Whale Optimization-Driven Framework for Enhanced Traffic Control in Software-Defined Networking Environment

Software-defined networking (SDN) is a new approach to design networks that allows software applications to manage them in an intelligent and centralized manner. It also changes the way networks operate by splitting the data layer and the control layer into independent layers. This enables manage, control and optimize network resources through software programming. SDN has several advantages, like improved network performance, greater flexibility, and simplified network management. The data layer and the control layer are separated the same in SDN. The control layer makes decisions and collects network information, whereas the data layer forwards network traffic. This separation allows put all of your network control in a centralized controller that can automatically configure and manage network devices according to network requirements. SDN centralizes control so you can obtain a global view of the network at once. As a result, makes it more efficient to control traffic and allocate resources. SDN architecture is composed of three standard tiers: the data, control, and application layers. The application layer consists of the software that communicates with the SDN controller to set rules and requirements for the network. The control layer is responsible for determining how to forward and route network traffic and do other network tasks. The data layer, on the other hand, is in charge of actually sending and receiving network packets based on what the control layer tells it to do. SDN also has a number of benefits over traditional networking methods [1-3].

• It lets network admins change and optimize network resources on the fly based on changing needs, which makes the network work better and uses resources more efficiently. This makes it very hard to balance the load on the network.

• Programmable SDN makes it possible to make new network apps and services. This helps networks quickly come up with new ideas and make changes.

Load balancing is an important aspect of SDN because it allows for efficient distribution of network traffic across multiple devices. In SDN, Artificial Intelligence (AI)-based load balancing approaches have been developed to enhance learning capabilities and enhance network performance. These approaches use AI algorithms to dynamically distribute network traffic across multiple devices, which provides efficient resource utilization and improved application availability. Since SDN is software-based, it is easy to change policies in SDN that are less prone to errors, and since it is programming-based, complex functions in the network can be developed in simpler ways. Numerous research studies have focused on SDN-based load balancing. These studies cover various aspects, including the design and analysis of algorithms, SDN-enhanced load balancing techniques for cloud systems, and a novel SDN approach to load balancing in data center networks. The writers present a platform for discussing load balancing and other SDN-related topics. Overall, SDN-based load balancing is a promising approach to improve network performance and resource utilization, and with the increasing adoption of SDN in various industries, further research and development in this field is expected to enhance the capabilities and effectiveness of SDN-based load balancing [4, 5].

A multi-method load balancing technique [6-9] has been devised for network architectures to accommodate the access capabilities of wireless mobile devices with limited connections and links [10, 11]. These cases and scenarios can equilibrate the user distribution across two distinct bands by utilizing wireless spectrum resources, primarily aimed at alleviating network congestion. Nevertheless, load balancing [12] does not account for task processing in fog networks. This study proposes a dynamic load balancing method, using graph partitioning and repartitioning, to create a specific approach for virtual machine nodes. This method then provides services to users by using graph clustering and partitioning. This can also increase the load on other nodes, potentially impacting overall system performance. Effective performance monitoring and failure management mechanisms are crucial to ensure service continuity. Additionally, there exists a trade-off between network strength and network capacity. In this study, we propose a load balancing model that can reduce the response time and improve the quality of service in SDN. The model uses a new optimization algorithm that effectively implements and uses the load distribution coefficient.

This paper is structured as follows: Section 2 examines the prior studies and investigations pertaining to the topic. Section 3 explains the concept and theoretical framework of software-defined networks, including the principles of load balancing optimization and the proposed algorithm. Following this, Section 4 describes the simulation of the proposed methodology. Finally, Section 5 analyzes the results and offers a conclusion.

3.1 Theoretical model of software-defined networking

In software-defined networks [28], there are three main parts: the application part, the SDN control part, and the infrastructure part, also called the data part. The SDN structure and architecture also have a part called "applications." This part includes many applications, such as those that manage network traffic, balance load, and deal with security. Through the North Bridge Interface (NBI), the application part transmits its network requirements and appropriate behavior to the control part. This architecture is shown in Figure 1.

1.png

Figure 1. General software-defined networking (SDN) architecture [28]

Control layer software-defined networks comprise a centralized controller that utilizes predefined logic to interpret the requirements of incoming transactions for the data links within the SDN. In the network, an administrator can separate the control and data parts and then can change network policies and make the network fit business by software instead of hardware, thereby fostering a flexible and scalable infrastructure. The control layer is linked to the data component via a Southbound Interface (SBI), which serves as the interface between the two parts. SDN controllers aggregate general data, including throughput, load, and end-to-end latency, and can oversee operations to enhance load balancing. Consequently, software-defined network controllers offer open-source programming interfaces to facilitate network connections, resource management, and other functions through software applications.

