network status and operation details. Therefore, the controller can globally detect available network resources and requirements. This paradigm creates a global view of the entire network, enabling global automatic management and control without needing to configure devices individually. The SND technology has several advantages:
Efficient utilization of network resources.
Enables development of programming-based solutions for network configuration and management.
Provides a perfect ecosystem for ML paradigm and intelligent applications.
Simple and improved network management, control, and data manipulation, since network administrators can remotely alter the network configuration and operation as response to dynamical changes in the network.
High speed, through optimal handling of the traffic load.
Adopts the virtualization technology, which allows running multiple applications over the same shared hardware.
Combining AI and SDN has been attracting researchers in recent years to develop network management and operation mechanisms. The SND architecture provides centralized control of network policies and enables administrators to effectively overcome problems with ML methods.
2.5.1 SDN Architecture
The architecture of SDN is comprised of three planes:
Data Plane: comprised of the forwarding devices, i.e., switches and virtual switches. Unlike distributed network architectures, in which switching and routing devices listen to events from neighboring network elements and make decisions based on a local view, switches and routers are responsible for forwarding, dropping and modifying packets based on policies received from the CP.
Control Plane: The CP is considered to be the brain of SDN. It can program network resources and dynamically update the rules of forwarding, in addition to making the management of the network flexible through the centralized controller. The centralized controller controls communication between switches and applications. On the other hand, the controller exposes the network status and summarizes the information to the application plane. Also, the CP translates the requirements from applications to specific policies and distributes them to devices. Further, it provides the basic functions needed by most network applications such as routing algorithms, network topology, device configuration, and state information notifications.
Application Plane: composed of network applications that define management and optimization policies to be applied on the network. Applications can get network state information from the controller and implement the needed control to change network behavior.
The inclusion of ML in SDN may require a new architectural structure that differs from the traditional of SDN. In [10], a new plane is proposed called the knowledge plane KP as shown in Figure 2.6. The KP hosts ML algorithms that use statistical learning to learn the network behavior. These algorithms contribute to decision-making. Hence, the KP in SDN communicates directly with the controller, which, in turn, asks the network elements to implement decisions.
The controller gets information from network devices through the OpenFlow protocol. A server is used to process information and run ML algorithms. The execution of recommended commands is the responsibility of the controller which is connected to the KP. On the top, the application plane is running to manage the network.
Figure 2.6 The SDN architecture with knowledge plane.
2.5.2 The OpenFlow Protocol
The open source OpenFlow protocol enables the realization of the SDN technology. It has been used for researching different protocol designs over existing hardware. OpenFlow is widely used for the communication between the control and the data planes. It is developed by the Open Networking Foundation (ONF). It is the interface between devices and the SDN controller, providing the rules for switching control features from network devices to the central controllable software. It has a controller and a switch, functioning as secure channel as shown in Figure 2.7. The controller can modify, discard, and send packets to the switch. The path of the packet is determined at the times of packet transmission. OpenFlow calculates the path and sends it to the switch, which stores it in the Flow Table. When a switch receives a packet, it looks up the flow table and sends it along the stored path [11]. The primary task of the switch is to exchange data using flow tables, which are controlled by the controller of the CP. This architecture simplifies the design of switches and reduces their tasks, because they have become just data deliverers, without being required to perform any of the control functions.
Figure 2.7 The OpenFlow architecture.
The implementation of an SDN controller can be centralized or distributed. In the centralized implementation, a single SDN controller centrally controls and manages all network devices, which would possibly lead to bottleneck. Distributed implementation of the SDN controller would overcome this issue. The CP may include multiple controllers, depending on the network size. This will help boosting the network performance.
2.5.3 SDN and ML
SDN has strengthened applying programmatic principles on network, allowing network administrators to have precise, flexible, and innovative control of the network and thus reducing operational expenses.
The SDN architecture provides an opportunity to more efficient application of cognitive network concepts in a centralized system, leading to self-aware networks. The adoption of SDN-based systems highly depends on their success in providing solutions to problems that could not be solved by traditional network architectures and protocols [12].
Applying ML techniques with SDN is considered to be effective for the following reasons [13]:
The recent advanced developments in computing and the accompanying advanced processors, thus creating a new opportunity to apply promising learning techniques.
It is well known that ML algorithms depend on data. The SDN controller has a holistic view on the network and is able to collect different network data, simplifying the application of ML algorithms.
Based on the ability of the SDN to act in real time and deal with historical data, ML techniques can create intelligence in the controller unit, by conducting data analysis relying on analyzed data in decision-making and thus improving the network and its services.
The programmatically feature of SDN can help to find optimal solutions to network problems such as configuration and resource allocation. Thus, ML algorithms can be implemented in real time.
2.6 Cognitive Radio Networks
The Federal Communications Commission (FCC) defines cognitive radio as: “a radio that can change its transmitter parameters based on interaction with the environment in which it operates”.
The main features of cognitive radio are as follows [14]:
Awareness: CR is aware of its surrounding environment through the sensing capability.
Intelligence: CR is a programmable intelligent wireless communication system capable of learning from information collected from the environment.
Adaptivity: CR adapts to the variation of the radio spectrum conditions and application requirements by dynamically reconfiguring its operational parameters.
Dynamic Spectrum Access has been proposed