EAGER: Leveraging Server Offload Features For High-Performance Network Function Implementations

网络功能（NFs）或中间盒以复杂的方式检查和修改网络数据包和流。例如，入侵检测系统（ids）、负载平衡器、缓存代理等。最近，在通用计算资源上运行的基于软件的NFs取代专用的NF硬件的趋势非常强烈，这种趋势被称为网络功能虚拟化（NFV）。NFV正在各种环境中扎根，包括服务提供商、电信、企业和数据中心网络。随着基于软件的中间件的设计和部署的急剧增加，因此了解如何最好地实现中间件软件变得至关重要，以便在不付出成本的情况下达到或超过硬件元素的性能和功能。该项目旨在验证一个假设，即现代服务器设备的新型硬件和虚拟化软件功能可以在有效的中间盒软件实现中发挥关键作用。特别是，现代服务器现在支持各种加速特性，特别是在它们的网络接口卡或网卡上，它们现在支持各种“卸载”特性。同样，在软件方面，也有一些创新，例如容器，它提供了比传统虚拟机更精简的虚拟化替代方案。在这个项目中，首席研究员试图了解是否有可能，以及如何以一种适当地在middlebox内部定制软件和各种内在硬件和软件加速特性和支持之间分割的方式，最佳地实现给定的middlebox功能。拟议的研究将评估在最先进的服务器硬件和虚拟化软件的情况下重新实施中间盒的可行性和局限性。探索将包括：对各种现代硬件/软件特性集的性能进行基于测量的研究；制作简单的中间盒原型，以最佳地利用可用功能；了解各种多租户/虚拟化设置中卸载特性的性能，特别关注干扰的影响；了解如何在NFV中最好地支持分布式中间件和服务链。智力优势：该项目将导致对服务器加速特性和现代虚拟化软件在有效实现未来软件中间件方面所带来的机遇和挑战的关键初步见解。该项目将导致测试中间件实现的基准套件。它还将产生一小类高性能中间盒的初始软件原型。这一初步探索的结果可以为进一步研究如何通过利用这些新出现的特性来整体重构数据中心软件基础设施提供信息。它还可以为研究提供适当的api和框架，以有效地支持未来的SDN+NFV场景。更广泛的影响：如果成功的话，这个探索性的工作将导致一个更大的、潜在的变革性的研究项目，它将围绕系统api，为所有未来的数据中心基础设施调用适当的硬件和软件功能。PI计划聘请博士后参与本研究，为博士后提供专业发展机会。这项研究将在很大程度上利用CloudLab进行实验。这项研究的结果将作为概要文件发布在CloudLab上，供其他研究人员在他们的工作中使用。这项研究也将紧密整合到PI Akella将于2015年秋季教授的云计算课程中，并整合到威斯康星大学麦迪逊分校的其他研究生和本科生系统和网络课程中。项目负责人将与相关行业合作伙伴合作，确保项目采用最先进的硬件和软件，并使行业了解项目的最新发现。

Network functions (NFs), or middleboxes, examine and modify network packets and flows in sophisticated ways. Examples include, intrusion detection systems (IDSs), load balancers, caching proxies, etc. Recently, there has been a strong push to replace dedicated NF hardware with software-based NFs running on generic compute resources---trend known as network functions virtualization (NFV). NFV is taking root in a variety of setting, including service provider, telco, enterprise and data center networks. With a corresponding steep increase in the design and deployment of software-based middleboxes it has therefore become crucial to understand how best to implement middlebox software, so as to meet or exceed the performance and functionality of hardware elements without the cost. The project aims to verify the assumption that the novel hardware and virtualization software capabilities of modern server equipment can play a crucial role in effective middlebox software implementation. In particular, modern servers now support a variety of acceleration features, in particular on their network interface cards, or NICs which now support a variety of "offload" features. Likewise, on the software front, there are innovations such as containers, which offer a leaner alternative to virtualization than traditional virtual machines. In this project the principal investigator seeks to understand whether it is possible, and how to, optimally implement a given middlebox's functionality in a manner that is split suitably between middlebox-internal custom software and various intrinsic hardware and software acceleration features and support. The proposed research will evaluate the feasibility and limits of middlebox re-implementation given state-of-the-art server hardware and virtualization software. The exploration will entail: a measurement-based study of the performance of various modern hardware/software feature sets; prototyping simple middleboxes to leverage the available features optimally; understanding performance of offload features in various multi-tenant/virtualized settings, focusing in particular on the impact of interference; and, understanding how best to support distributed middleboxes and service chaining in NFV. Intellectual Merit: This project will lead to key initial insights into the opportunities and challenges posed by server acceleration features and modern virtualization software in effectively realizing future software middleboxes. The project will lead to benchmark suites for testing middlebox implementations. It will also result in initial software prototypes of a small class of high performance middleboxes. The outcomes of this initial exploration can inform further research into how to restructure data center software infrastructure as a whole by leveraging these emerging features. It can also inform research into appropriate APIs and frameworks for effectively supporting future SDN+NFV scenarios. Broader Impacts: If successful, this exploratory work can lead to a much larger and potentially transformative research program, rooted around systematic APIs for invoking appropriate hardware and software features for all future data center infrastructures. The PI plans to engage a postdoctoral research in conducting this research, thus providing professional development opportunities for the postdoctoral researcher involved. The research will leverage CloudLab heavily for experimentation. Outcomes from this research will be made available as profiles on CloudLab for other researchers to leverage in their work. This research will also be tightly integrated into a Cloud Computing course that PI Akella will be teaching in the Fall of 2015, and into other graduate and undergraduate systems and networking courses at UW-Madison. The PI will engage with the relevant industry partners to ensure that the project employs state-of-the-art hardware and software and to keep the industry abreast of the project's latest findings.