NeTS: Small: A Programmable Network Data Plane for Resource Management in Datacenters

数据中心已经彻底改变了计算，允许小公司和研究团体利用数万台计算机的能力，大公司为数十亿用户提供广泛的服务。数据中心网络体系结构的研究对于以可接受的成本满足许多应用程序的严格性能要求非常重要。现代数据中心网络的一个关键方面是它们在网络内使用复杂的资源管理算法，如果没有这些算法，完成许多数据密集型计算任务将非常昂贵甚至不可能。 在过去的二十年里，研究人员、工程师和网络运营商一直在以健康的步伐开发新的资源管理方法。大部分工作需要更改网络路由器，其数据平面通常在硬件中实现。硬件缺乏灵活性，因此研究人员无法令人信服地展示他们的新想法，运营商也无法在实践中对其进行评估。目前可编程网络的两种方法-软件路由器和带有可编程元件的硬件路由器-都有缺点。到目前为止，软件路由器未能提供性能，至少比硬件慢一个数量级，而硬件路由器中的可编程元件无法提供方便的平台来表达大多数队列管理，调度和控制算法。为了解决这个问题，该提案引入了Flexplane，这是一种新的方式来构建能够支持各种资源管理方案的可编程网络数据平面。目标是支持在软件模拟器或软件路由器中以高级语言（如C++）指定的方案，但在真实的网络上以硬件速率运行它们。其关键思想是使用集中式仲裁器来模拟整个网络的抽象数据包行为，并将该行为反映到端点通过实际网络发送的真实的数据包上。使用Flexplane，数据包到达目的地的时间和报头模拟运行编程方案的路径上路由器。更广泛的影响：Flexplane对Facebook等公司的网络基础设施团队表现出了兴趣。这项研究的结果将直接使Facebook、谷歌、微软、Amazon.com等公司运营的网络中心受益。该教育计划包括将这项研究的结果引入本科课程，并为本科生提供在友好的编程环境中实现网络协议的机会，但让它们以高速率运行，这是迄今为止无法实现的能力。研究生阶段的教育目标是探索集中仲裁和数据平面可编程性的限制。 课程材料将通过麻省理工学院开放式课程和MITx MOOC广泛提供。

Datacenters have revolutionized computing, allowing small companies and research groups to harness the power of tens of thousands of computers, and large companies to provide a wide range of services to billions of users. Research on datacenter network architectures is important to meet the stringent performance requirements of many applications at acceptable cost. A key aspect of modern datacenter networks is their use of sophisticated resource management algorithms within the network, without which it would be prohibitively expensive or even impossible to accomplish many data-intensive computing tasks. Researchers, engineers, and network operators have been developing new methods for resource management at a healthy pace over the past two decades. Much of this work requires changes to network routers, whose data planes are typically implemented in hardware. Hardware is inflexible, so researchers cannot convincingly demonstrate their new ideas, and operators cannot evaluate them in practice. The two approaches available today for programmable networks---software routers and hardware routers with programmable elements---both have drawbacks. To date, software routers have failed to deliver on performance, being at least an order of magnitude slower than hardware, whereas programmable elements in hardware routers have not been able to provide a convenient platform to express most queue management, scheduling, and control algorithms. To address this problem, this proposal introduces Flexplane, a new way to architect a programmable network data plane capable of supporting a variety of resource management schemes. The goal is to support schemes specified in a high-level language like C++ as in a software simulator or a software router, but run them at hardware rates on real networks. The key idea is to use a centralized arbiter to emulate the behavior of an entire network with abstract packets, and reflect that behavior on to real packets sent by the endpoints over the actual network. With Flexplane, packets arrive at their destinations with the timings and headers that mimic on-path routers running the programmed schemes. Broader Impacts: Flexplane is of demonstrated interest to network infrastructure teams in companies like Facebook. The results of this research will directly benefit datacenters operated by companies like Facebook, Google, Microsoft, Amazon.com, and many other enterprises. The education plan includes the introduction of this research's findings into the undergraduate curriculum and offers undergraduates an opportunity to implement network protocols in a friendly programming environment, but have them run at high rates, a capability that was hitherto unavailable. The graduate-level education goals are to explore the limits of centralized arbitration and data-plane programmability. The course material will be made widely available through MIT OpenCourseWare and on the MITx MOOC.