Detecting and Predicting Models for QoS Optimization in SDN

Abstract

Recently, deep learning algorithms and software-defined networking technologies enabled traffic management to be more controllable in IP networking and mobile Internet to yield quality services to subscribers. Quality of service (QoS) needs more effort to optimize QoS performance. More specifically, elephant flow management is a critical task that needs further research since its heavy hit behavior leads to high CPU utilization, packet drops, high latency, packet overflow, and network congestion problems. For this purpose, we focused on elephant flow management since elephant flows are big flows that hinder good service delivery (QoS) on demand. Hence, elephant flow detection and early prediction techniques optimize QoS. In this regard, we employed DNN and CNN deep learning models to detect elephant flows, and the random forest model predicts elephant flows in the SDN. As a result of our experiments, the findings reveal that deep learning algorithms within the Ryu controller significantly outperform in detecting and predicting performance in order to yield good network throughput. This solution proves to be significant for QoS optimization in data centers.

This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

1. Introduction

Multimedia applications running on the Internet have grown radically. Despite the fabrication being interesting, their traffic output requires high bandwidth and good traffic management to deliver qualified voice, images, and streaming videos [1]. Multimedia applications generate heavy hit traffic, which hinders the delivery of quality services. The generated traffic from such applications is usually elephant flows that exceed a certain threshold per unit of time or duration [2].

Elephant flows are not numerous, but they can occupy a disproportionate share of the total bandwidth over a period of time. They can lead to problems such as high CPU utilization, packet drops, and high latency. Managing elephant flows is a prerequisite to delivering QoS and satisfies subscribers on demand [3]. QoS is about the overall performance of a computer network, evaluated by Internet service providers and the network users. QoS is the transfer of the data in accordance with data size, duration, and reserve network bandwidth.

Given the high rate of elephant flows in network traffic, their prediction, detection, and effective control are critically important to satisfy applications’ requirements. Furthermore, QoS can be optimized for high throughput, low delay, and minimum loss. To achieve good QoS, we used deep learning models for elephant detection and a random forest (RF) algorithm for predicting elephant flows within the software-defined network (SDN) [4].

Hence, the concern of this study is to provide high QoS throughout by controlling the delay, latency, bandwidth, and packet loss parameters, which allows the network operator to allocate available resources for real-time applications and optimize end-to-end QoS using SDN and deep learning model opportunities [1].

SDN is an essential networking architecture to automate networks and maximize network throughput and user experiences [2, 5]. Thus, QoS optimization is investigated in this new networking architecture aiming QoS (throughout) optimization by developing deep neural network (DNN) and convolutional neural network (CNN) models to detect elephant flows, and we employed RF to predict the occurrence of elephant flows and then forecast the probability of network congestions in advance [1]. QoS optimization is achieved due to the prediction of the elephant flows using the prediction model from the labeled dataset (1, 0).

SDN eliminates the need to navigate several switches, resulting in significant time and effort savings, which enhances network scalability and improves traffic management [6]. Hence, making the controller more dynamic and intelligent is crucial. Furthermore, integrating deep learning models in the Ryu controller becomes a potential mechanism to optimize network performance within the SDN.

This work integrated elephant detection and prediction models to optimize QoS throughput. To optimize QoS, we installed the Anaconda tool in the Ryu controller and developed an elephant prediction model on top of the Ryu controller. The SDN dataset is processed in the Ryu controller and bandwidth optimization was carried out using the Mininet emulator and Ryu controller [4].

We contributed to elephant detection and predicting models for QoS optimization using deep learning models and RF algorithms on top of the SDN environment. Optimizing QoS within the SDN using these models outperforms compared with the detection and prediction performance using the Naive Bayes algorithm. Predicting and detecting elephant flows optimize routes between the two nodes.

1.1. Motivation

The problem arises when the massive amount of Internet traffic from multimedia applications puts a great overhead on network capacity, specifically in data centers. This, in turn, impacts QoS negatively. To ensure optimal QoS, we need to search for a solution to control elephant flows that lead to handling congestion.

Predicting big traffic helps network planners and administrators provision their network capacity in a cost-effective manner while meeting network QoS requirements such as good throughput.

