Kafrelsheikh Journal of Information Sciences ISSN (Online): 2535-1478, ISSN (Print): 2537-0677 Volume 4, Issue 2, 2023, PP.

1–38

Software defined network traffic routing optimization: A systematic literature

review
Omar M. Mohamed
Department of Computer Science, Faculty of Science, Minia University, Egypt.
[email protected]
Tarek M. Mahmoud
Computer Science Department, Faculty of Computers and Artificial Intelligence, University of
Sadat City, Egypt. [email protected]
Abdelmgeid A. Ali
Department of Computer Science, Faculty of Science, Minia University, Egypt.
[email protected]

Abstract
The recent surge of interest in Software Defined Networking (SDN) technology is attributed to its
centralized administration and control approach, which enhances network management and
streamlines infrastructure maintenance. Despite its apparent sudden emergence, SDN is rooted in
a lineage of endeavors aimed at enhancing network programmability. SDN offers real-time
responsiveness and meets demanding high availability criteria. However, this novel paradigm
encounters various technological challenges, some intrinsic and others inherited from pre-existing
technologies. This study focuses on illuminating routing traffic concerns within the realm of SDN
and provides insights into the forthcoming challenges that confront this transformative network
model, encompassing both protocol and architecture perspectives. Additionally, we aim to explore
diverse extant solutions and mitigation strategies that tackle issues of SDN scalability, elasticity,
dependability, reliability, high availability, resiliency, and performance. This study entails a
systematic analysis of 16 scholarly articles addressing routing traffic matters in the context of
SDN. Through inductive analysis, this paper discerns and elucidates solutions for recurrently
highlighted issues within academic discourse.
Keywords—Software Defined Networking, SDN, Traffic, Load-Balancing, Traffic Optimization and
Path Selection

1. Introduction
Traditional routing protocols exhibit inefficiency within expansive and intricate networks,
primarily due to their reliance on information flooding and localized routing computations [1], [2]
. The Software-Defined Networking (SDN) paradigm effectively tackles these drawbacks by
centralizing the control plane, thereby facilitating enhanced information exchange and utilization.
This engenders a network architecture characterized by heightened flexibility and optimization,
aptly aligned with contemporary application requisites. Specifically, SDN decouples the control
plane from the data plane, yielding a more streamlined approach to packet management [3]. The
control plane assumes responsibility for routing determinations in an entity known as the SDN
controller [4], while the data plane undertakes the task of packet forwarding. This segregation
empowers the control plane with a holistic network perspective, enabling more judicious routing
choices [5]. The controller effectively manages flow control between the two planes by interacting
through Application Programming Interface (API) between the two plans passing by the
application layer as illustrated in Figure 1. The primary objectives of SDN are to simplify network
and service management, reduce costs, and improve adaptability. SDN employs the OpenFlow
protocol [6] to facilitate communication between the control and data planes. This strategic
arrangement empowers the control plane to optimize traffic routing at the level of each distinct
flow. As a consequence of these inherent benefits, SDN emerges as a propitious technology for
elevating the efficiency and scalability of expansive, intricate networks.

Figure 1. SDN paradigm architecture

SDN enables the utilization of two primary Application Programming Interfaces (APIs), namely
Southbound APIs and Northbound APIs, which facilitate bidirectional communication and
interaction with the control and data planes. Employing Northbound APIs within an SDN
controller enables programmatic control over the network infrastructure, while Southbound APIs
serve as the intermediary links between control and forwarding components [7]. Ye contemporary
data landscape is characterized by high volumes of video traffic, the proliferation of large data
centers, and the mobility of network users. Yese factors can lead to significant traffic congestion
and performance problems for network operators. In addition, data center operators are facing
challenges due to the rapid growth of server and virtual machine deployments, as well as the
increasing amount of server-to-server communication traffic. SDNs can provide a centralized
control plane that can optimize network traffic flows and dynamically adapt to changing conditions
[8]. Yis can help to improve network performance and reliability and reduce the risk of traffic
congestion. While software-defined networking (SDN) offers many advantages for improving
network performance, its architecture also introduces some challenges. Ye structure of the paper
is organized as follows. Section 2 discusses the background of SDN traffic routing issues and
strategies. Section 3 discusses the Related literature where Research pertaining to traffic load
balancing and traffic resilience is analyzed in sub-section 3.1 and sub- section 3.2 review the
schemes designed for traffic optimization and path selection. Section 4 explores a comparison of
evaluation tools and performance results including the limitations. Finally, section 5 concludes
this paper and introduces future works.

2. SDN Traffic Routing issue and strategies

2.1. SDN Traffic Routing issue

The challenges and issues that are considered a paramount importance of traffic routing in the
SDN environment are the following: The following are the most important challenges and issues
of traffic routing in SDN environments [9]:
- Scalability remains vital: The capacity of SDN networks to handle extensive device counts
and traffic flows is crucial. This poses a challenge as SDN controllers must manage substantial
data and swiftly determine routes.
- Real-time adaptation to topology shifts: SDN networks must promptly respond to dynamic
alterations in network structure. This is imperative since SDN controllers must swiftly reroute
traffic to bypass failed connections or congested nodes.
- Meeting diverse QoS demands: Adhering to distinct application QoS requirements is essential
for SDN networks. SDN controllers must effectively prioritize and route traffic, aligning with
each application's distinct needs.
- Elevated security and privacy considerations: Ensuring the security and confidentiality of user
data within SDN networks is paramount. The centralized control and data of SDN networks
create susceptibility to potential breaches.
- Optimal resource utilization: Efficiently leveraging network resources is a core requirement
for SDN networks. SDN controllers must formulate routing decisions that minimize
congestion and maximize bandwidth deployment.
Despite these challenges, SDN is a promising technology with the potential to revolutionize the
way networks are managed. As the technology matures, these challenges will likely be
addressed, and SDN will become a more viable solution for a wide range of networks.

2.2. SDN Traffic Routing strategies

Software-Defined Networking (SDN) employs diverse strategies for traffic routing, the common
strategies shown in figure 2 and categorized as follows [10]:
 Figure 2. SDN common traffic routing strategies

Static Routing: Involves manually set routes, suitable for stable traffic patterns.
Dynamic Routing: Adapts routes dynamically using algorithms, excelling in dynamic traffic
scenarios.
Hybrid Routing: Merges traditional and SDN-based approaches for combined benefits.
Load Balancing: Distributes traffic evenly across paths to enhance resource usage and avert
congestion.
Traffic Engineering: Intelligently manages traffic flows to optimize performance and resource
allocation.
Multi-path Routing: Improves network dependability by employing multiple routes
simultaneously. This includes Equal-Cost Multipath (ECMP), which uniformly spreads traffic
across equitably costed paths to optimize resource usage and balance loads, and Unequal-Cost
Multipath (UCMP), which allocates distinct weights to paths according to their capacities,
latencies, or expenses.
Machine Learning-based Routing: Utilizes machine learning for optimized routing decisions
based on historical data and changing network conditions as predictive routing and reinforcement
learning routing.
Quality of Service (QoS)-based Routing: Prioritizes traffic according to QoS needs like latency
and packet loss.

3. Related literature

This paper provides an extensive survey of academic literature concerning the enhancement of
traffic routing within the realm of Software-Defined Networking (SDN). The survey is organized
into two primary sections. The first section critically reviews studies focused on traffic load
balancing and traffic resilience, which aim to evenly distribute traffic across multiple links or paths
to enhance network performance and maintain traffic delivery even during link failures. The
second section analyzes research pertaining to traffic optimization and path selection, which is
concerned with finding the best paths for traffic to take in order to minimize delay, maximize
throughput, or minimize cost.

3.1.Traffic Load Balancing & Traffic Resilient

A method of load balancing based on server response time (LBBSRT) [11], aims to solve the
problem of load balancing in the server cluster based on the server response time. LBBSRT
chooses the server with minimum response time to provide services to users. The server response
time is defined as the interval from accepting user requests to responding to user requests.
Traditional schemes do not incorporate the server response times while balancing the server loads.
The system model was formulated within the OpenFlow environment, illustrated in Figure 3.

Figure 3. LBBSRT System model

LBBSRT uses the control plane to obtain the server response times accurately and effectively. The
system is composed of two algorithms: one for real-time measuring of response time for each
server in which, the server response time is obtained by parsing the Packet_in message which is
sent by the switch. While the other for handling user requests, the controller handles the ARP
messages (ARP_broadcast) that are sent by users and handle the user request by sending an ARP
reply packet to users and then selecting the server with minimum or stable response time according
to the obtained server data. Because the server response time directly reflects the server load
capability, selecting a server based on the response times helps to send user requests to the servers
operating under minimum server load to extract maximum performance. The longer the response
time is, the higher the corresponding load is. The suggested LBBSRT method effectively leverages
server resources, resulting in superior load-balancing outcomes and has been shown to have
significant advantages in terms of overall response times and load balancing when contrasted with
conventional Round Robin and Random strategies.
André and Fernando introduced a fault-tolerant controller framework for Software-Defined
Networking (SDN) called RAMA [12]. The novelty of the solution lies in Rama not requiring
changes to OpenFlow nor to the underlying hardware, allowing immediate deployment. The Rama
controller framework adopts a primary/backup model to ensure SDN controller fault tolerance.
RAMA, has a high-level architecture with OpenFlow-enabled switches, controllers managing the
switches, and a coordination service. The model includes a primary controller and multiple backup
controllers to tolerate faults. The primary/backup model allows for fault tolerance by electing a
new leader when the master controller fails. The coordination service ensures strong consistency
among controllers, but it becomes a system bottleneck due to the need for agreement between
replicas. The protocol presented aims to handle switches' state consistently in the presence of
faults. Furthermore, the RAMA controller framework guarantees three essential properties: (i)
events are processed precisely once by the controllers, (ii) all controllers process events in the
same order, ensuring they reach the same state, and (iii) switches process commands exactly once
using Open Flow bundles [13].
The RAMA controller framework uses a two-stage replication protocol ensuring the consistency
of the controller state. The first stage involves replicating the event to all replicas of the controller.
The second stage involves verifying that the event has been processed successfully by all replicas
of the controller as figure 4.