The SDN data component, referred to as the dispatch center, comprises numerous network elements, including switches, among others. The current devices reveal the data path selection or routers and direct network traffic through the network. In this network configuration, the switches function as transmission devices that direct the transmission and reception as defined by the routing controller protocol. The SBI consists of two parts: the SDN control component and the data component that controls all transmissions according to a schedule.

The distribution of computational tasks and load balancing criteria is intricate, as it must consider equipment efficiency and the complex communication overhead to minimize latency.

The undirected weighted SDN graph is shown in Figure 2. The current study is concentrated on load balancing technology as an intelligent routing protocol in SDNs. It emphasizes the requirement of reducing the response time. At the moment, the number of people using the Internet is constantly growing. This causes several web traffic, which in turn causes network congestion and packet loss. Despite this, managing server load is still very hard, and it often leads to unexpected service degradation.

2.png

Figure 2. Undirected software-defined networking (SDN) graph

Figure 3 illustrates an overview of the load balancing strategy. This paper proposes a novel SDN-based load balancing model that incorporates optimization principles as its primary objective. The scenario under consideration includes end devices, fog nodes, cloud servers, and a SDN controller. According to graph theory principles, the SDCFN topology is represented as an undirected weighted graph defined as G_R = {S_n, S_v}, where S_n = {S₁, S₂, …, S_k, SDN, cloud} denotes the set of vertices comprising virtual machines, the SDN controller, and the cloud server. Each node S_i ∈ S_n represents a virtual machine, While SDN refers to the control unit, the cloud represents the cloud server. The edge group is defined as follows: S_v = {eS₁, eS₂, …, eS_i, eS_j, …, eS_k−1, eS_k, …, eS₃, cs, eS₄, cloud}, where Tl_i and Tl_j stand for the communication link between two nodes. The cloud computing center is considered as a distributed computing node with its own processing capabilities for the improvement of the system performance via cloud computing. Similarly, CS_i shows how much computing power node S_i has, and TlS_i and TlS_j show how long it takes for nodes to talk to each other. When task L is sent from the end user to the switch, it is broken up into several smaller tasks, which are called L_i.

3.png

Figure 3. Load balancing layout with software-defined networking (SDN) [29]

The subtasks are then forwarded to fog and cloud nodes for processing. The distribution node is involved in the task. Finally, all incomplete findings are aggregated and delivered to the end consumers. The overall latency investigated in this study is expressed in Eq. (1). TlS_i,cloud denotes the transmission overhead between fog node S_i and cloud node, while S_i and S_j represent the fog nodes engaged in the communication.

$t\left(L_i\right)=\max \left(\frac{u_i}{C_{S_i}}\right)+T l_{S_i, S_j}, i, j=1,2, \ldots, k$         (1)

3.2 Whale Optimization Algorithm

3.2.1 Solution coding

This study suggests a new load-balancing model to cut down on delays. It does so by creating an optimization model that quickly figures out a load distribution coefficient a (time allocation (M) of the task in question, L_i).

A novel enhanced optimization algorithm is proposed for this optimal selection. The solution provided for the proposed algorithm is illustrated in Figures 3 and 4.

4.png

Figure 4. Solution encryption

The primary aim of the proposed method is articulated in Eq. (2), In this context, "t" (L_iCloud) refers to the latency of the cloud server. Based on the encoding process, the interface L_i: i = 1,..., cloud is assigned to each time slot M₁, M₂,..., M_N. When L_i2 is scheduled to run in time slot M₁, it does so first. Then, the task for time slot M₂ runs. This allocation process reduces the time required by distributing the workload. The suggested optimization method regularly alters the task assigned to each slot, Mⱼ = 1, 2, ... N.

Due to the diversity of the allocation steps, based on the time fitness function.

$F=\max \left\{t\left(L_{i_1}\right), t\left(L_{i_2}\right), \ldots, t\left(L_{i_N}\right)\right\}$        (2)

3.3 Traditional Whale Optimization Algorithm

In this section, we first discuss how the proposed method is derived.

