                                                                 Computer Standards & Interfaces 97 (2026) 104109


                                                                      Contents lists available at ScienceDirect


                                                          Computer Standards & Interfaces
                                                              journal homepage: www.elsevier.com/locate/csi




DEA-GAO: A two-stage approach optimal controller placement in
software-defined networks using data envelopment analysis
Fatemeh Dashti , Ali Ghaffari * , Nahideh Derakhshanfard , Shiva Taghipoureivazi
Department of computer engineering, Ta.C., Islamic azad university, Tabriz, Iran




A R T I C L E I N F O                                       A B S T R A C T

Keywords:                                                   Software-defined networks (SDN) have demonstrated considerable benefits in various practical domains by
Software-defined networking (SDN)                           decoupling the control plane from the data plane, thus facilitating programmable network management. This
Controller placement problem                                paper presents a two-stage approach for solving the problem of controller placement called DEA-GAO. In the first
Delay
                                                            stage, this strategy assumes the SDN network as a graph and using Data Envelopment Analysis (DEA) and relying
load balancing
reliability
                                                            on graph centrality metrics such as closeness centrality, betweenness centrality, and eigenvector centrality,
                                                            calculates the efficiency of nodes to determine the optimal locations for deploying controllers. In the second
                                                            stage, to allocate switches to controllers, the proposed strategy employs the Green Anaconda Optimization al­
                                                            gorithm (GAO) to achieve an optimal allocation while considering network parameters such as average delay,
                                                            load balancing, and reliability. Finally, to assess the efficacy of the proposed methodology, it is juxtaposed with
                                                            three extant methods utilizing diverse datasets from the Internet Topology Zoo. The experimental findings
                                                            indicate that the proposed approach significantly surpasses the existing methods, specifically the hybrid RDMCP-
                                                            PSO algorithm, heuristic CPP algorithm and PSO algorithm in terms of both average delay (8.8 %, 28.8 % and
                                                            22.2 % respectively) and controller utilization (1.5 %, 7.3 % and 32 % respectively).




1. Introduction                                                                              processing capacity, more so concerning scalability, performance, and
                                                                                             reliability. To mitigate these challenges, the use of multiple controllers is
    The Software-Defined Networking (SDN) paradigm is a sophisticated                        a feasible solution. The multi-controller approach splits the network into
model supporting dynamism, automation, flexibility, and centralized                          smaller domains with every one of them being handled by a specific
control of the underlying network that distinguishes it from other con­                      controller. The set of controllers controlling these regions communicates
ventional networking models. The new paradigm has intrinsic merits                           with one another through their west/eastbound interfaces [1].
like enhanced performance and reduced latency [1]. In SDN, there is a                            For SDN multi-controller networks, the number of deployed con­
clear distinction between the control plane, which decides on routing,                       trollers is based on the network size and the amount of traffic [2,4]. The
and the forwarding plane, which is responsible solely for packet for­                        major challenges with SDN multi-controller structures involve the
warding based on routing instructions it receives from the control plane                     number of deployed controllers and the placement of the deployed
(Fig. 1). The planes communicate with each other via Application Pro­                        controllers on distributed switches, typically the controller placement
gramming Interfaces (APIs), e.g., Representational State Transfer                            problem (Fig. 2). This problem aims to determine the most suitable lo­
(REST) on the northbound interface and OpenFlow on the southbound                            cations for SDN controllers to achieve different goals like the reduction
interface [2,3]. While separating control and data planes offers technical                   of latency, efficiency in terms of power, load balancing, and enhanced
benefits, it also involves several challenges, including how many con­                       reliability [2,4]. The CAPEX and OPEX of the network are greatly
trollers are required, optimally where they must be placed, and how data                     influenced by the number of installed controllers. Hence, the best
plane devices are to be assigned to controllers [1,3].                                       number of controllers and the determination of their most optimal lo­
    With the increasing scale of Software-Defined Networking (SDN)                           cations are also of prime concern in SDN multi-controller networks [4].
infrastructures, the implementation of one controller to handle the                              The ideal placement of these controllers directly affects different
entire network has presented more challenges to the SDN controller’s                         performance parameters such as latency, reliability, and fault tolerance



 * Corresponding author.
   E-mail address: A.ghaffari@iaut.ac.ir (A. Ghaffari).

https://doi.org/10.1016/j.csi.2025.104109
Received 12 August 2025; Received in revised form 13 September 2025; Accepted 28 November 2025
Available online 2 December 2025
0920-5489/© 2025 Elsevier B.V. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
F. Dashti et al.                                                                                                 Computer Standards & Interfaces 97 (2026) 104109


which are extremely crucial to ensure a robust network infrastructure. In         closeness, betweenness, eigenvector, dispersion, information, and po­
real-world environments, this problem is further exacerbated by the               tential influence as parameters, thereby defining each graph node as a
need to consider possible controller failures that can bring network              DMU within the DEA framework. The outcome of this methodology
operations to a standstill [5,6]. The resilient and functional SDN topol­         delineates the locations of the network controllers. Subsequently, in the
ogy should not only be designed to perform optimally under typical                second stage of the proposed methodology, we endeavored to optimally
operating conditions but also to handle effectively the scenario of one or        allocate the network switches to the identified controllers by utilizing
more controller failures while minimizing interruption and maintaining            the Green Anaconda optimization algorithm, thus providing an optimal
continuity of services [7]. This need arises from an overall objective of         resolution for the controller placement conundrum in SDN networks,
creating Software-Defined Networks that are not only operationally                which we implemented in Python and simulated within the Mininet
effective and responsive but also inherently resilient and fault-tolerant         environment.
to the extent of surviving all types of failures without noticeable                   The main contributions of our research are delineated as follows:
degradation of service [7,8]. The need to address this issue stems from
the fact that the situation in SDN research today calls for adult and              • Considering the Software Defined Networking (SDN) infrastructure
highly sophisticated answers to the knotty issue of controller placement,            as a comprehensive graph and meticulously calculating the various
particularly against the backdrop of the dynamic and generally unpre­                centrality metrics that correspond to each node present within the
dictable nature of network traffic and the inherent risks of component               network topology.
failures [7]. Numerous solutions have been developed to address this               • Defining each distinct node within this intricate network as a
multi-dimensional problem based on different network parameters and                  Decision-Making Unit (DMU) and subsequently calculating the
as per different strategies, such as latency-aware models, traffic load              operational efficiency of each DMU to select the most optimal loca­
balancing strategies, cost-effective schemes, and overhead-aware con­                tion for the deployment of a controller by employing the Data
trol plane solutions [9]. Network reliability-based solutions, quality of            Envelopment Analysis (DEA) method.
service (QoS) concerns, and multi-objective optimization methodologies             • Formulating the challenge of assigning network switches to their
are also present that attempt to balance a set of conflicting parameters in          respective controllers is a sophisticated optimization problem that
controller placement [9]. Some researchers employed clustering-based                 necessitates a thorough analysis and strategic approach to find the
approaches, partitioning the entire network into multiple small clus­                most effective solution.
ters, each managed by independent controllers, and ultimately needing              • Utilizing the Green Anaconda Optimization (GAO) methodology to
synchronization among all of the controllers to offer centralized man­               achieve the optimal allocation of network switches to their corre­
agement of the entire network infrastructure [8].                                    sponding controllers, whereby each allocation scenario is concep­
    In this paper, we initially employ Data Envelopment Analysis (DEA)               tualized as a population of anacondas and progresses through
to ascertain the positioning of controllers across all nodes within the              various evolutionary stages to ultimately arrive at the most satis­
network, each of which possesses the potential capability to be desig­               factory solution to the stated optimization problem.
nated as a controller. Indeed, the application of the DEA methodology in
addressing the controller placement dilemma constitutes a robust and                  The subsequent sections of this manuscript are organized as follows.
empirical strategy for navigating the intricacies associated with this            Section 2 reviews pertinent literature that has employed evolutionary
matter, and through the utilization of this approach, the optimal loca­           algorithms in addressing the controller placement issue. Section 3 pro­
tions for the controllers can be discerned while accommodating various            vides a comprehensive exposition of the proposed DEA-GAO method­
and often conflicting criteria. The DEA methodology assesses the effi­            ology, incorporating the foundational principles of Data Envelopment
ciency of each unit within the Decision-Making Unit (DMU) framework               Analysis (DEA) and the GAO algorithm. Section 4 assesses the effec­
by delineating inputs and outputs corresponding to each unit, thus                tiveness of the proposed approach. Finally, we synthesize our findings
enabling the prioritization of units based on their efficiency, which fa­         and conclusions in Section 5.
cilitates the selection of several high-efficiency units [10]. To establish
the input and output parameters, we conceptualized the                            2. Related works
Software-Defined Networking (SDN) architecture as a graph and
employed graph centrality metrics, including degree centrality,                      All past research has considered the challenges of the distributed




                                                Fig. 1. SDN architecture vs. traditional network architecture.

                                                                              2
F. Dashti et al.                                                                                                  Computer Standards & Interfaces 97 (2026) 104109


