An Adaptive Symmetrical Load Balancing Scheme for Next Generation Wireless Networks

Abstract

In dense Wi-Fi networks, achieving load balancing is critical to optimize network utilization and provide equitable network consumption among the users. Traditional Wi-Fi networks have issues in attaining effective load balancing. Software-Defined Networking (SDN) has presented a viable solution by isolating the data plane and control plane, enabling more agile and cost-effective networks. In this paper we put forward an Adaptive Symmetrical Load Balancing (ASLB) scheme to ensure fairness of load symmetry in Software Defined Wi-Fi Networks (SD-Wi-Fi), while also optimizing the flows transition process using an Analytical Hierarchal Process (AHP). User activity is monitored by access points (APs), which operate under OpenFlow standards. Three essential features, packet volume, packet category and delay hindrance, are used for flow assignment to various controllers. The controllers are arranged in two tiers, universal and regional controllers. The universal controller (UC) handles the workload statistics of regional controllers (RC) in the form of clusters. Extensive simulations using OMNeT++ simulator are performed. The performance parameters taken into consideration are throughput, delay, packet loss rate, network transition count and workload distribution. Our findings demonstrate that the ASLB technique effectively optimizes the network utilization and ensures equitable network consumption among the end users. The proposed scheme outperforms the Mean Probe Delay scheme (MPD), Channel Measurement-based Access Selection scheme (CMAS), Received Signal Strength Indicator-based scheme (RSSI) and Distributed Antenna Selection scheme (DASA), being 40% higher in throughput and 25% lower in delay.

1. Introduction

SDN is a well-known network paradigm which furnishes an encouraging perspective to answer distinct problems in wireless networks [1]. Wi-Fi networks and cellular phones have become common in today’s lifestyle universally [2]. The vulnerability of Wi-Fi systems has risen drastically, particularly in communal regions such as shopping malls, universities and airports [3]. The personal and communal Wi-Fi networks are executed in an improvised way and vie for an unlicensed spectrum which can result in a low network efficiency as the APs overlap in coverage due to severe obstruction [3]. In today’s dispersed network structure, it is challenging to set up and handle the unplanned installed APs [4]. There is an immediate requirement for coordinated regulation for compact Wi-Fi networks [5]. Cisco-dominated CAPWAP protocols are commonly used to regulate distributed APs infrastructure [6].

SDN is a pioneer network structure designed to create a programmable, effortlessly controllable, and governable network [7]. SDN is logically managed by a centralized network controller that has a comprehensive perspective of the entire network and can facilitate the implementation of network management protocol [8]. SDN has been recommended to improve Wi-Fi networks and tackle ordinary issues such as obstruction, load balancing, potential management and other related issues [9]. The inaugural configurable southbound interface in SDN, known as OpenFlow, is created for Ethernet switches and does not have provisions for wireless traffic [10]. Numerous studies have used SDN for wireless local area networks (WLANs) [11].

Research has also been conducted to integrate SDN into high density Wi-Fi networks to achieve programmability [12]. In a high density Wi-Fi network scenario, each wireless device (WD) connects with the AP that has the highest RSSI [13]. Thus, many WDs may connect with the same AP with the highest RSSI, while other APs might not be utilized adequately [14]. Such disproportionate traffic patterns may have a substantial adverse effect on network efficacy. At the same time, if numerous users link to the same AP, the contention will be more intense and the throughput will decrease significantly [15]. Across the span of several decades, numerous load balancing strategies have been recommended to improve the association policy in such a way that WDs can select the most suitable AP to connect to the network [16]. The decentralized nature of existing Wi-Fi networks makes it impossible to centrally manage user imbalance distribution [17].

