Scalability of Software Defined Network Presented by Lin Zhou&Lei Zhang

Agenda Background of our project The Design of our project

Brief Introduction to SDN Distributed Data plane and Control plane Data plane:enquiring,forwarding Control plane:management,route Networklization Intellectualization Virtualization Computationlization Networklization Intellectualization Virtualization Computationlization

The Origin of the Scalability Problems in SDN Early benchmarks on NOX, which showed it could only handle 30,000 flow initiations per second while maintaining a sub-10 ms flow install time.

Current Solutions to the Probelm DIFANE DevoFlow HyperFlow Kandoo

DIFANE Model

DIFANE Advantages: (i) DIFANE achieves small delay for the first packet of a flow by always keeping packets in the fast path. (ii)DIFANE achieves significantly higher throughput than NOX. Disadvantages: (i)A number of authority switches are needed for the large networks we evaluated. (ii)DIFANE does not address the issue of global visibility of flow states and statistics.

DevoFlow Reducing the number of flows that interact with the control-plane. By pushing responsibility over most flows to switches and adding efficient statistics collection mechanisms to identify significant flows, which are the only flows managed by the central controller

DevoFlow Advantage: It can reduce the load of the controller so that it will enlarge the scalability of the controller. Disadvantages: (i)How many flows would be sufficient to achieve the desired results in different environments still is a question. (ii)It is hard to build a efficient statistics collection mechanisms.

HyperFlow Logically centralized Physically distributed Does not require any changes to the OpenFlow standard

HyperFlow Model

HyperFlow Advantages: (i)Enables network operators deploy any number of controllers to tune the performance of the control plane based on their needs. (ii)Keeps the network control logic centralized and localizes all decisions to each controller to minimize control plane response time. Disadvantage: HyperFlow doesn't influence the number of the switches of one controller.

Kandoo Kandoo creates a two-level structure for controllers: (i) Local controllers execute local applications as close as possible to switches (ii) A logically centralized root controller runs non-local control applications.

Kandoo Model

Kandoo Advantages: Preserving scalability without changing switches. Good at dealing with local flows. Disadvantages: Can not help any control applications that require networkwide state

The Design of our project Goals a controller system can serve as many switches as possible like Hyperflow a root controller with network-view state can serve as many switches as possible like Kandoo

The Design of our project Model Design--HMKH

Details of the HMKH Assumptions 1. The communications are all wireless and wire- less communication detail is not the coverage of this report. 2. Any switches controlled by a root controller in a site is in the control range of that root con- troller. 3. The direct neighbours of any root controller are within the communication range of the root controller.

The Design of our project Implementations of HMKH Initiation Periodic Communications and Failure-free Mechanism Network Change Information

Implementations of HMKH Initiation root controller broadcast to get its local controllers local controllers broadcast to get its switches

Implementations of HMKH Periodic Communications and Failure-free Mechanism root controllers periodically broadcast local controller once getting this will respond local controller failure root controller failure

Implementations of HMKH Network-view State Synchronization -switches failure -split a network -interconnect networks

The Design of our project traditional SDN with one controller incapable if n*m exceeds k Kandoo with one root controller incapable if n*m*p exceeds k Hyperflow with n*m/k root controllers and each serve k/m switches capable Our Model with n*m*p/k root controllers and each serve k/(m*p) switches Evaluations of HMKH the process ability for controller is k msgs/sec n switches,each sends m msgs/sec and p percent of them need network-view state

Conclusions The scalability lies not only in how many switches a controller system can serve,but also how many switches a controller with network-view state can serve(overhead) The number of root controllers and local controllers should be estimated or calculated given the requests from switches to minimize the overhead while satisfy the requirements

Thanks and QA!