controller’s strategy in two main parts such as: (i) identifying the             system; a nature-inspired metaheuristic technique Particle Swarm
optimal number of controllers necessary for a specified network topol­           Optimization (PSO) is later employed for its resolution. With strategic
ogy such that each controller is neither overburdened nor underutilized,         placement of controllers, this adaptive method dramatically reduces
and (ii) optimizing the placement of controllers, as this has far-reaching       network latency over what would have occurred using the previous
implications for various factors such as flow setup latency, resilience,         method and builds upon switch assignment techniques. By so doing, we
and load distribution [11,12].                                                   believe that not only will this innovative approach better the operation
    Because this paper employs a biologically inspired evolutionary al­          of SDN (Software-Defined Networking), but also it sets standards for
gorithm known as the Green Anaconda Optimization Algorithm (GAO),                later controller deployment methodologies of tomorrow.
this section will delve exclusively into the evolutionary-based methods              Chunlin Li et al. have come up with a dependable controller place­
to optimize the quantity, positioning, and distribution of controllers           ment framework in the form of an optimization-based delay and load
within the Software-Defined Networking (SDN) milieu. These method­               consideration [13]. Their Load Optimization Controller Placement
ologies are categorized as bio-inspired techniques and Genetic Algo­             (LOCP) algorithm marginally performs better than K-means, Simulated
rithms (Table 1).                                                                Annealing (SA), and Equal-Cost Multi-Path (ECMP) algorithms in
                                                                                 propagation delay, queuing delay, and the efficiency of load balancing
                                                                                 in massive networks, achieving an average of 18.36 % improvement in
2.1. BIO-inspired CPP
                                                                                 load balancing along with optimizing propagation and queuing delays
                                                                                 simultaneously. In addition, they proposed a resource allocation algo­
    Numerous educational computing methodologies for SDN-oriented
                                                                                 rithm with task delay and reliability parameters along with a resource
Controller Placement Problems (CPP) emphasize the development of
                                                                                 allocation strategy employing a heuristic ant colony algorithm. This
hybrid optimization algorithms. Notable algorithms in this domain
                                                                                 solution addresses the problem of intelligently allocating resources of
include Particle Swarm Optimization (PSO) [12], Ant Colony (ACO)
                                                                                 Mobile Edge Computing (MEC) servers when the server cannot provide
[13], and others, summarized in the first part of Table 1.
                                                                                 task reliability in one channel to minimize average task delay. The
    Thalapala et al. introduced a novel method called Whale Optimiza­
                                                                                 experimental results establish that, relative to existing procedures, the
tion Controller Placement (WOCP) to carefully place the minimum
                                                                                 resource allocation method employing the Heuristic Ant Colony Algo­
number of controllers in an SDN architecture [11]. The WOCP is
                                                                                 rithm (HACA) effectively optimizes the workload distribution issue and
inspired by the mathematical model resembling the spiral bubble-net
                                                                                 substantially reduces user resource response time and the mean
feeding behavior of humpback whales to conduct optimization. By
                                                                                 completion delay at the cost of reliability parameters.
taking advantage of an objective function, this method aims to minimize
                                                                                     In their latest contribution to this area, Tahmasebi et al. have
the average latency of the network via the optimal placement of con­
                                                                                 formulated the placement problem of a controller as an optimization
trollers. They reduce the average network delay by 18 % in MLCP
                                                                                 problem with reliability, fault tolerance, and performance metrics
compared to the traditional method, MLCP, by reducing the latency
                                                                                 against latency, synchronization overhead, and deployment cost con­
between the switch node and the controller.
                                                                                 cerns [14]. They employed a nature-inspired, population-based
    Singh et al [12] focused on communication latency and reliable
                                                                                 meta-heuristic algorithm known as SYNCOP for solving the optimization
communication as the dual dimensions of the Controller Placement
                                                                                 problem with acceptable results.
Problem (CPP). The primary object is to suppress the average latency
                                                                                     Abeer A.Z. et al. employed a state-of-the-art heuristic method of
over a network. The test field will be planted with robust network
                                                                                 current standards to solve the Controller Placement Problem (CPP) with
models that can still perform (although less effectively) when n-1 out of
                                                                                 the view to optimize switch distribution and reconfiguration to near
n controllers are nonoperational. The form this work is about to take can
                                                                                 controllers in addition to their originally assigned controllers [15]. Their
be seen at its start. By use of an efficient algorithm called “Capacitated
                                                                                 algorithm computes optimal controller placement, minimum required
Controllers Arrangement (CCA)” to abstract the problem participants




                                         Fig. 2. Controller placement problem with different number of controllers.

                                                                             3
F. Dashti et al.                                                                                                          Computer Standards & Interfaces 97 (2026) 104109


Table 1
Evolutionary-based controller placement methods.
  Evolutionary Type     Ref.   Method                            Implementation                                           Evaluation metrics                      Year

  Bio-Inspired CPP      [11]   WOCP: Whale optimization          Simulation: Python programming                           Latency                                 2025
                                                                  Hardware: Intel i3 11th generation processor            Reliability
                                                                 Topology: Nsfnet, Agis, Geant
  ​                     [12]   PSO                               Simulation: MATLAB.                                      Latency                                 2024
                                                                  Hardware: Intel 1–7 processor,16 GB RAM, 1.6 series
                                                                 Graphic card
                                                                 Topology: Internet topology zoo
  ​                     [13]   Ant colony                        Topology: Internet2 OSE3, Janetbackbon                   load balancing                          2022
                                                                                                                          delay
                                                                                                                          Reliability
  ​                     [14]   Cuckoo                            Simulation: Java language                                Reliability                             2021
                                                                 Hardware: Mac OS, Core i5 2.9 GHz, and 8 GB memory       fault tolerance
                                                                 Topology: No standard topology                           latency
                                                                                                                          synchronization overhead
                                                                                                                          deployment cost.
  ​                     [15]   RDMCP-PSO                         Simulation: MATLAB.                                      Reliability                             2024
                                                                  Hardware: Intel Core i7-G10 processor and 16 GB RAM     load balancing
                                                                 Topology: Usnet, Iris-net, Chinanet, Bellcanada, Agis,   delay
                                                                 Aarnet
  ​                     [16]   APSO                              simulation: MATLAB.                                      Delay                                   2018
                                                                 Hardware: Intel ® core ™ i5–700 U CPU at 2.50-GHz        delay jitter
                                                                                                                          link traffic
                                                                                                                          reliability
                                                                                                                          controller load
                                                                                                                          economic Expense
  ​                     [17]   SSOA with chaotic maps            simulation: MATLAB.                                      Time utilization                        2019
                                                                  Hardware: Intel core i5 processor,                      cost utility
                                                                 Topology: Internet topology zoo
  ​                     [18]   PSO                               simulation: MATLAB.                                      SC and CC delay                         2017
                                                                 Hardware: ISP network with system 351 from               controller load
                                                                 Rocketfuel                                               imbalance
                                                                                                                          number of controllers
                                                                                                                          switch assignment
  ​                     [19]   FFA                               simulation: MATLAB.                                      SC and CC latency                       2018
                                                                 Hardware: Intel Core i5 with 4-Core processors           multi-path connectivity between the
                                                                                                                          switch and controller
  ​                     [20]   APFO based MOCP schema            simulation: MATLAB.                                      SC and CC and propagation latency       2018
                                                                 Topology: Internet2 OSE3, Tinet, RF-2                    Reliability
                                                                                                                          load balance among controllers
  ​                     [21]   VBO                               simulation: MATLAB, python programming                   Latency                                 2020
                                                                 Topology: Abilene, Savvis, Biznet, Internet2 OS3E        load balance (capitated and
                                                                                                                          incapacitated)
  Genetic Algorithm     [22]   Genetic Algorithm Hybridized by   simulation: MATLAB.                                      Propagation delay                       2019
   Based CPP                   AHP technique                     Hardware: 4-GH Intel Corei7 machine                      hop count
                                                                 Topology: Internet topology zoo                          link utilization
  ​                     [23]   Kuhn-Munkres algorithm            Hardware: Intel Core i7 (2.19 GHz)                       Propagation delay                       2018
                               Genetic algorithm                 Topology: TA2, GERMANY50 and INDIA35                     controllers load and balancing
  ​                     [24]   MHNSGA                            Simulation: MATLAB.                                      SC and CC latency                       2018
                                                                 Hardware: 4-GH Intel Corei7 machine                      Load balancing
                                                                 Topology: Internet topology zoo
  ​                     [25]   MHNSGA-2                          Simulation: MATLAB.                                      SC and CC latency                       2018
                                                                 Hardware: 4-GH Intel Corei7, CPU 4GH                     Load balancing
                                                                 Topology: Internet topology zoo
  ​                     [26]   Genetic Algorithm (BIP)           Simulation: MATLAB.                                      Energy consumption                      2017
                                                                 Hardware:2 Intel Xeon, E5–2430 processors with 8         Control paths delay
                                                                 cores, equipped with 16 GB of RAM                        controller load
                                                                 Topology: Abilene, Janos-us, Pioro, Zib
  ​                     [27]   MOGA (MHNSGA-2, PSO)              Topology: Internet topology zoo                          load balancing                          2021
                                                                                                                          delay



controllers, and the estimated assignment cost on failure with minimal                networking (SDSN) to enhance performance using dynamic CPP (DCPP)
possible cost. In their technical paper, the Robust Dynamic                           and static CPP (SCPP) techniques [16]. Both techniques utilize a
Multi-Controller Placement (RDMCP) framework was integrated with a                    multi-objective optimization model solved by the APSO heuristic algo­
metaheuristic Particle Swarm Optimization (PSO) algorithm to develop                  rithm, with improved performance metrics compared to traditional
a hybrid approach that improves the optimization of the objective                     SDSN, SCPP having lower economic costs, and DCPP offering greater
function in terms of cost and reliability. The efficiency of their hybrid             flexibility.
RDMCP-PSO was then benchmarked through comprehensive experi­                              Ateya et al. introduced a latency- and cost-conscious CPP scheme for
mental evaluations and compared with a variety of benchmark algo­                     large-scale SDN networks based on a dynamic optimization algorithm
rithms. Results indicate that the proposed hybrid method significantly                for the efficient use of controllers and distributed switches [17]. Their
enhances network performance with several controllers employed and                    SSOA, improved with chaotic maps, showed enhanced performance in
balancing loads compared to the standalone heuristic CPP algorithm.                   minimization of latency and deployment cost in simulations over game
    Wu et al. suggested a CPP model for software-defined satellite                    theory and meta-heuristic methods, yielding spectacular reliability


                                                                                  4
F. Dashti et al.                                                                                                  Computer Standards & Interfaces 97 (2026) 104109


