The Traffic Management Mechanisms Of Data Center Network Based On SDN

As the basic support platform of data centers,data center network(DCN)carries different kinds of key businesses.In DCNs,business traffic distributes unevenly and appears the co-existence of elephant flows and mice flows and obvious locality and dynamics,which makes traffic management on DCNs facing severe challenges.The implement cycle of traditional network traffic engineering technologies is always very long so that these technologies are not suitable for the high throughput and low delay environment of DCNs.To meet ever growing business needs,it is very important to research efficient traffic management technologies to improve the QoS of DCNs.Although current traffic management research on DCNs has achieved a certain amount of progress,there still exist the following technique challenges.①In DCNs,due to the discrepancies between a tremendous number of servers,the hardware-based load balancers become performance bottlenecks.The ECMP algorithm widely adopted in DCNs only statically figures out the routing information and ignores link loads,resulting in the hash collision of elephant flows,which aggravates the traffic distribution unevenness.②Under the many-to-one transmission pattern of DCNs,a lot of concurrent flows easily let buffers of the bottleneck switch overflow and trigger the TCP timeout retransmission,giving rise to throughput collapse.③In DCNs,interactive applications own strict time limits.Fair share-based transmission layer protocols cannot distinguish data flows with different size and urgency so that delay-sensitive flows have to experience longer transmission time and miss their deadlines.④Data center provides different kinds of services and theses services make DCNs facing situations like high concurrent visiting traffic and unascertained visiting sources,which bring DCNs a lot of security threats.SDN provides new ideas for carrying out traffic management research on DCNs.SDN realizes the control functions of traditional network devices into the centralized controller and provides a global network view.The controller uses protocols like OpenFlow to obtain distribution and load of network traffic and by calculating some performance metrics such as throughput,delay and queue length,the controller can not only adjust scheduling strategies according to traffic variation,but also precisely locate network congestion and take some actions to maintain high service level of DCNs.For above technique challenges,this paper utilizes SDN’s superiorities to launch deep research on normal and abnormal traffic management in DCNs and present some traffic management algorithms from aspects like load balancing,congestion control,transmission control and security defense.Specifically,innovative research contents are as follows.(1)To address the problem that load balancing algorithms employed by load balancers fail to evaluate real loads of servers precisely and cannot dispatch tasks efficiently,resulting in unbalances of network traffic distribution,a dynamic feedback-based load balancing algorithm for server cluster is proposed.The algorithm periodically collects each server’s performance statistics and selects the targeted server by calculating weight.Meanwhile,this algorithm gets global network load with SDN OpenFlow protocol and classifies flows using a two-phase elephant recognition algorithm.For identified elephant flows and mice flows,the algorithm leverages a flow classification-based routing scheme to forward them.The expriemental results show that the algorithm effectively balances loads of server cluster,noticeably raises network throughput and bandwidth utilization and reduces finish time and average latency of flows.(2)To solve the issue that a large number of concurrent flows compete bottleneck bandwidth resources and make the bottleneck switch overflow and incur TCP Incast problem under the many-to-one transmission mode of DCN,a TCP congestion control algorithm based on multi-constraint routing and path congestion level is presented.Compared with terminal-based network congestion control methods,this algorithm uses the SDN global view to monitor network loads and to reduce the possibility that links congest in the course of transmission,it works out an optimal forwarding path with constraint conditions,namely available bandwidth and balanced degree of link bandwidth of a path.Besides,the algorithm obtains congestion level of paths of data flows via periodically watching instantaneous queue of switches and explicitly notifies terminals with congestion control messages in order to generate rapid congestion feedback.Terminals adjust their congestion windows according to the path congestion level and thus alleviate network congestion.The experimental results show that the proposed algorithm can effectively reduce frequency of timeout retransmission and shorten RTTs,improving network throughput and bandwidth utilization.(3)To resolve the problem that the fair share-based bandwidth allocation strategy used by transmission layer protocols in DCNs makes bandwidth-sensitive flows get a high bandwidth usage and cannot ensure that delay-sensitive flows finish transmission within their deadlines,a time-constrained flow transmission mechanism based on a flow’s deadline and size is presented.DCTCP and D2TCP employ a single ECN flag to evaluate network congestion level,but bring queue oscillation.To improve this phenomenon,the mechanism uses SDN OpenFlow protocol to monitor network link load and takes a bottleneck link utilization rate-based double ECN marking method to control the packet queue and smooth network throughput.Considering shortcomings in the congestion control of D2TCP,the mechanism uses the remaining completion time and size of a flow to calculate the penalty factor of the flow’s congestion window in order that the congestion window of a more urgent flow can decrease much less.The experimental results show that the proposed mechanism can fairly allocate bandwidth for delay-aware flows and observably shorten flow completion time,raising the deadline completion rate.(4)To deal with the problem that infrastructures of DCNs have security defense bugs and are lack of the ability of dynamic collaboration and fast disposal recovery,all of which incur ever-increasing cyber-attacks on DCNs,a hybrid strategies-based DDoS attack defense mechanism is proposed.The mechanism uses OpenFlow flow table,SNMP and a cyber-attack dataset to defend IP spoofing attacks and TCP SYN flood attacks.For IP spoofing,this mechanism employs flow table matching and a tree-based intrusion prevention rule library to validate a packet’s authenticity and legitimacy.Besides,it adopts available host resource-based anomaly detection algorithm to examine suspicious network behavior.For TCP SYN flood,the mechanism first uses an unitary Gaussian distribution anomaly detection algorithm to do early detection for data flows,and second,it do further operation with an anomaly detection algorithm based on entropy and network self-similarity index.Finally,the mechanism uses a DiffServ model-based attack mitigation strategy to limit abnormal traffic’s bandwidth usage.The experimental results show that the proposed mechanism not only validates a packet and defends those well-known attacks,but also signally reduces TCP half-open connections,alleviating attack influences on the targeted host.