Date Published: May 31, 2019
Publisher: Public Library of Science
Author(s): Jehad Ali, Byeong-hee Roh, Seungwoon Lee, Jun Huang.
The software-defined networking (SDN) paradigm has simplified the management of computer networks by decoupling data and control planes. Moreover, the separation of the data and control planes has transitioned network complexity from traditional devices to controllers; therefore, controllers have become indispensable entities in SDN. Controllers have multiple features and direct the network from a central point and respond to updates to topological changes. However, the supportive capability of these features is strong in one controller but weak in another. Due to several controllers and each controller having a set of features, selecting an optimal SDN controller can be considered to be a multi-criteria decision-making (MCDM) problem. Herein, a two-step approach is proposed for SDN controller selection. First, the controllers are ranked with analytical network process (ANP) according to their qualitative features which influence the performance of these controllers and then a performance comparison is performed to check for the QoS improvement. The controller with a high-weight value from the feature-based comparison is quantitatively analysed by experimental analysis. The main contribution of this paper is checking the applicability of the ANP for controller selection in SDN considering its features and performance analysis in real-world Internet and Brite topologies. The simulation results show that the controller computed through the proposed approach outperforms the controller selected with existing approaches. The selection of an optimum controller with ANP results in a reduction of topology discovery time and delay in the normal and traffic load scenario. Similarly, an increase in throughput with a reasonable utilization of the central processing unit (CPU) is observed for the proposed controller.
The Contemporary computer networks have been revolutionized owing to the ease of programmability, innovation, flexibility, and centralized management spearheading concepts broached by SDN. Its abstraction has reduced the complexity of traditional network devices by shifting the distributed control logic from these devices to a central controller, i.e., separating data and control planes. The data flow is controlled by the centralized controller logic along with the applications running and interacting with the controller through its northbound interface and the data plane devices. This control is managed by a protocol known as OpenFlow .
In the literature, different approaches have been used for SDN controller selection. These approaches can be broadly classified into three categories. The first category involves comparing controllers based on their features, the second compares controllers based on their performance, and the third is a hybrid approach. The hybrid approach selects an optimum controller by combining the results of a feature and performance-based comparison. These approaches are discussed below.
The software-defined networking (SDN) is composed of data, control and management planes. However, the control plane plays the main role because it manages the data plane which is the actual network topology. Therefore, SDN controllers and their features are important for the performance of SDN. Each controller has several features. In this section, ten features are discussed which influence the performance of the SDN.
The proposed approach for controller selection is based on the qualitative and quantitative analysis of the SDN controllers. A bird’s-eye view of the proposed method is illustrated in Fig 1. First, ANP is applied for the qualitative feature-based comparison of the SDN controllers. The ANP sorts the controllers with the provided feature set controllers by calculating weights for each controller. Further, the quantitative analysis of the high-weight controller is performed through several simulations in Mininet. The procedure to choose the optimum SDN controller is described below:
The ANP was proposed by saaty . It can be applied to the quantitative and qualitative data about a network. It can evaluate the feedback and dependency relationship between the criteria and alternatives. The ANP process for SDN controller selection is described in this subsection. The general procedure  for applying the ANP is given below.
Fig 8 shows the experimental framework used for the performance evaluation. The performance analysis was conducted for C2 controller computed using the proposed approach, i.e. ANP. A performance comparison was made with the controller (RYU) calculated through AHP. First, the network topologies considered for experiments were converted to Mininet environment. The source and destination pairs of routers were selected in each topology and shortest path discovery time was calculated. i.e. the time a controller took for discovering the shortest path. After this, two hosts were attached with source and destination. Furthermore, the delay was computed for the request and response time taken by a controller for that path in the normal and a traffic generation scenario. Likewise, the throughput and CPU utilization were recorded for each controller. The experiments were performed for both controllers. Experimental scenario design and performance evaluation are discussed below;
The objective of this study was to select the optimum SDN controller regarding its features and the performance analysis of the controller in real-world and Brite network topologies. Because the controller selection process was based on multiple features, which include platform support, southbound interface, northbound interface, modularity, and productivity, it was considered to be an MCDM problem. Therefore, the ANP approach was used to solve this problem. The objectives were identified first, and criteria parameters were established based on which the proposed controller was computed using ANP model. Next, a pairwise comparison matrix was created to compare every element in the criterion cluster with every alternative in the alternative cluster, and vice versa. The final resultant matrix, known as a limit matrix, prioritizes the alternatives. Thus, controller with high-priority value was proposed for further quantitative analysis. The results from the limit super-matrix showed that C2 controller provides the optimum features, therefore its performance was validated in Mininet.