PiBuffer：面向数据中心的 OpenFlow 流缓存管理模型

摘要

基于 OpenFlow 的 SDN（Software Defined Networking）技术在数据中心中得到广泛研究和应用，如何缓解集中的控制平面成为网络性能的瓶颈是其中的研究热点．OpenFlow 规范提出，当数据平面有缓存能力时，未命中的报文仅需发送少量摘要信息至控制器触发规则下发，从而减少控制平面与数据平面的通信负载．然而，现有的缓存模型采用报文粒度的缓存方式，使得同一条流的多个未命中报文会被送至控制器造成额外的通信负载，而且交换机处理报文的顺序会导致流内报文乱序，从而降低通信的性能．针对上述问题，该文提出了一种支持流内报文保序的 OpenFlow 交换机流缓存管理模型．通过基于流粒度的未命中报文缓存方式，进一步减少控制平面与数据平面的通信开销．通过设计流动作预处理机制，实现同一条流内报文传输保序．该文分别基于软件交换机 OFSoftSwitch与硬件网络实验平台 NetMagic 对该流缓存管理模型进行了原型系统验证．%The technology of OpenFlow-based SDN (Software Defined Networking)is widely studied and deployed in datacenter network.How to alleviate the centralized control plane becoming the bottleneck of network performance is the hotspot of research.OpenFlow specification points out that the switch can buffer the mis-matched packets and send only abstracts of them to the controller for flow rules,so as to reduce the communication load between the control plane and data plane.However,existing OpenFlow switches adopt the model of packet-granularity buffer, which cause extra communication loads by sending many packets in the same flow to the controller. In addition,the packet buffer affects the sequence of packet processing and incurs packet out-ordering,which will degrade the performance.Aiming at these problems,a novel model of flow buffer management in the OpenFlow switch is proposed.In order to further reduce the communication load,the mis-matched packets are buffered in flow-granularity.By designing the mechanism of flow action pre-processing,packets in the same flow are kept transmitting in order.We validate the correctness of the model by implementing the flow buffer on software switch OFSoftSwitch and NetMagic network experimental hardware platform.