Mobile Networks and Applications

https://doi.org/10.1007/s11036-023-02202-x

A Stochastic Performance Model and Mobility Analysis

in the Integrated Cloud-Fog-Edge Computing System
Yönal Kırsal1

Accepted: 5 July 2023

© The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2023

Abstract
In line with digital transformation and the rapid progress on the Internet, technology has led to new hierarchical infrastructures
for seamless connectivity. The integrated cloud-fog-edge computing system is a hierarchical infrastructure that provides better
operational connectivity for users at all network ends. Lack of mobility management, the requirement for high-speed internet
connectivity, limited bandwidth, and security issues are some of the issues that result in poor Quality of Service (QoS) in
such environments. Mobility or seamless connectivity of users is one of the main challenging factors in the performance
evaluation of an integrated cloud-fog-edge computing system. Hence, to obtain better QoS and support seamless connectivity,
new and efficient modelling approaches are necessary. Thus, a stochastic performance model is considered in this paper for
QoS analysis. The performance analysis is done to determine the operational space of the proposed model. This gives an idea
of how to design and develop more efficient architectures for such systems. The mean queue length, throughput, and mean
response time results are examined for performance evaluation. A discrete event simulation is also presented to validate the
proposed analytical model.

Keywords Mobile computing · Mobility analysis · Queuing theory · Performance analysis

1 Introduction edge devices [7]. It is seen as an evolution of cloud comput-

ing. Fog computing allows users to carry out their storage,
During the last decades, the ubiquity of high-capacity net- communication, and application services at a distributed
works, the development and design of computers and mobile level. Fog computing does not work on a cloud. It works on
equipment, as well as the adoption of virtualization, service- the network edge, so it is faster. In addition, fog computing
purpose architecture, autonomous computing and service supports Internet of Things (IoT) applications that demand
computing, have driven massive growth in an integrated real-time responses, especially when the fog computing is
cloud-fog-edge computing system [1, 2]. closer to the end-user [8]. In today’s world of cutting-edge
Cloud computing is the common name for internet- technology, IoT has been integrated into users’ lives and is
based computing services that allow subscribers to provide now monitored by this technology. Cloud storage is the back-
resources that can be used at any time and anywhere through bone of the IoT, and now users store/share a very huge amount
the internet [3]. However, there are certain issues causing of data. The future of big data proceeds to edge computing
latency, such as a lack of mobility management, the require- [9].
ment for high-speed internet connectivity, limited bandwidth, The philosophy behind cloud-fog-edge computing is to
and security issues [4–6]. Hence, fog computing is one of build better operational connectivity between the server core
the main solutions to these challenges. Fog computing is an and remote application users at both ends. The cloud-fog-
architecture that broadens services offered by the cloud to edge computing architecture supports high mobility and
has less demand for bandwidth usage since tasks are col-
lected at certain points instead of being sent directly over
B Yönal Kırsal
cloud channels. Hence, the new architecture is faster, sup-
[email protected]
ports mobility along with location awareness, and has a
1 Electrical and Electronics Engineering Department, very low delay. Thus, recently, the design and modelling of
Faculty of Engineering, European University of Lefke, the cloud-fog-edge computing architecture have seen more
Lefke TR-10 Mersin, North Cyprus