enhancement and computation efficiency.                                          (MHNSGA-II) to optimize resource allocation for large networks [25].
    Liao and Leung suggested a multi-objective genetic algorithm                 The extension constitutes a multi-objective optimization model capable
(MOGA) with a PSO-based mutation operator, which was found to be of              of supporting measures of latency as well as load balancing using op­
improved performance in accuracy and convergence rate for single and             erations such as greedy initialization and an efficient Pareto finder to
multiple objectives in large search spaces [18].                                 establish Pareto optimal solutions. Hu et al. employ a binary integer
    Sahoo et al. proposed a Controller Placement Problem (CPP) strategy          programming model blended with the genetic controller placement al­
to minimize single link failures and at the same time minimize                   gorithm (IGCPA) to reduce the energy consumption of network control
communication delays to enhance the performance of SDN [19]. They                traffic with solutions very similar to the BIP with little extra link utili­
considered CPP as a multi-objective combinatorial optimization prob­             zation [26].
lem and utilized population-based meta-heuristics such as Particle                   Huang et al. presented an extensive framework for tackling the
Swarm Optimization (PSO) and Firefly Algorithm (FFA), which yielded              controller placement and scheduling problem (CPSP), amalgamating
optimal controller placement settings and switch-controller mappings.            both challenges into a constrained optimization model that improves the
Simulations discovered notable average delay reductions under situa­             efficiency of the control plane while reducing network latency [27]. The
tions of single link failures.                                                   proficient resolution of the controller scheduling problem (CPS) was
    In a recent paper, Zhang et al. proposed a complete CPP framework            accomplished via a gradient-descent-based scheduling algorithm, which
designed to optimize controller placement, switch-controller assign­             optimized the distribution of requests among controllers, while the
ment, and the management ability of each controller in large-scale SDNs          controller placement problem (CPP) was addressed using a
[20]. They formulated a multi-objective optimization controller place­           Clustering-based Genetic Algorithm with Cooperative Clusters
ment (MOCP) problem and solved it using a bio-inspired heuristic al­             (CGA-CC) to strategically position controllers within non-overlapping
gorithm known as Adaptive Bacterial Foraging Optimization (ABFO)                 sub-networks and minimize the search space. Simulations illustrated
that employed modified computational rules and chemotaxis. Perfor­               that the GD-based algorithm resulted in a 37.5 % decrease in response
mance assessment in three network topologies revealed that ABFO                  time, while CGA-CC outperformed existing algorithms in expansive
greatly improved reliability, reduced latency, and enhanced load                 environments concerning computational costs, response times, and
balancing in comparison to the Particle Swarm Algorithm (PSA), with              control plane efficacy.
significant measures indicating higher reliability, load distribution, and
reduced latency.                                                                 3. Proposed method
    In a simultaneous study, Singh et al. proposed a heuristic CPP
method for WAN based on SDN with significantly low average latency                   To solve the problem of controller placement in the SDN network, we
between controllers and switches thus improving overall SDN perfor­              used a hybrid method in which we first examined the network nodes for
mance [21]. They proposed a novel optimization algorithm,                        selection as controller locations and attempted to sort the nodes based
Varna-Based Optimization (VBO), which best characterizes switches and            on their performance in the role of the network controller. Considering
controllers as dynamic particles, and proved its superiority in simula­          that various factors contribute to performance calculation and the
tions on Matlab over famous algorithms such as PSO and Jaya.                     impact of each factor on the overall network performance differs, we
                                                                                 employed Data Envelopment Analysis (DEA) for this purpose. With this
2.2. GENETIC algorithm based CPP                                                 method, it is easy to select any number of controllers with the highest
                                                                                 efficiency and priority from among the network nodes. The next issue is
    The genetic algorithm evolution-based CPP planning schedules are             how to allocate switches to the selected controllers; to solve this prob­
elaborated fully in the second portion of Table 1, and this is followed by       lem, we also utilized the Green Anaconda Optimization method for the
an extensive explanation.                                                        best allocations (Fig. 3). We will discuss the details of each stage of the
    Jalili et al. also specified a Switch to Controller Assignment Process       proposed method in detail below.
that will enhance Quality of Service (QoS) in Software Defined Networks
(SDN) based on an Analytic Hierarchy Process (AHP) for multi-criteria            3.1. First stage of the proposed method (Determining controller locations)
decision-making and optimal controller assignment [22]. Besides that,
they also suggested a Genetic Algorithm for Controller Placement                    In the first stage, we use the Data Envelopment Analysis (DEA)
(GA-CP) to solve the Controller Placement Problem (CPP) with param­              method to select controller locations from among all the nodes in the
eters such as controller load, latency, hop count, and link utilization,         SDN network. The goal of using this method is to select high-
with simulations being more effective in load balancing than                     performance and optimal nodes as controllers, as the DEA method,
latency-based assignments.                                                       due to its nature and mechanism, can easily calculate the performance of
    Yuan et al. introduced the Controller Placement Problem (CPP)                each node as a controller based on the characteristics of the network and
model efficiently and effectively to solve placement and assignment is­
sues of Wide Area Networks (WAN) [23]. The model employs Genetic
and Kuhn-Munkres algorithms for optimal solutions with minimum
average propagation delay between switches and controllers. Experi­
ments on various network topologies indicated significant performance
enhancement in terms of propagation delay minimization and better
load balancing, subject to the condition that CPP solutions are contin­
gent upon other fundamental problems of the network such as resource
scheduling and security.
    Ahmadi and Khorramizade presented a multi-objective controller
placement strategy with a global strategy in an efficient evolutionary
algorithm adaptation, the Multi-Start Hybrid Non-Dominated Sorting
Genetic Algorithm (MHNSGA), and performance metrics such as SC-
latency, CC-latency, load balance, and reliability for link failure as a
part of the objective function [24].
    In subsequent research, Jalili et al. extended the MHNSGA algorithm
[24] with a heuristic approach referred to as Multi-Start Hybrid NSGA-II                        Fig. 3. High-level view of proposed DEA-GAO.

                                                                             5
F. Dashti et al.                                                                                                      Computer Standards & Interfaces 97 (2026) 104109


the parameters of each network node. Ultimately, we will have a sorted
                                                                                   maxθ,λ θ                                                                       (1)
list of nodes based on their performance, allowing us to easily select the
desired number of controllers for our network from the top of the list.            n
                                                                                   ∑
We will first provide a brief explanation of the DEA method and then                     xj λij ≤ xi0 , ∀i
discuss how to use it in determining the optimal locations for the                 j=1

controllers.
                                                                                   n
                                                                                   ∑
                                                                                         yj λrj ≤ θyr0 , ∀r
3.1.1. Data envelopment analysis (DEA) and its application to optimization         j=1
problems
    Data Envelopment Analysis (DEA) constitutes a non-parametric                   n
                                                                                   ∑
technique employing mathematical programming to systematically                           λj = 10λj ≥, ∀j
                                                                                   j=1
assess the relative effectiveness of a collection of decision-making units
that exhibit analogous inputs and outputs [10]. This article is a brief                  Where:
introduction to this technique developed first by Charnes, Cooper, and
Rhodes all in 1978, now finding wide application with trends for using it           • xij input value for node j for criteria that need to be minimized, such
growing among engineering, economics, management sciences as well                     as the average distance from other nodes or network latency
as computer networks.                                                               • yrj output value for node j for criteria that need to be maximized,
    The DEA method shows the optimal Units of Efficiency Frontier, with               such as eigenvector centrality which indicates the influence of the
all others below being individually inefficient relative points.                      node in the network
Conversely, points further from this frontier are even less efficient in            • λj weight of nodes in the convex combination to define node
their own right and so on. The two main DEA models are [10]:                          efficiency
                                                                                    • θ the efficiency level of the node under consideration
  - CCR Modal                                                                       • The first constraints ensure that the inputs do not exceed the speci­
                                                                                      fied amount for the node in question.
    Assuming that the production function has constant returns to scale             • The second constraints ensure that the outputs are at least equal to
(CRS)–meaning changes in inputs will cause proportionate changes in                   the specified amount.
outputs. The Charnes-Cooper-Rhodes model (CCR model for short) is                   • The third constraint is specific to the BCC model and makes the re­
structured based on fractional linear programming, attempting to                      turn to scale variable (VRS).
maximize weighted inputs-to-output ratios as now determined for each
decision-making unit [10].                                                             Regarding the selection of parameters for the DEA method, we used
                                                                                   graph structure centrality criteria [28] based on the SDN network
  - BCC Model                                                                      structure. These criteria rank nodes based on their importance. Below,
                                                                                   we briefly explain each of them:
    This model, developed by Banker, Charnes, and Cooper, considers
variable returns to scale (VRS), allowing decision-making units to have                  i. Degree Centrality
increasing or decreasing returns under different conditions [10]. The
BCC model models variable returns by adding a free variable to the CCR                 Degree centrality constitutes one of the most fundamental and sig­
model. One of the advantages of the DEA method is that it makes the best           nificant measures of centrality within the realm of network analysis.
decision for resource allocation and optimizing system performance                 This metric quantifies the number of direct connections associated with
solely based on its data, without the need for prior information about             a given node within the network and serves to illustrate the extent of
input-output functions. This characteristic has led us to leverage it to use       connectivity of that node to other nodes within the network [28]. Degree
it as a powerful tool in solving the problem of controller placement in            centrality for a node i in an unweighted and undirected graph is formally
SDN networks.                                                                      delineated as follows:
                                                                                   CD (i) = d(i) = |{j ∈ (i, j)|V ∈ E}|                                           (2)
3.1.2. Determining the location of controllers using DEA method
    The network performance in SDN networks is decided by optimal
                                                                                   where:
controller placement, and the Data Envelopment Analysis (DEA) method
is a multi-criteria decision-making tool that enables us to compare the
                                                                                    • CD(i) is the degree centrality value of node i.
efficiency of various locations of controllers based on various criteria.
                                                                                    • d(i) is the degree of node I (the number of edges connected to it).
Through this method, nodes that have the most suitable balance among
                                                                                    • V is the set of nodes in the network.
various parameters can be selected as controllers. The biggest advantage
                                                                                    • E is the set of edges in the network.
of DEA compared to traditional methods is that it uses more than one
effective criterion at a time instead of based on a single criterion and
                                                                                       For weighted graphs, the degree of any node is usually calculated by
defines the relative efficiency of alternatives even without any pre-
                                                                                   the sum of the weights of all edges connected to that node.
decided weights.                                                                            ∑
    In this context, each node of a network is considered a DMU unit, and          CD (i) =    Wij                                                      (3)
the BCC model of Data Envelopment Analysis (DEA) has been used to                             jϵN(i)

find the most effective controller placement in SDN networks. This
                                                                                      It is where Wij represents the weight of the edge connecting node i
model takes into account variable returns to scale, and since in an SDN
                                                                                   and node j.
network, a variation of one input parameter does not impact the output
                                                                                      Given the importance of selecting the best position for controllers in
parameter proportionally, constant returns to scale hypothesis of the
                                                                                   SDN networks, we use the degree of centrality criterion to place the
CCR model can lead to erroneous unit efficiency estimates. Here, BCC
                                                                                   controller because:
with greater flexibility and structural and parameter accounting scale
differences is determined to be the suitable model for the node efficiency
analysis and optimal controller placement. The mathematical form of
this model is presented as follows.