Therefore, early identification of elephant flows helps to forecast traffic that can analyze the impact of forecasted traffic on the QoS parameters, helps identify optimal links, and finally suggests an optimal routing strategy that can improve the overall network health [7].

2. Related work

Network optimization has been practiced to improve network performance. Network optimization can be explored in various architectural designs and methodologies. For example, load balancing is one of the techniques to optimize network performance by minimizing packet loss, jitter, congestion, latency, and managing bandwidth utilization [7]. Traffic management can be leveraged to sustain using predictive quality of service (QoS) [8]. Proactive elephant flow prediction can map applications and maximize throughput.

In our work, we employed network architecture (SDN) and a deep learning model to detect elephant flows for congestion presentation before the multipath routing algorithm starts optimal path identification. The authors in the paper [9] optimized QoS using reinforcement learning for online QoS routing in SDN-based networks.

Software-defined networks are state-of-the-art network architectures that manage and monitor the network. It is an emerging paradigm that promises to dramatically simplify the network management system and enable innovation [9]. It is a more flexible and programmable architecture comprising AI components for dynamic traffic management. Unlike legacy network architectures (e.g., OSI and TCPIP), SDN separates the network into a control plane and a data plane [10]. The separation of the network’s control logic from underlying routers and switches is logically centralized in software-based controllers (the control plane) and network devices turn into simple packet forwarding devices (the data plane) that can be programmed globally on each hardware unit via an open interface (southAPI) OpenFlow [9]. The control layer and the application layer are connected to the northAPI protocol, as demonstrated in Figure 1.

[figure(s) omitted; refer to PDF]

The control plane takes the responsibility of traffic flow decisions. Ryu, POX, and Floodlight are some examples of controllers. On the other hand, the infrastructure layer comprises of the data plane (OpenSwitches) and its traffic-forwarding capability [7]. The application concerns the traffic from various multimedia and real-time applications, which results in elephant flows. Elephant flow detection using SDN enables the provision of good QoS based on various traffic classes, namely, VoIP, video streaming, and file transfer using Mininet and Ryu controller [4].

Recently, SDN with machine learning (ML) has become an emerging and promising solution for QoS optimization. Thereby, we integrate a deep learning model in the control plane, specifically within the Ryu controller to detect the occurrence of heavy hit traffic. Elephant flow prediction in advance becomes possible using more suitable prediction algorithms such as RF, and the Ryu controller detects big flows from flooded UDP flows [11]. Thus, we develop a deep learning model in the SDN environment and integrate a deep learning model in the Ryu controller. In the paper [14], the deep learning model yields good network traffic management and optimizes QoS. Detecting elephant flows with the help of a deep learning model dynamically helps to prevent traffic congestion from emerging Internet multimedia applications [9]. In previous works [9, 12], decision tree (J48), Naive Bayes, RF, XGBoost, and GBM were used to detect a number of flows in SDN for taking swift actions to avoid congestion or reroute the flow in the network. The RF model compares the relationship between a dependent traffic class variable (elephants and mice) and independent variables (network traffic inputs) [4].

However, their model development was using Weka, which is outside of the SDN environment. Unlike the works of the authors in [9, 12], we developed and deployed a deep learning model in the Ryu controller using a Mininet emulator. The summaries of related papers are presented in Table 1.

Table 1

Summary of related literature.

Reference	Author	Year	Traffic dataset	Approach	Key contributions
					SDN	Deep learning	QoS	Elephant
[7]	Askar et al.	2022	SDN data	Best effort	✓	X	✓	x
[11]	Babbar and Rani	2022	Generate data using Iperf	Best effort	✓	X	✓	✓
[4].	Pacharakit and et al.	2022	VoIP and video streaming	Best effort	✓	X	✓	✓
[12]	Mondala et al.	2021	TCP and UDP flows	Machine learning	✓	X	✓	✓
[13]	Zhang et al.	2022	Online generated data	Machine learning	✓	X	✓	x
[14]	Polat et al.	2022	TCP and UDP flows	Deep learning	✓	✓	x	x
Our paper	Getahun and et al,	2023	SDN dataset	Deep learning and random forest	✓	✓	✓	✓