123
 Mobile Networks and Applications

attraction for improving the quality of service (QoS) [10– in real-time optimisation and decision-making processes.
14]. Benchmarking, simulations, and analytical modelling Therefore, using a two-stage open queueing tandem net-
are the three main approaches for performance evaluation work in the fog-edge subsystem can lead to more realistic
[15]. The development of benchmarks is very costly for such results due to its ability to capture certain characteristics
systems. Simulators imitate the operation of real systems. and dynamics of the proposed model. Hence, the two-stage
The IoTSim-Edge [9], FogNetSim ++ [16], PureEdgeSim open queueing tandem network is considered in this paper
[17], MobFogSim [18], Edge-Fog cloud [19], and iFogSim to obtain more realistic QoS outputs. In addition, regardless
[20] are some popularly used simulation software that can of the fact that there have been many studies on QoS analy-
manage IoT devices, fog-edge and cloud networks. How- sis, resource management, and power consumption, mobility
ever, the development and design of complex systems are issues in both edge and fog networks have not been consid-
time-consuming in simulation environments. In addition, and ered and examined in many studies.
most importantly, the simulation time is the main drawback Thus, in this paper, the analytic models and the solu-
of the simulations. The analysis and considered cases pre- tion approaches are employed to obtain better QoS and
sented in [21, 22] and [23] clearly showed that the simulation find operational space for such an environment in terms of
time is significantly high compared to the proposed ana- mobility. The cloud-fog-edge computing architecture suffers
lytical models. Hence, the analytical modelling approaches from delays and data loss caused by mobility and/or syn-
are mostly preferred for QoS analysis due to their shorter chronization. For this reason, the quality of the matching
processing times, even though the simulations are more processes on all devices as well as networks is important
accurate than the analytic models. Moreover, analytical mod- in such environments. Delivering data/information to the
elling approaches help analyse and optimise the performance, mobile user is a challenging task when the user loses the
resource allocation, and decision-making processes in inte- connection due to mobility. Hence, mobility is a key fac-
grated cloud-fog-edge computing environments. tor for seamless connectivity. The integrated cloud-fog-edge
The model presented in [24] focuses on evaluating and computing system is a complex system. It is subject to a
optimising the performance of a cloud-fog-edge comput- range of performance characteristics that can be difficult
ing infrastructure using queueing theory. However, mobility to model accurately using benchmarking and/or simulation
issues have not been considered at all. In addition, a one- approaches. Therefore, such complex architecture requires a
dimensional continuous-time Markov chain (CTMC) was stochastic performance model to accurately predict system
used for performance evaluation. Similarly, in [25] one- behaviour. On the other hand, some of the key character-
dimensional CTMC was applied for performance analysis istics of the integrated cloud-fog-edge computing system
for only cloud networks. The transitions between fog and that necessitate the use of stochastic performance models
edge networks have not been considered. On the other includes such realistic modelling, scalability, predictive and
hand, in [26] a generic CTMC model was employed for sensitivity analysis etc. Stochastic performance models can
QoS analysis to base resource availability models for fog help address these challenges by using probabilistic methods
computing environments. However, all of the studies used to model system performance. Thus, stochastic performance
one-dimensional CTMC for modelling. In fact, in real-world modelling and analysis of mobility help to explore the pos-
fog-edge-cloud systems, there is often dependency and coor- sibility of designing and developing an effective solution for
dination between the fog and edge layers. The two-stage open QoS enhancement in such complex infrastructure.
queueing tandem network captures this interaction, where This paper presents analytical models and solution
the output of the edge network serves as the input for the approaches to address the challenges mentioned earlier,
fog layer. Similarly, the output of the fog network serves specifically pertaining to performance analysis in an inte-
as the input for the cloud layer. This enables the model to grated cloud-fog-edge architecture. The aim is to get more
represent the handover/migration of tasks or data between practical QoS results by taking into account the mobility
the fog, edge, and cloud layers, taking into account any issues of the users for each network using a two-stage open
delays, queues, or dependencies that may exist in the sys- queueing tandem network in the fog-edge subsystem. The
tem. In other words, many existing analytical models in such proposed model’s primary focus and principal contribution
systems make simplified assumptions about system param- are to obtain an operational space that provides uninter-
eters, workload patterns, or resource availability that may rupted connections and ubiquitous mobility for users in such
not accurately reflect complex real-world scenarios. More environments. Therefore, the proposed model helps identify
importantly, the existing models may lack the flexibility to performance bottlenecks and optimise QoS outputs through
adapt to dynamic changes in the systems, network conditions, decision-making processes to enhance system performance.
or resource availability. This can limit their effectiveness In addition, by analysing the proposed model, cost reduction

123
Mobile Networks and Applications

is also possible. On the other hand, the proposed model presented in this paper are flexible and efficient. Thus, it can
assists in evaluating the scalability and resource utilisation be easily mapped into similar systems.
of such systems. By analysing the operational spaces, QoS To summarize, the main contributions of this paper are
outputs, and system parameters, the proposed model enables
the development of models that capture the dynamic nature • A stochastic performance model was developed for
of the system, considering real-world workload patterns and mobility analysis of the considered integrated cloud-fog-
resource characteristics. Furthermore, to demonstrate the edge computing architecture. To obtain more realistic
accuracy and efficacy of the proposed approach, the QoS results, this model used a two-stage open queueing tan-
outcomes obtained from the analytical model are compared dem network in the fog-edge subsystem;
with the results obtained through discrete event simulation • The proposed model’s operational capabilities were eval-
(DES). The DES was adopted for the scenario considered and uated through performance analysis, specifically focus-
implemented in the C++ [27]. The simulation programmes ing on its mobility-related aspects. Through the per-
developed are also validated by using well-known queuing formance evaluation, the impact of varying degrees of
theory models such as M/M/1, M/M/c, and M/M/c/L, as well mobility on system performance metrics, such as mean
as results from the literature [21–23]. The primary purpose queue length, throughput, and mean response time, was
of DES is to validate the proposed work; however, it can also quantified. This analysis provided valuable insights into
be employed to evaluate the QoS of systems, as discussed in the proposed model’s effectiveness, facilitating the opti-
this paper, as it replicates real-world scenarios instead of rely- misation of system parameters and enhancing its overall
ing on the Markov models presented herein. Consequently, performance under mobile conditions;
DES serves as a crucial reference point for comparison in • A simulation was also presented to validate the proposed
this research. In addition, most of the parameters used in analytical model and demonstrate the model’s applica-
[24] are also considered in this paper since the system is bility.
analysed under realistic conditions. However, other parame-
ters are also taken into account from the relevant literature. The remainder of this paper is organized as follows:
[21–23, 25, 28–33]. Section 2 introduces the related studies on modelling and
The proposed model considers a two-stage open queu- performance analysis in such a system. Section 3 describes
ing model for the edge-fog networks and a M/M/c queueing the proposed models and the solution approaches. Section 4
model for the cloud network. Therefore, the proposed model presents numerical results and discussions of the analysis.
is a hybrid and decentralized system. All networks in the Finally, Section 5 provides the conclusions and future work.
system have their own characteristics (service rate, number
of CPUs in each network, queue capacity, etc.) and operate
independently. However, the transition between networks is 2 Related works
possible in the proposed model. For instance, the through-
put of the fog system is the input of the cloud subsystem. In Implementation of the real-world cloud-fog-edge computing
the proposed integrated cloud-fog-edge computing system, systems is very costly to design and perform QoS analysis
no single component dominates the whole system. Instead, since many networks, IoT devices, hardware equipment, etc.
the system is designed to leverage the strengths of each are needed to incorporate [37].
component to provide a unified and optimized computing The proposed work in [38] addressed the challenge of
environment. These interactive input/output strategies and efficient virtual machine (VM) scheduling in cloud environ-
mobility analysis give the opportunity to get more realistic ments with heterogeneous resources and mobility factors.
QoS results as well as to obtain operational space for such The proposed online VM scheduling scheme (OSEC) utilised
systems. Furthermore, the exact spectral expansion solution reinforcement learning theory, dividing the scheduling pro-
approach is used together with the M/M/c queueing model cess into VM allocation and migration. Experimental results
for performance evaluation. The exact spectral expansion demonstrated that the proposed method significantly reduces
solution approach is a powerful technique used to solve energy consumption, VM execution costs, and service level
Markov models that arise in the performance and dependabil- agreement (SLA) violations compared to state-of-the-art
ity modelling of discrete event systems. It offers a rigorous algorithms. A large number of mobile and edge applica-
mathematical framework for analysing the behaviour of com- tion (APPs) recommendation frameworks for connected and
plex systems. This approach has been successfully applied autonomous vehicles (CAVs) were presented in [39] for
to a wide range of practical system models in the literature edge computing. The proposed framework utilised machine
[21, 34–36]. The analytical models and solution approaches learning techniques, including joint matrix factorization