                                                                               6
F. Dashti et al.                                                                                                  Computer Standards & Interfaces 97 (2026) 104109


 • Minimize communication delay: Typically, nodes with a higher de­              connections come out from this node (switch), but this standard also
   gree of centrality are located in positions that afford them faster           reflects the weight of connections that the connected vertex carries [28].
   access to other nodes on the network.                                         By this, if a switch is connected to some other more important and
 • Increasing network reliability: Choosing high-degree nodes for con­           influential switches it will still have a sizeable degree of centrality. This
   trollers can help guard against network failure because those nodes           means that if a switch is connected to more important and influential
   are connected to many switches.                                               switches, it will sooner or later be more important itself. The eigenvector
 • Enhancing load system efficiency: Controllers placed on nodes with            centrality of each switch i is defined as follows:
   high-centrality degrees can bear a heavier workload this way and
                                                                                        1∑
   have more switches.                                                           xi =            xi Aij                                                       (6)
                                                                                        λ jϵN(i)
    ii. Closeness Centrality
                                                                                 where:
   Closeness centrality is a measure of how close a node is to other
nodes in the network [28]. Shorter paths mean faster access points                • xi is the eigenvector centrality value of switch i.
among nodes, which can be calculated using this index.                            • Aij is the matrix entry at (i,j) in the adjacency matrix A, showing a
   The closeness centrality of a node i is defined by the formula (4):              connection or no connection between switches i and j.
                                                                                  • N(i) is the switch connected to i.
                   1                                                              • λ is the largest eigenvalue of the matrix A.
Cc (i) = ∑                                                            (4)
               j d(i, j)
                                                                                     Some switches have high eigenvector centrality values and are
    Where:
                                                                                 connected to influential nodes. Controllers located there naturally
                                                                                 improve network performance. When you use this criterion in
 • CC(i) denotes the closeness centrality of node i.
                                                                                 conjunction with other measures of centrality such as betweenness,
 • d(i,j) indicates the shortest distance between nodes i and j.
                                                                                 closeness or node strength for finding the optimal location for control­
                                                                                 lers in a network, it is easiest and quickest because switches with high
    In a problem of controller placement for software-defined networks
                                                                                 eigenvector centrality are already among the most important nodes in
or a problem in a system with such requirements, for example, closeness
                                                                                 your graph.
centrality can serve as one of the criteria determining the optimal
location for controllers. Controllers located near areas of high closeness
                                                                                    v. Katz Centrality
are said to have quickly close communication with switches in the
network and therefore low latency values.
                                                                                     Katz Centrality is one of the metrics for measuring the importance of
                                                                                 nodes in a network that takes into account not only the number of direct
    iii. Betweenness Centrality:
                                                                                 connections (similar to degree centrality) but also indirect connections.
                                                                                 This metric evaluates the significance of a node not only by the number
    Betweenness measures how important a node is to routing infor­
                                                                                 of its immediate neighbors but also through its connections with other
mation through a network. This metric tells you how many times a node
                                                                                 pivotal nodes in the network [28].
falls on the shortest path between pairs of other nodes [28]. The
                                                                                     Katz Centrality for each node i is calculated as follows:
betweenness centrality of a node i is defined as follows:
                                                                                           ∑
            ∑ σ st (i)                                                           Ck (i) = α Aij Ck (j) + β                                             (7)
CB (i) =                                                              (5)                         j
            s∕
             =i∕
               =t
                       σst
                                                                                 where:
    Where:

                                                                                  • CK(i) is the Katz centrality value for node i.
 • CB(i) is identified as the betweenness centrality attributed to node i.
                                                                                  • Aij is the element (i,j) in the adjacency matrix A (1 if there is an edge
 • σst denotes the total count of shortest paths connecting nodes s and t.
                                                                                    and 0 otherwise).
 • σst(i) represents the number of shortest paths between nodes s and t
                                                                                  • α is a decay factor for the effect of longer paths (usually chosen to be
   that traverse through node i.
                                                                                    less than the largest eigenvalue of matrix A to ensure convergence).
                                                                                  • β is a positive constant that ensures even nodes without direct con­
    Nodes with higher values of CB(i) are key points for network infor­
                                                                                    nections have a positive centrality value.
mation transfer and they play a significant role in communication. Since
a large amount of information flow passes through them, their posi­
                                                                                    This formula indicates that Katz’s centrality considers not only the
tioning in a network can be critical for network efficiency. Nodes with
                                                                                 number of direct connections of a node but also accounts for indirect
high betweenness can form bottlenecks in a network, with serious
                                                                                 paths with a decay factor.
consequences for efficiency.
                                                                                    In the problem of controller placement in SDN, Katz centrality can
    In a problem of controller placement for software-defined networks
                                                                                 help select nodes that not only have direct connections with other
or a problem in a system with such requirements, nodes with higher
                                                                                 important nodes but also lie on key information paths. This feature al­
betweenness values are assumed to be the best places for locating new
                                                                                 lows controllers chosen based on this metric to have a high capability in
controllers. These nodes act like hubs of data flow within the network.
                                                                                 managing network traffic and reducing latency.
Placing a controller in such nodes would, it is hoped, speed up com­
munications and lead to improved network efficiency.
                                                                                    vi. Information Centrality
    iv. Eigenvector Center Location
                                                                                    Information Centrality is a measure of a node’s importance in a
                                                                                 network. It involves examining not just the length of this node and that
   In the problem of controller placement in SDN networks, eigenvector
                                                                                 path, but all ways that information can be transferred through paths
centrality can be used as one of the criteria for determining the optimal
                                                                                 connecting these two nodes as well. The assessment of the significance of
location of controllers. It not only takes into account how many direct
                                                                                 a node within the network takes into account its role as a transfer point

                                                                             7
F. Dashti et al.                                                                                                   Computer Standards & Interfaces 97 (2026) 104109


for information between all other nodes in the net [28].                          3.2. Second stage of the proposed method (Allocating switches to selected
     The information centrality of node i is calculated using the average         controllers)
transmission capability of information along paths between that node
and other nodes. This metric is based on the inverse of the sum of path               In the first stage of the proposed method, among all the network
lengths between nodes, so nodes that have shorter paths to other nodes            nodes, considering the graph centrality criteria, the efficiency values of
will be valued higher.                                                            the nodes were calculated as candidates for the controller locations
     Information Centrality of node i in general can be calculated through        using the DEA method, and they were sorted based on this. Now, by
the following formula:                                                            determining the number of controllers, we can easily select the nodes
          ∑                                                                       with high-efficiency values from the top of the sorted list for the final
CI (i) =     dij                                                        (8)       locations of the controllers. In the second stage of the proposed method,
            i∕
             =j
                                                                                  it is necessary to identify the switches that each controller will be
where:                                                                            responsible for managing. To this end, we use the Green Anaconda
                                                                                  optimization method to allocate each switch from among the selected
 • dij is the length of the path between nodes i and j.                           controllers to a controller in such a way that the overall productivity of
                                                                                  the network reaches its highest level. In the following, while explaining
    You can see from this formula that how close node i is to other nodes         the details of the GAO method, we will elaborate on the micro-
in the net will greatly influence its information centrality.                     operations of allocating switches to controllers.
    In SDN networks, information centrality can be used to determine                  In the year 2023, a novel metaheuristic algorithm known as Green
the best place for a controller. High-value nodes tend to be found in             Anaconda Optimization (GAO) was promulgated within the domain of
positions that allow for quick access to various parts of the net more            optimization problem resolution, emulating the innate behavioral pat­
efficiently spreading information around. This reduces latency that oc­           terns of the green anaconda and has recently attracted considerable
curs when data is transferred between switches and controllers and                scholarly interest across a multitude of issues, particularly in the context
therefore improves network performance generally.                                 of complex and nonlinear challenges [29]. This algorithm exhibits sig­
                                                                                  nificant efficacy attributed to its innovative designs and distinctive
    vii. Influence Potential Centrality                                           characteristics in addressing optimization problems laden with
                                                                                  numerous obstacles, particularly in intricate scenarios such as the allo­
     Influence Potential Centrality is a network analysis metric that             cation of controllers in Software-Defined Networking (SDN). The foun­
measures the ability of a node to influence other nodes. This metric takes        dational inspiration for GAO is derived from the mechanisms employed
into account the assumption that nodes located in strategic locations of          by male green anacondas for locating material species during the
the network would be capable of disseminating their information, ideas,           reproductive season, in conjunction with the predation strategies
or influence more quickly and widely in the network [28]. Influence               exhibited by the species [29]. The mathematical formalization of GAO is
Potential Centrality is typically derived from such measures as Eigen­            predicated on the simulation of these dual strategies of the green
vector Centrality or Influence Spread Models. It is quantified based on           anaconda, delineated into two phases: exploration and exploitation. The
the quantity and quality of the connections that a node has in the                empirical results presented by the authors of the manuscript indicate
network.                                                                          that the GAO methodology possesses substantial capabilities in explo­
     One of the popular estimation approaches to Influence Potential              ration, exploitation, and establishing a harmonious balance between
Centrality applies the following function:                                        optimization algorithms, outperforming its competing counterparts. In
           ∑                                                                      light of the aforementioned considerations, it was resolved that after the
CIP (i) =     wij CIP (j)                                              (9)        identification of controller locations via the Data Envelopment Analysis
            jϵN(i)                                                                (DEA) method, the second phase would incorporate the green anaconda
                                                                                  methodology, augmented by specific modifications for the allocation of
where:                                                                            switches to controllers. A comprehensive examination of the particulars
                                                                                  associated with this phase will be conducted.
 • CIP(i) is the value of Influence Potential Centrality for node i.
 • N(i) is the set of neighboring nodes of node i.                                3.2.1. Initial setup of the algorithm
 • wij is a weight that represents the degree of influence of node i on              The Green Anaconda algorithm constitutes a population-based
   node j.                                                                        metaheuristic approach wherein the green anacondas function as the
                                                                                  constituents of the population. From a mathematical perspective, each
    In SDN networks, the controller’s optimal placement is very impor­            green anaconda is conceptualized as a candidate solution to the opti­
tant as it has a direct impact on network performance. With the Influ­            mization problem, with its location within the search space representing
ence Potential Centrality, the nodes that play the most significant role in       the values of the decision variables [29]. Consequently, each green
information spreading in the network can be found.                                anaconda can be effectively represented through an array (Fig. 4). Given
    This characteristic has the potential to minimize communication               that in the initial stage of the proposed methodology, employing the
latency between switches and controllers and enhance network perfor­              DEA algorithm, out of n nodes within the network, b nodes have been
mance by choosing nodes that spread information extensively and                   designated as controller locations, it follows that there are n-b = a nodes
rapidly, and also by optimizing control coverage and minimizing failure           necessitating assignment to a controller. Hence, the dimensionality of
points in the network. Therefore, using Influence Potential Centrality in         the array symbolizing the anaconda must correspond to the total num­
controller placement can lead to the design of a more efficient SDN               ber of network switches, denoted as a, where each element of the array
network.                                                                          signifies a switch within the network, with its content encapsulating the
    After sorting the network nodes based on their efficiency as con­             index of the controller allocated to the respective switch (Fig. 4).
trollers using the Data Envelopment Analysis method, it is time to                Furthermore, the collective population of green anacondas, composed of
allocate switches to the top k controllers on the sorted list, which we
accomplished using the Green Anaconda optimization method, and we
will explain this in detail in the next section.


                                                                                                Fig. 4. Each population member (an anaconda).