In this paper, we address the load balancing issues in high density Wi-Fi networks by leveraging SDN to boost network performance. By utilizing a network controller that acquires a comprehensive overview of the network, we create and apply load balancing algorithms that incorporate user-side and AP-side data for a symmetrical load balance in the data plane and control plane. In the past the SDN was applied on wired networks and extending it to wireless networks was cumbersome. We set up and test two SDN-based wireless experimental platforms to assess the proposed algorithms using OMNeT++ which supports OpenFlow. The first algorithm makes use of AHP and clustering to enhance load balancing in control plane and the second algorithm makes use of the AP data received by the SDN controller to optimize load on the APs in the data plane. The APs report the information to the SDN controller. In our multi-controller design a UC is designated for overseeing other RCs in terms of processors usage, workload, and response duration. When a new flow arrives at the switch, the switch examines the initial flow packet and categorizes it into various classifications. The switch requests the UC to calculate the built-in relevant flow guidelines. The UC may delegate the appeal to an RC for flow processing. The UC organizes the regional controllers into three categories, including over-utilized clusters (OUC), under-utilized clusters (UUC), and non-utilized clusters (NUC). The UC employs the proposed ASLB scheme to determine the state of the RCs and dispatch flow requests to the most suitable RC to compute flow policies. The policies are then deployed to the switch to achieve quick request processing and a low response time. The UC also keeps track of the load information of the APs; once an AP becomes overloaded the UC shifts the WD to a least-loaded AP. Implementation of SDN into Wi-Fi networks sets a tradeoff between performance and programmability. An SDN controller requires an expert level of programming skills to be monitored before reaching an optimal throughput performance. To the best of our knowledge this is the first standardized method of achieving load balancing in the data plane and control plane of a SD-Wi-Fi. The major contributions of the study are:

• We propose an algorithm to balance the wireless traffic load in the data plane of an SD-WiFi.
• We propose an algorithm to balance the wireless traffic load in the control plane of an SD-WiFi.
• We simulate the high density SD-Wi-Fi in an OMNeT++ simulator supported by OpenFlow standards.

The related work on control plane load balancing and data plane load balancing is discussed in Section 2. The methodology, including two algorithms, is explained in Section 3. The experimental details are described in Section 4. Performance evaluation is presented in Section 5 and finally we conclude the paper in Section 6. The table of abbreviations and symbols used in the study are shown in

Table 1. List of Abbreviations.

Abbreviations	Description
ASLB	Adaptive Symmetrical Load Balancing Scheme
SDN	Software Define Networking
AP	Access Point
UC	Universal Controller
RC	Regional Controller
CAPWAP	Control and Provisioning of Wireless Access Points
AC	Access Controller
WLANS	Wireless Local Area Networks
WD	Wireless Device
RSSI	Received Signal Strength Indicator
DCF	Distributed Coordinated Function
OUC	Over-Utilized Controller
UUC	Under-Utilized Controller
NUC	Non-Utilized Controller
AHP	Analytical Hierarchical Process
ACH	Adaptive connection and Hand off
LIFO	Last In First Out
FIFO	First In First Out
COLBAS	Co-operative Load Balancing Scheme
HFA	HoneyBee Foraging Algorithm
GA	Genetic Algorithm
AT	Active time
PT	Passive time
MPD	Mean Probe Delay
CMAS	Channel Measurement-based Access Selection
RRM	Round-Robin Mechanism
WRR	Weighted Round-Robin Approach
VoIP	Voice over Internet Protocol
ESS	Extended Services Set
BSS’s	Basic Service Sets
DASA	Distributed Antenna Selection Algorithm
DL-SINR	Downlink Signal-to-Interference-plus-Noise Ratio

and

Table 2. List of Symbols.

Symbols	Full Form
PC	Packet Category
L${}_{\iota }$	Limit of Flow
$\eta$	Load of Regional Controller
X${}_{\eta }$	Central Processor Usage
Y${}_{\eta }$	Memory Usage
Z${}_{\eta }$	Hard Drive Usage
Ap	Amount of Packet Processing
Pb	Packets in Buffers
B	Burden Status
S${}_{\eta }$	System Parameter
C${}_{\eta }$	Consumer Parameter
CI	Consistency Index
$\lambda$	Eigenvalue
CI/RI	Consistency Ratio
RI	Random Index
Vol${}_{\mathrm{d}}$	Volume of Data of Aps
n${}_{\mathrm{d}}$	Number of Users AP can Maintain
Us${}_{\mathrm{d}}$	Users Currently connected to AP
pdp${}_{\mathrm{d}}$	Packet Distortion Proportion
B${}_{\mathrm{d}}$	Burden data
msir${}_{\mathrm{d}}$	Mean Signal Inaccuracy Rate
cpc${}_{\mathrm{d}}$	Central Processing Consuming
mu${}_{\mathrm{d}}$	Memory Utilization
Q	Utilization Level

, respectively.