3. Methodology

3.1. QoS Optimization Techniques

The SDN architecture has been enacted in the data center network to control traffic. SDN uses southbound OpenFlow protocol to interface OpenSwitches with the controller. The northbound interface communicates network applications with the controller. The software-defined controlling mechanism handles network traffic in a dynamic way. Furthermore, the SDN controller enables optimal link selection or topology changes on top of network topology [15]. Amongst controllers, Ryu is one of the Python-based controllers that supports various OpenFlow protocols. It works with the Mininet network emulator [4]. Due to the diversity of SDN applications and the use of the Ryu controller, it has become one of the choice of best-fitted controllers for QoS performance. The Ryu controller makes decisions regarding network QoS with the support of a deep learning model [16]. Deep learning models were developed using Anaconda and integrated into the Ryu controller given the SDN dataset. The SDN dataset was labeled as elephant flows and mice flow. Then, the elephant traffic detection and prediction were possible based on this labeled dataset on top of the Anaconda platform. TensorFlow and Keras framework were installed on the Anaconda before integrating the Anaconda into the Ryu controller.

To test to what extent the prediction model detects and predicts elephant flows, traffic generation is accomplished using to what extent the detection and prediction model detects and predicts elephant flows, and traffic generation is accomplished using the Iperf tool. Iperf generates, measures traffic, and informs the presence or forecast probability of elephant flows’ existence in advance,

Ryu monitors hosts and switches since the controller obtains the globalized network view. We designed linear network topology (hosts [h], switches [s], and links) in Mininet with the same number of switch scenarios. The custom topology is demonstrated in Figure 2.

[figure(s) omitted; refer to PDF]

The workflow for the elephant flow prediction includes model construction, validation, and prediction based on the labeled SDN dataset. The Anaconda deep learning platform was used to develop the prediction model.

The QoS parameter (throughput) is considered to evaluate the QoS optimization with the help of the traffic prediction model. The deep learning module contains a deep learning–based RF model in the Ryu controller. We also developed and described a Naïve Bayes’ traffic prediction model in Ryu for comparison with deep learning–based RF models.

For emulation purposes, Mininet, Iperf, ping, and Ryu controllers were employed during experimentation. An evaluation of Ryu controller performance was conducted and investigated in terms of QoS parameters and throughput using UDP traffic. The Ryu controller performed better in throughput, delay, and jitter than the POX controller [17]. Thus, we choose the Ryu controller as we require higher throughput for our experimentation. To measure the throughput on top of the RF prediction model, Iperf is employed to generate traffic with elephant or mice parameters. Iperf has server and client functionality to create data flows and measures the performance from the end-to-end network [11]. In the case of UDP measurements, Iperf measures a default bandwidth of 1 Mbps for the links. We can change the bandwidth to 10 Mbps using the –b option during the experimentation.

3.2. RF

The RF algorithm is one of the most robust mathematical methods for deep learning. It is applicable for problems which have a sequence of actions’ behavior and to optimize the overall output function [18]. We, therefore, suggest an RF model for solving the problem of traffic prediction (elephants) for more QoS optimization.

3.3. Performance Evaluation Metrics

Multimedia applications and their traffic are considered in the QoS class definition because “elephant” flow prediction and detection are vital for healthy transmission [19]. UDP data streams were used to measure the throughput of a network (QoS). Multimedia traffic prefers UDP transmission over TCP traffic flow. UDP traffic flows are floodable, which leads to network congestion. Therefore, it is the most important critical traffic issue to prevent congestion and save the bottleneck link [20]. Throughput is defined as the number of packets processed through the network in a unit of time. The throughput is calculated a followss:

  throughput = total data sent/delivery time.

3.4. Emulation Setup and Tools

To set up the network simulation and network environment, it is necessary to install many software tools. The details of software tools of each element/system used in this simulation are shown in Table 2.

Table 2

Tools used.