Figure 4. RAMA event processing structure [12]

RAMA offers significant advantages as it ensures consistent command and event processing,
providing equivalent strong assurances as Ravana [14] without the need for any modifications to
switches or the OpenFlow protocol. This quality makes RAMA a highly effective facilitator for
the seamless implementation of fault-tolerant SDN solutions. RAMA is a robust SDN controller
platform designed for fault tolerance, providing equivalent strong assurances as Ravana without
the need for any modifications to switches or the OpenFlow protocol. On the other hand, RAMA
does have some drawbacks compared to Ravana. The implementation of Rama results in higher
costs due to increased network message exchanges and the introduction of additional mechanisms
like bundles, which add to the overall overhead of the solution. Despite these drawbacks, the
performance impact is relatively minor, and Rama's core value proposition of ensuring consistent
command and event processing without requiring modifications to switches or the OpenFlow
protocol remains compelling. Consequently, Rama remains a valuable enabler for the immediate
adoption of fault-tolerant SDN solutions.
[15] Introduced a mechanism which aims to achieve efficiency by reducing host overhead and
preventing packet reordering, without being restricted to a particular version of OpenFlow
supported by the used controllers and switches. The proposed mechanism was assessed using
MultiPath Transmission Control Protocol (MPTCP) [16] . The study explores the integration of
SDN components, such as the SDN controller and virtual switch, within end-hosts to improve
network performance. The MPTCP connection functions as a thin intermediary between the
application and TCP layers. It enables the creation, management, and termination of TCP sub-
flows, which start with the exchange of SYN, SYNACK, and ACK messages, as shown in Figure
5. The MPTCP protocol utilizes several types of messages, such as MP_CAPABLE,
ADD_ADDR/REMOVE_ADDR, DATA_FIN, FIN flag, RST/FIN, and MP_FASTCLOSE, to
establish and terminate connections and ensure backward compatibility.

Figure 5. MPTCP Operations [15].

MPTCP connection must support the throughput of all flows without unfairly affecting normal
TCP flows in the network. Additionally, both ends of the connection must be controlled during the
connection's lifetime for the successful deployment of MPTCP. The introduced mechanism
architecture setup involves a switch that controls the host's external network interfaces (eth0 and
eth1). Two virtual network interfaces (veth0 and veth1) are connected, where veth0 is linked to
the switch, and veth1 is an internal gateway for application traffic. The switch employs match-
action rules set by the controller to implement traffic load balancing, selecting the external
interface (eth0 or eth1) through which packets are forwarded as illustrated in figure 6.

Figure 6. The architecture of host-based network load-balancing

mechanism using OpenFlow [15].
This mechanism optimizes traffic distribution and load balancing within the network by
performing the following steps:
• Packet Processing: The controller receives packets from the switch.
• TCP SYN Check: The controller checks if the received packet is a TCP SYN packet.
• Interface Assignment: If the packet is a TCP SYN packet, the controller employs a
weighted round-robin load-balancing algorithm to assign an appropriate interface to the
new flow.
• OpenFlow Rule Creation: Once the interface is determined, the controller creates an
OpenFlow rule that matches TCP/IP packets and the specific source port of the received
packet.
• Forwarding Instruction: The controller instructs the switch to forward the packet through
the chosen interface as the output action.
• Rule Installation: Finally, the controller installs the newly created OpenFlow rule on the
switch.
By carrying out these steps, the introduced mechanism enables dynamic interface assignment
and traffic redirection based on OpenFlow rules, facilitating effective and efficient
management of network flows. Implementing the proposed mechanism in existing SDN-based
load-balancing approaches, it demonstrates excellent performance, not only when using single-
link capacity but also using Multipath TCP (MPTCP) approaches.
Hamza et al. [17] Proposed the multiple threshold load balance (MTLB) switch migration scheme
that aims to solve load imbalance and prevent controller overload. MTLB categorizes the load into
various progressive levels that serve as the foundation for switching the migration process in cases
where the load of a controller is dissimilar from that of others. This results in the threshold value
being modified dynamically. The threshold value is subject to change based on the load status,
with a dynamic approach that can shift from one value to another based on the average load. MTLB
utilizes a trigger factor instead of periodic updates to update load information among controllers.
The scheme initially categorizes the load into appropriate threshold levels for synchronization and
migration handling. When a controller's load exceeds or approaches the threshold, it notifies the
others to update their load information, reducing unwanted overhead from load information
synchronization and handling the migration effectively. If there is a significant difference in load
levels between controllers, the scheme executes a switch migration with careful consideration of
the emigrant switch and target controller as mentioned in figure 7.
 Figure 7. The architecture of the MTLB-Distributed SDN Model [17].

The scheme has three stages modules: checking for updates, detecting load imbalance, and
selecting the suitable switch and controller for migration shown in figure 8.

Figure 8. The MTLB scheme flowchart [17].

The system employs four status levels and three load thresholds. The highest threshold,
"Overload," signifies that the controller has reached its full capacity and prompts immediate
migration. Controllers in the "Highly Loaded" status can still operate for a limited duration but
necessitate migration if other controllers are idle or in a normal state. "Normal" controllers are
equipped to handle unexpected scenarios, while "Idle" controllers have the highest priority in
receiving switches. The system's load management is thus designed to ensure efficient resource
utilization and prevent performance bottlenecks. The MTLB scheme outperforms all other
schemes with its high average throughput of approximately 1300 and 5000 packet/s. It also excels
in having the lowest average packet delay compared to other schemes, while SMDM and EASM
fall in between, and DDS experiences significant fluctuations and the highest peak delay. In terms
of migration cost, SMDM has the highest, while EASM and DDS show similar costs due to their
migration decision similarities. On the other hand, the MTLB scheme boasts the lowest migration
cost and packet loss. The MTLB scheme is also superior in terms of communication overhead,
being the lowest among all schemes. SMDM and EASM rank in the middle, while DDS exhibits
high fluctuations based on traffic load. The MTLB scheme's effectiveness lies in its use of
controller load status for efficient load information dissemination among controllers. Overall, the
MTLB scheme offers the best performance and efficiency among the evaluated schemes.
To address the challenges of communication delay between controllers and switches, as well as
inter-controller communication issues resulting from link failures in the network, Chunlin et.al
[18] introduced a novel model based on task latency and dynamic constraints. The heuristic ant
colony algorithm (HACA) [19] is employed for dynamically allocating computational resources,
considering factors such as the traffic volume of each controller, available resources, the distance
between controllers and switches, and the time delay between controllers caused by the
communication delay occurring when controllers interact with each other. The proposed model
leverages two key aspects: first, the development of a dependable controller placement method
that optimizes latency and load considerations, improving the load optimization multi-controller
placement (LOCP) algorithm. Secondly, the formulation of a resource allocation algorithm that
takes into account task latency and reliability constraints, using the HACA. Improvement
operation of a multi-controller placement (LOCP) algorithm illustrated in figure 9: (1) User
devices connect to the edge computing layer via network access points (e.g., wireless access points,
base stations) to access services. (2) A multi-access edge computing (MEC) server [20] located
near the base station provides computing, and storage resources, and collects/analyzes information
from end devices, reducing data and enhancing network service quality. (3) The MEC server
connects to a local SDN controller through an OpenFlow switch for efficient network traffic and
resource management. (4) A global controller oversees the local SDN controller, updating data
matching and processing rules through operation status exchange. (5) Controllers communicate
using an east-west interface, ensuring seamless coordination.

Figure 9. Dynamic controller placement and resource allocation in an SDN-based multi-access edge computing
system [18].
By adopting this architecture, the approach optimizes network performance and resource
utilization through edge computing and SDN-based control, resulting in enhanced user experiences
and efficient network services. Regarding the controller placement problem, the research focused
on three primary performance metrics. Firstly, it examines the delay between the controller and
the switch. Secondly, it analyzes the time delay between controllers, due to inter-controller
communication. Lastly, the research addresses load balancing to ensure equitable distribution of
processing tasks among controllers. The study addressed the controller placement problem,
focusing on delay and load optimization, while also considering network link connectivity.
Experimental results demonstrate the improved LOCP algorithm's efficacy in achieving a balanced
network load and reducing network overhead, particularly in small and medium-sized networks
when compared to existing approaches. In addition, the improved HACA algorithm' solves the
problem of work resource load allocation and effectively reduces the user resource response time
as well as the average completion delay under the premise of ensuring the reliability constraint.
Jehad and Byeong-hee [21] Proposed a mathematical decision-making framework by calculating
the optimal controller in terms of its features that enhance the performance of the Software-Defined
Internet-of-Things (SD-IoT) using an analytical network decision-making process (ANDP) model
[22]. The controller selection technique is based on a qualitative and quantitative examination of
SDN controllers for SD-IoT. They determined ten considered characteristics of the controllers for
the IoT environment listed in table 1. ANDP is used to determine the high-weight SD-IoT
controller by computing weights for each controller, and then ANDP ranks the controllers with the
best feature set for SD-IoT among others.
Calculate Controller Weights after applying the comparison matrix that is the outcome of all
judgments of the controllers’ supporting features which are important in SDN based on the 10
characteristics listed in the following table.
Table 1. List of features for SD-IoT performance evaluation [21].
Serial# Notation Name Description
1 B1 OpenFlow-support OpenFlow version1.0–1.5
2 B2 GUI Web based or Python-based
3 B3 NB-API support REST-API
4 B4 Clustering support To ensure reliability and performance
5 B5 Openstack networking Enabling different network technologies via quantum API
6 B6 Synchronization State synchronization of the clusters
7 B7 Flow requests handling The capability to handle the flow requests
8 B8 Scalability Adoptability in the extended networks
9 B0 Platform support Windows, Mac, Linux
10 B10 Efficient energy management The ability to utilize energy efficiently

The evaluation of controllers’ supporting features is represented according to the level of support
in which G1 indicates extremely low support and G4 denotes very strong support. G2 indicates
medium support, but G3 only reveals strong support, where the characteristics evaluation score
from G1 to G4. After that comes the comparison stage of Controllers regarding their Features for
SD-IoT, finally, the controller weights were calculated. This model presented a novel controller
selection approach for SD-IoT environments, based on the Analytical Network Process (ANDP)
model, is evaluated in terms of delay, throughput, CPU utilization, and reliability. Figure 10
 illustrates the ANDP model for paired comparisons in selecting an SD-IoT controller. The figure
represents the ranking model of the ANDP, comprising a features cluster (top one) and an
alternatives cluster (bottom one). Additionally, a circular line indicates the interdependency among
these features.

Figure10. The ANDP model for controller selection in SD-IoT [21].

The proposed model was compared with previous benchmark schemes, namely AHP [23] and EB-
TOPSIS [24], through a series of experiments. Key Findings achieved that the proposed
controller:
- Reduces delay in various traffic scenarios.
- Increases throughput while efficiently utilizing the CPU.
- Exhibits enhanced reliability during link failure recovery.
Jehad et.al. [25] Introduced ESCALB, a load-balancing scheme designed for multi-domain SDN-
enabled IoT networks (SD-IoT). Its main goal is to efficiently migrate switches to controllers with
available resources in a dynamic manner. ESCALB uses a hierarchical model for a control plane
consisting of multiple domain controllers (DCs) and a global control (GC) plane. The GC plane
comprises four sub-modules as figure 11, including the Load Calculation Module (LCM) and ANP
Module (ANPM), which monitor load status by receiving information from the Distributed Control
Plane (DCP) and rank controllers based on CPU usage, Flow Requests Capacity FRC, memory
utilization, and the number of attached switches. In addition, the ANP Module (ANPM) utilizes
the Analytic Network Process (ANP) model to prioritize slave controllers in the DCP. ANPM
employs a mathematical procedure with a 9-point scale matrix to rank controllers, where 1
indicates equal importance and 9 signifies extreme significance. The Switches Migration Module
(SMM) initiates switch migration to slave controllers if the master controller's load exceeds a
predefined threshold. Flows Forwarding and Updating Module (FFUM) collaborates with ANPM
and SMM to prioritize slave controllers, migrate switches, and forward flow requests to controllers
with higher weights.
 Figure 11. Proposed framework for load balancing using ANP module with SD-IoT [25].