2. Related Work

2.1. Load Balancing in Control Plane

A software-defined networking (SDN)-enabled WLAN system architecture was created with a two-level flexible load equalization approach [18]. Load equalization functions on the controller side were implemented resulting in scalability at the control plane. The affiliation indexes preserved are detached using a policy of LIFO (Last-In-First-Out) [19,20]. Consequently, the first consumer in line has to endure a longer wait time than the last user in line. The incoming streams are categorized based on priority, disregarding the sequence in which the data packets have arrived. Research has suggested a collaborative burden equalization plan for hierarchical SDN regulators, where a top-level regulator is designated to manage the flow requests of other regulators [21]. In line with the adjustment in traffic circumstances, the guidelines are correctly established and reinstated to achieve burden equalization. The top-level regulator is accountable for connecting with other typical regulators, allowing for the creation of a flexible burden equalization multi-regulator structure. The scheme ignores the latency factor. Controllers are organized in clusters, and whenever a cluster attains the utmost burden limit, a proportion of the burden is transferred to the controller with lighter workloads [22]. In situations of overcapacity, the clusters are reconfigured. However, the study was unable to devise a method for determining the burden on each controller. The scheme we propose diminishes the waiting time in request processing by prioritizing the data streams.

SDN was integrated into the cloud domain to distribute the workload through implementation of the honeybee foraging algorithm using planned installations [23,24]. The selection process is based on First-In-First-Out, whereas the algorithm is employed for decision-making. The burden at each node is anticipated based on CPU utilization details and the user is allocated accordingly. A Finite State Machine (FSM) and guideline are devised [25]. The FSM was utilized to align the states of controllers. Throughput performance was not evaluated. Load equalization was accomplished through utilization of a genetic algorithm [26]. The GA entails a selection process, crossover and mutation for estimating the optimal load balancing among the controllers. Although these measures enable the detection of the burden on the controllers, they fall short of equalizing the load when additional incoming requests are received. The scheme we propose demonstrates effective processing of data packets, even in the presence of a large number of requests. The round-robin mechanism was one of the prevalent techniques employed for load balancing [27]. The Weighted Round-Robin approach (WRR) was used to distribute the workload in SDN. The WRR scheme allocates fixed weight values to servers based on their capability; as a result, the distribution of load using these weighted values is not effective for managing a large number of flow requests. Our proposed scheme, on the other hand, is a flexible algorithm architecture that uses ASLB and clustering to balance the load in the control plane.

2.2. Load Balancing in Data Plane

Distributed load balancing in wireless fidelity (Wi-Fi) networks has been extensively studied in the prior decades [28,29]. A dynamic distributed load-balancing algorithm using a specific affiliation control approach is designed [30]. The researchers expressed the distributed stabilizing complication in a comparable optimization structure by utilizing stochastic approach [31]. It is noteworthy that only AP data are utilized in the aforementioned design. In the SDN design, the network controller acquires the global understanding of the network, opening up innovative opportunities for scalable administration. Not long ago Wi-Fi networks have expanded with SDN capabilities [32]. Researchers have integrated Mininet with NS-3 to develop an Open Net simulator for managing simulations for SD-Wi-Fi [33]. The Mininet-Wi-Fi tool was equipped to support genuine network assessment for load balancing at data plane [34]. DCF access protocol in this emulator made it more difficult to conduct studies on uplink network throughput when several users were sharing a single channel. SD-Wi-Fi prototype systems are typically built using software-defined wireless network testbeds [35]. It should be noted that the testbeds are only suitable for implementing network layout with small-scale experiments and are not applicable for highly dense Wi-Fi network scenarios with a significant number of users and APs.