123
 Mobile Networks and Applications

models and hidden relationship mining, to leverage neglected queuing model was proposed in [44] for cloud computing to
information and learn latent features for effective APPs rec- decrease the average waiting time and server utilization. An
ommendation. Experimental results using real-world data open-source analytical queuing model called FogQN based
demonstrate that the proposed framework outperforms other on open multi-class queuing networks was proposed in [32]
existing methods in all tested scenarios. In addition, the pro- for fog computing environments. The proposed model anal-
posed method in [40] utilised a privacy-oriented approach to ysed the cost and average response time of requests arriving
resist attacks from privacy attackers with prior knowledge of at a system. In [26] a continuous-time Markov chain (CTMC)
green communication and networking. It formulated the task model was used to develop a stochastic model for resource
offloading decision problem as a Markov decision process availability in fog computing environments. The proposed
(MDP) and proposed a deep reinforcement learning (DRL) model was validated using simulations. The results demon-
method to solve it, resulting in offloading policies that reduce strated that the proposed model minimizes service provider
privacy loss, energy consumption, and time delays. How- violations in terms of availability. However, all of the above
ever, none of the studies considered a stochastic performance studies considered only one of the networks (i.e., cloud, fog,
model to predict the integrated edge-fog-cloud computing and/or edge networks) for the QoS analysis. The transitions
system. between the networks and the mobility effects of users on
As mentioned before, queuing theory-based QoS anal- each network in an integrated environment have not been
ysis has received much attention in such environments. A considered. Thus, the two-stage open queueing tandem net-
detailed survey research study was done in [28] on the works are necessary to consider dynamic transitions in such
modelling of cloud computing using the queueing theory. an architecture.
The performance analysis of cloud networks considering In [24] a Markov queueing model was presented for per-
resource sharing of virtual machines was presented in [29]. formance evaluation and optimization in an integrated cloud-
The blocking probability, mean response time results, and a fog-edge environment. The proposed model mainly focused
numerical example were presented for performance analysis. on maximizing social welfare considering a computational
A dynamic resource management scheme was proposed in resource allocation method. A sports activity monitoring sys-
[41] to achieve energy efficiency in cloud computing. The tem was proposed in [45] using the mathematical modelling
proposed approach used queueing theory with time anal- approach in a cloud-fog-edge computing environment for
ysis, integer programming, and control theory techniques a sports person. The queueing theory approach was used
to improve energy savings. In addition, in [42] a hierarchi- to evaluate the minimal computer resources required for
cal management framework for load balancing and resource service-level contracts. The big data visualization method
allocation in cloud computing system was also proposed. was also proposed for athletic fitness tracking systems utiliz-
Another load balancing framework was proposed in [30] ing IoT. The results showed that the proposed approach can
for power management of cloud data centers. The proposed be used for reliability and cost savings. In addition, a queue-
system also used queueing theory to build the simulation ing model was proposed to obtain the performance output
model. In addition, in [25] a queueing theory based model parameters of different fog and cloud networks for an e-health
was proposed to determine the response time of the cloud net- patient monitoring system in [33]. However, mobility issues
work. Cloud networks were modelled using an open Jackson have not been considered in any of the studies at all. In [46]
network model for QoS evaluation. However, only cloud net- a user mobility prediction and decision-making architecture,
works have been considered for the QoS analysis. In addition, Mobi-IoST was proposed for the cloud-fog-edge environ-
the proposed models mainly used a one-dimensional Markov ment. A simulation was used for the analysis. The results
chain for performance evaluation. Similarly, in [31] a queu- demonstrated that the proposed work reduced the delay and
ing theory based modelling approach was presented for fog power consumption in fog computing. However, only the fog
computing networks. The proposed model also used a closed network was considered and analysed.
Jackson queueing theory to obtain performance results. In Therefore, this paper aims to investigate the performance
[10], a queueing theory-based modelling approach for an of the integrated cloud-fog-edge computing architecture,
edge computing-based system was introduced. An emer- considering mobility issues, using a two-stage open queueing
gency management support system was examined as a case tandem network. The proposed edge-fog subsystem solu-
study. In [43] an M/G/C/K queuing model was proposed to tion is two-dimensional. Hence, the exact spectral expansion
maximise profit in cloud computing. New analytical formulas solution approach is employed for the proposed subsystem.
were derived to evaluate the mean waiting time and delay of According to our knowledge, this is the only study that
the new arrivals. The results showed that the expected profit presents a detailed analysis of the operational space of such
in one unit of time for the cloud was obtained. A M/M/C environments. By analysing the operational spaces and QoS