3.3.1 Derivation

Whales are the largest mammals globally and among the most remarkable creatures on the planet. An adult whale may reach lengths of up to 30 meters and weigh as much as 180 tons. There are seven primary species of these colossal animals: killer, minke, sei, humpback, right, finback, and blue. Whales are recognized as predators. They remain awake as they must surface to breathe. In fact, only half of their brain sleeps when they are resting. T Whales are regarded as the most emotionally intelligent animals, which is an intriguing fact. Whales possess unique cells in a particular region of their brain that are also present in the human brain [30, 31]. These cells are called Von Economo neurons (also known as spindle neurons). These neurons are responsible for judgment, emotions, and social behavior in humans. In other words, Von Economo neurons distinguish higher mammals from other creatures. Whales have twice as many of these cells as humans, which is considered a contributing factor to their high cognitive capacity. Research has shown that whales are able to think, learn, judge and communicate. They have emotional depth similar to that of humans, but at a much lower level of intelligence. It has been noted that whales, especially orcas, can develop their own dialects. Another interesting thing about whales is their social behaviour. They live alone and in groups. But they are mostly seen in groups. Some whales, like killer whales can live their whole lives within one family. One of the largest toothless whales is the humpback whale (scientific name Megaptera novaeangliae). An adult humpback whale is about the size of a school bus. Their preferred prey is krill and small fish. The most interesting thing about humpback whales is their unique hunting method. This feeding method is called bubble-net feeding [32]. Humpback whales prefer to hunt krill or small fish at the surface of the water. This feeding method has been observed to produce bubbles in a circular or figure-of-nine pattern. Before 2011, this behavior had only been observed on the surface of the water. The paper [33] investigated this behavior using signal sensors. The authors of the paper observed 9 humpback whales performing bubble-net feeding 300 times. They observed two maneuvers that were accompanied by bubble production, which they called the “upward twist” and the “double loop.” In the first maneuver, the humpback whale dives about 12 m underwater and begins to produce bubbles in a spiral pattern around the prey and swims upward. The second maneuver consists of three stages: the central loop, the lobtail, and the capture loop. More detailed information about these behaviors can be found in [33-38]. It is worth noting here that bubble-net feeding has only been observed in humpback whales. In this paper, the spiral bubble-net feeding maneuver is mathematically modeled for optimization.

A comprehensive survey of the Whale Optimization Algorithm (WOA) and its numerous variants and application domains is provided in the study [39], which situates the present improvement within the broader body of WOA research.

3.3.2 Mathematical model of Whale Optimizer Algorithm

This section presents the mathematical model of prey encirclement, bubble-net feeding maneuver, and prey search.

3.3.3 Prey encirclement

The humpback whale can discern the position of prey and encircle it. Given that the location of the optimal point within the search space is unknown a priori, the whale optimizer algorithm posits that the best current candidate solution represents the target prey or the point nearest to the optimal point. Once the optimal search agent is identified, the remaining search agents endeavor to adjust their positions in relation to the optimal search agent. This behavior is shown by the following relations [31, 40]

$D=\left|C \cdot X^*(t)-X(t)\right|$       (3)

$X(t+1)=X^*(t)-A \cdot D$     (4)

where, t is the current iteration, A and C are the coefficient vectors, X* is the position vector of the best solution obtained so far, X is the position vector, | | is the absolute value and the operation is point-to-point multiplication and it is essential to note that if a better solution exists, X* should be updated at each iteration [31, 40]. The vectors A and C are calculated as:

$A=2 a \cdot r-a$      (5)

$C=2 \cdot r$      (6)

where, a decreases linearly from 2 to 0 during the iterations (both in the exploration and extraction phases) and r is a random vector in the interval [0,1].

3.3.4 Bubble network attack method (extraction phase)

To mathematically simulate the behavior of the bubble network for humpback whales, the following two approaches are suggested:

1. Minimizing the siege mechanism: This behavior is achieved by decreasing the value of a in Eq. (5). Note that the fluctuation of the interval A also decreases as a decrease. In other words, A is a random value in the interval [-a, a], where a decreases from 2 to 0 during the iterations. By placing the random value for A in the interval [-1,1], the new location of the search agent can be anywhere between the agent’s original location and the current best location [31, 40].
2. Spiral Position Update: Firstly, the distance between the whale (X, Y) and the prey (X*, Y*) is calculated. Then, a spiral relation is created between the whale and prey positioned to simulate the spiral movement of a humpback whale as shown in the following section.

