软件定义网络理论模型、体系结构及其控制机理

Current Internet is composed of proprietary network devices (including the underlying hardware and upper layer software) manufactured by specific device vendors, and these devices are “black-boxes” to the users and only open to network operators. .Network management has to be done through limitedly exposed and nonstandard interfaces provided by device manufacturers without a fully customized and unified control over the traffic forwarding behaviors. This “black-box” design contributed to fast network device deployment in the past to build large-scale networks. However, since the emergence of rich applications and their customized requirements on network performance in recent days, the “black-box” design hinders the Internet from quick evolving, contrast to the fast development of other computer systems, e.g., database systems, operating systems. More precisely, the “black-box” design not only increases the difficulty for network operators to conduct configurations, but also prevents the Internet from becoming controllable, more cost-efficient and more customizable. Therefore, future Internet is eagerly looking forward to some open network architectures..Software-Defined Networking (SDN), recently proposed as a clean-slate future Internet architecture, embraces this fundamental transition with an innovative perspective -- SDN allows network operators to manage network services though abstraction of lower-level functionality. This is done by decoupling the control plane and data plane. The control plane is dedicated to routing decisions, while the data plane is responsible for packet forwarding. This new network architecture exposes the data plane hardware by API abstraction to make the devices more configurable. However, it also brings new challenges to design such a hardware-based substrate to support rich yet constantly growing high-level applications, especially with the ever-increasing network bandwidth. The exposure of the data plane hardware and the centralized controller may cause new security issues as well. Therefore, large-scale SDN deployment still requires solid model-level verification, elaborate design-level innovation plus implementation-level re-engineering..This proposal aims to address the modeling, architecture and operating principles on SDN. Specially, we plan to study the following technical details: 1) high-speed flow table lookup and update with small memory footprint; 2) programmable and re-organizable data plane to support elastic high-level applications; 3) SDN security issues; 4) memory-efficient SDN measurement; 5) key issues related to the SDN network architecture; 6) fundamental theory and modeling on SDN. The research results will provide potential technical support to improve the performance of future SDN networks and devices..

传统网络设备从底层硬件到上层软件均由设备厂商遵循网络协议和标准独立实现。这种近乎“黑盒”式的网络设备开发方式造成了互联网的僵化，严重阻碍了互联网的快速发展。针对这一状况，软件定义网络（SDN）最近被提出，其核心思想是通过将网络设备的控制平面和转发平面分离，合理抽象并开放转发平面的API接口，利用控制平面软件对转发平面硬件的行为进行管控。这种分离的新架构给网络设备数据平面的高速硬件接口设计、可编程性和可扩展性设计等带来了新的机遇和挑战。新的体系结构也面临新的安全问题。本申请以软件定义网络为背景，旨在探索SDN网络的体系结构及其控制机理。具体地，本申请计划研究以下内容：1）高速大容量流表的查找和更新；2）数据平面硬件资源可编程、可重构；3）高效、轻载的网络测量机制；4）网络体系结构中的若干关键技术问题；5）SDN网络关键技术的原型验证系统。研究成果将对改进SDN网络及其设备的性能提供技术支持。

软件定义网络的思想和初期的技术已经被工业界采用，例如以TCAM为查找引擎的流表设计方案已经在数据中心网络的SDN交换机中广泛使用。但是这种方式存在着一些技术缺陷，如随着流表规模的不断增大，以TCAM为查找引擎的方案的成本日趋增高，功耗随之增大，更新效率亟待提高。针对这种情况，本课题研究了两类解决方案：. 1）在使用TCAM作为高速查找缓存的架构中使用无重叠的规则替换策略。一方面将高速缓存中流量的查找命中率最大化，另一方面发明了一种新的完全去依赖规则的算法，将单条插入的规则的覆盖空间最大化。这样，更新规则只需要插入/换出一条规则，可做到节省成本、降低功耗、提高查找速率的目的；. 2）在一般的TCAM查找结构中，采用一种顺序栈数据结构代替动态规划（DP）的方案，在做到减少更新移动的计算时间的同时降低规则实际物理移动的次数，提高规则更新的性能。. 精确流测量是SDN集中控制的基础。目前的网络流测量方案在高速情况下，很难做到准确计数（同时统计分组的个数和流的字节数）。针对这一情况，课题组提出了cache协助的高速流统计方案，并在理论上进行了创新：. 1）提出了CASE每流测量方案，能够做到精确流计数，速率达到200Gbps worst case情况下的线速测量；. 2）扩展了Morris的简单计数理论到任意两个正整数的近似计数情况，可保证最大误差的解析式上界。这一理论成果具有普适性，可推广应用到大数据和数据库等领域。. 在简化布隆过滤器计算方面开展了研究工作，取得成果：. 1）提出了一种仅需一次哈希计算的布隆过滤器求解理论，以一个哈希函数和k个取模运算代替了标准布隆过滤器中的k个哈希函数计算，计算代价近似为标准布隆过滤器计算代价的1/k。在降低哈希计算量的同时，单哈希布隆过滤器保持和标准布隆过滤器几乎相同的假阳性概率；. 2）设计了一种基于SIMD的并行布隆过滤器加速计算技术，使得超快速布隆过滤器可提升查询速度2~3倍。. 提出了一种基于cache的流表结构FlowShadow，解决OVS存在的更新一致性问题，能在流表更新的同时，实现快速交换。