基于可编程硬件的高度可扩展虚拟网络功能的研究

论文总字数：62845字

摘 要

网络功能虚拟化（Network Function Virtualization，NFV）是支撑下一代运营商网络和数据中心网络的关键技术。为满足未来网络对高性能和高灵活性的需求，实现低延时高扩展的NFV系统已成为亟待解决的关键性挑战问题。本课题采用创新性方法解决现有NFV系统中存在的报文处理延时高和跨机扩展开销大的问题，为其在未来网络中的部署和应用提供重要的理论基础和方法支持。主要研究内容包括：

针对NFV系统普遍存在的报文处理延时高、网络功能服务链冗余等问题，提出了模块化网络功能服务链（Modular Service Function Chain， MSFC）设计。以最优编排和单机内有效扩展为目标，提出了高性能、资源高效的模块化网络功能服务链框架CoCo。CoCo使用0-1二元规划对编排问题建模，提出了性能敏感的编排算法；创新性地提出了旁推扩展策略来保证单机扩展时的性能，并设计了自动化调度器处优化资源分配。实验结果显示，与对比方案相比，CoCo能够减少2.46倍的报文传输延时，并将单机扩展时的性能提升了45.5%。

针对带状态网络功能跨机扩展时正确性无法保证、性能差等问题，本课题结合可编程数据平面带来的新能力，创新性地提出基于可编程数据平面、状态逻辑分离的高性能NFV跨机扩展架构ScaleNF。为了支持数据中心架顶交换机百万并发流量的状态管理，本课题设计了一个精度无损的、无延时开销的数据平面交换机内存管理机制与状态存储结构OTSL。为了低开销透明化地将遗留的传统网络功能程序迁移到现有框架下，本课题设计了自动化将现有基于C语言的网络功能程序转化为ScaleNF架构下的编译器。实验结果表明，ScaleNF能够全面满足NFV系统对跨机扩展的高性能、低开销、正确性保证三大需求；OTSL通过寻找最优的摘要长度，能支持百万并发真实流量的正确的状态管理；同时，OTSL将对比工作下中网络功能的未利用空间从41%降至最低2%。

最后，本课题结合DPDK高速报文处理套件、Docker容器技术，借助可编程交换机Toffino、智能网卡和通用服务器完成了高性能高可单机跨机扩展NFV系统的原型实现和组网，实验结果表明，该系统与传统方案相比，具有更低的报文处理延时和更高的资源利用率；具有敏锐的自动扩展能力，在百万并发流量下能够维持扩展时与运行时的高性能以及正确扩展。

本论文有效地探讨了NFV系统中延时和扩展问题，有效地验证了方案的实际能力，对推动未来高性能、高可扩展性网络体系结构的发展具有重大意义。

关键词：网络功能虚拟化、可编程数据平面、模块化技术

Research on High Scalable NFV System Exploiting Programmable Hardware

Abstract

Network Function Virtualization (NFV) is a key technology supporting next-generation carrier networks and data center networks. In order to meet the demand for high performance and high flexibility in the future network, achieving a low-latency and scalable NFV system has become a key challenge to be solved. This paper adopts innovative methods to solve the problem of high packets processing delay and heavy cross-machine scaling cost existing in existing NFV systems, and provides important theoretical basis and method support for deployment and application in future networks. The main work includes:

For the ubiquitous packet processing delays of NFV systems and the redundancy of network function service chain, a Modular Network Function Service Chain (MSFC) design is proposed. With the goal of optimal orchestration and effective scaling up within a single machine, a high-performance and resource-efficient MSFC framework CoCo is proposed. CoCo uses 0-1 binary programming to model orchestration problems and proposes a performance-sensitive orchestration algorithm; innovatively proposes a push-aside scaling up strategy to guarantee the performance of single-machine scaling, and designs an automated scheduler to handle resource allocation for modules. Experimental results show that CoCo can reduce the packet transmission delay by 2.46 times compared with related work, and enhance the performance when scaling up by 45.5%.

For the problem of violated correctness and poor performance in cross-machine scaling with stateful network functions, this paper combines the new capabilities brought by the programmable data plane and innovatively proposes a stateless high-performance NFV scaling framework ScaleNF. In order to support the state management in the top-of-rack switch in data center with millions of concurrent traffic, an error-free and instant switch memory management mechanism and state storage structure OTSL implemented in data plane are designed. In order to enable low-overhead and transparent migration of traditionally legacy network function programs into ScaleNF, a compiler is proposed which automatically transforms the existing C-based network function program into ScaleNF architecture. The experimental results show that ScaleNF can fully meet the three requirements of NFV systems for high performance, low overhead, and correctness of cross-machine scaling; By finding the optimal digest length, OTSL can support the correct state management of millions of concurrent real-world traffic as well as reduce the unutilized space of the network functions from 41% (under comparison work) to a minimum of 2%.

Finally, this paper adopts the DPDK high-speed packets processing tools and Docker container technology, based on programmable switch Toffino, and SmartNIC, to complete the prototype implementation of a high-performance and scalable NFV system. The experimental results show that the system, compared with traditional solutions, has lower packets processing delay and higher resource utilization; meanwhile the system has a keen automatic scaling capability, and can maintain high performance and correctness no matter when scaling or at runtime.

This paper effectively discusses the delay and scaling issues in NFV systems, and effectively verifies the actual capabilities of its solution. It has significant implications for the development of future high-performance and high-scalability network architectures.

Keywords: Network Function Virtualization, Programmable Data Plane, Modular Technology

目 录

摘 要 I

Abstract II

第一章 绪 论 1

1.1 课题背景和意义 1

1.2 国内外研究现状以及发展动态分析 2

1.2.1 基于软件的 NFV 网络功能延时高 2

1.2.2 NFV系统可扩展性差 2

1.2.3 可编程数据平面带来的新功能 3

1.3 论文研究目标和内容 4

1.3.1 研究目标 4

1.3.2 研究内容 4

1.4 研究方案 5

1.4.1 高性能模块化服务链系统CoCo 5

1.4.2 高可扩展ScaleNF架构 6

1.5 论文特色和贡献 7

1.6 论文主要内容和安排 7

1.7 本章小结 8

第二章 相关技术研究 9

2.1 虚拟网络功能性能相关技术 9

2.1.1 报文IO层 9

2.1.2 CPU性能隔离 9

2.2 虚拟网络功能扩展相关技术 10

2.2.1 可编程交换机软硬件 10

2.2.2 虚拟网络功能程序分析 11

2.3 本章小结 11

第三章 模块化NFV系统研究 12

3.1 报文处理抽象 12

3.1.1 处理模块管道 12

3.1.2 约简聚集多个处理管道 14

3.2 性能敏感的MSFC布局 16

3.2.1 报文传输代价分析 16

3.2.2 资源分析 16

3.2.3 MSFC布局算法 17

3.3最优独立扩展 18

3.3.1 旁推扩展 19

3.3.2 贪心扩展 20

剩余内容已隐藏，请支付后下载全文，论文总字数：62845字