$D^{\prime}=\left|X^*(t)-X(t)\right|$      (7)

$X(t+1)=D^{\prime} \cdot e^{b l} \cdot \cos (2 \pi l)+X^*(t)$      (8)

where, D′ = |X*(t) − X(t)| denotes the distance of the ith whale to the prey (the best solution obtained so far), b is a constant defining the logarithmic convolution shape, l is a random number in the interval [1, -1], and. is the point-to-point multiplication. Note that the humpback whales swim around the prey in a shrinking circle along the convolutional path. To model this synchronous behavior, we assume that there is a 50% probability that one of the encirclement or convolutional mechanism models is selected to update the whales’ positions during the optimization. The mathematical model is as follows:

$\vec{X}(t+1)=\left\{\begin{array}{l}\vec{X}^*(t)-\vec{A} \cdot \vec{D}, \text { if }, p<0.5 \\ \vec{D}^* \cdot e^{b l} \cdot \cos (2 \pi l)+\vec{X}^*(t), \text { if }, p \geq 0.5\end{array}\right.$        (9)

where, p is a random number in the range [0,1]. In addition to the bubble network method, humpback whales act randomly to search for prey. The mathematical model of the search is as follows.

3.3.5 Exploration phase (search for prey)

The same method, which involves altering the vector A, can be employed to seek prey (exploration). Actually, humpback whales conduct their searches at random in accordance with the positions of their peers. Consequently, we employ A with values greater than one or less than -1 to force the search agent to deviate from the reference whale. In contrast to the extraction phase, in the exploration phase we update the search agent's position with respect to a random search agent (instead of choosing the best search agent so far). This mechanism and A > 1 emphasize exploration and allow the whale's optimization algorithm to perform a global search. The mathematical model is as follows [31, 40]:

$D=\left|C \cdot X_{\mathrm{rand}}-X\right|$       (10)

$X(t+1)=X_{\mathrm{rand}}-A \cdot D$     (11)

where, X_rand is a random position vector (random whale) selected from the current population. The WOA begins with a collection of random responses. At each iteration, the search agents update their positions relative to a random search agent and the best response obtained thus far. To complete the exploration and extraction operation, the parameter a is reduced from 2 to 0. When | A | >= 1, the random search agent is chosen, and when | A | < 1, the best search agent is chosen to update the search agent positions. Depending on the value of p, the whale algorithm can switch between rotational and torsional motion. Finally, the whale algorithm terminates by satisfying the termination criterion.

3.4 Proposed algorithm

Despite the whale algorithm demonstrating competitive performance regarding reduced parameters and local non-optimality, A slow convergence speed is an obstacle for this. This study proposes an advanced algorithm by making ingenious changes to the traditional WOA to improve the convergence rate. In this version of WOA, the probabilities of spiral update and bait search are the same, i.e., th = 0.5. In real scenarios, they must be dynamic and depend on the quality of the improved solutions. If it was highly successful in the previous iteration, the probability of using the spiral update process should increase. The spiral update process should be terminated if it yields inadequate solutions. This paper tackles the issue by providing a dynamic threshold contingent upon the rate of enhancement in the fitness function value. The methodology underlying the proposed algorithm is as follows:

In the conventional WOA, the transition between spiral exploitation and prey-search behavior is governed by a fixed probability threshold of 0.5. However, in dynamic SDN-fog environments, a static threshold may not adequately reflect the current search progress. To address this limitation, the proposed improved Whale Optimization Algorithm (iWOA) introduces an adaptive threshold that is updated at each iteration based on the proportion of solutions whose fitness improves relative to the previous iteration. Specifically, the initial threshold is set to (th₀ = 0.5), and the threshold is then updated according to Eq. (12), where (N_t) denotes the number of improved whales and (Pop) represents the population size. This adaptive rule makes the balance between exploration and exploitation responsive to the optimization progress, thereby accelerating convergence and improving load-distribution quality.

threshold$_t=\min(1, \max(0, th_0-\frac{N_t}{Pop}))$       (12)

The use of the reference value 0.5 preserves consistency with the original WOA, while the adaptive term modifies this baseline according to the observed improvement rate.

In the proposed method, p is chosen randomly and compared with the dynamic threshold obtained from Eq. (12) instead of using the fixed threshold of 0.5 as in the traditional WOA.

In this section, the results of the NS-2 studies of Section 4 are presented and discussed. The performance of the proposed SDN-fog framework based on the iWOA has been compared with three existing methods:

(1) A cloud-only offloading strategy, where all tasks go straight to the centralized cloud server, which results in higher latencies of longer communication path.