Several investigations have been carried out to enhance the performance of SD-Wi-Fi [36]. The researchers constructed the Ethanol experimentation platform to execute a load-sensitive admission control design by amalgamating information from both the client side and AP side [37]. The work ensured the potential of distributing the workload among access points in SD-Wi-Fi. The researchers also presented load distribution techniques for data plane in SDN but their research was restricted to few APs [38,39]. The efficiency of load distribution techniques is yet to be evaluated for SD-WiFi. In this paper, we are inspired to examine the workload distribution issues for high density SD-WiFi to facilitate the adaptive network environment.

3. Methodology

The overall architecture proposed in the research is depicted in Figure 1. The application plane carries the ASLB scheme to achieve load balancing in data and control plane. The control plane has multiple SDN controllers. The control plane has two tiers, namely the UC and RCs. The UC is responsible to monitor the RCs and shift the flows to the most underloaded RC. AHP is used to filter the flows and the most delay sensitive flow is forwarded first to the UC to minimized delay and enhance throughput at the user end. The data plane has all the forwarding devices. The UC keeps track of load reports of all the APs. Once the AP becomes overloaded the controller using OpenFlow deassociates the WD and re-associates it to the least-loaded AP.

The proposed ASLB works in two planes as depicted in Figure 2. Symmetrical load balancing at the control plane and data plane is achieved side by side. Symmetry of load balancing is achieved, as the load among the APs in the data plane and load among the controllers in the control plane is balanced.

In the multi-controller SDN architecture, a UC is designated to supervise other RCs in terms of processing time, memory and response time. Upon the arrival of a new flow at the switch, the switch inspects the initial flow packet and categorizes it into various classes. The switch requests the UC to calculate the corresponding flow strategies. The UC may assign the request to a RC for flow processing. The UC is responsible for organizing the regional controllers into three types of clusters comprising OUC, ULC and NUC.

The UC employs the ASLB technique to distribute the flow requests to relevant RCs for computing flow policies and then deploys the policies back at the switch to achieve a high processing rate and low response time for the requests. The ASLB is adaptive to any network wise traffic condition. The symmetry of load is maintained among the APs in the data plane and multiple controllers in the control plane. In the past, AHP technique has been utilized in numerous applications such as planning, resource allocation, and priority setting for precise decision making and selecting the optimal SDN controller from a list of controllers [40]. In the data plane the ASLBs mutually analyze the handoff process and access information of both AP-side and WD-side data to attain load balancing.

3.1. Algorithm 1

Algorithm 1 is designed to perform the load balancing in the control plane. The load among the controllers is balanced through clustering and AHP to achieve the highest throughput and minimum delay.

Algorithm 1 Control Plane Load Balancing

1: Flows classified into AT or PT through AHP.  2: if AT (VoIP, Live Streaming, Video conferencing then  3:     PC == 1 and lowest L${}_{\mathrm{i}}$ then  4:     Priority level == Highest  5: else  6:     if PT (file transfer, email, web browsing etc) then  7:         PC == 0 and L${}_{\mathrm{i}}$ is greater  8:         Priority level == Lowest  9:     end if 10:     RCs report load status to UC 11:     RCs are classified into OUC, UUC and NUC by UC 12:     Load of RC is divided on the basis of 5 parameters X$\eta$, Y$\eta$, Z$\eta$, Pp${}_{\eta }$ and Pb${}_{\eta }$. 13:     Check the ${}_{\eta }$ at every 5 s. 14:     Compute Equation (2) for the the load status 15:     Compute Equations (3) and (4) to calculate the processor used, memory used, hard disk used, packets in processing and Packets in the buffer. 16:     Load is evenly distributed among the RCs by the UC. 17: end if

3.1.1. Analytical Hierarchical Process

To ensure fairness, the AHP technique is used as a multi-dimensional decision-making process which is capable of identifying irregularities. The reason for applying the AHP method is that it enables us to make decisions based on multiple criteria, where it may be challenging to compare two criteria directly. A hierarchy shown in Figure 3, is created for the incoming flow packets, considering three criteria such as packet volume, packet category and delay hindrance. In AHP, numerical ratings are assigned based on packet preferences.