The primary objective of the GC plan is to optimize network performance and resource utilization
by utilizing ANP-based ranking, switch migration, and load-balancing mechanisms. The proposed
scheme was compared with previous benchmark schemes, namely SASLB [26], DLB [27], SCLB
[28], and SMLB [29]. ESCALB's effectiveness resides in its adeptness at intelligently selecting
the most appropriate controller for load distribution, resulting in enhanced performance within
SDN environments.
Table 2 outlines a comprehensive overview of studies pertaining to traffic load balancing and
resilience, encompassing details regarding description, objectives, and traffic routing methods. In
addition, Table 3 offers a comparative analysis of decision-making criteria and Implementation
levels within the Data Plane and/or Control Plane.
Table 2. Traffic Load Balancing and Traffic Resiliency: Proposals description, objectives, and Techniques Method
Employed.
Reference Proposal description Objectives Techniques
Hong et. al. [11] LBBSRT chooses the server Enhancing the load balancing effect Least
characterized by minimum response by reducing the server response Response
time in order to provide services to time. Time Load
the users. Balancing
André and Rama controller framework is a Prevent the controller from Fault tolerance
Fernando [12] promising approach to ensuring the becoming a single point of failure,
fault tolerance of SDN controllers. which necessitates the integration of
the switch state into the fault-
tolerant SDN framework.
Anees et. al. Network traffic control mechanism Improving Load Balancing and Least
[15] achieves efficiency by reducing host enhancing network efficiency by Connection
overhead and preventing packet dynamically selecting the most Load
reordering. suitable network interface for each Balancing
traffic flow.
Hamza et al. MTLB is an efficient load balancing Employing a switch migration Clustering
[17] scheme for managing the load scheme to address load imbalance
distribution among distributed and prevent controller overload
Software-Defined Networking involves categorizing the load into
(SDN) controllers. multiple progressive tiers and
dynamically adapting the threshold
value.
 Chunlin et.al. A novel dynamic controller Efficiently determine the optimal Clustering
[18] placement method that focuses on locations for placing controllers
optimizing both delay and load within a network to solve the multi-
factors by improving the multi- controller placement problem,
controller placement (LOCP) and enhancing the overall network
ant colony algorithms. service quality performance.
Jehad and Optimal SDN controller selection is Enhancing throughput while Clustering
Byeong-hee [21] critical to ensure optimal network efficiently utilizing the central
usage, leading to improved Quality processing unit (CPU) and
of Service (QoS) in SD-IoT, which minimizing recovery latency during
involves assessing its attributes and network failures. Improving latency
validating its performance within in both normal and heavy traffic
the SD-IoT environment. scenarios
Jehad et.al. [25] A multi-criteria decision-making Demonstrating the problem of static Adaptive Load
based slave controller selection slave controller assignment to Balancing
strategy for SDN in IoT networks ensure effective migration of SDN
with dynamic switch migration. switches.

Table 3. Traffic Load Balancing and Traffic Resiliency: Decision-Making Criteria and Implementation in the Data
Plane and /or Control Plane.
Reference Control Data plane Decision-Making based on
plane
Hong et. al. [11] The controller selects the server with minimum
ü ü or stable response time according to obtained
server data.
André and Fernando Controller primary/backup Decisions and
[12] ü Coordination Service Decisions
Anees et. al. [15] enables dynamic interface assignment and traffic
ü redirection based on OpenFlow rules.
Hamza et al. [17] The dynamic threshold value can be changed
ü depending on the average load status.

Chunlin et.al. [18] Calculating the optimal placement of multiple

ü controllers based on the network topology,
traffic load, and available computational
resources.
Jehad and Byeong-hee Calculate Controller Weights utilizing the all-
[21] ü considered judgments of SD-IoT controllers'
supporting features.
Jehad et.al. [25] Monitors the control plane in real-time,
ü acquiring load information to assess and
prioritize slave controllers and ensure successful
switch migration.

3.2) Traffic Optimization and Path Selection

A novel traffic-aware QoS control mechanism for SDN-based virtualized networks was introduced
[30], based on the single rate three color marker (srTCM) proposed by the Internet Engineering
Task Force (IETF) [31]. This mechanism combines srTCM (Single rate three color marker) with
two novel global token buckets (srTCM+ GTB) to effectively meter and mark packets from each
virtual network, ensuring QoS in IP networks. Initially, the mechanism employs srTCM, which is
based on Differentiated Services (DiffServ) principles [32], to classify and manage network traffic.
DiffServ enables the provision of low latency for critical network traffic, such as voice or
streaming media, while offering simple best-effort service to non-critical services like web traffic
or file transfers. By using srTCM, packets within a stream are metered and categorized into three
traffic parameters and add two Token Buckets (TB). The three considered traffic parameters are
committed information rate (CIR), committed burst size (CBS), and excess burst size (EBS), which
are respectively marked as green, yellow, or red.
IF packet size:
< CBS, then packet marked green; packets can be forwarded immediately.
<CBS & < EBS, then packet marked yellow; packets can be forwarded or dropped
according to the link state.
Otherwise, then packet is marked red; packets will be dropped directly.
Two Token Buckets (TB): A token is added to the C-bucket every 1/CIR s.
If No. of Tokens in C-bucket > CBS, the extra token is added to E-bucket, then If No. of Tokens
in E-bucket > EBS, then discard extra token as illustrated in figure 12.

Figure 12. Token refill scenario of srTCM +GTB [30]

The proposed architecture of an autonomic managed network is illustrated in figure 13. A

virtualization plane is introduced between the control plane and data plane to efficiently allocate
the physical network based on the unique requirements of each virtual network. The virtual
networks are categorized into different service levels using the differentiated services (DiffServ)
approach. This involves color-marking packets according to the link status and inserting them into
corresponding queues with different precedence using a weighted random early detection (WRED)
queue [33].
 Figure 13. The architecture of autonomic managed network [30].

An autonomic manager is responsible for monitoring, collecting, and analyzing traffic and resource
usage of both virtual and physical networks. It then dynamically redistributes resources among all
virtual networks through a resource manager. To optimize resource utilization during periods of
light traffic in some virtual networks, the srTCM and marking algorithm are extended. Two
additional buckets, known as the global C bucket and global E bucket, are introduced. Similar to
private token buckets of each virtual network, these are not automatically refilled, and their depths
are not fixed. This adaptive approach ensures efficient resource allocation and enhances overall
network performance. The performance evaluation was conducted by analyzing the TCP and UDP
performance of the proposed srTCM+ GTB method compared to CBBPM. The proposed approach
employs a traffic-aware quality-of-service control mechanism to enhance the quality of service for
virtual machines in software-defined networking-based virtualized networks. The findings
demonstrate the efficacy of this mechanism because srTCM+ GTB is better able to control
congestion and ensure that all virtual networks receive their fair share of bandwidth which
guarantees their isolation.
Junjie et.al. [34] Proposed a collaboration approach between MPTCP (Multipath TCP) and SR
(Segment Routing) is aimed at addressing the resource consumption challenges in SDN-based
Data Center Networks (DCNs) [35] [36]. The approach enables the simultaneous use of multiple
paths for data transmission using MPTCP and provides a flexible and scalable mechanism for
forwarding packets based on a segment identifier using Segment Routing (SR). In a large DCN or
during a peak period, when a large flow arrives, it is divided into multiple subflows and transmitted
using the MPTCP protocol. The SDN controller then allocates these subflows to specific
transmission paths and maps each path into an SR path. This innovative method effectively
manages traffic and reduces the demand for storage resources. The architecture proposed through
the mentioned collaboration adopts a comprehensive four-layer approach for the DCN to ensure
enhanced clarity and description. This four-layer architecture offers a comprehensive and well-
structured approach to designing the DCN, incorporating advanced technologies such as SR and
MPTCP while leveraging the capabilities of SDN for efficient traffic management. MPTCP offers
a traffic-splitting capability, enabling the division of a flow into multiple subflows. This ensures
that the flow can be simultaneously transmitted via multiple paths between peers. MPTCP proves
beneficial in increasing the throughput of DCNs while reducing the likelihood of network failures,
packet losses, and delays. In case of unavailability or poor performance of a path, MPTCP allows
the flow to be rerouted through an alternative path. MPTCP architecture comprises four essential
components: path management, packet scheduling, subflow interface, and congestion control.
Regarding the SR method, it operates using a series of segments as an ordered list of routing
instructions for packets. SR necessitates Traffic Engineering (TE) decisions for the entire network.
The transmission path is represented by two types of segments: node segments and adjacency
segments. Node segments serve as unique identifiers for nodes within the network domain, and
they possess global significance, enabling other nodes to transmit packets based on these
identifications, often using Open Shortest Path First (OSPF) as the default protocol.
On the other hand, adjacency segments represent node-local interfaces that differ from node
segments and are locally significant. These segments facilitate the transmission of packets to
specific adjacent nodes through the associated interfaces.
A transmission example of MPTCP and SR is shown in figure 14.

Figure 14. A transmission example of MPTCP and SR [34].