(2) A greedy nearest-fog assignment, which sends tasks to the nearest available fog node without any optimization, which can result in load imbalances and resource inefficiencies.

(3) The standard classic WOA, which uses a fixed probability threshold (p = 0.5) for switching between spiral updates and prey search, which lacks the dynamic adaptation mechanism that iWOA has. The comparative analysis highlights the efficiency of the dynamic threshold of iWOA, which is modified according to rate of solution enhancement. This adaptability facilitates more precise load distribution adjustments, thereby enhancing the system's responsiveness to network fluctuations.

This approach also accelerates convergence, which helps avoid the common issue of getting stuck in local optima. Standard deviations are typically 5-10% and are not displayed in the tables for the sake of brevity. But when they are relevant, they are addressed. The results are averaged over 40 simulation runs for each configuration. We analyze fluctuations in essential parameters: task arrival rates (spanning from 8 to 40 tasks/s, reflecting low to high load conditions) and the quantity of fog nodes (ranging from 15 to 60, evaluating scalability in diverse environments). Adaptation is limited by the speed of change.

5.1 Average end-to-end latency

The average end-to-end latency is the sum of the elapsed time between the creation of a task at an end-user or IoT node and the transmission of the computation results. It encompasses return path overheads, processing times at fog or cloud nodes, queuing times, and transmission delays. Applications that require immediate response, such as self-driving systems or real-time IoT monitoring, necessitate reduced latency. The proposed iWOA is highly effective due to its dynamic optimization of load distribution coefficients, which ensures that tasks are dispatched to underutilized fog nodes that are located closer to users. This reduces congestion and long-distance cloud transfers. The average end-to-end latency as a function of task arrival rate is reported in Table 1, and its dependence on the number of fog nodes is reported in Table 2.

Table 1. Average end-to-end latency (s) vs. task arrival rate

Table 2. Average end-to-end latency (s) vs. number of fog nodes

Overall, iWOA's dynamic threshold provides superior latency reductions (15–30% vs. classic WOA) by enhancing exploration-exploitation balance, especially in high-load or large-scale scenarios where classic WOA's static p = 0.5 leads to delayed optimizations.

5.2 Average resource utilization

The average resource utilization measures how efficiently edge resources are used by avoiding capacity wastage or overload, which can lead to failures. This is determined by calculating the average percentage of computing capacity (MIPS) utilized by all fog nodes during the simulation. To prevent hotspots, the ideal utilization values should range from 70% to 90% and be balanced accordingly. The iWOA utilizes optimized coefficients to uniform distribution, effectively avoiding the imbalances that can occur due to the greedy and gradual encouragement of adaptations of traditional WOA.

Table 3 presents the average resource utilization against task arrival rate, and Table 4 presents it against the number of fog nodes.

Table 3. Average resource utilization (%) vs. task arrival rate

Table 4. Average resource usage (%) compared to the number of fog nodes

Overall, iWOA's dynamic threshold provides superior latency reductions (15–30% vs. classic WOA) by enhancing exploration-exploitation balance, especially in high-load or large-scale scenarios where classic WOA's static p = 0.5 leads to delayed optimizations. As reported in Tables 3 and 4, iWOA maintains balanced fog utilization within the target 75–88% band under high demand, avoiding both persistent overload and resource underutilization, whereas the greedy and classic-WOA baselines drift lower as load or scale increases.

5.3 Task success rate

The task success rate is the proportion of tasks completed and delivered on time, defined based on five times the latency of the cloud-only baseline. This indicates the system's reliability despite challenges such as delays, interruptions, or high traffic load. For mission-critical applications, high success rates (exceeding 90%) are required. Adaptive balancing in iWOA avoids task accumulation, thus improving overall success. The task success rate versus task arrival rate is given in Table 5, and versus the number of fog nodes in Table 6.

Table 5. Task success rate (%) vs. task arrival rate

Table 6. Task success rate (%) vs. number of fog nodes

As Tables 5 and 6 show, the proposed iWOA consistently outperformed the standard WOA algorithm, showing improvements of 10% to 30% across different evaluation metrics. The observed improvement is attributed to the use of an adaptive thresholding approach. This method enhances the system's adaptability and enables a faster response to dynamic conditions. Consequently, the proposed framework offers a effective approach for supporting emerging applications that require both low latency and high efficiency. Further research could explore the impact of node mobility limitations or energy consumption in fog computing environments.