123
Mobile Networks and Applications

outputs, it leads to the development of more accurate and effi- edge computing node is connected to a fog computing node
cient models for such architectures that convey the dynamic belonging to the fog layer. Similar to edge computing nodes,
nature of the systems. The throughput of the fog network fog computing nodes serve users/tasks coming from the avail-
is dynamically calculated considering the mobility rates in able edge computing nodes. All fog nodes are linked to cloud
the edge-fog subsystem and entered as an external arrival computing at the cloud layer. The cloud provides services to
rate to the cloud network. Thus, the operational space can be all users in the coverage area of the proposed network.
defined, which gives an insight into the idea of designing and The service nodes are typically located at the edge of the
developing such an integrated system. Hence, the proposed network, closer to the end-user devices in an edge computing
model gives a complete picture of how to obtain the most system. This allows for faster processing of data and reduced
promising QoS output results for such an architecture. latency. This is important for applications that require real-
time data processing. Therefore, the service time of the edge
nodes is considered higher than that of the fog nodes in this
3 Model description paper. In contrast, in a fog computing system, service nodes
may be located further away from the end-user devices. How-
This section presents the proposed models and the solution ever, it is still closer than in traditional cloud computing
approaches for QoS analysis in an integrated edge-fog-cloud infrastructures. In addition, edge computing is more scalable
computing system. A queuing model is used to improve the when it comes to managing large numbers of edge devices,
QoS in the proposed system when a user switches to a new while fog computing can provide more scalability when it
network or joins a network for better service due to mobility. comes to managing complex distributed applications that
In addition, this paper also presents the solution approaches span multiple levels of the network hierarchy. Therefore,
for the proposed model. Figure 1 shows the infrastructure the queue capacity of the edge network is also considered
of the integrated edge-fog-cloud computing system, which larger than that of the fog network. The main motivation
is based on the scenario described in [24]. This paper takes for this paper lies in the need to optimize the operational
into account mobility issues at each layer of the system. The spaces of the overall architecture for maximum performance
users/IoT devices connect to the edge computing nodes at and QoS, considering all network characteristics. Consider-
the edge layer for communication. The system has a number ing each subsystem alone and/or removing the relation of
of edge computing nodes that provide service to all users. nearest neighbour transitions may not have a convenient way
It is assumed that each computing node is located at the to obtain realistic QoS results for such systems. Hence, a net-
base station (BS) in the cellular network. Moreover, each work of queues can be used to obtain better QoS results rather