Table 3. Priority Preferences.

Priority Level	Packets Status
Low Priority	0
Equal Priority	1
Small Priority	2
Medium Priority	3
Large Priority	4

shows the typical numeric evaluation used. The evaluation table leads to the calculation of the matrix consistency index (CI) which is defined in Equation (1).

$$CI=\frac{{\lambda }_{\max }-n}{n-1}.$$

(1)

AHP is used as a method for prioritizing flow requests in an SD-Wi-Fi architecture that implements load balancing. The method involves a process to rank flow requests according to their significance. The numerical ranking of the alternatives chosen and their order of preference are represented by the alternatives maximal Eigenvalue ($\lambda$max) in the judgment matrix. To assess the consistency of the judgment matrix, the consistency ratio $(CI/RI)$ is computed, where $RI$ is the random index representing the number of entities involved in the correlation. Once the flow requests have been ranked, they are assigned to the RCs based on their priority. The highest priority request is allocated first, followed by sequentially ranked requests. By following this procedure, the suggested SDN architecture with load balancing is guaranteed to be capable of managing incoming flow requests effectively, even when the workload on the controllers rises as a result of the increase in requests.

3.1.2. Clustering

Two types of traffic are taken into account when examining the prioritization of data: active time traffic (AT) and passive time traffic (PT). It is important to note that the method proposed here can be extended to multiple types of traffic streams with minor adjustments. AT streams could be VoIP sessions, live video streaming sessions, online gaming, video conferencing and so on. These types of streams are given high priority because they are sensitive to delays. PT streams could be file transfer sessions, email, web browsing and so on, which can tolerate delays, and thus they can be processed with lower priorities. The queue for AT streams hold the high priority requests, and the queue for PT streams hold the low priority requests. Let $L_i$ indicate the limit of flow. $PC$ represents the packet category for flow, which can be either 0 (PT) or 1 (AT). The flow request with the lowest $L_i$ and $PC$ = 1 will be given the top precedence for processing by the UC. If a flow appeals a higher $L_i$ than other requests and the $PC$ = 0, the precedence for processing will be low.

The UC oversees the RCs and is responsible for making decisions regarding sending the flow and relocating them. The UC creates three distinct clusters: such as the OUC, UUC, and NUC. The OUC controllers are tasked with processing a larger number of flows. The UUC controllers have fewer flows, meaning they can process any new incoming flow and the NUC controllers do not have any flows to process. The load of a RC $\eta$ is determined based on 5 parameters: central processor usage X${\eta }_{}$, memory usage Y${\eta }_{}$, hard drive usage Z${\eta }_{}$, the amount of packets being processed AP${}_p$ and the number of packets waiting in the buffer P$b_{}$. Among these parameters the first three are the system-based parameters and the rest two are client-based parameters.

The RC constantly updates its load status to the UC. The load status is refreshed at regular intervals (i.e., every 5 s).

$$B_{\eta }=(S_{\eta },C_{\eta }).$$

(2)

The current load status of the regional controller B is directly dependent on S${}_{\eta }$ and C${\eta }_{}$, where S${\eta }_{}$ represents the system parameters and C${\eta }_{}$ represents the client parameters of controller $\eta$. The modifications in the measurements may influence the load status of the RCs. The measurements are expressed as Equations (2) and (3).

$$S_{\eta }=(X_{\eta }+Y_{\eta }+Z_{\eta }).$$

(3)

By looking at Equation (3), S$\eta$ is determined by sum of triads, where X$\eta$ indicates how much processor memory is occupied, Y$\eta$ indicates how much RAM is being used and Z$\eta$ denotes how much hard drive is utilized. Equation (4) represents the C$\eta$ which calculate the load due to forwarding devices, where Pp${}_{\eta }$ indicates packets in processing phase and Pb${}_{\eta }$ indicates packets in the buffer. The measurement is dependent on the current WDs connected to the APs. With an increase in the number of WDs, the load contribution from client-based measurements also proportionately increase.

$$C_{\eta }=(P{p}_{\eta }+P{b}_{\eta }).$$

(4)

Using the load information provided by each RC, the UC assigns each RC to an appropriate cluster. The data packets from the OUC are then directed to the corresponding UTC, reducing the processing load on the RC and minimizing the waiting times of the packets in the queue. Based on the regular updates of the load, the clustering of RCs is enhanced and the workload is evenly distributed, ensuring efficient processing of incoming data packets.