Systems	Details
Operating system	Lubuntu 20.04
SDN controller	RYU 4.3.4
Switches	Open vSwitch 1.3
Network emulator	Mininet 2.2.2
Processor and memory	Core I7 and 8 GB memory
Programming language	Python 3.7.3
OpenFlow	OpenFlow 1.3
Iperf	Iperf 3.1.3 to generate traffic
WinSCP	WinSCP 6.1.2 to transfer file from window to Mininet
Oracle VM VirtualBox	VirtualBox 7.0.12
Xming	Xming 6.9.0.31
PuTTY	PuTTY 0.79 for SSH and telnet client on windows platform.
Anaconda	Anaconda 2.6.0
Spyder	Spyder 3.11

3.5. SDN Dataset

Elephant and mice network traffic were considered during model formulation in the SDN, which is presented in Table 3.

Table 3

Elephant and mice flow heuristics.

Label	Descriptions
Elephants	Elephant flows ≥ 15 packets and takes ≥ 10 second duration
Mice	Elephant flows < 15 packets(500 byte) and takes < 10 second duration

3.6. Evaluation Metrics

The traffic prediction model performances were evaluated using accuracy, and the QoS parameter (throughput) was considered to evaluate the traffic prediction model.

4. Results and Discussion

4.1. Environmental Setup

SDN traffic is differentiated between elephant and mice flows using ML algorithms. The Naïve Bayes classification algorithm was employed and compared with the RF algorithms. The experimentation was conducted on the Mininet emulator, Ryu controller, and TensorFlow on the Anaconda deep learning tool.

For elephant flow prediction, we flooded different flows and set the duration to 10 s during the dataset class assignment. The predictive elephant model provides better QoS to flows of small sizes by deprioritizing elephant flows.

We tested the prediction model to enable network performance using the Iperf tool. We got a promising elephant prediction performance.

We then compare the performance with best-effort QoS and deep learning–based QoS mechanism under the same network topology within the Mininet environment and Ryu controlling utilization. An experimental evaluation was accomplished using throughput on UDP traffic. We run the elephant detection model using Processor Intel(R) Core (TM) i7-4500U CPU @ 1.80 GHz, 2401 MHz, 2 Core(s), 4 logical processor(s), and 8 GB RAM.

4.2. Result Findings

We present a QoS prediction model, predicting the probability of elephant flows and the maximum achievable throughput. To this end, the deep learning and RF models were in the Ryu controller. We also developed and described a Naïve Bayes traffic prediction model in Ryu for comparison with the RF models. The comparison is represented in terms of accuracy and model performance seed. The confusion matrix is represented by the number of flow instances in Naive Bayes and RF models.

The elephant traffic prediction model using the Naïve Bayes algorithm yields 71.46% accuracy within 0:01:57 s.

4.2.1. DNN

The model performance increases from 95.60% to 99.98% accuracy. This indicates that the model was well-trained and performed with high accuracy except for small errors. As presented in Figure 3, the validation accuracy increases from the start and finishes with almost similar performance.

[figure(s) omitted; refer to PDF]

4.2.2. CNN

Both the training and validation performance scores were 98.28% and 98.23% in terms of f accuracy, respectively, as presented in Figure 4. However, the validation accuracy becomes lesser when we compare it with training accuracy. Dropout has the greatest impact on validation accuracy whereas the hidden layer size appears to have no impact on validation accuracy [21].

[figure(s) omitted; refer to PDF]

4.2.3. Prediction Using the RF Model

The elephant traffic prediction model using the RF algorithm outperforms and yields 100% within 9 s, as presented in Figure 5. RF is the most successful tactic to conquer the dilemma of training overfitting than DNN and CNN. Every one of the decision trees from the RF is assembled on a subset of the traffic information, which is reached by sampling with replacement in the initial SDN dataset [22]. Thus, the developed model truly predicts elephant flows with evidence of detecting elephant flows within 0:01:11 s.

[figure(s) omitted; refer to PDF]

4.3. Confusion Matrix Analysis and Algorithm Comparison

Elephant flow predictive analysis is conducted using RF and Naive Bayes algorithms [23]. To check the performance of a traffic behavior prediction assuming QoS requirement RF and Naive Bayes models, the confusion matrix is deployed. We summarized the number of correctly and incorrectly predicted traffic. Almost all items were predicted in RFs as our class label assignments (elephant and mice flows). Meanwhile, the confusion matrix from Naive Bayes states that many mice instances are predicted to be elephant flows mistakenly.

