon credits to provide carbon and performance management for distributed cyberinfrastructure. Our multi-faceted research and education plan will introduce key sustainability principles to both system design and pedagogy, and contribute to a sustainable digital world.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.