                                                                              8
F. Dashti et al.                                                                                                     Computer Standards & Interfaces 97 (2026) 104109


these arrays, can be systematically represented through a matrix by as               network, we first determine the load parameter of each link by calcu­
follows.                                                                             lating the number of packets that pass through the link per unit of time,
                                                                                     and with the help of the formula (14), the load balancing of the links is
                                                                                     obtained.
                                                                          (10)              √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
                                                                                                z (                                             )
                                                                                               ∑        1
                                                                                     LBL =                  ∗ (Loadl − Load)2                             (14)
                                                                                             2


                                                                                               l=1
                                                                                                         z
    Where:
                                                                                     where:
 • S: Set of switches S={S1,S2,…,Sa}
                                                                                      • LBL: Link Load Balancing in the Network
 • C: Set of controllers C={C1,C2,…,Cb}
                                                                                      • LoadL: A load of each link (the number of packets transmitted over
 • a: Number of switches
                                                                                        the link per unit of time).
 • b: Number of controllers
 • I: a member of the set {1,2,…,b}                                                   • load: Average load of Controllers
 • N: Number of anacondas                                                             • z: Number of links
 • Ai: An anaconda (a member of the population)
 • A: Anacondas (population members) A={A1,A2,…,AN}                                      Finally, the load balancing parameter is obtained by averaging the
                                                                                     link load balancing and the controller load balancing according to for­
   At the initiation of the algorithm’s execution, it is imperative to               mula (15). In this paper, the impact of two factors, link load balancing
randomly ascertain the initial position of each green anaconda within                and controller load balancing, on network efficiency is considered equal.
the search space. Since the content of each anaconda’s variables de­                 If one of these two factors takes precedence over the other, their impact
termines its position in the search space, we must randomly assign each              on the final load balancing calculation can easily be varied by applying
switch to a controller, and we do this using the formula (11).                       coefficients.

Ai[t] = 1 + ri[t] ∗ (b − 1)                                               (11)       Load Balance = (LBL + LBc )/2                                             (15)

where:                                                                                - Delay

 • i = 1,2,…,N                                                                          To calculate the delay between the controller and the switch, we
 • t = 1,2,…,a                                                                       employ Dijkstra’s algorithm to determine the shortest path connecting
                                                                                     them, and then we use the delay of this path as the minimum delay
    Based on the proposed values associated with each green anaconda                 between the controller and the switch in the objective function.
for the decision variables, the objective function of the problem becomes
evaluable. We defined the objective function as a combination of three                - Reliability
parameters: the delay between the controller and the switch, the
bandwidth, and the reliability of the path, with coefficients α, β, and γ.              To calculate the reliability parameter, we also utilize formula (16),
Fitness(Ai ) = Average [α ∗ Load Balance + β ∗ Delay + γ ∗ Reliability]              such that by multiplying the reliability of each link in the path between
                                                                          (12)       the controller and the switch and then averaging the reliability of all
                                                                                     paths, the value of this parameter is obtained.
                                                                                                   ∑P (                                 )
                                                                                                    p=1 R pathp IR pathp = Πl=1 R lonk
                                                                                                                              L
  - Load Balancing
                                                                                     Reliability =                                                         (16)
                                                                                                                    P
   The load balancing parameter pertains to the equilibrium of loads
across controllers as well as the load distribution among network links.             where:
As you know, controllers in the SDN network are responsible for man­
aging and processing requests from switches. To examine the load                      • P: Number of paths between the controller and the switch
balancing of controllers, we first need to calculate the load on each                 • L: Number of links in the path
controller, namely the number of requests or the volume of traffic pro­               • R_path: Path reliability
cessed by the controller, and then obtain the load balancing of the                   • R_link: Link reliability
controllers using the formula (13). It is evident that in this formula, the
value of LBc close to zero indicates a uniform distribution of the load                  Min-max normalization is used in this paper to scale the values of
among the controllers.                                                               input variables with different ranges, i.e., delay, load balancing, and
      √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅                      reliability, to a uniform range [0,1]. This is because scaling these values
      √ b (                                                )                         to a common range allows each variable an equal contribution to the
      √2 ∑        1
LBc = √               ∗ (Loadc − Load)2                               (13)           overall fitness function. The general formula for Min-Max normalization
         c=1
                  b
                                                                                     is given as follows [30]:

where:                                                                                          x − xmin
                                                                                     xnorm =                                                                   (17)
                                                                                               xmax − xmin
 • LBc: Load balancing of controllers in the network                                    Where:
 • Loadc: A load of each controller (number of requests or volume of
   traffic processed).                                                                • x represents the original value of the variable
 • load: Average load of Controllers                                                  • xmin and xmax denote the minimum and maximum values of the
 • b: Number of controllers                                                             variable
                                                                                      • xnorm signifies the normalized value within the interval [0,1]
    Now, if we want to examine the load balancing of links in the


                                                                                 9
F. Dashti et al.                                                                                                   Computer Standards & Interfaces 97 (2026) 104109


    For variables where lower values signify superior performance (such            problem-solving domain to circumvent entrapment within local optima
as delay), an inverse normalization approach is implemented [30]:                  [29]. In the context of the controller placement problem in
           xmax − x                                                                Software-Defined Networking (SDN) architectures, each green anaconda
xnorm =                                                               (18)         delineates the status of switch allocation to controllers, indicating that
          xmax − xmin
                                                                                   the implementation of each designated anaconda as an array results in
   This guarantees that elevated normalized values consistently reflect            each cell corresponding to a network switch housing the identifier of the
enhanced fitness contributions across all evaluated criteria.                      controller to which the switch is assigned. In essence, each anaconda
   Now, after fully explaining how to calculate the values of the                  symbolizes an articulated scenario for the configuration of the SDN
objective function, this set of values can be represented mathematically           network concerning the clustering of network switches and their allo­
using a matrix according to formula (19).                                          cation to controllers. Conversely, the value of the objective function
 ⎡
   Fitness1
            ⎤         ⎡
                        Fitness(A1 )
                                     ⎤                                             associated with each anaconda signifies the efficacy of the scenario it
 ⎢     ⋮    ⎥         ⎢       ⋮      ⎥                                             proposes. Within this phase, for each anaconda, alternative anacondas
 ⎢          ⎥         ⎢              ⎥
F⎢                                                                                 that exhibit superior objective function values—thereby presenting
 ⎢ Fitness2 ⎥N × 1 = ⎢ Fitness(A2 ) ⎥N × 1                         (19)
            ⎥         ⎢              ⎥
 ⎣     ⋮    ⎦         ⎣       ⋮      ⎦                                             more favorable network scenarios—are selected as female candidates. In
   Fitnessn             Fitness(AN )                                               these instances, the allocation of switches to controllers is orchestrated
                                                                                   in such a manner that the overall proposed scenario yields a more ad­
where:                                                                             vantageous outcome, thereby aiming to emulate enhanced scenarios to
                                                                                   realize optimal switch-to-controller assignments at the network level.
 • F: Calculated objective function vector                                         After the identification of the female candidate anacondas, one candi­
 • Fitness (Ai): The calculated objective function for the ith green               date is selected at random to facilitate the movement of the anaconda
   anaconda                                                                        toward it. The movement of the anaconda towards the identified female
 • Selection of coefficients for the objective function parameters                 anaconda in our framework signifies the emulation of the assignments
                                                                                   proposed by the female anaconda, contingent upon the enhancement of
    We used sensitivity analysis to evaluate the impact of the coefficients        the current state of the existing anaconda. The specifics of this proce­
on the parameters of the objective function on the final results of the            dural methodology will be expounded upon subsequently.
model. This method helps us find the optimal values for these co­                      To mathematically represent this process, the design of the GAO
efficients and understand which parameter (delay, load balancing, and              assumes that for each green anaconda, the members of the GAO popu­
reliability) has the most significant impact on the output, and assigning          lation that possess superior objective function values are designated as
a higher coefficient to that parameter leads to a more desirable result in         the female counterparts of the green anaconda. In other words, among
the model output. For this purpose, in the simulation phase, we first              all conceivable scenarios for the SDN network within the population
selected initial values for the coefficients. In each execution cycle, we          space, the scenarios wherein the set of female candidate anacondas for
kept one coefficient constant and ran our algorithm multiple times for             each green anaconda is delineated are determined according to the
different values of the other two coefficients, recording the output re­           formula (20).
sults obtained each time, ultimately, by comparing the results, we sta­                       {                                 }
bilized the best values as the final values of the coefficients.                   CA(Az ) = Ay : Fitnessy > Fitnessz and y ∕
                                                                                                                            =z                           (20)
    To ascertain the optimal placement of controllers within the                      Where:
Software-Defined Networking (SDN) framework, the individual corre­
sponding to the highest objective value is designated as the optimal                • y, z ∈ {1,2,…,N}
member (ideal candidate solution) through a comparative analysis of                 • CA(Az): the collection of female anacondas that qualify as candidates
the computed values for the objective function. Given that in each                    for the z-th anaconda.
iteration of the Green Anaconda Optimization (GAO) algorithm, the                   • y: the index of the green anaconda within the GAO population matrix
positions of the green anacondas, along with the corresponding objec­                 and the corresponding element’s index in the objective function
tive function values, are recalibrated, the optimal member must be also               vector, which exhibits a superior objective function value relative to
revised accordingly.                                                                  the z-th green anaconda.

3.2.2. Algorithmic mathematical model                                                  In the natural environment, the secretion of pheromones by female
    In the context of the GAO design, the positioning of the green ana­            anacondas plays a crucial role in attracting male counterparts, thereby
condas within the search domain is modified through the simulation of              influencing their movement towards the source [29]. To replicate the
their behavior in two distinct phases aimed at both exploration and                pheromone concentration, the objective function values, specifically the
exploitation during the search procedure [29]. We will also try to                 fitness values associated with each anaconda’s proposed solution for the
develop a simplifying mathematical model here that will facilitate the             problem of switch allocation to controllers, are utilized. In essence, the
assignment of switches to controllers in both steps of the GAO algorithm.          more efficient the process of switch assignment to controllers is
                                                                                   perceived to be for an anaconda, from the perspective of selective
3.2.3. Exploration phase                                                           criteria, the greater the likelihood of that anaconda being recognized as
   The process of discovering the Green Anaconda Optimization                      a dominant female. Thus, within the operational framework of the
method is based on the behavior of female green anacondas during the               method, a member’s superior objective function value correlates posi­
reproductive process in nature [29]. In this aspect, in the same way as            tively with its likelihood of being selected by the green anaconda. The
the natural behavior of the female green anacondas, e.g., pheromonal               pheromone concentration probability function for the female species
secretion to attract male counterparts, the processes of this phase are            associated with each member of GAO is determined via formula (21).
performed [29]. At the initial phase of the GAO algorithm, which is the                                               (    )         (       )
exploration stage, the position coordinates of green anacondas are                                (    )       Fitness Au,v − Fitness Au,max
                                                                                   Probabilityp.c. Au,v = ∑nu {         (    )         (       )}       (21)
modified through strategies used by male animals in the search for fe­                                       k=1 Fitness Au,k − Fitness Au,max
male mates and approaching them at the time of mating. This strategic
                                                                                      Where:
approach engenders substantial alterations in the positional coordinates
of green anacondas within the search space, thus underscoring the
                                                                                    • u ∈ {1,2,…,N}
GAO’s capacity for global exploration and precise examination of the