In the example, the controller selects three paths through the network for a packet to take. These
paths are {A-B-E-G-H-J}, {A-D-G-J}, and {A-C-FG-I-J}. The SR technology then expresses
these paths as a segment label list. The segment label list is a list of the segments that the packet
must traverse in order to reach its destination. When the packet receives the segment label list from
the controller, it follows the instructions in the list to complete the transmission process. The
segment label list is a more efficient way to represent a path through the network than a traditional
routing table. This is because the segment label list only includes the segments that the packet must
traverse, while a routing table includes all of the possible paths through the network. As a result,
the segment label list can be used to select the best path through the network for a packet, even if
the network topology changes. In summary, the proposed MPTCP & SR showcases superior
performance in terms of throughput maintenance and average link utilization under varying
transmission demands. The dynamic path adjustment capability contributes to more efficient load
balancing and response management, leading to more stable and improved network performance.
Shengru et. al. [37] introduced a forwarding technique that's independent of protocols, efficient in
bandwidth usage and saves flow-table space. It utilizes POF-FIS, a protocol-oblivious forwarding
instruction set. To implement POSR in an SDN network system, packet format is designed, and
packet processing pipelines are developed to support unicast, multicast, and link failure recovery.
Packet format design
In consequence of the independent nature for POF protocol [38], the packet format design for
POSR allows for a dedicated approach to accommodate the network scenario without the need to
reuse existing protocol packet fields. The configuration of POSR packet header fields is illustrated
in Figure 15.
 Figure 15. Header format design of POSR unicast Figure 16. Header format design of POSR multicast
packets [37] packets [37]

Specifically, the source routing header fields are introduced between the Ethernet and IP headers,
and the Ethernet header Type field is defined as "0x0908" to signify a POSR packet. The Time-
to-live (TTL) field indicates the remaining hops of the packet, while the Port field stores the
designated output port on the switch for a given hop. To encode the routing path, a POSR packet
incorporates multiple Port fields, where each intermediate switch extracts the outermost Port field
to determine the packet's designated output port. For enabling POSR-based multicast, the POSR
header's Port field is substituted with a VPort field, as depicted in Figure 16. The VPort field
consists of multiple bits, with the first bit serving as the Fork Flag, indicating whether the
corresponding switch acts as a fork node on the packet's multicast tree. The remaining bits in the
VPort field represent the Group Label, which is assigned for each active multicast session. A fork
node signifies a switch from which multiple branches originate within a multicast tree.

POSR Packet Processing Procedure

Upon the arrival of the first packet of a flow at the ingress POF switch, the switch sends a PacketIn
message to the controller as no flow entry has been set up for the flow. The controller then
calculates the flow path and sets up a flow entry in the ingress switch, which directs the switch to
encode path information in the flow's packets using the POSR format. Given that the packet
processing procedure for all POSR packets is identical in any intermediate switch, including
destination switches, flow entries can be installed in all POF switches during network initialization,
allowing POSR packets to share them.

FAST LINK FAILURE RECOVERY WITH POSR

To enable fast failover (FF) in POF switches, two entries are included in each FF group table to
monitor switch port status and determine the backup path segment based on the packet's output
port. The first entry in the FF group table forwards packets normally when the output port is up. If
the output port is down, the second entry initiates POSR-based link failure recovery using the
routing instructions of the backup path segment to replace those of the broken link in the headers
of affected POSR packets, as illustrated in figure 17.

Figure. 17. POSR-based fast link failure recovery [37].

By leveraging the FF group table in POF switches, the system proactively monitors switch port
status and identifies any port failures quickly. This approach enables affected packets to be swiftly
rerouted through the backup path segment, minimizing network disruptions, and enhancing overall
reliability in case of link failures. As the data transfer rate increases, OF-SP experiences a faster
escalation in packet loss. The findings of this research suggest that POSR has the potential to be a
viable solution for addressing the scalability challenges of SDN. POSR's ability to provide efficient
source routing without relying on specific protocols allows for greater flexibility and adaptability,
making it a promising option for large-scale SDN deployments.

Conserving energy is crucial not just from a financial and ecological standpoint, but also for
ensuring the sustainable expansion of the Internet. This is because the delivery of power to and the
removal of heat from massive data centers present significant challenges [39].
[40] The academic work introduced a heuristic scheme named Exact Path Control (EPC) designed
for the incremental deployment of SDN switches in hybrid SDNs. EPC involves flow-level explicit
path control to power off redundant links and switches in the network, thereby conserving energy.
EPC algorithm procedures
The algorithm functions as follows: Firstly, it sorts all flows based on priority and then reroutes
them one by one until all requests are arranged. Then, it calculates the energy savings by shutting
down the active edges and switches. Additionally, it sorts all links according to traffic volume in
ascending order and judges whether each link and switch can be turned off. It also checks whether
a specific link can be powered off. The algorithm must ensure that the forwarding path's delay does
not exceed the maximum delay and that the volume on each link is within its capacity limit. To
enhance Network Congestion Avoidance (NCA), they proposed upgrading the least key nodes
incrementally that all flows must pass through. This is achieved by SDN switches rerouting packets
based on multiple MPLS labels that contain forwarding port numbers of switches on the route.
Fine-grained flow scheduling is crucial for energy conservation, and encapsulating MPLS labels
can help reroute flows and power off idle links and switches. When selecting switches to update,
we consider the topologies' structure. Two optimization techniques have been identified: (1)
isolating nodes without flows passing through them and retaining nodes with flows, and (2)
separating specific sub-topologies from the original topology and selecting key nodes in the sub-
topologies as SDN switches.
The provided academic content demonstrates an example of utilizing MPLS labels to achieve
energy savings in a network as shown in figure 18. The scenario involves two flows, f1 from h1 to
h3, and f2 from h2 to h3. The controller selects switch s2 to install flow entries for rerouting flow
f1. Upon entering SDN switch s2, packets of flow f1 are encapsulated with 3 MPLS labels, each
indicating a forwarding port along its route. As the packets traverse through the switch, they pop
up one MPLS label at a time, determining their forwarding port. This rerouting results in a new
path for flow f1, leading to energy savings by shutting down switches s3, s4, and the corresponding
links.
 Figure 18. An example of Using MPLS label to achieve network energy saving [40]

The novel aspect of this approach is that, unlike traditional network MPLS, there is no need to
maintain a path state in the forwarding path beyond the ingress node. This is because packets are
now routed based on the list of labels they carry, enabling a more efficient and energy-saving
routing strategy.
Wu et. al. addressed two critical issues in multi-path algorithms: scheduling efficiency and end-
to-end delay fluctuations, which can impact video transmission quality. To tackle these challenges
in media streaming applications, a dynamic and adaptive multi-path routing algorithm (DAMR) is
proposed, utilizing SDN for centralized routing computations and real-time network state updates
[41]. The key module of DAMR is the routing module, responsible for aggregating bandwidth
based on monitoring information to ensure reliable transmission and avoid data packet loss due to
link failures or congestion. By dynamically allocating data flows to multiple effective paths
between nodes, DAMR optimizes network resource utilization, reducing congestion, packet loss
rates, and end-to-end delays.
Three main algorithms constitute DAMR:
The network topology update algorithm optimizes the network topology and updates link
information. The network routing algorithm handles packets in packets sent by OpenFlow switches
and makes forwarding decisions based on routing decisions, acquiring QoS parameters from the
monitor module for optimal path calculations. The packet loss rate algorithm ensures real-time
measurement through periodic updates and calculates packet loss rates for each flow to determine
optimal routes. The flowchart designed for the proposed algorithm is shown in Figure 19. Upon
program initiation, the controller first checks for an optimal path to find, periodically querying the
QoS parameters of the underlying network at 5-second intervals from the perspective of
OpenvSwitch. When a packet reaches the controller, it calculates all accessible paths and their
available bandwidths. The path with the maximum available bandwidth becomes the flow
transmission path. If this path meets the delay constraint, it becomes the optimal path, and video
streaming flows remain unchanged. However, if the path fails to meet the delay constraint, the
controller utilizes the DAMR algorithm to find the optimal path. If no such optimal path exists due
to unmet delay constraints, no rerouting occurs. Nevertheless, when an optimal path is found, the
controller acquires QoS information from the monitor module and calculates link weights between
node pairs. The link-state status information of each link on the flow's path is then updated. By
utilizing the topology management module from the floodlight controller, link-state information,
including connected nodes and ports, is stored in clusters. By traversing each node and its
connected nodes, the full path from the source to the destination node is obtained, ensuring the
optimal path that meets the delay constraint is achieved.

Figure 19. The flowchart design of the DAMR algorithm [41]

DAMR algorithm is compared with a single-path algorithm (e.g., Dijkstra and Bellman-Ford) that
selects a subset of paths for video streaming based on additive cost properties like hop counts. In
contrast, the ECMP algorithm [42]distributes traffic equally among multiple equal-cost paths,
involving all links in the network. Experimental results show DAMR overcomes computational
overhead, effectively adapting to dynamic network changes, and improving link utilization and
user service quality. DAMR utilizes OpenFlow centralized control to optimize resource allocation
and aggregate bandwidth resources by employing optimization theory.
Muteb and Abdelmounaam presented POX-PLUS [43] is an upgraded version of the commonly
used POX controller in SDN networks, which includes a novel routing module called DRAPSP
(Dynamic Routing based on All Pairs Shortest Paths). DRAPSP is responsible for calculating and
effectively managing the shortest routes between nodes in the SDN. POX-PLUS consists of three
primary components as shown in figure 20:
• DR-APSP: This module is the main routing component that operates on our dynamic routing
all-pairs shortest path (APSP) algorithm [44]
• POX Controller is the original POX controller that has been modified to take on the additional
responsibility of updating DR-APSP with information regarding the initial topology, as well as
any subsequent modifications to the network.
• Data Structures are utilized to store information related to the topology and its shortest paths.
These data structures comprise the Topology Graph, Hosts, and a forest of shortest paths trees
(SPT).

Figure 20. POX-PLUS Architecture [43].

The controller detects any link event in the network utilizing link deletion and link insertion
algorithms. When detecting a link deletion, DR-APSP starts a loop over all shortest path trees
(SPTs), that consists of two main phases. In the first phase, the algorithm identifies the affected
switches in each SPT, and subsequently removes affected flows that selected the deleted edge, for
each of the affected switches. Then the controller will apply a fresh flow rule in response to
subsequent requests. In the second phase update the current SPT for future requests using APSP
approach. The Link Insertion algorithm is triggered by a link insertion event, following which it
begins a loop over all vertices. The loop comprises two primary steps: the first step entails updating
the current SPT, while the second step involves removing impacted flows from affected switches.
The measured performance parameters include:
"init time" for initializing flow tables at all switches, measuring the duration of data packet
transmission between every host pair in the network. "Add time" indicates the time taken by an
application to transmit a data packet between any two hosts after incorporating new network links.
"Delete time" represents the time required for an application to send a data packet between any
two hosts following the removal of network links. "Selected paths time" measures the duration
taken by an application to transmit a data packet between any two hosts after configuring flow
tables at all switches.
DR-APSP module is compared to the three routing schemes offered by the POX controller (l2
learning, l2 multi, and l3 learning). DR-APSP is the sole method that preserves the shortest paths
and flow rule tables without recomputing paths. The research shows that l2 learning and l3 learning
involve a high cost due to a short idle time for flows in flow tables, which is necessary for fast
recovery.
Yi-Ren et. al. [45] provided a solution that helps to solve a traffic engineering (TE) problem of
SDN in terms of throughput and delay. This solution developed a reinforcement learning routing
algorithm (RL-Routing) which predicts the future behaviour of the underlying network using an
RL agent to learn the optimal routing paths in a network. The RL agent interacts with the network
by selecting actions (routing paths) based on the current state of the network and suggests better
routing paths between switches. The RL-Routing application comprises two main modules are
shown in figure 21. The first module, the Network Monitoring Module (NMM), utilizes both
passive and active network measurements to collect crucial information about network devices,
including link delay, throughput, and port speed. This data is utilized to represent states and
compute rewards. The second module, known as the Action Translator Module (ATM), converts
the selected action by the agent into a series of appropriate OpenFlow messages.