3.2. Algorithm 2

Since user flow requests are not likely to be similar, initially separating them at the data plane produces better outcomes. Therefore, classifying flow requests before they reach the controllers can lead to better load balancing efficiency. The volume of data of the AP is kept as a triplet of ordered pairs, namely vol${}_d$ = (n${}_d$, us${}_d$, pdp${}_d$), where n${}_d$ represents the greatest number of WDs that an AP can maintain in its association table, us${}_d$ denotes the numerals of WDs currently associated to the AP and pdp${}_d$ represents the average packet distortion proportion of the APs linked wireless connections. The APs load is also saved as three tuples, namely bd = (msir${}_d$, cpc${}_d$, mu${}_d$), where msir${}_d$ is represented as the mean signal inaccuracy rate, cpc${}_d$ is the central processor consumption and mu${}_d$ is the memory utilization of APs as per wireless connections. We utilize an assessment formula to calculate, which represents Jains fairness index $J_f$ as;

$$J_f=\frac{{\sum }_{k=1}^{mb}\frac{usd}{nd}{(max\left(bd\right))}^2}{nd{\sum }_{k=1}^{mb}\frac{usd}{nd}{(max\left(bd\right))}^2}$$

(5)

Equation (5) shows that the total load of the extended services set (ESS) is squared in the numerator, while the denominator consists of the multiple of the number of APs and the aggregate of each APs load squared. Therefore, $J_f==1$ if each of the APs are uniformly loaded. The mb in Equation (6) is the number of basic service sets (BSSs) in an ESS. The controller computes the mean utilization level of the BSS since APs are likely to exhibit varying load levels.

$$Q=\frac{{\sum }_{k=1}^{mb}\frac{usd}{nd}max\left(bd\right)}{mb}$$

(6)

The magnitude of Q varies from 0 to 100, and a value of 0 ($Q=0$) or 100 ($Q=100$) indicates the minimum or maximum utilization level, respectively. The controller employed the mean load level to classify the APs into overloaded and underloaded categories. Algorithm 2 define the steps involved.

Algorithm 2 Data Plane Load Balancing

1: APs operating in the same ESS. 2: Retrieve MAC address of WDs. 3: List of WDs. 4: AP updates load reports to the controller periodically 5: Controller computes Equations (5) and (6) 6: Categorize the APs as either AP load${}_{\max }$ or AP load${}_{\min }$ 7: if AP is categorized as AP load${}_{\max }$ and exceeds the average threshold for load then 8:     Reassociate the WD to the APload${}_{\min }$ 9: end if

4. Simulation Setup

We have executed the simulations of the SD-Wi-Fi architecture using OMNeT++ as shown in Figure 4. A high density SD-Wi-Fi is illustrated where the purple circles denote the controllers and applications installed. The green messages are the flow requests generated by the wireless devices/mobile nodes. The format for the flow request is the same used by physical layer and medium access control (MAC) layer of OMNeT++. Where MN denotes the mobile nodes. The APs are connected to OpenFlow enabled switches. OMNeT++ is a modular, open-source, and extensible simulation framework used for building discrete event simulations of various network protocols, architectures, and algorithms. It is written in C++ and is based on the event-driven simulation paradigm, where events are processed in sequential order. The IDE allows users to configure various simulation parameters, such as simulation time, packet generation rate, and routing protocol parameters. It provides real-time visualization of the simulation progress, including network traffic, node behavior, and packet flow. It also allows users to analyze simulation results using built-in tools.

For the analysis of the proposed ASLB, we have analyzed delay, throughput, packet loss rate, average number of handoffs and workload optimization parameters. Our network topology is formed using one UC and three RCs with OpenFlow switches and APs. We have used 100 WDs in simulation. The simulation topology is depicted in Figure 5. All simulations are performed on a HP PC with windows 11, 16 GB of RAM and a AMD Ryzen 5 5600U with Radeon Graphics @ 2.30 GHz processor.