Fig. 1 The edge-fog-cloud network model considered

123
 Mobile Networks and Applications

than the product form solution. The novel queuing model the number of CPUs in the cloud subsystem with an infi-
considered in this paper is shown in Fig. 2 for an integrated nite queue capacity. If a task requests a service and finds all
edge-fog-cloud computing system. Such an architecture can the CPUs busy at a subsystem, it will join the queue asso-
solve similar problems due to the mobility that arises when ciated with the subsystem or migrate to available CPUs in
the resource is shared. In addition, in an integrated model a different subsystem. Subsequently, when one of the CPUs
for system performance analysis, mobility issues should becomes idle, a task from the queue will be selected for ser-
also be considered with realistic behaviour in such an vice according to first-come-first-serve (FCFS) discipline. In
environment. addition, the task may move to another subsystem for further
In other words, the two-stage open queuing model is a service after completion of service at the CPU. The notations
mathematical model used to analyze and optimize queuing used in the proposed model are given in Table 1.
systems. Thus, the proposed model in Fig. 2 is particularly
useful when there are two distinct networks in the queuing 3.1 Edge-fog subsystem
process, such as an edge network and a fog network area. The
proposed two-stage open queuing model can be divided into A two-stage open queuing tandem network is taken into
two stages: the edge network stage and the fog network stage. account for an integrated edge-fog subsystem, as shown in
Users arrive at the edge network and receive service if all Fig. 2. The system has multiple CPUs at both the fog and
nodes are available. If not, users wait in a queue until a CPU edge networks with respected service rates, μ F and μ E . Each
becomes available. In addition, mobile users can hand over CPU has an independent and identically distributed exponen-
to the fog network without getting a service and/or after get- tial service time distribution with means 1/μ E and 1/μ F for
ting a service. Please note that the proposed model assumes edge and fog networks, respectively. The arrival rates are λ E
that users arrive at the edge network according to a Poisson and λ F for the edge and the fog networks, respectively. How-
process and that the service times follow an exponential dis- ever, according to the assumptions and reference scenario in
tribution. Similarly, in the fog network stage, the users are [24] only λ E is also considered an external arrival in this
served by one or more available CPUs. Mobile users can also paper. The arrival process of the edge network is Poisson. m 1
move to the cloud network without getting a service and/or and m 2 are mobility rates at the edge and the fog networks,
after getting a service. An edge network has E number of respectively. Mobility rates in the proposed model refer to the
CPUs associated with it. Similarly, F is the number of CPUs rates at which mobile users move between different queues or
in the fog network for the edge-fog computing subsystem in service nodes within the network. In other words, these rates
the proposed model. Q E and Q F are the queue capacities are commonly referred to as transition rates, and they play a
of the edge and fog networks, respectively. Moreover, C is crucial role in determining the behavior and performance of

Fig. 2 The proposed model

considered

123
Mobile Networks and Applications

Table 1 Summary of symbols 3.1.1 The proposed Markov model

Symbol Definition
The edge-fog subsystem can be modelled as a stochastic pro-
E Number of CPUs in the edge network cess. The subsystem states are specified by pairs (s0 , s1 ),
F Number of CPUs in the fog network s0 ≥ 0, s1 ≥ 0, where si (i=0,1) is the number of tasks at
C Number of CPUs in the cloud network CPU in the steady-state.
QE Queue size of the edge network The state changes occur on the completion of service of
QF Queue size of the fog network any CPUs or on external arrivals. Hence, the transition of all
QC Queue size (∞) of the cloud network inter-arrival, service and mobility rates are shown in Fig. 3.
LE System capacity of edge network, L E = E + Q E The model presented in Fig. 3 can be explained using a two-
LF System capacity of fog network, L F = F + Q F dimensional CTMC, where the number of tasks in the edge
λE Mean arrival rate of tasks in the edge network network and the fog network are represented by x and y,
λF Mean arrival rate of tasks in the fog network respectively. The state space of this model is defined by the
λC Mean arrival rate of tasks in the cloud network set (0,1,. . .,L E ) x (0,1,. . .,L F ). Whenever a task arrives at the
μE Mean service rate of tasks in the edge network edge network and the CPU of the edge network is available,
μF Mean service rate of tasks in the fog network the state (x,y) transitions to (x+1,y) with a rate of λ E . Con-
μC Mean service rate of tasks in the cloud network versely, the state (x,y) transitions to (x,y+1) with a rate of λ F .
m1 Mobility rate in the edge network
The fog network CPUs process requests with a rate of yμ F
and as they do so, the state (x,y) transitions to (x,y-1). The
m2 Mobility rate in the fog network
service of the edge network occupied by the new request will
M Threshold value defined in the spectral expansion
method result in the transition from (x,y) to (x-1,y) with a rate of xμ E .
α Transition rate used in matrix A, α= μ E (1 − m 1 ) In addition, the state (x,y+1) transitions changes to (x+1,y)
β Transition rate used in matrix C/C j , β=μ F (1 − with a rate of (y+1)μ F m 2 and vice versa (x+1,y) transitions
m2) to (x,y+1) with a rate of (x+1)μ E m 1 due to the mobility,
Pi, j The steady-state probabilities in the edge-fog sub- when the CPUs of edge and fog networks are available. Let
system Ps0 ,s1 be the joint probability of s0 tasks at the fog layer and
M Q L E /M Q L F The mean queue length of edge and fog networks s1 tasks at the fog layer in the steady-state. The exact spectral
M Q LC The mean queue length of cloud network expansion solution approach is applied to obtain the steady-
γ E /γ F The throughput of edge and fog networks state probabilities of the proposed edge-fog subsystem. Thus,
M RTE /M RTF The mean response time of edge and fog networks the state of the proposed edge-fog subsystem at time t can be
C Mean arrival rate of tasks in the cloud network expressed as X (t) and Y (t), identifying the number of tasks
Pn The steady-state probabilities in the cloud subsys- present at a time t for the edge and the fog layers, respec-
tem tively. The proposed model consists that X(t)=0, 1, . . ., L E ,
ρ The traffic intensity in the cloud network and Y(t)=0, 1, . . ., L F , t ≥0, which represents the number
Wq The mean waiting line in the cloud network of tasks in the edge and fog layers including the one(s) in
W The mean queue length in the cloud network service, respectively.
Lq System capacity of cloud network