The traffic prediction time of RF is 9 s. When we compare the model running time, Naïve Bayes, DNN, CNN, and RF, RF is found to be well-performing and is fast enough. Specifically, the RF model perfoms 100% in accuracy, whereas the Naive Bayes model performs less as it yields 71.46% in accuracy. RF predicts elephant flows accurately. The DNN algorithm performs with 99.98% accuracy, which is a promising algorithm. We also detect elephant flows using a CNN and we found that CNN performs with 98.23% accuracy, which is also a well-performing algorithm except for some training dropout challenges [21].

To summarize, we present an algorithm comparison in Figure 5.

4.3.1. Elephant Prediction and QoS Optimization

The predictability of the traffic model was evaluated using throughput on top of connected hosts under Mininet topology and Ryu controller. First, the reachability of hosts was checked across each host using the ping command. The transfer of 11 packets from Host 1 (10.0.0.1) to Host 18 (10.0.0.18) is shown as an example in Figure 6.

[figure(s) omitted; refer to PDF]

The client hosts were also connected to servers for downloading and uploading files. The connection for downloading was tested on Host 6 using the Xterm tool, i.e., 135 packets of a file (Index.html) were downloaded.

The connectivity is also proved by uploading a file from client Host 6 to the server Host 17 using the Xterm tool.

Once we checked the reachability and connectivity of topology, we measured the performance of the network in a long Ryu environment and deep learning model. We run Host 1 Iperf -u -c 10.0.0.17 -i 10 -b 10m -t 30 at Mininet. As a result, throughputs were recorded using a RF algorithm.

For the UDP test, we configured the maximum bandwidth limit to a large value (-b10m) so that we can measure the actual bandwidth of the custom topology. As indicated in Table 4, the measured UDP throughput was 9.85 Mbits/sec in the first 10 s, which is very close to the capacity of the links in the network (10 Mbits/sec). This test was conducted using UDP packets because UDP flows do not have transmission control. UDP traffic is flooded from client Host 1 (10.0.0.1) to Host 17 (10.0.0.17) within 10-time intervals (duration) by setting the bandwidth of 10 Mbits/sec of data transmission capacity. The model predicts 11.7 MBytes, 11.9 MBytes, and 11.9 Mbytes as elephant flows since each of these flows are greater than 15 packets. We can entail that each mice flow requires less than 15 packets, approximately [24], and each packet contains 500 bytes [25]. Mice flows have a size of 10 KB in OpenSwitches of data centers [26] and an average duration of 10 s [26].

Table 4

UDP flows.

ID	Interval in sec	Transfer in MBytes	Bandwidth in Mbits (sec)
1	0.0–10.0	11.7	9.85
2	10.0–20.0	11.9	9.99
3	20.0–30.0	11.9	9.99
4	0.0–30.0	35.6	9.94

Therefore, a mice flow contains 7500 bytes to be transferred in a second. A mice flow also takes 10 s at maximum to transfer 75,000 bytes on average.

When we compare 75,000 bytes with the first-round data, 11.7 Mbytes (11700000 bytes), the latter is much larger than 75,000 bytes. Therefore, we can infer that this flow is an elephant flow. The prediction model predicts this flow as an elephant flow, as the prediction model performs with 100% accuracy. For this purpose, we run and train a model using the RF algorithm in the Ryu controller. The prediction performance of the model is described using the confusion matrix in Figure 7 and displayed its elephant flow identification capability.

[figure(s) omitted; refer to PDF]

Three flow transfer predictions were undertaken on the three throughputs 11.7 MBytes, 11.9 MBytes, and 11.9 Mbytes.

Network throughput of 11.7 MBytes, 11.9 MBytes, and 11.9 MBytes and a total throughput of 35.6 MBytes were observed over the 30 s. All forecasted traffic was predicted as elephant flows. Hence, we can conclude that the QoS optimization goal is achieved using the RF algorithm over Naïve Bayes prediction. Furthermore, a comparative analysis of these two classifiers shows that maximum accuracy can be achieved by the RF methods [27] The deep learning–based elephant flow detection and prediction optimize QoS in the SDN since controlling elephant flow results in good QoS improvement. Detecting and predicting the existence and probability of elephant flows conserve CPU resources and manage other competing flows, and congestion causes issues such as increased latency or packet drops.