Figure 21. The components of RL-Routing [45].

These messages are employed to update the flow tables of switches when configuring a new path.
To prevent Packet-In messages from being sent to the controller, the ATM transmits these messages
from the last switch of the path to the first switch. Finally, the old rules in the switches of the
previous path are deleted. RL-Routing has the potential to address scalability issues in routing by
automating the process of path selection and reducing the need for manual configuration and
maintenance. However, the effectiveness of RL-Routing in addressing scalability issues can
depend on various factors, such as the size and complexity of the network and the performance
metrics being evaluated. The performance of RL-Routing was evaluated on three well-known
network topologies: Fat-tree, NSF Network (NSFNet), and Advanced Research Projects Agency
Network (ARPANet), comparing RL-Routing with two widely used baseline solutions: Open
Shortest Path First (OSPF) and Least Loaded routing algorithm (LL). The evaluation metrics are
the reward function, which is a score computed using network throughput and delay, and the
utilization rate, which is calculated in the destination switch. The reward function can be adjusted
to optimize either upward or downward network throughput. Gururaj et.al. introduced the
resource-efficient multicast tree construction model (REMTC) [46] that uses Dijkstra’s Shortest
Path algorithm for initial tree formation, identifies a multicast path, and processes the Shortest
Path Tree to reduce the overall hop count and path cost. Aims to reduce tree alteration using more
common paths to reach the devices by tree optimization algorithm which enables the dynamic join
and leaves of participating devices. In this model, the multicast tree construction algorithm
continuously monitors the network and user requirements and adapts the multicast tree
accordingly. For example, if a link becomes congested or unavailable, the algorithm can
dynamically reroute traffic to other available links to maximize network efficiency and meet user
requirements. REMTC model seeks to construct a tree comprised of nodes and links with the
lowest possible values for both total path cost and total hop count. The REMTC algorithm is
designed to identify the optimal values for both of these factors during the initial construction of
the multicast tree. The trunk path that connects the maximum number of end devices in the shortest
path tree (SPT) is then selected as it improves route stability. The construction of the multicast tree
utilizes level information to enhance resource utilization. The tree optimization algorithm works
as follows, the collection of participant information is performed by the SDN controller during the
establishment of a session between end devices and a server. The dynamic joining/leaving of the
participants from/to the multicast group is managed through Session Initiation Protocol (SIP) and
Session Description Protocol (SDP) messages [47]. To evaluate the bandwidth availability, the
SDN controller centrally computes the link cost based on switch statistics gathered at regular
intervals, including the number of packets transmitted and received by each switch port. Dijkstra's
shortest-path tree algorithm uses this link metric to construct the shortest-path tree, which is then
utilized to determine the trunk path and level formation. The performance of REMTC was
evaluated by comparing it with SPT, ST, BAERA [48], and OBSTA [49] for the following metrics:
- The "number of links" refers to the edges that are utilized in the multicast tree to forward the
multicast data from the source node to the participating nodes.
- "Processing Latency" represents the amount of time it takes for the algorithm to construct a
multicast tree for the set of participants.
- The "Rerouting Cost" is the delay that is incurred when constructing or rearranging the
multicast tree due to a dynamic join or leave of a participant.
The REMTC model is capable of maintaining stable bandwidth consumption during dynamic join
and leave events, which is not achieved by ST and SPT methods. The REMTC model only modifies
the multicast tree to accommodate dynamic participants, thereby avoiding changes to the
forwarding table of SDN switches and achieving multicast tree stability.
Majda et.al. presented a new model called Video Streaming Adaptive QoS Routing with Resource
Reservation (VQoSRR) for Software Defined Networking (SDN) networks [50]. This model is
designed to enhance the quality of video streaming services in SDN networks by offering adaptive
QoS routing and resource reservation. The VQoSRR model employs a centralized management
architecture to offer adaptive QoS routing and resource reservation tailored to video streaming
services. By utilizing a feedback mechanism, the VQoSRR model dynamically adjusts QoS
routing according to network conditions. Network state information, including bandwidth
availability, delay, and packet loss, is considered to determine the optimal path for video streaming
traffic. Furthermore, the model ensures resource reservations to guarantee sufficient resources for
video streaming services. To achieve this, the VQoSRR model divides the network into multiple
virtual networks and allocates resources to each one. It employs a bandwidth reservation
mechanism to secure the required bandwidth for video streaming traffic and a packet scheduling
mechanism to prioritize video streaming traffic over other types of traffic. Upon initiation of a new
video stream flow by the server, the switch sends the first packet copy to the controller's QoS
routing manager for determining the routing path. The VQoSRR employs two routing paths to
balance between frequent dynamic network state updates and reduced routing computation
overheads. One path is used as the primary route for the current flow, while the other serves as an
alternative path for potential rerouting. These two paths are selected based on a weighted graph,
where link characteristics such as packet loss rate and available bandwidth are determined using
QoS routing manager algorithms with the aid of topology manager and statistics collector modules.
In addition, this study introduced the Dynamic Traffic Rerouting Algorithm (DR-RA), which
adapts to network conditions to ensure video QoS compliance, ultimately enhancing video quality.
DR-RA updates the route's cache and reroutes traffic using alternative paths or generating new
ones. The algorithm operates periodically, reading network statistics at predefined intervals. In
case the current path violates the flow's QoS requirements, its flow entries are deleted from the
switch, and an alternative route is examined. If the alternative path also fails to meet quality
requirements, the algorithm calculates another path. Employing the alternative path offers several
advantages:
- Reduces time spent recalculating the routing path by utilizing the alternative route instead
of rerunning the routing algorithm.
- Improves response times for flow installation.
- Enhances network resilience by providing two paths, which can be beneficial in the event
of a path failure.
Experiments showed that VQoSRR improved user perception of video quality and offered better
control over routing and resource allocation and the videos delivered with DR-RA had higher
quality than those without DR-RA or VQoS-RR.
Table 4 outlines a comprehensive overview of studies pertaining to Traffic Optimization and Path
Selection, encompassing details regarding description, objectives, and traffic routing methods. In
addition, Table 5 offers a comparative analysis of decision-making criteria and Implementation
levels within the Data Plane and/or Control Plane.
Table 4. Traffic Optimization and Path Selection: Proposals description, Objectives, and Techniques method
Employed.
Reference Proposal description Objectives Techniques
Jiameng and srTCM+ GTB is a traffic-aware To provide a traffic-aware Traffic-Aware Multi-
Sang-Hwa [30] quality-of-service control quality-of-service control Path Routing with
mechanism for software-defined mechanism for software- (QoS)-based Routing
networking-based virtualized defined networking-based
networks. virtualized networks.
Junjie et.al. Develop a collaborative traffic Providing a better traffic Hybrid Multi-Path
[34] transmission mechanism management solution, which is Routing
utilizing MPTCP and SR to still effective in DCNs’ peak
address resource consumption hours.
issues in an SDN-based DCN.
Shengru et. al. Design protocol oblivious source Addressing the issue of Source-Based Multi-
[37] routing (POSR) by proposing a scalability in SDN to Path Routing
protocol-agnostic, bandwidth- significantly mitigate flow-table
optimized, and flow-table- consumption, minimize path
conserving packet forwarding setup latency, and expedite link
approach. failure recovery.
Xuya et. al. [40] Introduces a heuristic approach Attain energy efficiency Traffic-Aware Multi-
for the gradual deployment of through the proficient rerouting Path Routing with
SDN switches in hybrid SDNs to of flows and the optimal Energy-Efficient
intelligently power off redundant
 links and switches in the network shutdown of idle links and
to effectively conserve energy switches whenever feasible.
resources.
Wu et. al. [41] DAMR is a dynamic and Improving network Dynamic Multi-Path
adaptive multi-path routing performance by dynamically Routing with
algorithm that addresses packet selecting multiple paths while (QoS)-based Routing
loss, time delay, and bandwidth considering network congestion
constraints in multimedia and link quality.
applications through centralized
routing computations and real-
time network state updates.
Muteb and DR-APSP is an efficient routing Preserving the integrity of the Dynamic Routing
Abdelmounaam module devised to elevate the shortest paths and flow rule
[43] POX controller version to POX- tables following an update
PLUS, thereby facilitating the operation, obviating the
computation and perpetual necessity of recomputing paths.
maintenance of inter-node paths
within the Software-Defined
Networking (SDN) framework.
Yi-Ren et. al. RL-Routing represents a Mitigate scalability challenges Machine Learning-
[45] proficient approach to routing in routing through the based Routing
within SDN, employing automation of path selection,
Reinforcement Learning (RL) to thereby minimizing the reliance
predict the future behaviour of on manual configuration and
the underlying network and upkeep.
suggests optimal routing paths
between switches.
Gururaj et.al. The REMTC algorithm is a Improving network resources Traffic Engineering
[46] resource-efficient multicast tree utilization and communication
construction model that through Constructing a stable
constructs a reliable and scalable multicast tree with more
multicast tree for multiple common paths to reach
receivers. This, in turn, multicast participants.
optimizes resource consumption
and enhances bandwidth
efficiency.
Majda et.al. VQoSRR is a novel model Improving the quality of video Quality of Service
[50] utilizing queue mechanisms to streaming services in SDN (QoS)-based Routing
meet bandwidth guarantees for networks by providing adaptive
video traffic, while also offering QoS routing and resource
two routing paths between the reservation.
source and destination to cater to
multiple QoE constraints.
Table 5. Traffic Optimization and Path Selection: Decision-Making Criteria and Implementation in the Data Plane
and /or Control Plane.
Reference Control Data plane Decision Making based on
plane
Jiameng and Sang- The traffic load, bandwidth and latency of each
Hwa [30] ü virtual machine.