                                                                              10
F. Dashti et al.                                                                                                           Computer Standards & Interfaces 97 (2026) 104109


 • v ∈ {1,2,…,nv}                                                                     • Fitness(Au-new): the value of the objective function for the newly
 • nu: the total count of female anaconda candidates for the u-th                       assigned value of the u-th anaconda.
   anaconda.
 • Probabilityp.c. (Au,v): the likelihood of pheromone concentration of
   the v-th female anaconda relative to the u-th anaconda.                           3.2.3.1. Exploitation phase. The subsequent phase of the GAO algorithm
 • Fitness(Au): an array encompassing the objective function values of               emulates the instinctual hunting behavior of the anaconda species.
   the female anaconda candidates pertinent to the u-th anaconda.                    Consequently, the positions of the individuals within the population are
 • Fitness(Au,v): the objective function value corresponding to the y-th             recalibrated by the predatory strategies exhibited by the green anaconda
   female anaconda candidate for the u-th anaconda.                                  [29]. This methodological approach results in marginal adjustments in
 • Fitness(Au,max): the maximum objective function value among the                   the spatial positions of the green anacondas within the search domain,
   female anaconda candidates pertinent to the u-th anaconda.                        thereby demonstrating GAO’s proficiency in capitalizing on localized
                                                                                     searches for superior solutions in proximity to the identified solutions
    In the GAO framework, it is posited that the green anaconda engages              [29]. In the context of switch allocation to SDN network controllers,
in a random selection of one candidate female anaconda and proceeds                  each anaconda epitomizes a specific networking scenario. During this
towards it. To emulate this selection mechanism, the cumulative prob­                phase, we implement a slight modification to the proposed scenario,
ability function of the candidate female anacondas is initially computed             which entails the reassignment of switches for each controller, with the
through formula (22).                                                                stipulation of enhancing the overall network performance in terms of
                        (   )                 (    )                                 latency metrics, load distribution, and reliability.
Cumulative Probability Au,v = Probabilityp.c. Au,v
                                                         (       )                       To emulate the predatory strategy and the spatial realignment of the
                                + Cumulative Probability Au,v− 1    (22)             population members concerning the prey that has come within their
                      (    )                                                         vicinity, an initial random position is generated proximate to each green
Cumulative Probability Au,1 = 0                                                      anaconda utilizing the formula (26). It is imperative to note that in this
                                                                                     formula, our objective is to revise the value of each element within the
    Where:
                                                                                     anaconda array, which entails altering the controller designation such
                                                                                     that the relevant switch in that specific element is allocated to the new
 • u ϵ {1,2,…,N}
                                                                                     controller. Subsequently, by formula (25), if the objective function value
 • v ϵ {2,…,nv}
                                                                                     at this revised position exhibits an improvement, the modification of the
 • nv: the quantity of candidate female anacondas corresponding to the
                                                                                     green anaconda’s position is regarded as valid.
   u-th anaconda.
                                                                                                                                b− 1
    Subsequently, by comparing the cumulative probability function                   Au− new [t] = Au [t] + (1 − 2∗ ru [t]) ∗                                        (26)
                                                                                                                                  x
values against a random variable (r) distributed normally within the
                                                                                         Where:
interval [0,n], where n denotes the highest cumulative probability value
of the female anacondas, the chosen female species for the green
                                                                                      • u ϵ {1,2,…,N}
anaconda is ascertained following formula (23).
                                                                                      • t ϵ {1,2,…,a}
                                         (       )
SA(u) = CA(u, v) : Cumulative Probability Au,v− 1 < r                                 • x: algorithm iteration counter and xϵ {1,2,…,T}
                               (    )                                                 • N: total count of anacondas
      < Cumulative Probability Au,v                                 (23)
                                                                                      • a: total count of switches
    Where:                                                                            • b: total count of controllers
                                                                                      • T: maximum allowable iterations of the algorithm
 • SA(u): the identified female species for the u-th anaconda.                        • Au-new: updated value for the u-th anaconda
                                                                                      • Au-new[t]: newly assigned controller value corresponding to the t-th
    Upon the identification of the selected female anaconda and the                     switch in the u-th anaconda.
simulation of the green anaconda’s movement in its direction, a sto­                  • ru[t]: stochastic number within the range [0,1]
chastic position within the search domain for the green anaconda is
computed utilizing the formula (24). If the objective function value at              3.3. Overall algorithm framework
this newly calculated position demonstrates an improvement, as indi­
cated by formula (25), the position of the corresponding green anaconda                  In the first stage of the algorithm, we receive information about an
is revised to this new location; otherwise, it retains its previous position.        SDN network, including the number of nodes, the number of links, the
                                                                                     delay of each link, the reliability of each link, the number of controllers,
Au− new [t] = Au [t] + ru [t] ∗ (SAu [t] −   Iu [t]∗ Au [t])            (24)
                                                                                     the capacity of the controllers, and the location of each controller, which
       {
           Au− new if Fitness(Au− new) > Fitness(Au )                                is obtained in the first stage of the proposed method with the help of the
Au =                                                                    (25)         DEA algorithm, as inputs to the problem. We then specify the population
                           Au , else
                                                                                     size and the number of iterations for the Green Anaconda optimization
    Where:                                                                           algorithm. To determine the number of iterations for the algorithm,
                                                                                     various policies can be considered: reaching a specific value of the
 • u ϵ {1,2,…,N}                                                                     objective function, a fixed number of iterations, minor changes in the
 • t ϵ {1,2,…,a}                                                                     value of the objective function for several consecutive iterations, and
 • N: the total count of anacondas                                                   execution time. After obtaining the initial information, we generate the
 • a: the total count of switches                                                    population size, and an anaconda member with a random scenario for
 • Au-new: the updated value about the u-th anaconda                                 switch allocation to the controller (assigning each switch randomly to a
 • Au-new[t]: the corresponding controller associated with the t-th                  controller).
   switch of the u-th anaconda.                                                          Given that the number of switches under the control of each
 • SAu[t]: the corresponding controller of the t-th switch within the                controller is limited and defined at the beginning of the algorithm, in
   selected female anaconda, relative to the u-th anaconda.                          every stage of the algorithm (initial population generation, exploration
 • Iu[t]: a stochastic value from the set {− 1,+1}                                   phase, or exploitation phase) where a change in the allocation of

                                                                                11
F. Dashti et al.                                                                                                                         Computer Standards & Interfaces 97 (2026) 104109


switches to controllers occurs, the capacity of each controller must be                           first, in the exploration phase (simulating the mating of the green
checked before finalizing the new scenario. If the number of assigned                             anaconda), each proposed scenario is modeled by one anaconda based
switches exceeds the capacity threshold of the controller, we must apply                          on another anaconda’s scenario with a better objective value, altering
a policy to adjust the scenario. In the proposed method, a function called                        how it allocates its switches to the controller, and if necessary, the
"restoration function" is designed, which, upon receiving a network                               scenario is repaired and then updated. Then, in the exploitation phase
scenario, discovers overloaded controllers, meaning controllers with an                           (simulating prey hunting by the anaconda), a random location close to
overflow of switches, and adjusts them. The adjustment process involves                           the member is created, and if the member’s objective function improves,
randomly selecting switches from the set of allocated switches of that                            a transition to the new location occurs, meaning slight changes are made
controller, equal to the number of excess switches, and assigning them to                         to the introduced scenario of each anaconda, provided that it improves
the nearest controller with available capacity.                                                   its objective function, and the repair function is called again to ensure
    In implementing this function, we utilized a delay matrix of di­                              compliance with the controller thresholds. Finally, at the end of the
mensions b * b (where b is the number of controllers) that contains the                           iteration, we save the information of the best anaconda obtained by
delay between each pair of controllers. Another use of this matrix is                             updating the best anaconda variable and proceed to another iteration of
during the determination of a backup controller. When the primary                                 the algorithm. At the end of the last iteration of the algorithm, the best
controller fails, all switches under its control are assigned to the nearest                      anaconda obtained is presented as the solution to the problem. Algo­
controller with available capacity using this delay matrix. After creating                        rithm 1 represents the pseudocode of the problem, and Fig. 5 displays
the initial population of the algorithm, at each iteration of the algorithm,                      the overall flowchart of the proposed method.


Algorithm 1
Pseudocode switch assignment.
  Start algorithm

  1. Input network parameters: the number of nodes, the number of links, the delay and reliability of each link, the number of controllers, the capacity of controllers and the location of
     each controller.
  2. Define the objective function.
  Fitness(Ai ) = Average [α ∗Load Balance + β ∗Delay + γ ∗Reliability]
  Load Balance = (LBL + LBc )/2
           √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
              ∑b (1                                                     )
  LBc =                          ∗ (Loadc − Load)2
            2

                   c=1 b
           √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
              ∑z ( 1                                                   )
  LBL =                          ∗ (Loadl − Load)2
            2

                   l=1 z
                    ∑P (                                                     L
                                                                                     )
                          p=1 R pathp R Ipathp = Πl=1 R link
  Reliability =
                                                           P
  Delay: Calculated based on Dijkstra ’s shortest path algorithm.
  3. Set GAO population size (The number of different scenarios of each generation of the population) and iterations.
  4. Generate the initial population (Basic random scenarios).
  Ai [t]= 1 + ri[t] * (b - 1)
  5. For x = 1 to T
  6. For i = 1 to N
  7. Exploration phase:
  8. Begin
  9. Determine the candidate female anacondas (Scenarios with better objective function (.
               {                                                                }
  CA(Az ) = Ay : Fitnessy > Fitnessz and y ∕                                  =z
  10. Calculate the probability of pheromone concentration of candidate anacondas.
                                                              (         )          (    )
                    (         )                 Fitness Au,v − Fitness Au,max
  Probabilityp.c. Au,v = ∑nu {                                     (         )        (   )}
                                             k=1 Fitness Au,k − Fitness Au,max
  11. Calculate cumulative probability of candidate anacondas.
                                       (         )                               (   )       ( )
  Cumulative Probability Au,v = Probabilityp.c. Au,v + Cumulative Probability Au,v− 1
                                       (          )
  Cumulative Probability Au,1 = 0
  12. Get selected female anaconda randomly.
  SA(u)=CA(u,v): Cumulative_Probability(Au,v-1) < r < Cumulative_Probability(Au,v)
  13. Choose a new location to move to the selected Anaconda.
  Au− new [t] = Au [t] + ru [t] ∗ (SAu [t] − Iu [t] ∗ Au [t])
  14. Calculate objective function of new location.
  15. Update ith anaconda scenario.
            A           if Fitness(Au− new) > Fitness(Au )
  Au = { u− new
                                       Au , else
  16. Repair ith anaconda scenario.
  17. End
  18. Exploitation phase:
  19. Begin
  20. Calculate the new location for ith anaconda.
                                                                  b− 1
  Au− new [t] = Au [t] + (1 − 2 ∗ ru [t]) ∗
                                                                       x
  21. Update ith anaconda scenario.
            A           if Fitness(Au− new) > Fitness(Au )
  Au = { u− new
                                       Au , else
  22. Repair ith anaconda scenario.
  23. End
  17. End.
  18. Save the best scenario
  19. End.
  20. Output the obtained best scenario.
  End algorithm


                                                                                             12
F. Dashti et al.                                                           Computer Standards & Interfaces 97 (2026) 104109




                   Fig. 5. The overall flowchart of the proposed method.