3.1.2 The exact spectral expansion approach

the proposed architecture. m 1 is typically expressed as tran- The exact spectral expansion technique has gained recogni-
sition probabilities in the edge network, which represent the tion as a promising approach for effectively modelling the
probability that a mobile user moves from the edge network performance and dependability of discrete event systems.
to the fog network in the next time interval. On the other It offers a solution for a specific class of Markov models
hand, a mobile user can move to a different network and/or commonly encountered in practical system models. Various
complete the service at a (1-m 1 ) rate as shown in Fig. 2. applications and outcomes have been reported, including per-
Hence, the output of the edge network, labelled μ E m 1 is the formability modelling of multiprocessors [35], networks of
input of the fog network. Similarly, at the fog network, the queues with unreliable servers [21], and other real-world sys-
users may be routed to the cloud network with a probabil- tems [36]. While initial conceptual foundations were known
ity of m 2 or they may leave the system due to the mobility earlier, a highly efficient solution algorithm for exact spectral
with a probability of (1-m 2 ). Likewise, the output of the fog expansion was introduced in [34]. In this section, we delve
network, labelled μ F m 2 is the input of the cloud system. into the mathematical principles underlying the exact spectral

123
 Mobile Networks and Applications

Fig. 3 The state transition diagram of edge-fog subsystem

expansion method and present the corresponding solution is divided into two different regions for j ≤ 1 and 1 < j ≤
algorithm. L F . When using spectral expansion solution, the service rates
The exact spectral expansion solution approach has three are often assumed to be invariant with respect to the num-
matrices as A, B and C. These matrices define the transition ber of tasks in the system. In order to determine this point
rates of the proposed model. More details of the exact spectral of independence, a threshold value M is typically employed.
expansion solution approach and the definitions, formation Hence, the M is taken as M=L F . If the number of tasks in the
etc. of the matrices can be found in [21, 34, 35]. Thus, the proposed model is less than the number of available CPUs,
matrix A shows instantaneous transition rates from the state a CPU is assigned for a task. Moreover, if the number of
(i,j) to state (s,j) with zeros on the main diagonal. Hence, the tasks is greater than the number of available CPUs, all of
matrix A/A j is given below and has purely lateral transitions. the available CPUs are assigned to incoming tasks. There-
λ E , μ E , E, Q E , and α are the parameters of matrix A/A j . fore, the downward transition rate is chosen as the minimum
On the other hand, one-step upward and downward transi- number of tasks and the number of available CPUs with the
tions are interpreted in matrices B and C. Both matrices are corresponding mobility rates. The transition matrices of the
also presented below. However, the matrix C depends on the proposed edge-fog subsystem are of size (L E +1) x (L E +1)
number of tasks for j=0, 1, . . ., L F according to the definition where α = μ E (1 − m 1 ), β = μ F (1 − m 2 ).
in the spectral expansion solution. Therefore, the matrix C j

123
Mobile Networks and Applications

⎛ ⎞
0 λE 0 0 0 0 0 0
⎜μ λE 0 ⎟
⎜ E 0 0 0 0 0 ⎟
⎜ .. ⎟
⎜ 0 2μ . 0 ⎟
⎜ E 0 0 0 0 ⎟
⎜ .. ⎟
⎜ . ⎟
⎜ 0 0 0 λE 0 0 0 ⎟
A = Aj = ⎜ ⎟ (1)
⎜ 0 0 0 Eμ E 0 λE 0 0 ⎟
⎜ ⎟
⎜ .. ⎟
⎜ 0 0 0 0 Eα + α 0 . 0 ⎟
⎜ ⎟
⎜ .. ⎟
⎝ 0 0 0 0 0 . 0 λE ⎠
0 0 0 0 0 0 Eα + Q E α 0
⎛ ⎞
λF 0 0 0 0 0 0
⎜ . ⎟
⎜ μE m1 . . 0 0 0 0 0 ⎟
⎜ ⎟
⎜ .. ⎟
⎜ 0 . λF 0 0 0 0 ⎟
⎜ ⎟
⎜
B = Bj = ⎜ 0 0 Eμ E m 1 λF 0 0 0 ⎟ (2)
⎟
⎜ .. ⎟
⎜ 0 0 0 Eμ E m 1 + μ E m 1 . 0 0 ⎟
⎜ ⎟
⎜ .. ⎟
⎝ 0 0 0 0 . λF 0 ⎠
0 0 0 0 0 (E + Q E )μ E m 1 λ F
⎛ ⎞
0 Fμ F m 2 + Q F μ F m 2 0 0
⎜0 Fβ + Q F β Fμ F m 2 + Q F μ F m 2 0 ⎟
⎜ ⎟
C=⎜ . ⎟ (3)
⎝0 0 . . Fμ F m 2 + Q F μ F m 2 ⎠
0 0 0 Fβ + Q F β

j≤1
Let’s specify certain diagonal matrices of size (L E + 1) ×
⎛ ⎞ (L E + 1) as follows:
μF μF m2 0 0
⎜ 0 2μ F 2μ F m 2 0 ⎟
⎜ ⎟ E +1
L E +1
L
Cj = ⎜ . ⎟ (4)
⎝ 0 0 . . Fμ F m 2 ⎠ D Aj (i, i) = A j (i, k); D A (i, i) = A(i, k); (7)
0 0 0 Fμ F k=0 k=0