Junjie et.al. [34] Dividing the main flow into mul;ple subflows, and
transmi>ng using MPTCP, while implemen;ng an
ü ü ordered list of rou;ng instruc;ons for packets
using SR.
Shengru et. al. [37] Redesign the source packet header format and
ü design a packet processing pipeline that is
op;mized for unicast and mul;cast traffic.
Xuya et. al. [40] Sor;ng all flows based on their priority and then
sequen;ally rerou;ng them un;l all requests are
ü organized. This entails selec;ng SDN switches for
deployment based on flow requirements and the
associated switch costs.
Wu et. al. [41] Calcula;ng the aggregated bandwidth using real-
ü ü ;me monitoring informa;on and a periodic
scheduling strategy, taking into considera;on both
congested and uncongested scenarios.
Muteb and Detec;on of network link events, which
Abdelmounaam subsequently trigger the removal of all impacted
[43] ü and obsolete flow rules that do not contribute to
any of the established shortest paths following the
update opera;on.
Yi-Ren et. al. [45] The RL agent learns to select paths that op;mize
network performance metrics, such as latency,
ü throughput, and packet loss. The agent an;cipates
future network behaviour to propose improved
rou;ng paths between switches.
Gururaj et.al. [46]
ü ü During the tree forma;on phase, the algorithm
calculates the shortest path based on link cost.
Majda et.al. [50] Dividing the network into mul;ple virtual
ü ü networks and alloca;ng resources to each virtual
network. A bandwidth reserva;on mechanism
reserves the required bandwidth for video
streaming traffic, and a packet scheduling
mechanism priori;zes video streaming traffic over
other traffic types.

4) Evaluation Tools and Performance Results

4.1) Evaluation Tools
Table 6 provides a comprehensive overview of simulation model tools, datasets, and metrics used
to evaluate the proposed research in traffic load balancing and traffic resilience. The table is
organized by simulation model tool, followed by dataset and metrics [51].

Table 6. Summarizes the simulation model tools, dataset and metrics used to evaluate traffic load balancing and traffic
resilience.
Reference simulation model tools Dataset Measured metrics
Hong et. al. Mininet [52], Floodlight 12:30 clients Server response ;me, CPU
[11] controller [53] and
OpenvSwitch [54].
u;liza;on and Memory
u;liza;on.
André and Mininet, OpenvSwitch, Two controllers and 2:64 Throughput and failover ;me.
Fernando [12] ZooKeeper 3.4.8 (REF) and switches
iperf.

Anees et. al. Mininet, GNS3, Ryu controller, single network interface Throughput and file download
[15] OpenvSwitch and Linux traffic and multiple network
control tool ‘tc’.
;me vs different file sizes.
interfaces
Hamza et al. Mininet and Floodlight Two real network Throughput - Packet loss -
[17] controller [53] topologies, Cernet: 36 Response ;me - Migra;on
nodes & 53 links. and
cost Overhead.
DFN:58 nodes & 87
links.
Chunlin et.al. Mininet, Ryu controller [53] Internet2 OS3E: 34 nodes Controller-to- switch delay -
[18] and 42 links. Propaga;on delay between
Janetbackbon: 22 nodes controllers - Average task
and 35 links.
comple;on ;me – Load
balancing degree.
Jehad and Mininet and OpenvSwitch linear topology up to 500 Throughput - E2E Delay – CPU
Byeong-hee sensor nodes u;liza;on - recovery latency in
[21]
case of link failure
Jehad et.al. Mininet, ONOS controller [55], Network Topology E2E latency - communica;on
[25] iperf [56]traffic generation. DFN: 58 switches, 87 cost - load curve for
links. OS3E: 34 switches,
controllers- Migra;on ;me -
42 links. RedIris: 19
switches, 32 links. Response ;me - CPU
Interoute:110 switches, u;liza;on – Jain’s fairness
149 links. Abilene: 11 index (JFI).
switches ,1 link.

Table 7 presents a comprehensive overview of simulation model tools, datasets, and metrics
utilized for evaluating the proposed research in traffic optimization and path selection. The table
is organized based on the simulation model tool, followed by the dataset and metrics.
Table 7. Summarizes the simulation model tools, dataset and metrics used to evaluate traffic optimization and path
selection.
Reference Simulation model tools Dataset Measured metrics
Jiameng and MiniNet, OVX virtualization TCP and UDP TCP & UDP throughputs and
Sang-Hwa [30] platform and iperf performance for randomly data arrival rate
three virtual networks
Junjie et.al. NS-3.26 [57] 4k fat tree topology with Throughput, average link
[34] 20 switches and 64 hosts utilization
Shengru et. al. POF switches, POF controller 14 POF switches Throughput, Path Setup Latency,
[37] and iPerf. Packet Loss Ratio for Link
 Failure and Average Failure
Recovery Time
Xuya et. al. [40] ISP 1755 and the ISP Energy saving ratio – Number of
3967 network topologies controlled flow - Upgrading cost
[37,38]
Wu et. al. [41] Mininet, Foodlight1.2 Custom network topology Throughput -Time delay – Packet
controller and OpenvSwitch with 100 Mbps for all loss rate
links.
Muteb and Mininet, POX controller [55] SYNTH1: 100 switches, CPU time, init time, add time,
Abdelmounaam 11 hosts, and 129 delete time and deleted paths
[43] randomly generated links. time.
AS: 100 switches, 2
hosts, and 205
randomly generated links.

Yi-Ren et. al. Mininet, Ryu controller Fat-tree, NSFNet, and Reward is a score computed -
[45] ARPANet network Utilization rate is calculated in
topologies the destination switch

Gururaj et.al. Mininet 250, 500, 750,1000,2000 number of links, processing

[46] and 3000 nodes with latency and rerouting cost
10,25,50,75 and 100
Destinations

Majda et.al. Mininet, Floodlight controller Three video types: Structural Similarity Index Metric
[50] and Iperf Video 1, SD, Video 2, HD (SSIM), Mean Opinion Score
And Video 3, HD (MOS) and Video Multimethod
Assessment Fusion (VMAF)

4.2) Performance results analysis

This section presents a comprehensive overview of the compared schemes, demonstrated results,
and drawbacks and limitations of the proposed traffic load balancing & traffic resilience solution
and traffic optimization and path selection solution as shown in table 8 and table 9 respectively.
Table 8. Describes the demonstrated results, and drawbacks of the proposed traffic load balancing and traffic
resilience solution.
Reference Compared The demonstrated results Drawbacks& limitations
with
Hong et. al. Round Robin, LBBSRT has significant advantages A slight difference in memory
[11] Random compared to other schemes in terms of u9liza9on at 50% and CPU u9liza9on
the overall response 9mes and achieves a at 75% for the servers that had been
much be<er effect of load balancing. experimented.
Overcomes the drawbacks of tradi9onal Using a server response 9me as the
methods, including high cost, low main metric for load balancing might
reliability, and poor extensibility. not be universally suitable for all
network environments and
applica9ons.
André and Ravana Rama, a resilient SDN controller plaPorm, Network overhead due to increased
Fernando [12] offers comparable robustness as Ravana, network message exchanges. Need
rather, it performed be<er throughput, to evaluate the impact of network
ensuring fault tolerance without latency.
 necessita9ng modifica9ons to switches or
the OpenFlow protocol.
Anees et. al. MPTCP a significant 55% increase in achieved Network Dependency:
approach and throughput compared to the tradi9onal Any network failures or
[15]
tradi6onal single network approach, surpassing the communica9on issues between the
WRR, RR and
performance of Mul9path TCP (MPTCP) end hosts and the controller could
MBW
approach.
by 10%. This improvement can be disrupt traffic control and lead to
a<ributed to the Weighted Round Robin service interrup9ons due to the
(WRR) method, which allocates TCP flows proposed mechanism based on the
based on computed link weights, external network interfaces
resul9ng in be<er performance than modifica9on of the data plane
MPTCP when using a single network switch.
interface.
Hamza et al. DDS, SMDM, The MTLB scheme stands out among all - Further studies are required to
[17] EASM [ other evaluated schemes with its high precisely determine the op9mal
average throughput. intervals and iden9fy efficiency
It also achieves the lowest average levels.
packet delay, outperforming SMDM and - The current load distribu9on
EASM. operates reac9vely, triggered only
MTLB scheme demonstrates the lowest when thresholds are exceeded,
migra9on cost, packet loss and poten9ally resul9ng in spikes in
communica9on overhead controller load and performance
degrada9on.
- Addi9onally, the scheme does not
consider the variability in flow
processing 9mes, where certain
flows demand more processing than
others, leading to uneven load
distribu9on.
Chunlin et.al. K-means, SA - The improvement of LOCP algorithm - Consider more metrics to be
[18] and ECMP. slightly outperforms in propaga9on delay, evaluated like network throughput. -
LBRA, CARA queuing delay, and load balancing degree - The experimental setup lacks
and SCA
in large-scale networks, resul9ng in an considera9on for crucial factors such
average improvement of 18.36% in load as security concerns, the influence of
balancing while ensuring a lower mobile devices' mo9on trajectory,
propaga9on delay and queuing delay. and the task priority of mobile
-The average user comple9on delay devices on MEC servers.
growth rate of the HACA improved - An implementa9on in the case of
algorithm is lower than other algorithms, network failures is required to
including 50.39% lower than that of the evaluate the reliability of algorithms'
LBRA and 32.29% lower than that of the improvement.
SCA algorithm.
Jehad and AHP and EB- The results show that the proposed Focusing on controller selec9on may
Byeong-hee TOPSIS controller outperforms AHP and EB- limit the applicability of the scheme
[21] TOPSIS strategies in terms of delay to a specific aspect of SD-IoT, and it
reduc9on in flow request management may not address other important
and load balancing features. aspects such as security or energy
The proposed controller maintains a efficiency.
consistent throughput, exhibits a faster
start, and demonstrates be<er CPU
u9liza9on, even with increased traffic.
The recovery latency 9me of the
proposed controller is smaller than that
 of the other strategies in case of link
failures.
Jehad et.al. SASLB, DLB, The response 9me results indicate With a rising number of nodes and
SCLB and ESCALB's efficacy, a<ributed to its links in the network of the five
[25]
SMLB efficient slave controller selec9on. topologies, the JFI, is becoming
ESCALB consistently outperforms other lower. However, ESCALB strategy
schemes in terms of end-to-end (E2E) preserves it as near to 1 as possible,
latency for all network topologies. On the it may cause ineffec9ve packet
other hand, SASLB demonstrates lower distribu9on among controllers.
communica9on costs but suffers from ESCALB did not achieve op9mal CPU
uneven packet distribu9on among SDN u9liza9on
controllers, unlike ESCALB.

