NeTS: Medium: Collaborative Research: Network Configuration Synthesis: A Path to Practical Deployment

社会的所有部门都依赖于正常运行的计算机网络。例如，每天都有数以百万计的公民通过云订购处方药、支付电费、预订酒店、购买食品杂货，并参与成千上万的在线活动。但是，如果传递信息的网络宕机，这些服务都将无法工作。此外，现代企业、医疗保健、军队和政府与普通公民一样依赖可靠的网络。许多网络中断是由操作员手动（并且不正确地）编程“配置文件”引起的，该文件管理网络设备转发信息的方式。虽然配置文件允许的灵活性是必不可少的，但网络中断通常是由操作员在每个网络设备上使用数百个低级指令来创建网络范围的策略造成的。由于即使是很小的配置更改也会对全局造成巨大影响，因此许多组织需要花费数周的时间来审核即使是很小的更改，这限制了他们有效响应流量波动、业务机会、安全威胁和硬件故障的能力。这些问题的自然解决方案——类似于过去几十年程序员从机器码转向Java的软件开发编程语言的趋势——是定义更健壮、更高级的编程语言来实现网络策略。然而，要想在工业环境中大规模部署新语言，还需要克服一些技术和务实的障碍。特别是，对于许多期望的网络策略，现有的全网策略语言的表达能力不够，通常需要大规模迁移到新的网络平台。因此，这个项目的首要目标是克服阻碍高级网络编程语言实际部署的技术挑战。该项目正在开发有效支持和增量部署高级网络策略所需的核心技术。该项目利用与两家主要云提供商的连接，在实际工业网络上测试生成的语言和系统，确定采用的实际障碍，并最终在可能的情况下部署该技术。该项目基于pi最近在丙烷上的工作，丙烷是一种新的网络编程语言，允许用户描述域内和域间路由的端到端路径，以及为行业标准BGP协议生成配置的编译器。该项目的成果将以多种方式扩展丙烷，以支持实际部署：首先，用户将能够声明设备角色（例如，顶架交换机）以及与之相关的连接不变量，以实现简洁的规范。新的编译器将在存在此类声明的情况下验证策略的安全属性，并生成参数化模板，使编译器输出对操作符更容易理解。其次，用户将使用一种新的声明性语言指定管理运输成本的金融合同，编译器将通过生成符合目标的优化策略来自动优化路线。第三，Butane编译器将瞄准并利用异构后端协议和平台的优势。第四，工具将帮助网络运营商从现有的低级配置中推断出新的高级配置，并验证新配置是否等同于旧配置。最后，Butane将支持混合模式（传统和高级网络操作），因此工程师可以随着时间的推移慢慢迁移他们的网络，并在一小部分实时流量上测试部分部署。

All sectors of society depend on properly functioning computer networks. For example, every day, millions of citizens order prescription drug refills, pay their electricity bills, book hotels, shop for groceries, and participate in thousands more activities online, through the cloud. But none of these services will work if the networks that deliver information are down. Moreover, modern business, healthcare, the military and the government are just as dependent on reliable networks as everyday citizens. Many network outages are caused by operators manually (and incorrectly) programming the 'configuration files' that manage the ways that network devices forward information. While the flexibility allowed by configuration files is essential, network outages are often caused by operators using hundreds of low-level directives at each network device to create network-wide policy. Because the global consequences of making even small configuration changes is so drastic, many organizations take several weeks to audit even small changes, limiting their ability to respond effectively to traffic fluctuations, business opportunities, security threats and hardware failures. A natural solution to these problems -- analogous to the trend in programming languages for software development over the last several decades as programmers have moved from machine code to Java -- is to define more robust, higher-level programming languages for implementing network policies. However, there are technical and pragmatic hurdles to surmount before it will be possible to deploy new languages in industrial settings on a large scale. In particular, existing network-wide policy languages are not expressive enough for many desired network policies and often require wholesale migration to new networking platforms. Hence, the overarching goal of this project is to surmount the technical challenges that impede practical deployment of high-level network programming languages. The project is developing the core technology necessary to efficiently support and incrementally deploy high-level network policies. The project leverages connections to two major cloud providers as a means to test the resulting languages and systems on real industrial networks, identify pragmatic barriers to adoption, and ultimately deploy the technology where possible.The project builds on the PIs' recent work on Propane, a new network programming language that allows users to describe end-to-end paths for intra- and inter-domain routing, along with a compiler that produces configurations for the industry-standard BGP protocol. The results of this project will extend Propane in several ways to support practical deployment: First, users will be able to declare device roles (e.g., top-of-rack switch) and the connectivity invariants related to them to enable concise specifications. A new compiler will verify safety properties of policies in the presence of such declarations and generate parameterized templates that make compiler outputs more intelligible for operators. Second, users will specify financial contracts that govern transit costs using a new declarative language and the compiler will optimize routes automatically by generating refined policies that meet objectives. Third, the Butane compiler will target and exploit the benefits of heterogeneous back-end protocols and platforms. Fourth, tools will help network operators infer new high-level configurations from existing low-level configurations and to verify that new configurations are equivalent to old ones. Finally, Butane will support mixed mode (legacy- and high-level network operations) so engineers can migrate their networks slowly over time and test partial deployment on small fractions of their live traffic.