                                            13
F. Dashti et al.                                                                                                   Computer Standards & Interfaces 97 (2026) 104109


4. Performance evaluation and results                                              the GAO algorithm to allocate other nodes to the selected controllers.
                                                                                   This algorithm has two phases: exploration and exploitation, and during
   This section first describes the detailed simulation settings in Section        each phase, the population members (anacondas), each representing a
4.1 and then presents the simulation results of the proposed method in             network scenario, attempt to improve their fitness by moving toward the
Section 4.2 and finally evaluates the performance of our proposal with             optimal location in the solution space.
the previous three related methods in several metrics in Section 4.3.                  In the simulation, we monitored and tracked the number of move­
                                                                                   ments of the anacondas, each of which represents an attempt to find a
4.1. Simulation environment                                                        solution to the controller placement problem, and we obtained the re­
                                                                                   sults presented in Table 3. This table contains simulation information for
    Our simulation framework is established on a physical server (Intel            different numbers of controllers and executions, as well as for various
(R) Xeon(R) Platinum 8180, CPU 2.50 GHz, Ubuntu 20.04, 64 GB RAM).                 network topologies. In each execution, a fixed number of generations,
In addition, the Python programming language is employed to execute                equal to 150 generations, was considered, and for each network topol­
the proposed algorithms. The development environment utilized is Py­               ogy, the average number of anaconda movements was calculated and
thon 3.13, with a primary reliance on the numpy and matplotlib li­                 recorded. These movements occurred only if the new location of the
braries, alongside the ONOS controller throughout the simulation                   anaconda improved compared to its current location, meaning that
processes.                                                                         better performance was achieved in the resulting scenario. Some of the
                                                                                   resulting scenarios were repetitive, meaning they were produced in the
4.2. Topology information                                                          current generation or previous generations, while others were new so­
                                                                                   lutions for the problem of allocating switches to controllers.
    The simulation and planning model are executed and assessed across                 Fig. 6 displays the average number of new scenarios obtained during
six topologies selected from the Internet Topology Zoo [31]. These to­             various executions in the studied networks.
pologies include Aarnet, Agis, USnet, Chinanet, Bellcanada, and Iris_net               As you know, the movements of the anacondas and consequently the
(refer to Table 2). To affirm the efficacy and optimization of the pro­            generation of new network scenarios with a better objective function
posed methodology across diverse scenarios, these topologies were                  value ultimately led to the improvement of network parameters such as
selected based on their size, ranging from a limited number of nodes to            average delay, and this improvement trend is observable in Fig. 7.
those incorporating a substantial quantity of nodes.
                                                                                   4.6. Performance evaluation
4.3. Simulation parameters
                                                                                       The goal of our algorithm is to provide an optimal scenario for the
                                                                                   SDN network with the best controller placement and the best assignment
    The network parameter values are derived from [15], featuring a
                                                                                   of switches to controllers to achieve the lowest latency, the highest load
processing controller capacity of 7.8 k request packets/s. The demand
                                                                                   balance, and the highest reliability and the solution that has been
requested spans from [100 to 400] k req/s, while the access bandwidth
                                                                                   devised to failures (links and nodes) in the network is to assign the node
is set at 10 Gbps. The proposed algorithm is executed for 150 iterations,
                                                                                   to the closest controller with capacity using the Dijkstra algorithm. In
initiating with a population of 50 members, and the placement scheme
                                                                                   the simulation phase, the system components only need to be tested
yielding optimal performance is subsequently selected.
                                                                                   against failures during operation. That means the overload of frequent
                                                                                   control failures relates proportionally to the overall estimation of reli­
4.4. Simulation comparison                                                         ability. A comparison is made between the four algorithms: PSO [32],
                                                                                   RDMCP-PSO [15]f, heuristic CPP [33], and our proposed DEA-GAO, to
    The proposed algorithm underwent comparison with three distinct                show the average delay, controller number, and utilization of various
methodologies: the original Particle Swarm Optimization (PSO) algo­                networks.
rithm as referenced in [32], a hybrid RDMCP-PSO algorithm drawn from                   The reliability results highlighted two instances, namely before and
[15], and the heuristic CPP algorithm documented in the research pre­              after the controller’s failure for each algorithm (illustrated in Fig. 8).
sented in [33].                                                                    The average latencies exhibited a notable decrease as the number of
                                                                                   controllers increased. Fig. 8-a illustrates the average latency preceding
4.5. Simulation results                                                            the failure of the reliable link across various networks. The DEA-GAO
                                                                                   demonstrates the lowest average delay among these networks. It be­
    In the simulation, each network under study is defined as a graph,             comes clear that, for each network, the delay is contingent upon the
with the network nodes as graph nodes and the links between the                    number of nodes and their respective distances from the controllers.
network nodes as graph edges. Centrality measures are calculated for the           Consequently, the introduction of an additional controller may facilitate
entire graph, and based on these measures and with the help of the DEA             a reduction in delay, contingent upon the network’s scale. In summary,
method, the nodes are ranked as candidates for being controllers. For              among the six assessed topologies, the DEA-GAO methodology demon­
each execution of the second stage of the algorithm, which is the                  strates a 28.8 % enhancement relative to the Heuristic CPP approach, a
assignment of other nodes to controllers, the number of controllers is             22.2 % enhancement compared to the PSO methodology, and an 8.8 %
first determined, and the specified number of high-performance nodes               enhancement about the RDMCP-PSO approach.
obtained in the first stage are selected as controllers. Then, we utilized             Conversely, Fig. 8-b depicts the average delay after failure for the
                                                                                   same networks. Therefore, in terms of reliability, the overall delay es­
Table 2                                                                            calates across all four methodologies due to diminished resource
Network data information.                                                          scheduling efficiency. Furthermore, the latency following failure is
  Network name                        nodes                         edges          elevated as switches allocate time to reacquire access to the subsequent
                                                                                   controller. Note that our algorithm performs better than the others in
  Aarnet                              19                            24
  Agis                                25                            30             terms of latency and it has an improvement of 30.7 % over the Heuristic
  Usnet                               139                           195            CPP method, 21.5 % over the PSO method, and 2.9 % over the RDMCP-
  Chinanet                            42                            66             PSO method.
  Bellcanada                          48                            65                 The load balancing and utilization of SDN networks achieved by
  Iris_net                            102                           64
                                                                                   different algorithms are shown in Fig. 9. The utilization of the

                                                                              14
F. Dashti et al.                                                                                                      Computer Standards & Interfaces 97 (2026) 104109


Table 3
Average of anaconda moves during different phases of the GAO algorithm.
                   Exploration Phase                                                        Exploitation Phase

  # iteration      Network name                                                             Network name

                   Aarnet      Agis     Bellcanada     Chinanet      Iris_net    Usnet      Aarnet       Agis    Bellcanada      Chinanet        Iris_net     Usnet

  1–10             64          91       169            122           179         242        25           40      82              77              94           122
  11–20            62          89       166            121           175         237        23           35      73              69              83           109
  21–30            44          63       117            83            123         166        27           42      85              77              96           126
  31–40            55          79       145            105           154         208        28           44      90              81              101          132
  41–50            45          64       117            83            123         167        29           45      92              83              103          136
  51–60            33          48       87             62            92          125        18           28      57              53              65           85
  61–70            26          0        69             50            73          98         8            13      27              26              31           40
  71–80            35          50       92             68            98          132        6            10      23              24              27           35
  81–90            31          45       83             61            88          119        3            6       14              16              17           22
  91–100           22          31       58             43            62          84         4            6       13              14              15           20
  101–110          16          23       42             30            44          60         3            5       12              12              14           18
  111–120          10          14       26             19            28          37         3            5       11              10              12           16
  121–130          5           8        14             0             15          20         2            3       7               0               8            11
  131–140          6           0        0              0             0           0          4            0       0               0               0            0
  141–150          0           0        0              0             0           0          0            0       0               0               0            0




                                               Fig. 6. Average of new scenarios created per population generation.




                                       Fig. 7. The improvement trend of the average delay metric over different generations.


controller’s index in normal operation reflects the network performance              CPP, PSO, and RDMCP-PSO methodologies, respectively, across a
or efficiency of the resource scheduling for all controllers with the set of         comprehensive assessment of 6 distinct topologies.
switches and it is obtained by dividing the total request rate of each                   As a concluding evaluation, the optimal number of controllers across
controller’s switches by the total processing capacity of the controllers.           all four methodologies is compared and depicted in Fig. 10. Different
The findings indicate that the suggested approach demonstrates en­                   parameters have been assigned to the controllers for these algorithms. It
hancements of 7.3 %, 32 %, and 1.5 % in comparison to the Heuristic                  is observed that the selection of controllers varies according to the node


                                                                                15
F. Dashti et al.                                                                                                Computer Standards & Interfaces 97 (2026) 104109




                                                   Fig. 8. The average delay over various topologies.