Table 9. Describes the demonstrated results, and drawbacks of the proposed traffic optimization and path selection
solution.
Reference Compared The demonstrated results Drawbacks& Limitations
with
Jiameng and CBBPM - TCP throughput achieves significant - Sensitive to network latency.
Sang-Hwa [30] improvements in VN1 (160%), VN2 - Difficult to scale to large networks.
(36%), and VN3 (17%). - It is not as well-documented as
- VN1 shows 30%-100% better UDP some other bandwidth management
throughput using the proposed solutions.
srTCM+ GTB compared to CBBPM.
Similarly, VN2 and VN3 experience
approximately 4%-25% improvements
with proposed srTCM+ GTB over
CBBPM.
- Packet loss demonstrates the superior
performance of srTCM+ GTB.
Junjie et.al. SingleTCP Proposed MPTCP & SR demonstrated a Increased overhead by increasing the
[34] and MPTCP positive correlation with increasing packet header size. Need a Multi-
transmission demand, outperforming controller implementation ensures
SingleTCP and MPTCP, which showed high scalability for DCNs.
slower increases and throughput
reduction, respectively.
the proposed method demonstrated
higher average link utilization compared
to MPTCP, achieved through dynamic
path adjustment, leading to better load
distribution and reduced delays.
Shengru et. al. -
Traditional - Achieves 100% receiving throughput - Improve the performance of POF
[37] OpenFlow- with fewer flow entries. Switch to make process packets
based OF-- - The ability to share flow entries on more efficient and work smoothly
SP. intermediate switches reduces the for switches equipped with 10GbE
number of flow entries required. NICs.
- - Achieves shorter path setup latency, - May increase SDN management
especially under higher traffic loads. complexity for network admins,
- - Outperforms OF-SP in terms of packet and introduce computational or
loss ratio. memory overhead, affecting SDN
- Achieves shorter average failure performance.
recovery time than OpenFlow.
Xuya et. al. [40] PLSP and - EPC achieves effective control over Focusing only on the energy-
EA-FA. 95% of flows with only 10% of the saving aspect of hybrid SDN
upgrading cost, saving an additional networks, it may not provide a
 10% of total power consumption comprehensive analysis of all the
compared to existing solutions. factors that affect the performance
- The EPC consistently outperforms both of these networks.
PLSP and EA-FA in energy-saving
ratio.
Wu et. al. [41] - DAMR outperforms ECMP and single- - The primary emphasis is on
path methods in congested scenarios. unicast algorithms to ensure end-
- DAMR provides smooth, jitter-free to-end Quality of Service (QoS)
performance, increased bandwidth, and while not addressing multicast
improved system throughput. algorithms.
- 35%:70% improvement in quality-of- - Extend the research to incorporate
service with DAMR. varying priorities for different
flows, including prioritizing
specific flows and implementing
diverse QoS strategies to meet
bandwidth or delay constraints.
Muteb and l2 learning, DR-APSP outperforms other approaches - DR-APSP's routing approach
Abdelmounaam l2 multi, and in terms of speed by a factor of 4 to 10, based on the APSP algorithm may
[43] l3 learning while also being more cost-effective in not be optimal for all network
terms of recomputing shortest paths. types and traffic patterns due to its
limited versatility. Additionally, its
computationally intensive dynamic
routing algorithm demands
significant processing power and
memory, potentially compromising
performance in large networks.
Yi-Ren et. al. OSPF and - RL-Routing obtains higher rewards on Deploy RL-Routing in a real
[45] LL. all three network topologies. network environment and evaluate
- RL-Routing minimizes the file it on other topologies.
transmission time on all three network
topologies.
- RL-Routing avoids congested paths.
- Hosts re-transfer fewer packets with
RL-Routing than with baseline
solutions.
Gururaj et.al. SPT, ST, - REMTC constructs an optimal, - Applying a machine learning-
[46] BAERA, scalable, and stable multicast based approach to predict optimal
and topology. paths and adapt to changing
OBSTA - REMTC selects the trunk path based network conditions.
on the device count, unlike other - The adaptation of REMTC for
methods that rely on optimization dynamic traffic in SDN requires
techniques. the processing of real-time
- REMTC's rerouting time is notably network data and the construction
quicker than ST, taking only 20-25% of multicast trees in real time.
of the time. This improved efficiency This can be computationally
is due to the trunk path algorithm, expensive, especially in large and
which prioritizes common links to complex networks, and may not
reach multicast participants. scale well.
- REMTC's distinctive features
contribute to reducing the alteration
time during dynamic join and leave
events.
Majda et.al. No - VQoS-RR increased SSIM quality - Not covered the impact of
[50] VQoSRR, values for SD and HD videos. network failures and on the
VQoS-R and - VQoSRR improved average MOS by performance of the VQoSRR
VQoS-RR 36.0% for SD and 56% for HD. model.
 Subjective MOS evaluations better- - The experimental results are
reflected differences compared to limited to a specific set of video
objective SSIM when comparing durations. Further analysis is
results with the proposed VQoSRR. needed to assess the performance
- DR-RA achieved the highest HD of the proposed model for long-
SSIM at 7 seconds and the best SD duration videos and different
result at 5 seconds, suggesting longer resolutions.
intervals for HD rerouting. - The integration of the proposed
- DR-RA achieved high-quality videos model with the multicast
with improved SSIM average values. transmission is a promising
direction for enhancing smart city
tools and applications,
particularly for surveillance
systems in hospitals and civil
defence organizations.
- Extend the research to investigate
the scalability of the proposed
model in large-scale network
environments.

5) Conclusion and Future Directions

In conclusion, Software-Defined Networks (SDN) have wrought a transformative impact on
network management through the segregation of the control layer and forwarding devices, yielding
centralized oversight and adaptable traffic control. This paper has highlighted the significance of
Load Balancing & Resilient traffic and traffic optimization in the context of Traffic Routing within
SDN. By means of a comprehensive evaluation and juxtaposition of scholarly discourse, a diverse
array of techniques and resolutions for traffic routing within SDN has been explored. Nonetheless,
it is paramount to recognize that the domain of SDN is characterized by perpetual evolution,
ushering in novel challenges and prospects in tandem with the progress of SDN development. The
establishment of standardized SDN components and the assimilation of protocols tailored to SDN
constitute pivotal undertakings, indispensable for surmounting legacy network-related issues.
Subsequent research ought to center around the control plane, endeavouring to conceive innovative
controller solutions, which function as the linchpin of the SDN architecture. Mitigating the control
plane's susceptibility as a solitary point of failure and effectuating robust security measures stand
as imperative requisites. Additionally, the integration of high availability (HA) mechanisms and
performance enhancement strategies to adhere to the stipulations of service level agreements
(SLAs) and to efficaciously furnish services should be of paramount concern. In summation, this
exposition furnishes a comprehensive survey of traffic routing methodologies within SDN.
Nonetheless, the road ahead remains replete with imperatives encompassing standardization,
control plane augmentation, fortification of security protocols, and optimization of performance.
It is incumbent upon forthcoming research endeavours to address these challenges, in order to
further advance the capabilities and reliability of SDN networks.
REFERENCES

[1] D. M. Casas-Velasco, O. M. C. Rendon and N. L. S. d. Fonseca, "DRSIR: A Deep Reinforcement

Learning Approach for Rou;ng in So_ware-Defined Networking," in IEEE Transac*ons on Network
and Service Management, 2022.

[2] I. Wahid, S. Tanvir, A. Hameed and M. Ahmad, "So_ware-Defined Networks and Named Data
Networks in Vehicular Ad Hoc Network Rou;ng: Compara;ve Study and Future Direc;ons,"
Security and Communica*on Networks, vol. 2022, pp. 1-14, 2022.

[3] H. Leqing, "How to Realize the Smooth Transi;on From Tradi;onal Network Architecture to SDN,"
in 2020 5th Interna*onal Conference on Mechanical, Control and Computer Engineering (ICMCCE),
Harbin, China, 2020.

[4] Y. Li, D. Zhang, J. Taheri and K. Li, "SDN components and OpenFlow," in Big Data and SoGware
Defined Networks, London, IET Digital Library, 2018, p. 49–68.

[5] A. Nayyer, A. K. Sharma and L. K. Awasthi, "Issues with Rou;ng in So_ware Defined Networks," in
2020 Sixth Interna*onal Conference on Parallel, Distributed and Grid Compu*ng (PDGC),
Waknaghat, India, 2020.

[6] ONF, "Open Networking Founda;on," 15 April 2015. [Online]. Available:

hlps://opennetworking.org/wp-content/uploads/2013/07/OpenFlow1.3.4TestSpecifica;on-
Basic.pdf. [Accessed 22 October 2022].

[7] A. Alghamdi, D. Paul and E. Sadgrove, "A RESTful Northbound Interface for Applica;ons in So_ware
Defined Networks," in 17th Interna*onal Conference on Web Informa*on Systems and
Technologies (WEBIST 2021), 2021.

[8] M. Jammal, T. Singh, A. Shami, R. Asal and Y. Li, "So_ware defined networking: State of the art and
research challenges," Computer Networks, vol. 72, pp. 74-98, 2014.

[9] Q. Waseem, S. S. Alshamrani, K. Nisar, W. I. Sofiah, W. Din and A. S. Alghamdi, "So_ware-Defined

Network (SDN) Forensic. Symmetry," Future Technology, vol. 13, no. 5, pp. 767 - 785, 2021.

[10] A. Hodaei and S. Babaie1, "A Survey on Traffic Management in So_ware-Defined Networks:
Challenges, Effec;ve Approaches, and Poten;al Measures," Wireless Personal Communica*ons,
vol. 118, no. 2, p. 507–1534, 2021.

[11] H. Zhong and J. C. Yaming Fang, "Reprint of ‘‘LBBSRT: An efficient SDN load balancing scheme based
on server response ;me’’," Future Genera*on Computer Systems, vol. 80, p. 409–416, 2018.

[12] A. Mantas and F. M. V. Ramos, "Rama: Controller Fault Tolerance in So_ware-Defined Networking
Made Prac;cal," 2019.
[13] H. Wang, L. X. H. Zhu, W. Xie and G. Lu, "Modeling and Verifying OpenFlow Scheduled Bundle
Mechanism Using CSP," in 2018 IEEE 42nd Annual Computer SoGware and Applica*ons Conference
(COMPSAC), Tokyo, Japan, 2018.

[14] N. Kala, H. Zhang, M. Freedman and J. Rexford, "Ravana: Controller fault-tolerance so_ware-
defined networking," in In Proceedings of the 1st ACM SIGCOMM Symposium on SoGware Defined
Networking Research, 2015.

[15] A. Al-Najjar, F. H. Khan and M. Portmann, "Network traffic control for mul;-homed endhosts via
SDN," IET Communica*ons, vol. 14, no. 19, pp. 3312-3323, 2020.

[16] M. Scharf and A. Ford, "Mul;path TCP (MPTCP) Applica;on Interface Considera;ons," RFC6897,
IETF, 2013.

[17] A. Mokhtar, X. Di, Y. Zhou, Z. M. A. Alzubair Hassan and S. Musa, "Mul;ple-level threshold load
balancing in distributed SDN controllers," Computer Networks, vol. 198, pp. 108369 - 108385,
2021.

[18] C. Li, K. Jiang and Y. Luo, "Dynamic placement of mul;ple controllers based on SDN and alloca;on
of computa;onal resources based on heuris;c ant colony algorithm," Knowledge-Based Systems,
vol. 241, pp. 108330 - 108348, 2022.

