Article Preview
TopIntroduction
The vertically integrated structure of the traditional network scheme has a high degree of complexity (Benson, Akella, & Maltz, 2009; Ghodsi et al., 2011; Kim & Feamster, 2013), because the control and data planes intersect in each network forwarding device. This situation results both in mostly static architecture and flexibility towards innovation. Furthermore, one configuration error may cause failure over a large part of the network for a long time.
One of the most popular alternatives to traditional network architecture is Software Defined Networking (SDN). Among other network architectures, SDN has come into prominence as a programmable network. Programmability is basically attained by separating the two crucial network planes: control and data. With SDN, not only vendors and manufacturers, but also users are able to program the network controller API, which is abstracted from the individual network devices and the communication layer (Open Networking Foundation, 2012). SDN can be more comprehensively defined as a framework that is comprised of abstracted and centralized controller devices and a vendor-independent forwarding plane. In addition to the abstraction and synchronization of data among a diverse set of policy and devices, SDN provides comprehensive automation, which is an indirect result of programmability.
One of the several use-cases of SDN architecture is network management that can be derived from the control plane’s feature of overseeing the whole network. This makes it possible to monitor and measure network parameters without any additional cost. Furthermore, it allows the traffic to be managed dynamically based on monitored parameters.
Unlike traditional network, which requires separate equipment on network devices to employ congestion control mechanism, SDN manages traffic via network applications implemented on the top of the control layer. The studies of congestion and traffic control in SDN are primarily based on deep packet inspection (Bouet, Leguay, & Conan, 2013; Bremler-Barr, Harchol, Hay, & Koral, 2014) or efficient use of flows and their tables (McCormick, Kelly, Plante, Gunning, & Ashwood-Smith, 2014; Mo & Walrand, 2000; Schwabe & Karl, 2014). Those methods require performing the calibrations on the system in order to adapt them to the network’s traffic characteristics.
A comprehensive analysis of SDN traffic control was presented in Akyildiz et al. (2014). In the study, the scope of SDN traffic engineering was divided into four categories: flow management, topology update, fault tolerance, and traffic analysis characterization. The predictive model proposed in this paper can be a guide to develop traffic and congestion control applications that follow any of these categorized approaches. The proposition is independent of variety of the traffic. It is a general template; therefore, it is simpler to adapt and have a usage on different topologies and areas of application.
In this paper, OMNET++, an open source simulation environment, is used to simulate SDN traffic and build a network congestion prediction system using the statistical data collected on the control plane. To achieve this, Minimum Redundancy Maximum Relevance (mRMR) dimensionality reduction technique is implemented to determine the most relative features affecting the congestion. The predictive model proposed in this paper is essentially a dynamic model of a logical SDN system. This model is an important guide to build network applications on congestion avoidance for SDN.
In the following sections, it is further investigated into the SDN and OpenFlow protocol as well as the Nonlinear Autoregressive Exogenous Neural Network (NARX) which is used as the machine learning method to predict the SDN congestion. In Background Section, SDN capabilities and the architecture are reviewed. The congestion control system is explained in detail in Material and Methods, and the results of the experimental work are presented in Experimental Results. Finally, in Conclusion, several findings, constraints, and future plans of this research are discussed.