Classificação de regras de um controlador sdn utilizando redes neurais artificiais

Abstract

The new paradigm of Software Defined Networking (SDN) establishes the separation of control and data planes. In this type of network, such separation implies the insertion of another element of the network, the controller. The standard communication protocol known as OpenFlow (OF), allows routing elements to provide a programming interface for the network professional. In addition, OF allows the network administrator to extend the access and control of the query table used by the hardware. This control model determines the next step of each package, being built through programmable rules. Faced with these new concepts and applications generated by software-defined networks, it became necessary to measure, test and evaluate equipment that supports this technology standard. The controller acting on the SDN switch directly modifies the profile of the generated traffic, since it is accompanied by static and dynamic rules influences (inserted in the controller) that generate, in many cases, hesitations to the network professional about daily diagnoses. A classification tool is very useful in this context of static and dynamic rules, acting on a wide traffic and full of variable values. The use of an Artificial Neural Network (ANN), provides excellent results for the classification of rules observing the context of functionalities of a layer. Thus, an ANN was proposed as a classification model of rules of the SDN controller according to the effective traffic parameters, generated by two specific tools BWPING and OSTINATO. In addition, the traffic generated was based on four types of protocols: ICMP, TCP, UDP and HTTP. The controller used was the POX and the function of the rules was applied on the Link layer covering aspects of routing. Some parameters such as round-trip time, delay, bandwidth, number of packets, and flow table rules served as the basis for ANN feeding. The modeling and simulation were performed through a lab environment with network equipment and also using the Mininet virtualization for the SDN. The results of the experiments showed satisfactory performance of the neural network, reaching about ninety per cent accuracy.