CAREER: A Platform for Per-Packet AI using Heterogeneous Data Planes

现代网络基础设施（CI）为我们日常生活的许多方面（如医疗保健、金融、电力和通信）提供动力。而且，展望未来，随着新的用例（例如，增强/虚拟现实，自动运输和远程手术）和用户（例如，自动驾驶汽车和物联网设备）进入技术领域，我们对这种基础设施的依赖将会增加。为了满足这种不断发展的环境的严格安全性和性能需求，云数据中心网络和系统（构成CI的支柱）必须快速有效地适应和分配其（异构）资源。然而，这样做要求（计算密集型）网络管理和控制决策以快速和智能的方式以线路速率应用于每个数据包。不幸的是，目前可用的主流解决方案既不够快，也不够智能，无法满足这些需求。该提案旨在通过开发一个整体平台来弥合速度和智能之间的差距，该平台允许数据中心运营商直接在网络内以线路速率执行每包人工智能驱动的决策。该提案提出了三个逐步连接的研究方向：(1)为每包人工智能设计新颖的数据平面架构，(2)实现用于表达人工智能目标（和模型）的高级声明性框架，最后，(3)开发一套每包人工智能应用程序，以建立对所提议平台效用的信心。这是一个全新的范式，融合了多个学科（机器学习、网络和架构），从而为机器学习研究人员和架构师打开了一条途径，让他们拥有跨学科的知识，与网络设计师一起工作，实现每包人工智能的全部潜力。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。

Modern cyberinfrastructure (CI) powers many aspects of our day-to-day life (such as healthcare, finance, electricity, and communication). And, moving forward, our reliance on this infrastructure will grow even more as new use cases (e.g., augmented/virtual reality, autonomous transportation, and remote surgery) and users (e.g., self-driving cars and IoT devices) enter the technological landscape. To meet the strict security and performance requirements of this evolving landscape, the cloud datacenter networks and systems, which form the backbone of CI, must adapt and allocate their (heterogeneous) resources quickly and efficiently. Doing so, however, demands that (compute-intensive) network management and control decisions are applied per packet at line rate in a fast-and-intelligent way. Unfortunately, the dominant solutions available today are neither fast nor intelligent enough to meet these requirements.This proposal aims to bridge this gap between speed and intelligence by developing a holistic platform that allows datacenter operators to execute per-packet AI-driven decisions directly within the network at line rate. The proposal lays out the research across three progressively connected thrusts: (1) designing novel data-plane architectures for per-packet AI, (2) implementing high-level, declarative frameworks for expressing AI objectives (and models), and, finally, (3) developing a suite of per-packet AI applications to build confidence in the utility of the proposed platform. It is a radically new paradigm that converges multiple disciplines (machine learning, networking, and architecture), thus opening pathways for machine-learning researchers and architects—with their cross-disciplinary knowledge—to work alongside network designers to realize the full potential of per-packet AI.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.