                                                       Fig. 9. The optimal controller utilization.


density within each network, which is influenced by the precision of the        locations for their installation represent significant challenges within
load-balancing mechanism. The DEA-GAO is capable of minimizing                  the multi-controller model, collectively referred to as the Controller
deployment costs between switches and controllers while incurring a             Placement Problem (CPP). The CPP constitutes an optimization dilemma
low delay cost. It guarantees an equitable distribution of controllers          aimed at minimizing latency while achieving a balanced network
throughout the network while simultaneously optimizing resource                 configuration. The significance of CPP is accentuated when the objec­
allocation to attain a satisfactory level of reliability.                       tives include enhancing reliability and diminishing delay. In this study, a
                                                                                two-stage methodology for the strategic placement of controllers and the
5. Conclusion and future work                                                   allocation of switches in Software-Defined Networking (SDN) environ­
                                                                                ments was articulated. In the initial stage, the BCC model of Data
    Determining the optimal quantity of controllers and the ideal               Envelopment Analysis (DEA) was employed in conjunction with graph


                                                                           16
F. Dashti et al.                                                                                                                     Computer Standards & Interfaces 97 (2026) 104109




                                                                 Fig. 10. The optimal number of controllers.


centrality metrics to ascertain nodes exhibiting the highest efficiency for                    [3] M.A. Ouamri, T. Alharbi, D. Singh, Z. Sylia, A comprehensive survey on software-
                                                                                                   defined wide area network (SD-WAN): principles, opportunities and future
controller deployment. Subsequently, in the second stage, the optimal
                                                                                                   challenges, J. Supercomput. 81 (1) (2025 Jan) 1–47.
allocation of switches to controllers was executed utilizing the Green                         [4] T. Darwish, T.A. Alhaj, F.A. Elhaj, Controller placement in software defined
Anaconda Optimization algorithm. The findings indicate that the pro­                               emerging networks: a review and future directions, Telecommun. Syst. 88 (1)
posed hybrid approach (DEA-GAO) markedly enhances network per­                                     (2025 Mar) 1–33.
                                                                                               [5] R. Chaudhary, G.S. Aujla, N. Kumar, P.K. Chouhan, A comprehensive survey on
formance in terms of average latency and controller utilization                                    software-defined networking for smart communities, Int. J. Commun. Syst. 38 (1)
compared to the three preceding methodologies (The average                                         (2025 Jan 10) e5296.
improvement in the delay parameter compared to previous methods is                             [6] M. Kolli, M. Karre, A comprehensive study on optimization algorithms for
                                                                                                   controller placement in software-defined networking, in: 2025 3rd International
20 % and the average improvement in the controller utilization                                     Conference on Intelligent Data Communication Technologies and Internet of
parameter compared to previous methods is 13.6 %.).                                                Things (IDCIoT), IEEE, 2025 Feb 5, pp. 596–602.
    In prospective research endeavors, the evolution of the proposed                           [7] Q. Wu, J. Fang, J. Zeng, J. Wen, F. Luo, Monte Carlo simulation-based robust
                                                                                                   workflow scheduling for spot instances in cloud environments, Tsinghua Sci.
model will encompass dynamic network variables such as fluctuations in                             Technol. 29 (1) (2023 Aug 21) 112–126.
traffic, node failures, and load migration. Furthermore, the integration                       [8] Y. Shi, X. Hou, Z. Na, J. Zhou, N. Yu, S. Liu, L. Xin, G. Gao, Y. Liu, Bio-inspired
of the DEA approach with machine learning algorithms is anticipated to                             attachment mechanism of dynastes hercules: vertical climbing for on-orbit
                                                                                                   assembly legged robots, J. Bionic. Eng. 21 (1) (2024 Jan) 137–148.
substantially enhance decision-making precision and adaptability about                         [9] K. Li, L. Ji, S. Yang, H. Li, X. Liao, Couple-group consensus of
actual network conditions, which constitutes one of the authors’ ob­                               cooperative–competitive heterogeneous multiagent systems: a fully distributed
jectives in the continuation of their research.                                                    event-triggered and pinning control method, IEEe Trans. Cybern. 52 (6) (2020 Oct
                                                                                                   15) 4907–4915.
                                                                                              [10] A. Emrouznejad, E. Cabanda, Introduction to Data Envelopment Analysis and Its
CRediT authorship contribution statement                                                           applications. InHandbook of Research On Strategic Performance Management and
                                                                                                   Measurement Using Data Envelopment Analysis, IGI Global Scientific Publishing,
                                                                                                   2014, pp. 235–255.
    Fatemeh Dashti: Writing – review & editing, Writing – original                            [11] V.S. Thalapala, K.WOCP Guravaiah, Controller placement using whale
draft, Visualization, Validation, Software, Methodology, Investigation,                            optimization in SDN-WAN, Procedia Comput. Sci. 252 (2025 Jan 1) 883–892.
Formal analysis, Data curation, Conceptualization. Ali Ghaffari:                              [12] G.D. Singh, V. Tripathi, A. Dumka, R.S. Rathore, M. Bajaj, J. Escorcia-Gutierrez, N.
                                                                                                   O. Aljehane, V. Blazek, L. Prokop, A novel framework for capacitated SDN
Writing – review & editing, Validation, Supervision, Conceptualization.                            controller placement: balancing latency and reliability with PSO algorithm, Alex.
Nahideh Derakhshanfard: Writing – review & editing, Writing –                                      Eng. J. 87 (2024 Jan 1) 77–92.
original draft, Supervision. Shiva Taghipoureivazi: Writing – review &                        [13] C. Li, K. Jiang, Y. Luo, Dynamic placement of multiple controllers based on SDN
                                                                                                   and allocation of computational resources based on heuristic ant colony algorithm,
editing, Writing – original draft, Supervision.
                                                                                                   Knowl.-Based Syst. 241 (2022 Apr 6) 108330.
                                                                                              [14] S. Tahmasebi, N. Rasouli, A.H. Kashefi, E. Rezabeyk, H.R. Faragardi, SYNCOP: an
                                                                                                   evolutionary multi-objective placement of SDN controllers for optimizing cost and
Declaration of competing interest                                                                  network performance in WSNs, Comput. Netw. 185 (2021 Feb 11) 107727.
                                                                                              [15] A.A. Ibrahim, F. Hashim, A. Sali, N.K. Noordin, K. Navaie, S.M. Fadul, Reliability-
                                                                                                   aware swarm based multi-objective optimization for controller placement in
    The authors declare that they have no known competing financial                                distributed SDN architecture, Digit. Commun. Netw. 10 (5) (2024 Oct 1)
interests or personal relationships that could have appeared to influence                          1245–1257.
the work reported in this paper.                                                              [16] S. Wu, X. Chen, L. Yang, C. Fan, Y. Zhao, Dynamic and static controller placement
                                                                                                   in software-defined satellite networking, Acta Astronaut. 152 (2018 Nov 1) 49–58.
                                                                                              [17] A.A. Ateya, A. Muthanna, A. Vybornova, A.D. Algarni, A. Abuarqoub,
Data availability                                                                                  Y. Koucheryavy, A. Koucheryavy, Chaotic salp swarm algorithm for SDN multi-
                                                                                                   controller networks, Eng. Sci. Technol. Int. J. 22 (4) (2019 Aug 1) 1001–1012.
                                                                                              [18] L. Liao, V.C. Leung, Genetic algorithms with particle swarm optimization based
    Data will be made available on request.                                                        mutation for distributed controller placement in SDNs, in: 2017 IEEE conference on
                                                                                                   network function virtualization and software defined networks (NFV-SDN), IEEE,
                                                                                                   2017 Nov 6, pp. 1–6.
References
                                                                                              [19] K.S. Sahoo, D. Puthal, M.S. Obaidat, A. Sarkar, S.K. Mishra, B. Sahoo, On the
                                                                                                   placement of controllers in software-defined-WAN using meta-heuristic approach,
 [1] B. Sapkota, A. Ray, M.K. Yadav, B.R. Dawadi, S.R. Joshi, Machine learning-based               J. Syst. Softw. 145 (2018 Nov 1) 180–194.
     attack detection and mitigation with multi-controller placement optimization over        [20] B. Zhang, X. Wang, M. Huang, Multi-objective optimization controller placement
     SDN environment, J. Cybersecur. Priv. 5 (1) (2025 Mar 19) 10.                                 problem in internet-oriented software defined network, Comput. Commun. 123
 [2] A.M. Abdulghani, A. Abdullah, A.R. Rahiman, N.A. Hamid, B.O. Akram,                           (2018 Jun 1) 24–35.
     H. Raissouli, Navigating the Complexities of Controller Placement in SD-WANs: A
     Multi-Objective Perspective on Current Trends and Future Challenges, Computer
     Systems Science & Engineering, 2025 Jan 1, p. 49.


                                                                                         17
F. Dashti et al.                                                                                                                       Computer Standards & Interfaces 97 (2026) 104109

[21] A.K. Singh, S. Maurya, S. Srivastava, Varna-based optimization: a novel method for           [27] V. Huang, G. Chen, P. Zhang, H. Li, C. Hu, T. Pan, Q. Fu, A scalable approach to
     capacitated controller placement problem in SDN, Front. Comput. Sci. 14 (2020                     SDN control plane management: high utilization comes with low latency, IEEE
     Jun) 1–26.                                                                                        Trans. Netw. Serv. Manag. 17 (2) (2020 Feb 11) 682–695.
[22] A. Jalili, M. Keshtgari, R. Akbari, R. Javidan, Multi criteria analysis of controller        [28] N. Meshcheryakova, S. Shvydun, A comparative analysis of centrality measures in
     placement problem in software defined networks, Comput. Commun. 133 (2019                         complex networks, Autom. Remote Control 85 (8) (2024 Aug) 685–695.
     Jan 1) 115–128.                                                                              [29] M. Dehghani, P. Trojovský, O.P. Malik, Green anaconda optimization: a new bio-
[23] T. Yuan, X. Huang, M. Ma, J. Yuan, Balance-based SDN controller placement and                     inspired metaheuristic algorithm for solving optimization problems, Biomimetics.
     assignment with minimum weight matching, in: 2018 IEEE International                              8 (1) (2023 Mar 14) 121.
     Conference on Communications (ICC), IEEE, 2018 May 20, pp. 1–6.                              [30] Han Y., Jia Z., He S., Zhang Y., Wu Q. CNN+ Transformer based anomaly traffic
[24] V. Ahmadi, M. Khorramizadeh, An adaptive heuristic for multi-objective controller                 detection in uav networks for emergency rescue. arXiv prepr. arXiv:2503.20355.
     placement in software-defined networks, Comput. Electr. Eng. 66 (2018 Feb 1)                      2025 Mar 26.
     204–228.                                                                                     [31] Internet topology zoo, http://www.topology-zoo.org/.
[25] A. Jalili, M. Keshtgari, R. Akbari, Optimal controller placement in large scale              [32] L. Liao, V.C. Leung, Z. Li, H.C. Chao, Genetic algorithms with variant particle
     software defined networks based on modified NSGA-II, Appl. Intell. 48 (2018 Sep)                  swarm optimization based mutation for generic controller placement in software-
     2809–2823.                                                                                        defined networks, Symmetry 13 (7) (2021 Jun 24) 1133.
[26] Y. Hu, T. Luo, N.C. Beaulieu, C. Deng, The energy-aware controller placement                 [33] A.A. Ibrahim, F. Hashim, N.K. Noordin, A. Sali, K. Navaie, S.M. Fadul, Heuristic
     problem in software defined networks, IEEe Commun. Lett. 21 (4) (2016 Dec 28)                     resource allocation algorithm for controller placement in multi-control 5G based
     741–744.                                                                                          on SDN/NFV architecture, IEEe Access. 9 (2020 Dec 24) 2602–2617.




                                                                                             18
