The Controller Deployment And Networking In Software-defined Datacenter Networks

Currently,large-scale data centers are the key infrastructures for not only cloud services but also massively distributed computing.The intra-datacenter network has appeared as the dominate resource,which influences the performance of many datacenter applications.Recently,the introduction of software-defined networks(SDN)gives the freedom to refactor the network control plane,and then promote the innovations of network management and applications.The principles of SDN have attracted many attentions to design software-defined datacenter networks(SDDN).The control plane is traditionally implemented by a centralized controller,which has limited performance and scalability.Recent proposals bring distributed controllers to manage common networks.SDDN,however,lacks a scalable and reliable control plane,due to the critical coverage problem of controllers.Therefore,in this paper,the main contents and innovation points include the following aspects:(1)Minimal coverage problem of controllers in software-defined data centersThe minimal coverage problem of controllers in SDDN is to find the minimal number of sufficient controllers and their deployment locations such that each switch can be controlled by at least one controller.We prove that this problem is NP-hard and thus propose an approximation algorithm under the random regular topology.We further analysis the time complexity and the approximation ratio of the algorithm.The results of experiments show that our algorithm can considerably reduce the number of controllers,which are needed to control data centers.(2)Fault-tolerant coverage problem of controllers in software-defined datacenter networksTo against failures of controllers,we propose the minimal fault-tolerant coverage problem of controllers in SDDN,a NP-hard problem.A solution to fault-tolerant is to ensure that each switch is controlled by at least two controllers.In this way,each switch will be taken over by another controller when the master one fails.To tackle this problem,we further design an approximation algorithm under the random regular topology.The results of evaluations show that our algorithm can efficiently reduce the number of controllers and enhance the fault-tolerant capacity of the control plane.(3)Minimal cost synchronization among controllers in software-defined datacenter networksThe distributed control plane,based on multiple controllers,operates on a consistent global network view,so as to ensure correct control behaviors.This requires controllers to frequently synchronize their local network views.Thus,it is essential to reduce the synchronization overhead and delay.To tackle this issue,we model and solve the minimal cost synchronization problem.We then design a dedicated algorithm to significantly reduce the communication load,due to the state synchronization among controllers,and evaluate the performance of the algorithm.(4)The controller networking in software-defined datacenter networksThe controller network requires significantly higher availability and long-term survivability than the data plane,because it is the basis for recovering from data plane failures and monitoring critical hardware.To achieve the efficient and fast communications among controllers,it is essential to build the controller network.Compared to traditional data plane networks,the controller network has very different requirements.We firstly define the performance metrics of the controller network.Then,we compare the performance of typical topologies and propose the building and extending methods of an efficient controller network.In conclusion,this paper mainly researches and explores the controller deployment and networking problem.By applying the graph theory and the computational complexity theory,we propose two efficient methods to solve the minimal coverage problem and the minimal fault-tolerant coverage problem of controllers,respectively.After finishing the deployment of controllers,we further design an efficient information synchronization method among controllers.Meanwhile,to enhance the reliability of control plane,we also design the efficient controller topology,which isolates the control plane from the failures of the data plane.This paper has certain theory value and application value in the actual deployment and network construction of controllers in software-defined datacenter networks in real life.