Table 4. Emulation Variables.

Specifications	Value
Total Number of Controllers	4
Universal Controller	1
Regional Controllers	3
OpenFlow Switches	2
WIFI Access Points	8
WDs	100
Processors	4
RAM	16 GB
Packets Table Size	1000
Packets Arrival per Second	100
Active Flow per Second	50
Passive Flow per Second	50
Load Status Refresh every	5 s
Simulation Time	20 min

shows the important parameters taken into consideration during simulation of the proposed ASLB.

5. Results

In this section, output of our proposed ASLB is compared with with previous load balancing schemes. We have considered both AT and PT flow requests. The ASLB technique is compared with the previous schemes such as received signal strength indicator (RSSI), mean probe delay (MPD), channel-measurement-based access selection (CMAS) and distributed antenna selection algorithm (DASA).

5.1. Throughput

The graph shown in Figure 6 illustrates the throughput performance, which indicates that increasing the number of WDs results in an increase in throughput with respect to time. However, as more WDs are added, the rate at which throughput improves slows down gradually. This outcome aligns with the analysis that suggests the throughput in a Wi-Fi network that employs the distributed coordination function (DCF) reaches a saturation point or begins to decline as the traffic load increases. Traditional schemes depending on RSSI exhibit the poorest throughput performance because they select the destination AP based on only one parameter and fail to balance the load among the APs. As a result some APs become overloaded, leading to packet collisions and degraded throughput. In the traditional RSSI scheme, the WDs remain associated with a particular AP unless they move out of range or there is an issue with the APs RSSI value. The RSSI scheme contributes just 40% of the total throughput.

In contrast, the proposed ASLB balances the load among the APs and the controllers in a symmetrical way through a UC and selects the destination AP and RC that fulfills system and client parameter requirements. If an AP becomes overloaded, the UC prepares a de-association list for the overloaded APs and selects an underloaded AP as the destination for de-associated WD. This two-fold load balancing approach reduces contention and improves throughput performance. Additionally, the SDN controller de-associates the WD with the lowest RSSI values and associates them with the least loaded AP while considering packet loss rate, RSSI, and throughput. Consequently, this approach also enhances packet delivery rates.

Among the other schemes evaluated, CMAS scheme outperforms the DASA and MPD schemes. MPD performs the worst due to unnecessary probing frames used to estimate the load values. This approach introduces delays in transmitting and receiving the probing frames, which can lead to larger probing delays and packet collisions even in lightly loaded networks. In the DASA scheme, selecting an AP with the best DL-SINR improves the end-to-end aggregate throughput by improving the transmission rate between the AP and the WDs. In CMAS, the load among the APs is balanced by estimating the bandwidth using the DCF mechanism, while taking frame aggregation and hidden terminal problems into account. When compared to CMAS, DASA, MPD and RSSI schemes, ASLB optimizes the throughput by 10%, 14%, 23% and 35%, respectively.

5.2. Delay

Figure 7 displays the delays of the frames for five different schemes, including the proposed ASLB. The frame delay represents the average time taken from when the WDs contend for the channel until the AP successfully receives the frames. As the number of WDs increases, the frame delay also increases. This implies that there is a close relationship between the frame delay and the throughput. Lower throughput will result in a larger delay for the same packet generation rate. In comparison to CMAS, DASA, MPD and RSSI schemes, ASLB reduces the delay by 3%, 4%, 5% and 10%, respectively.

5.3. Packet Loss Rate

Figure 8 shows the performance of the packet loss rate. It can be seen that, as there are more WDs, the proposed ASLB, CMAS, and DASA schemes experience a minor increase in packet loss rate, however the MPD and RSSI schemes experience a rapid increase in packet loss rate. Approximately fewer packets are transmitted on average in the ASLB scheme than in the RSSI, MPD, DASA and CMAS schemes. Fewer packets are lost as a result of increased throughput and symmetrical load distribution among the APs and the controller in the ASLB. These outcomes demonstrate the effectiveness of the suggested ASLB for software-defined load balancing among numerous APs. When compared to CMAS, DASA, MPD and RSSI schemes, ASLB reduces the packet loss rate by 25%, 28%, 45% and 56%, respectively.