[19] C. Li, S. Liang, J. Zhang, Q.-e. Wang and Y. Luo, "Blockchain-based Data Trading in Edge-cloud
Compu;ng Environment," Informa*on Processing & Management, vol. 59, no. 1, pp. 102786-
102808, 2022.

[20] K. Antevski, C. J. Bernardos, L. Cominardi, A. d. l. Oliva and Alain Mourad, "On the integra;on of
NFV and MEC technologies: architecture analysis and benefits for edge robo;cs," Computer
Networks, vol. 175, pp. 107274-107291, 2020.

[21] J. Ali and B.-h. Roh, "A Novel Scheme for Controller Selec;on in So_ware-Defined Internet-of-
Things (SD-IoT).," Sensors, pp. 3591-3608, 2022.

[22] Z. C. S. O. O. Prince Boateng, "An Analy;cal Network Process model for risks priori;sa;on in
megaprojects," Interna*onal Journal of Project Management, vol. 33, no. 8, pp. 1795-1811, 2015.

[23] m. Belkadi and Y. Laaziz, "A Systema;c and Generic Method for Choosing A SDN Controller," Journal
of Computer Networks and Communica*ons, vol. 5, pp. 239-247, 2017.

[24] D. Kannan and R. Thiyagarajan, "Entropy based TOPSIS method for controller selec;on in so_ware
defined networking," oncurrency and Computa*on: Prac*ce and Experience, vol. 34, 2021.

[25] J. Ali, R. H. Jhaveri, M. Alswailim and Byeong-hee Roh, "ESCALB: An effec;ve slave controller
alloca;on-based load balancing scheme for mul;-domain SDN-enabled-IoT networks," Journal of
King Saud University - Computer and Informa*on Sciences, vol. 35, no. 6, pp. 101566 - 101577,
2023.
[26] A. Abdelaziz, A. T. Fong, A. Gani, U. Garba, S. Khan, A. Akhunzada, H. Talebian and K.-K. R. Choo,
"Distributed controller clustering in so_ware defined networks," PLoS ONE, vol. 12, no. 4, pp. 1-12,
2017.

[27] Y. Zhou, M. Zhu, L. Xiao, L. Ruan, W. Duan, D. Li, R. Liu and M. Zhu, "A Load Balancing Strategy of
SDN Controller Based on Distributed Decision," in 2014 IEEE 13th Interna*onal Conference on
Trust, Security and Privacy in Compu*ng and Communica*ons, Beijing, China, 2014.

[28] T. Hu, P. Yi, Z. Guo, J. Lan and Y. Hu, "Dynamic slave controller assignment for enhancing control
plane robustness in so_ware-defined networks," Future Genera*on Computer Systems, vol. 95, p.
681–693, 2019.

[29] S. K. Sagar, P. Deepak, T. Mayank, U. Muhammad, S. Bibhudala and W. Zhenyu, "ESMLB: Efficient
Switch Migra;on-Based Load Balancing for Mul;controller SDN in IoT," Internet of Things Journal,
vol. 7, no. 7, pp. 5852-5860, 2020.

[30] J. Shi and S.-H. Chung, "A traffic-aware quality-of-service control mechanism for so_ware-defined
networking-based virtualized networks," Interna*onal Journal of Distributed Sensor Networks, vol.
13, no. 3, 2017.

[31] J. Heinanen and R. Guerin, "A Single Rate Three Color Marker," RFC 2697, IETF, 1999.

[32] S. Blake, D. Black, M. Carlson, E. Davies, Z. Wang and W. Weiss, "An Architecture for Differen;ated
Services," RFC 2475, IETF,1998.

[33] L. B. Lim, L. Guan, A. Grigg, I. W. Phillips, X. Wang and I. U. Awan, "RED and WRED Performance
Analysis Based on Superposi;on of N MMBP Arrival Proccess," in 2010 24th IEEE Interna*onal
Conference on Advanced Informa*on Networking and Applica*ons, Perth, WA, Australia, 2010.

[34] J. PANG, G. XU and X. FU, "SDN-Based Data Center Networking With Collabora;on of Mul;path TCP
and Segment Rou;ng," IEEE Access, vol. 5, pp. 9764-9773, 2017.

[35] A. Ford, C. Raiciu, M. Handley and O. Bonaventure, "TCP Extensions for Mul;path Opera;on with
Mul;ple Addresses," RFC 6824, IETF, 2013.

[36] C. Filsfils, N. K. Nainar, C. Pignataro, J. C. Cardona and P. Francois, "The Segment Rou;ng
Architecture," in 2015 IEEE Global Communica*ons Conference (GLOBECOM), San Diego, CA, USA,
2015.

[37] S. Li, K. Han, F. Nirwan Ansari, Q. Bao, D. Hu, J. Liu, S. Yu and Z. Zhu, "Improving SDN Scalability with
Protocol-Oblivious Source Rou;ng: A System-Level Study," IEEE Transac*ons on Network and
Service Management, vol. 15, no. 1, pp. 275-288, March 2018.

[38] S. Li, K. Han, H. Huang, Q. Sun, J. Liu, S. Zhao and Z. Zhu, "SR-PVX: A Source Rou;ng Based Network
Virtualiza;on Hypervisor to Enable POF-FIS Programmability in vSDNs," IEEE Access, vol. 5, pp.
7659-7666.
[39] N. Huin, M. Rifai, F. Giroire, D. L. Pachec, G. Urvoy-Keller and J. Moulierac, "Bringing Energy Aware
Rou;ng Closer to Reality with SDN Hybrid Networks," in GLOBECOM 2017 - 2017 IEEE Global
Communica*ons Conference, Singapore, Singapore, 2017.

[40] X. Jia, Y. Jiang, Z. Guo, G. Shen and L. Wang, "Intelligent path control for energy-saving in hybrid
SDN networks," Computer Networks, vol. 131, pp. 65-76, 2018.

[41] W. Jiawei, Q. Xiuquan and N. Guoshun, "Dynamic and adap;ve mul;-path rou;ng algorithm based
on so_ware-defined network," Interna*onal Journal of Distributed Sensor Networks, vol. 14, no.
10, 2018.

[42] "Equal-Cost Mul;path Rou;ng in Data Center Network Based on So_ware Defined Network," in
2018 6th Interna*onal Conference on Informa*on and Communica*on Technology (ICoICT),
Bandung, Indonesia, 2018.

[43] M. Alshammari and A. Rezgui, "POX-PLUS: An SDN Controller with Dynamic Shortest Path Rou;ng,"
in IEEE 9th Interna*onal Conference on Cloud Networking (CloudNet), Piscataway, NJ, USA, 2020.

[44] M. Alshammari and A. Rezgui, "An All Pairs Shortest Path Algorithm for Dynamic Graphs,"
Interna*onal Journal of Mathema*cs and Computer Science, vol. 15, no. 1, p. 347–365, 2020.

[45] Y.-R. Chen, A. Rezapour, W.-G. Tzeng and S.-C. Tsai, "RL-Rou;ng: An SDN Rou;ng Algorithm Based
on Deep Reinforcement Learning," IEEE Transac*ons on Network Science and Engineering, vol. 7,
no. 4, pp. 3185-3199, 2020.

[46] G. Bijur, M. Ramakrishna and K. A. Kotegar, "Mul;cast tree construc;on algorithm for dynamic
traffic on so_ware defined networks," Scien*fic Reports, vol. 11, no. 1, 2021.

[47] M. Ramakrishna and A. K. Karunakar, "SIP and SDP based content adapta;on during real-;me video
streaming in Future Internets," Mul*media Tools and Applica*ons, vol. 76, no. 20, p. 21171–21191,
2017.

[48] L.-H. Huang, H.-J. Hung, C.-C. Lin and D.-N. Yang, "Scalable Steiner Tree for Mul;cast
Communica;ons in So_ware-Defined Networking," arXiv:1404.3454, 2014.

[49] S.-H. Chiang, J.-J. Kuo, S.-H. Shen, D.-N. Yang and W.-T. Chen, "Online Mul;cast Traffic Engineering
for So_ware-Defined Networks," in IEEE INFOCOM 2018 - IEEE Conference on Computer
Communica*ons, , Honolulu, HI, USA, 2018.

[50] M. O. Elbasheer, A. Aldegheishem, N. Alrajeh and J. Lloret, "Video Streaming Adap;ve QoS Rou;ng
with Resource Reserva;on (VQoSRR) Model for SDN Networks," Electronics, vol. 8, no. 11, pp. 1252
- 1275, 2022.

[51] M. S. Islam, M. Al-Mukhtar, M. R. K. Khan and M. Hossain, "A Survey on SDN and SDCN Traffic
Measurement: " Eng 4, no. 2: 1071-1115. hlps://doi.org/10.3390/eng4020063," Exis*ng
Approaches and Research Challenges, vol. 4, no. 2, pp. 1071-1115, 2023.
[52] N. Gupta, M. S. Maashi, S. Tanwar, S. Badotra, M. Aljebreen and S. Bharany, "A Compara;ve Study
of So_ware Defined Networking Controllers Using Mininet.," Electronics, vol. 11, no. 7, pp. 2715-
2752, 2022.

[53] Y. Li, X. Guo, X. Pang, B. Peng, X. Li and P. Zhang, "Performance Analysis of Floodlight and Ryu SDN
Controllers under Mininet Simulator," in 2020 IEEE/CIC Interna*onal Conference on
Communica*ons in China (ICCC Workshops), Chongqing, China, 2020.

[54] X. Zhang, K. Salama;an and G. Xie, "Improving Open Virtual Switch Performance Through Tuple
Merge Relaxa;on in So_ware Defined Networks," IEEE Transac*ons on Network and Service
Management, vol. 19, no. 3, pp. 2078-2091, 2022.

[55] A. L. Stancu, S. Halunga, A. Vulpe, G. Suciu, O. Fratu and E. C. Popovici, "A comparison between
several So_ware Defined Networking controllers," in," 2015 12th Interna*onal Conference on
Telecommunica*on in Modern Satellite, Cable and Broadcas*ng Services (TELSIKS), Nis, Serbia,
2015.

[56] J. Dugan, S. Elliol, B. A. Mah, J. Poskanzer and K. Prabhu, "iperf.fr," iperf, [Online]. Available:
hlps://iperf.fr/. [Accessed 2022].

[57] "Ns-3.26 Release Notes.," Available online: hlps://www.nsnam.org/releases/ns-3-26, accessed on

1/8/2023.

[58] Y. Zongying, L. Haoming and Z. Qiaoying, "The China Educa;on Research Network (CERNET) and
library services," MCB UP Ltd, 1998.

[59] L. Jing and O. Stephansson, "Discrete Fracture Network (DFN) Method," Developments in
Geotechnical Engineering, vol. 85, pp. 365-398, 2007.