When we compared with Naïve Bayes, DNN performs with 99.98% accuracy in detecting elephant flows, whereas RF is a potential algorithm to predict elephants, i.e., it performs with 100% accuracy.

We also show the QoS optimization by comparing the Ryu simple Switch13 and RF model performances. The latter optimizes QoS due to advanced congestion prediction capability, thereby reducing network delay, improving network link utilization, and improving network performance [28, 29].

Traffic flows save bandwidth in Mbits/sec when traffic management is monitored using a RF model. 9.94 Mbits/sec from the given bandwidth (10 Mbits/sec) was utilized using a RF model, which is presented in Table 4.

Whereas the best-effort algorithm of the Ryu controller utilizes 9.99 bandwidth in Mbits/sec, which consumes almost the total bandwidth as seen in Table 5. We can conclude that the path has a high probability of congestion problem.

Table 5

Bandwidth utilization using the best-effort approach (simple Switch13).

	Interval in sec	Transfer in MBytes	Bandwidth in Mbits (sec)
simple Switch13	0.0–10.0	11.9	9.99
	10.0–20.0	11.9	9.99
	20.0–30.0	11.9	9.99
	0.0–30.0	35.7	9.99

Thus, we can deduce that the RF model optimizes throughput due to the wise use of bandwidth. 9.94 Mbits/sec bandwidth is a widely used indicator of bandwidth than 9.99 Mbits/sec. 9.99 is closer to a congestion state since the whole bandwidth is ready to be consumed, and network congestion will happen.

The results show that the RF method performs well for elephant flow forecasts since the RF is the most stable and best-performing model than deep learning algorithms [30].

5. Conclusion and Recommendation

5.1. Conclusion

Multimedia applications create overhead on network capacity, specifically in data centers. To overcome the impacts of traffic congestion that affect QoS negatively, we employed deep learning models, DNN, and CNN to detect elephant flows, and a good elephant flow detection performance was gained.

Predicting elephant flows helps network planners and administrators to be aware of QoS constraints in advance. Early identification of elephant flows is used to analyze and forecast traffic to identify optimal links, and finally, an optimal routing strategy can be suggested that can improve the overall network health.

In this regard, the RF algorithm performs with 100% accuracy within 0:01:11 s. The RF algorithm is the promising elephant prediction algorithm than DNN, CNN, and Naïve Bayes.

The presented model acquired findings from a single dataset, which serves as a limitation of the model. Consequently, a distributed dataset with a large language model (transformer) can be examined in order to provide directions for future QoS enhancements. Furthermore, this work has limitations concerning DNN and CNN experimentation and application for QoS optimization.

5.2. Recommendation

It may fail to detect the elephant traffic in a multicontroller environment. So, these models shall better be evaluated in a multicontroller context. Memory, other limited resources, computing abilities, and a diversity of standards and protocols characterize the Internet of Things. Further traffic management research in the smart data centers, edge data centers, and content delivery networks (CDNs), and Internet exchange points (IXPs) will be vital to delivering good QoS to users. Furthermore, the experimentation requires testbeds on real physical Ryu controllers and OpenFlow on the data centers.

Author Contributions

All authors reviewed the manuscript and approved the findings. All authors have contributed their efforts. Correspondence authorship goes to Getahun Wassie and Jianguo Ding.

Funding

No funding was received for this manuscript.

Acknowledgments

The authors thank Dr.Million Meshessha who taught AI and machine learning courses to the first author (Getahun). Without their effort, the authors would not have been writers of this paper.

References

[1] N. Thazin, "Qos-Based Traffic Engineering in Software Defined," 2019. A Thesis Submitted to the University of Computer Studies, Yangon in Partial Fulfilment of the Requirements for the Degree of Doctor of Philosophy

[2] M. H. A. E. Al, "Flow-Aware Elephant Flow Detection for Software-Defined Networks," IEEE Access,DOI: 10.1109/Access.2020.2987977, 2020.

[3] CISCO, Cisco Elephant Flow Detection User Guide, 2023.

[4] A. A. B. G. Bassene, A Class-Based Adaptive Qos Control Scheme Adopting Optimization Technique over Wlan Sdn Architecture, 2020.

