DAFT:一种OpenFlow大规模流表区分存储与加速查找架构

软件定义网络作为一种数据转发与逻辑控制相解耦的创新网络范式,当采用OpenFlow协议进行大规模部署时,其数据平面的流表规模急剧增大,对OpenFlow交换机的流表存储资源和分组转发性能提出了严峻的挑战.对此,本文构建了一种OpenFlow大规模流表区分存储与加速查找架构DAFT.该架构根据流量分布特性将OpenFlow网络流区分为重要流和次要流,进而采用TCAM和SRAM分别存储其标识字段,并采用DRAM单独存储其内容字段,有效缓解OpenFlow流表存储资源紧张问题.针对重要流/次要流区分问题,在分析传统大象流/老鼠流区分方法的基础上,基于OpenFlow网络流的包成批特性,提出活跃流/空闲流区分方法,以提高TCAM命中率.针对SRAM流表查找性能瓶颈,利用掩码访问不均匀的特点,采用"往前移1"启发法自适应调整掩码顺序,以减少后续数据包的掩码失败探测次数;利用掩码探测多数会失败的特点,借助计数型布鲁姆过滤器预测元组查找失败结果,从而绕过对应的子流表遍历过程.最后,借助骨干网络流量样本,对本文所提DAFT流表架构的查找性能进行实验评估.实验结果表明:DAFT流表架构的TCAM命中率、SRAM平均查找长度和平均流表访问时间均明显优于传统的大象流/老鼠流架构,且稳定性强,有效提升了OpenFlow交换机的分组转发性能.

DAFT:A Differentiated Storage and Accelerated Lookup Architecture for Large-Scale Flow Tables in OpenFlow Networks

As a novel network paradigm,Software Defined Networking(SDN)decouples control logic functions from data forwarding devices,and introduces a separate control plane to manipulate underlying switches via southbound interfaces like OpenFlow.This paradigm paves the way for more flexible,programmable and innovative networking,and is widely regarded as one of the promising directions towards future Internet.When it comes to large-scale SDN deployments based on the OpenFlow protocol,this paradigm will bring a sharp increase to the size of flow tables in its data plane,primarily due to a multiplication effect of a large number of simultaneous flows in large-scale networks and an increasing number of flow table fields with the continuous version evolution of the OpenFlow protocol.Therefore,it will pose a serious challenge on flow table storage resources in OpenFlow switches,owing to the capacity limitation of the most prevalent flow table memory TCAM.To date,the most popular solution to the challenge is to accommodate flow tables by combining TCAM with other media such as SRAM.However,SRAM does not support wildcarding,and generally employs tuple space search to classify all flow entries into a small number of tuples in terms of masks and look up tuples one by one for a matched flow entry on each packet arrival,which results in the performance bottleneck of packet forwarding in OpenFlow switches.This paper is thus motivated to propose an efficient differentiated storage and accelerated lookup architecture for large-scale flow tables called DAFT.Based on differentiating packet flows into major ones and minor ones in terms of network traffic distribution,the DAFT architecture maintains the identification fields of major flows and minor flows respectively into TCAM and SRAM,and the content fields of both flows into DRAM,to effectively relieve the insufficiency of flow table storage resources.As for major/minor flow differentiation,we take an insight into the disadvantages of traditional elephant/mice flow differentiation,investigate into the impact of wildcards in match fields on the packet-in-batch feature within a packet flow in OpenFlow networks,and present active/idle flow differentiation based on the packet-in-batch feature to increase TCAM hit rates.After that,we mitigate the performance bottleneck of SRAM flow table lookups in two ways.On the one hand,we self-adaptively adjust mask order by applying the Move-Ahead-1 heuristic,to reduce the number of failed mask probing for subsequent packets,in virtue of the non-uniformity property of mask references.On the other hand,we predict tuple lookup failures by employing counting bloom filters(CBF),to bypass the traversal on the corresponding sub-flow tables,in terms of the property that most mask probing tends to fail.Finally,we evaluate the lookup performance of our proposed flow table architecture DAFT by experiments with backbone network traffic traces.The experimental results indicate that the DAFT flow table architecture obviously performs better and more robustly than the traditional one of elephant/mice flows,in terms of TCAM hit rates,average search length in SRAM and average access time on flow tables,which effectively promotes packet forwarding performance of OpenFlow switches.