5.4. Average Number of Handoffs

Figure 9 depicts the connection between the average number of handoffs and the WDs. The ratio of the total number of handoffs carried out by all WDs to the total number of WDs is known as the average number of handoffs. The average number of handoffs for all four approaches is almost the same for smaller numbers of WDs. As load and the number of WDs increase, the performance disparity between all four schemes widens. The MPD approach has a lower handoff among 100 WDs because additional probe-request packets aid in more precise estimation of each AP’s load. The ASLB that is being suggested has the fewest handoffs. This is because the ASLB can predict the order of an AP’s switch based on the activity of its WDs. The change will begin with the WDs with the weakest RSSI connections. As a result, there are fewer handoffs because the load on the APs remains balanced.

The WDs autonomously pick the handoffs in the MPD algorithm. The number of handoffs rises as a result of active wireless users contributing less to the handoffs. The universal controller, which chooses the order of the handoffs, monitors the handoffs in the proposed ASLB. In order to prevent handoffs from being repeated, the WDs are switched to the least-used AP when a particular AP becomes overused. In contrast to MPD system, DASA scheme chooses the best destination AP based on the APs SINR values. Higher handoffs times are the result of longer SINR calculation times. Accuracy provided by the SINR measurement results in fewer handoffs but more expensive probing overheads. The DASA scheme relies solely on the probe-request and probe-response between the APs and the WD’s for SINR measurement. The additional probing frames necessary for a precise estimate of the SINR are not taken into account. The CMAS technique uses the bandwidth that is available as the metric for identifying the handoffs. The number of handoffs for a high density Wi-Fi network cannot be entirely determined by this measure, which is only useful in calculating the contention in the Wi-Fi network. When compared to the CMAS, DASA and MPD schemes, ASLB reduces the packet loss rate by 26%, 29%, 35% and 44%, respectively.

5.5. Workload

The actual throughput for each AP is displayed in Figure 10 to demonstrate the efficacy of the load balancing methods. It is clear that the MPD- and RSSI-based systems display greater load variations amongst the APs and have poorer normalized throughput values. Due to the uneven load distribution among the APs, the same is true for DASA. The CMAS system displays better throughputs, but the loading disparity across the APs is also quite large. Due to the centralized SDN design, the proposed ASLB exhibits superior load distribution across all APs. Additionally, the ASLB network performance is enhanced by fewer computations. When compared to CMAS, DASA, MPD and RSSI schemes, ASLB optimizes the workload by 35%, 44%, 57% and 75%, respectively.

6. Conclusions

In this paper, we showcased the load distribution predicament for high density SD-WiFi for the data plane and control plane. We proposed a symmetrical load balancing scheme which comprises traffic shaping and cluster-based traffic classification for both the data plane and control plane. In the data plane we suggested a load distribution strategy, which progressively considers the AP side and the WD side details during load balancing. We built a simulation framework to appraise the load balancing methodologies. Load balancing is also achieved in control plane, to resolve the over loading issue in controllers. The control plane in two tier is managed by a UC. The UC manages all the regional controller in the clusters. AHP is employed to help prioritize the flows and assign packets to least loaded RCs. The extensive simulation runs demonstrate that by clustering and prioritizing packets, notable enhancements in performance can be attained when a symmetrical load balancing is achieved. The performance evaluation show that the ASLB algorithm accomplishes superior load distribution. ASLB has a greater throughput by 40% and lesser delay by 25% compared to the load balancing methods carried out by MPD, CMAS, RSSI and DASA schemes. The proposed ASLB can be implemented to attain better response for time-critical applications. We plan to induce artificial intelligence (AI) into APs to make association decisions themselves in the near future for health care. We are working to use the Mininet-Wi-Fi-NS3 emulation platform for AI enabled APs, which already has the SDN wireless extension libraries.