UniFaaS: A Unikernel-Based Serverless Operating System

无服务器计算，也被称为功能即服务（FaaS），是一种新兴的编程范式，在降低成本、数据中心效率、可扩展性等方面为租户（即应用程序开发人员）和提供商提供了显著的好处。凭借其真正的按需资源消耗和定价模式，以及租户从任何基础设施管理工作中解放出来的事实，无服务器有潜力完全实现云计算的核心承诺，专家们一致认为，它的使用将在未来几年飙升。无服务器计算是由两个关键概念实现的，这些概念由系统软件实现，确保功能的执行并在提供商的基础设施中运行：(1)在同一物理主机上运行的相互不信任的功能的数据和性能的隔离；(2)支持功能执行的系统软件的轻量级，即：该软件的低内存和磁盘占用以及快速调用时间的潜力。当前的无服务器基础设施在这两个概念上都不是最优的，因为它们使用了虚拟机（隔离良好但重量级）和容器（轻量级但存在一些严重的隔离问题）的组合。单内核是一种新的操作系统（OS）模型，在这种模型中，应用程序使用非常小的自定义操作系统层作为云中的最小虚拟机来执行。实际上，unikernels将虚拟机的强隔离性与容器级别的轻量级结合起来。这些特性使得单内核是运行无服务器基础设施系统软件的唯一合适的候选程序。我们建议探索在无服务器计算基础设施中使用单内核操作系统模型作为主要单元函数执行。我们注意到，尽管它提供了一些基本的好处，但单内核模型需要进化以完美地适应无服务器领域。主要问题是缺乏对重要特性的支持，即单内核内部隔离和多处理。这些缺点不仅仅是由于缺少实现，而是源于单内核操作系统模型的基本设计原则。因此，我们建议设计和实现UniFaaS，作为为无服务器计算量身定制的单内核操作系统模型的演变。UniFaaS旨在支持上述缺乏的特性，同时保持unkernels自然提供的隔离性和轻量级优势。UniFaaS将构建在现有的单内核（即OSv）之上。设计和开发工作将沿着三个主要途径进行：(1)新功能开发，特别是使用线程的多进程支持以及对无服务器计算的进一步专业化；(2)安全性增强，特别是使用现代硬件技术引入低开销的单内核内部隔离；(3)轻量化优化，通过各种方法进一步减少每个内核和每个函数的内存/磁盘占用以及启动/调用时间，以增加每个主机函数的密度。一旦构建了UniFaaS，我们将通过将其与使用虚拟机和容器的传统无服务器部署进行比较，来评估其安全性/隔离性、轻量级和性能。关于安全性，我们注意到当前评估隔离的度量（例如计算软件的代码行数）相当不精确，我们将开发基于可从不受信任的组件（应用程序代码，网络等）访问的可信代码（来宾内核和/或hypervisor/主机）数量的新方法。

Serverless computing, also know as Function as a Service (FaaS), is an emergingprogramming paradigm providing significant benefits for both the tenant (i.e.the application developer) and the provider in terms of costs reduction,data centre efficiency, scalability, etc. With its truly on-demand resourceconsumption and pricing model, as well as the fact that the tenant is relievedfrom any infrastructure management effort, serverless has the potential offully delivering on the core promises of cloud computing, and experts agreethat its usage will skyrocket in the years to come.Serverless computing is made possible by two crucial concepts implemented bythe systems software assuring the execution of functions and running in theprovider's infrastructure: (1) the isolation of the data and performance ofmutually untrusting functions running on the same physical host and (2) thelightweightness of the systems software supporting the execution of functions,i.e. the potential for low memory and disk footprint as well as fast invocationtimes for this software. The current serverless infrastructures are suboptimalregarding both concepts as they use a combination of virtual machines (wellisolated but heavyweight) and containers (lightweight but presenting someserious isolation concerns).The unikernel is a new Operating System (OS) model in which an application isexecuted with a very small custom operating system layer as minimal virtualmachine in the cloud. In effect, unikernels combine the strong isolation ofvirtual machines with a container-like level of lightweightness. Thesecharacteristics make that the unikernel is a uniquely fit candidate to run asserverless infrastructure systems software.We propose to explore the use of the unikernel OS model as the primary unit offunction execution in a serverless computing infrastructure. We note thatalthough it presents some fundamental benefits, the unikernel model needs toevolve to perfectly fit the serverless domain. The principal issue is the lackof support for important features, namely intra-unikernel isolation andmulti-processing. These shortcomings are not simply due to missingimplementations, but rather derive from fundamental design principles of theunikernel OS model. Hence, we propose to design and implement UniFaaS as an evolution of theunikernel OS model tailored for serverless computing. UniFaaS aims to supportthe aforementioned lacking features, while maintaining the isolation andlightweightness benefits that unikernels naturally offer. UniFaaS will be builton top of an existing unikernel, namely OSv. The design and development effortwill be made along 3 main avenues: (1) new functionalities development, inparticular multi-process support using threads as well as furtherspecialisation the towards serverless computing; (2) security enhancements, inparticular the introduction of low-overhead intra-unikernel isolation usingmodern hardware technologies; and (3) lightweightness optimisations to furtherreduce per-unikernels and per-function memory/disk footprints as well asboot/invocation time through various methods, in order to increase per-hostfunction density.Once UniFaaS is built, we will evaluate its security/isolation,lightweightness, and performance, by comparing it to traditional serverlessdeployments that use virtual machines and containers. Regarding security, wenote that the current metrics to assess isolation (such as counting the numberof lines of code of a software) are rather imprecise and we will develop anovel method based on the amount of trusted code (guest kernel and/orhypervisor/host) that can be reached from an untrusted component (applicationcode, network, etc.).