Research On Several Key Issues In Software Defined Data Center Networks

With the gradual maturity of cloud computing technology and its preliminary commercial applications, the data center as its core supporting infrastructure grows rapidly. In current data centers, cloud computing tasks usually require massive data flow interaction among commodity servers.and the data flow also presents many unique features compared to traditional network traffic, such as the round trip time is pretty small, many-to-one or many-to-many communication widely exist and so on. This makes those classical routing protocols with successful implementation experience in traditional networks can not play their roles in data center networks. Hence,put down the performance of data center networks. Furthermore, IaaS (Infrastructure as a Service), as a key business of data center, requires that the physical network resource can be dynamicly scheduled. Thus, the network resource such as network topology, address space, forwarding table space, link bandwidth, switch CPU (Central Processing Unit) should be virtualized. In addition, the limitation of the scale of layer 2 broadcast domain also need to be overcome due to virtual machine migration and some other particular business demand. The unique features and service requiremnts mentioned above bring many severe challenges to traditional data center networks. The Software-Defined Networking (SDN), as a rising network paradigm in recent years, has two key points. One is centralized management. SDN decuples the control plane from data plane of commodity switches physically, and moves the control logic to a remote centralized control center which is called SDN controller. Thus, the network resource management and allocation is easy to accomplish based on network-wide view. The other one is network programmability. The automation of network configuration and rapid deployment of new network protocol are enabled to meet operations’demands due to the northbound Application Programmability Interface (API) of SDN controller and standard southbound management interface between SDN controller and switches. Currently, Software Defined Data Center (SDDC) has attracted wide attention from academia and industry. In this desseration, we focues on the following three problems:1) multiple tenants network virtualization,2) flooding of broadcast packets, and 3) SDN controller placement. The research achievements are summaried as following:(1)A platform, NetKart, is designed to provide programmable virtual networks for tenants. We present NetKart, a SDN-based programmable multiple tenants network virtualization platform that enables tenants to create and manage their virtual networks. Each virtual network is assigned to an independent Network Operation System (NOS), and run its own network control software. The virtual network topology and addressing scheme of virtual network can be specified by tenants arbitrarily. Since NetKart logically decoupled virtual networks from underlying physical infrastructure, it gives tenants’NOS the illusion of full control over their own virtual networks. NetKart builds on the design of FlowVisor and enables the introduction of virtualization features such as topology virtualization, address space virtualization, control function virtualization, flow table space virtualization and so on. The evaluation result of this implementation shows that NetKart is capable of multiple dimensions network resource virtualization and solid isolation among virtual networks while incurring a modest overhead to the control channel. (2)FSDM, a floodless service discovery mechanism is designed. The flooding of broadcast packets has several drawbacks. Such as 1) the bulk of broadcast packets occupy a significant part of the network bandwidth resources.2) Spanning Tree Protocol (STP) has several drawbacks including too high traffic pressure of the root bridge, the waste of bandwidth of redundant links, and too long convergence time.3) The CPUs of end hosts can get overloaded by processing broadcast traffic received from the network, and 4) a large amount of precious flow table spaces are occupied. In a large-scale data center network, the traditional solution for handling broadcast packets in Ethernet faces more challenges. In this paper,we proposed FSDM, which takes advantage of network-wide view provided by SDN controller, FSDM uses ARP (Address Resolution Protocol) proxy to response ARP request directly via SDN controller, and uses DHCP relay to enable DHCP messages forwarding only among source hosts, SDN controller and DHCP server. Thus, broadcast packets do not need to flood in the whole network. The simulation results show that our model can significantly reduce network traffic and overhead of SDN control plane.(3)HTR, A heuristic algorithm for constructing the routing tree of SDN control plane is designed. In a Software Defined Data Center (SDDC) (e.g. Google data centers network), both the controller location and control messages routing tree have significant impact on the performance of SDN control plane (e.g. the average latency of control messages, the reliability of control messages routing tree). In this paper, a mathematical model is built, and algorithms, including the heuristic routing tree algorithm (HRT), the shortest path algorithm (SP) and greedy algorithm (GA), are proposed to optimize the performance of the average latency and routing tree reliability.The evaluation results show that our proposed HRT algorithm achieves the trade-off between control messages average latency and routing tree reliability, and performs obviously better than the shortest path algorithm and the greedy algorithm.