Design and Implementation of a Consistent Data Store for a Distributed SDN Control Plane

Design and Implementation of a Consistent Data Store for a Distributed SDN Control Plane
复制标题

分布式SDN控制平面一致数据存储的设计与实现

DOI:
10.1109/edcc.2016.12
复制
发表时间:
2016
期刊:
2016 12th European Dependable Computing Conference (EDCC)
影响因子:
--
通讯作者:
A. Bessani
A. Bessani
中科院分区:
--
文献类型:
--
作者:
F. Botelho;T. A. Ribeiro;P. Ferreira;Fernando M. V. Ramos;A. Bessani

文献摘要

被引文献

相似文献

可扩展和容错的分布式软件定义网络(SDN)控制器通常放弃网络状态的强一致性,而采用更有效的最终一致存储模型。这一决定主要是由于强一致性复制协议的性能开销(例如,Paxos,RAFT),这限制了网络应用程序的响应能力和可伸缩性。不幸的是,这种一致性的缺乏导致网络应用程序的复杂编程模型,并可能导致网络异常。在本文中,我们展示了控制平面一致性的缺乏如何导致网络问题,并提出了一种分布式SDN控制平面架构来解决这个问题。我们的模块化体系结构是由一个容错的数据存储,提供了强大的一致性属性所需的控制平面的透明分布的支持。为了处理这种设计的基本问题,我们采用了一些技术,专门为SDN优化数据存储性能。为了评估这些技术的影响,我们分析了三个真实的SDN应用程序与数据存储交互时产生的工作负载。我们的研究结果表明,与非优化设计相比,延迟和吞吐量分别提高了两到四倍。
Scalable and fault-tolerant distributed Software-Defined Networking (SDN) controllers usually give up strong consistency for the network state, adopting instead the more efficient eventually consistent storage model. This decision is mostly due to the performance overhead of the strongly consistent replication protocols (e.g., Paxos, RAFT), which limits the responsiveness and scalability of network applications. Unfortunately, this lack of consistency leads to a complex programming model for network applications and can lead to network anomalies. In this paper we show how the lack of control plane consistency can lead to network problems and propose a distributed SDN control plane architecture to address this issue. Our modular architecture is supported by a fault-tolerant data store that provides the strong consistency properties necessary for transparent distribution of the control plane. In order to deal with the fundamental concern of such design, we apply a number of techniques tailored to SDN for optimizing the data store performance. To evaluate the impact of these techniques we analyze the workloads generated by three real SDN applications as they interact with the data store. Our results show a two-to four-fold improvement in latency and throughput, respectively, when compared with a non-optimized design.