[5] B. K. Madhavi, Detection of Ddos Attacks on Sdn Network Using Machine Learning, 2022.

[6] M. Fallah, P. Mohammadi, M. NasiriFard, P. Salehpour, 2023.

[7] S. Askar, B. Hassan, T. A. Rashid, "SDN Based Fog Computing: A Review," International Journal of Science and Business, vol. 5 no. 6, 2021.

[8] A. Palaios, "Machine Learning for Qos Prediction in Vehicular Communication: Challenges and Solution Approaches," 2023. https://arxiv.org/abs/2302.11966

[9] A. M. Talpur, "Congestion Detection in Software Defined," Sustainable Communication Networks, 2017.

[10] S. Naik, A. Gawde, N. Nagori, International Journal of Innovations In Engineering Research And Technology, 2021.

[11] H. Babbar, S. Rani, IOP Conference Series: Materials Science and Engineering, 2021.

[12] P. K. Mondala, L. P. Aguirre Sancheza, E. Benedettoa, Y. Shen, M. Guo, "A Dynamic Network Traffic Classifier Using Supervised Ml for A Docker-Based Sdn Network," Connection Science, 2021.

[13] L. Zhang, Y. Lu, D. Zhang, H. Cheng, Security and Communication Networks, 2022.

[14] H. Polat, M. Türkoğlu, O. Polat, A. Şengür, "A Novel Approach for Accurate Detection of the Ddos Attacks in Sdn-Based Scada Systems Based on Deep Recurrent Neural Networks," Expert Systems With Applications, vol. 197, 2022.

[15] M. M. Raikar, S. M. Meena, M. Moin Mulla, N. S. Shetti, "Data Traffic Classification in Software Defined Networks (Sdn) Using Supervised-Learning," Procedia Computer Science, vol. 171, pp. 2750-2759, 2020.

[16] S. Askar, F. Keti, "Performance Evaluation of Different Sdn," International Journal of Business, 2022.

[17] S. J. Rashid, A. M. Alkababji, A. M. Khidhir, "Performance Evaluation of Software-Defined Networking Controllers in Wired and Wireless Networks," Telkomnika (Telecommunication Computing Electronics And Control), vol. 21 no. 1, pp. 49-59, 2023.

[18] Z. H. Choudhury, Vignan’s Foundation For Science, Technology And Research, 2023.

[19] P. Wang, S.-C. Lin, M. Luo, IEEE International Conference On Services Computing (Scc), 2016.

[20] M. P. Ramkumar, "Optimized Traffic Management in Software Defined Networking," International Conference On Advanced Computing, Machine Learning, Robotics And Internet Technologies, 2023.

[21] J. Barry-Straum, "An Evaluation of Training Size Impact on Validation Accuracy for Optimized Convolutional Neural Networks," Master of Science in Data Science, 2018.

[22] K. Mahesh, M. Laxmaiah, Annals Of RSCB, vol. 25 no. 5, 2021.

[23] V. Sofia, "Confusion Matrix-Based Feature Selection," Maics, vol. 710 no. 1, pp. 120-127, 2011.

[24] G. Tam, Cisco Application Centric Infrastructure, 2018.

[25] S. Adibi, "A Remote Interactive Non-repudiation Multimedia-Based M-Learning System," Telematics and Informatics, 2010.

[26] B. M. Vinicius, Journal of Internet Services And Applications, vol. 11, 2020.

[27] M. Waqas Nadeem, H. G. Goh, V. Ponnusamy, Y. Aun, "Ddos Detection in Sdn Using Machine Learning Techniques," Computers, Materials & Continua Tech Science Press, 2021.

[28] Q. Tang, H. Zhang, J. Dong, L. Zhang, "Elephant Flow Detection Mechanism in Sdn-Based Data Center Networks," Scientific Programming, 2020.

[29] D. Ó. Briain, "RYU Sdn Manual," 2020. https://www.obriain.com/training/sdn/RYU_Soft_Testbed_v2.0.pdf

[30] Y. Wang, L. Shi, Y. Hu, X. Hu, W. Song, L. Wang, "A Comprehensive Study of Deep Learning for Soil Moisture Prediction," State Key Laboratory of Water Resources Engineering and Management, 2024.