1 < j ≤ LF

⎛ ⎞
Fμ F m 2 + μ F m 2 Fβ + β 0 0
⎜ 0 Fμ F m 2 + 2μ F m 2 Fβ + β 0 ⎟
⎜ ⎟
Cj = ⎜ .. ⎟ (5)
⎝ 0 0 . Fβ + β ⎠
0 0 0 Fμ F m 2 + Q F μ F m 2

3.1.3 The steady-state solution E +1

L E +1
L
D Bj (i, i) = B j (i, k); D B (i, i) = B(i, k); (8)
Therefore, the steady-state solution can be applied to deter- k=0 k=0
mine Pi, j . The system steady-state probabilities can be E +1
L E +1
L
written as follows: D Cj (i, i) = C j (i, k); D C (i, i) = C(i, k); (9)
k=0 k=0
P(X (t)=i,Y (t)= j)
Pi, j =lim t→∞ ; 0 ≤ i ≤ L E + 1, 0 ≤ j ≤ 1 (6)

123
 Mobile Networks and Applications

and Q 0 = B, Q 1 = A − D A − D B − D C , Q 2 = C. Hence, performance metrics can be calculated. The mean queue

the definition of all state probabilities in a row is: length (M Q L E /M Q L F ), throughput (γ E /γ F ) and mean
response time (M RTE /M RTF ) of both edge and fog net-
v j = ( p0, j , p1, j , · · · , p L E +1, j ); j = 0, 1 (10) works are calculated respectively as follows:

Now, we can formulate the steady-state balance equations


LE LF 
LF 
LE
for limited queuing systems (0 ≤ j ≤ 1) as follows: M QLE = i Pi, j and M Q L F = j Pi, j (21)
i=0 j=0 j=0 i=0
v0 [D0A + D0B ] = v0 A0 + v1 C1 (11)

LE 
LF 
LF 
LE
v j [D Aj + D Bj + D Cj ] = v j−1 B j−1 + v j A j + v j+1 C j+1 ; γE = μi Pi, j and γ F = μj Pi, j (22)
0≤ j ≤1 (12) i=0 j=0 j=0 i=0

v L [D A + D C ] = v L−1 B + v L A (13) M QLE M QLF

M RTE = and M RTF = (23)
γE γE
Following is the normalization equation:
After obtaining the throughput of the fog network, the

L  E +1
1 L
mobility rate for the cloud subsystem can be calculated cor-
vje = P(i, j) = 1.0 (14)
respondingly. The cloud subsystem is regarded as a M/M/C
j=0 j=0 i=0
queue model similar to the model presented in [24]. How-
From the equations above, the following equation can be ever, the external arrival rates of the cloud subsystem are
written: dynamically evaluated according to the mobility rates in the
edge-fog subsystem. This is the main difference in this paper
v j Q 0 + v j+1 Q 1 + v j+2 Q 2 = 0; 0 ≤ j ≤ 1 (15) and leads to obtaining more realistic results.

Additionally, the characteristic matrix polynomial Q(λ) can 3.2 Cloud subsystem
be defined as follows:
We now turn our attention to the cloud subsystem. As stated
Q λ = Q 0 + Q 1 λ + Q 2 λ2 ; Q̄ β = Q 2 + Q 1 β + Q 0 β 2 ;(16) before, according to the assumptions made in this paper as
well as the relevant literature [22, 25, 28–31] the cloud sub-
where system can be modelled as a M/M/C queue system. Arrivals
in the edge-fog subsystem are Poisson with rate C where
 Q λ = 0; |Q λ | = 0; φ Q̄ β = 0; | Q̄ β | = 0; (17)
C = λC + γ F . λC represents generated tasks in the cloud
system, whereas γ F represents migrated tasks from other sub-
λ and  are eigenvalues and left-eigenvectors of Q λ . Sim-
systems. There are a total of C CPUs available. Each CPU
ilarly, β and φ are eigenvalues and left-eigenvectors of Q̄ β .
has an independent and identically distributed exponential
Hence, φ and β are vectors specified as:
service time distribution with a mean of 1/μC . As shown in
Fig. 4, the proposed cloud subsystem could be modelled as
φ = φ0 , φ1 , . . . , φ S+1 (18)
a birth-death process.
β = β0 , β1 , . . . , β S+1 (19) The birth rate is the overall arrival rate for all n in the
cloud subsystem, regardless of the number of tasks in the
E +1
L
j+1 2− j system. However, the rate of service completion depends on
Besides that, v j = (ak k λk + bk φk (i)βk ),0≤j≤ 1
the number of tasks in the system. All CPUs are in use if
k=0
and in the state probability form, there are C or more tasks in the system. The aggregate ser-
vice completion rate for the system is CμC since each CPU
E +1
L completes tasks at a rate of μC . Additionally, the total ser-
j−M+1 2− j
pi, j = (ak k λk + bk φk (i)βk ) 0≤ j ≤1 vice compilation rate for the system is nμC when there are
k=0 fewer than C tasks in the system and only n of the C CPUs
(20) are working. Consequently, μn can be expressed as:

Where λk (k = 0, 1, . . . , L E +1) and βk (k = 0, 1, . . . , L E + ⎧

1) are L E + 1 eigenvalues. As a result, after obtaining the ⎨ nμC (0 ≤ n < C)
μn = (24)
steady-state probabilities, Pi, js, a number of steady-state ⎩
CμC (n ≥ C)

123
Mobile Networks and Applications

Fig. 4 The state transition

diagram of cloud subsystem

The steady-state probabilities Pn can be calculated using Here, ρ = r /C < 1 is a prerequisite for the existence of
the prior theory created for birth-death processes as follows: a steady-state solution.In other words, the mean arrival rate
must be less than the mean maximum possible service rate
⎧
C
n
⎪
⎪ · P0 (0 ≤ n < C), of the system. Intuitively, this is what is anticipated. Thus,
⎨ n
n!μC
the QoS output measurements for the proposed cloud system
Pn = (25)
⎪
⎪ Cn can now be derived after obtaining all steady states. Since
⎩ · P0 (n ≥ C)
n
C n−C C!μC it is easier to compute the expected queue size (Q C ) than
the mean queue length (M Q L C ), this is the first step to be
Thus, it is clearly seen that Pn has the form of a Poisson taken into account. The evaluation of the Q C can be found
random variable for 0 ≤ n < C and the form of a geometric in Eq. A2 in the Appendix; hence, Q C is
random variable for n ≥ C. The normalization condition,
which states that the probabilities must add up to 1, can be rnρ
QC = P0 (29)
utilized to determine P0 as follows: C!(1 − ρ)2

∞
−1 To find M Q L C , Little’s formula is employed to find mean

C−1
C
n  Cn
P0 = + (26) waiting time Wq then Wq is used to find mean queue time
n
n!μC n
C n−C C!μC
n=0 n=C W=Wq +1/μC . Finally Little’s formula is employed again to
calculate M Q L C = C W . Thus, the Eqs. 30, 31 and 32 can
Let r=C /μC and ρ = r /C = C /CμC . Then P0 can be be used to calculate Wq , W , and M Q L C , respectively.
rewritten as :
Lq rC
∞
−1 Wq = = P0 (30)
 rn
C−1  rn  C!(CμC )(1 − ρ)2
P0 = + (27) 1 rC
n! C n−C C! W = + P0 (31)
n=0 n=C
μC C!(CμC )(1 − ρ)2
Furthermore, the infinite series in the above equation can be rCρ
M Q LC = r + P0 (32)
simplified. The simplification can be found in the Appendix C!(1 − ρ)2
in Eq. A1. As a result, the P0 can be written as

−1
rn  rn C−1
P0 = + (ρ = r /C < 1) (28) 4 Numerical results
C!(1 − ρ) n!
n=0
This section represents the numerical outcomes of the QoS
evaluation of cloud-fog-edge computing systems consider-
Table 2 Experimental Symbol Value ing mobility issues with integrated queuing systems. The
parameters used
capabilities of the proposed analytical model and solu-
E 10
tion approach are demonstrated through the presentation
LE 70
of numerical outcomes, including the mean queue length,
QE 60
μE 10 (tasks/sec)
F 13 Table 3 System Specifications
LF 52
Operating Windows 64-bit
QF 39 System
μF 5 (task/sec) Programming Matlab/C++
C 16 Platforms
QC ∞ CPU Intel(R) Core(TM) i7-363QM CPU @2.40 GHz
μC 5 (tasks/sec) RAM 16GB

123
 Mobile Networks and Applications

Fig. 5 M Q L E results with

various edge layer mobility rates

Fig. 6 M RTE results with

various edge layer mobility rates

Fig. 7 γ E results with various

edge layer mobility rates

123
Mobile Networks and Applications

Fig. 8 M Q L F results with

various edge layer mobility rates

Fig. 9 γ F results with various

edge layer mobility rates

Fig. 10 M RTF results with

various edge layer mobility rates

123
 Mobile Networks and Applications

Table 4 M Q L F results with various fog layer mobility rates (m 2 ) when mobility rate through the fog and cloud layers, respectively.
m 1 =0,3 Unless otherwise stated, the assumptions and parameters
M QLF used in the reference model [24] and various other mod-
λE m 2 =0,2 m 2 =0,4 m 2 =0,6 m 2 =0,8 els [21–23, 25, 28–33] were also considered in the analysis.
10 0,8207 0,7597 0,7051 0,6587
Thus, Table 2 shows the experimental parameters considered
in the analysis. In addition, some parameters were changed
20 1,6446 1,5179 1,4106 1,3189
according to the analysis. For example, the probability of
30 2,4685 2,2785 2,1171 1,9792
the mobility rates m 1 and m 2 varies between 0.1 and 0.9.
40 3,286 3,0375 2,8241 2,6382
The numerical analysis in this study was carried out on a
50 4,1098 3,7983 3,529 3,2972
PC, as shown in Table 3. The proposed analytical model was
60 4,9312 4,5571 4,2353 3,9544
validated using DES to demonstrate its accuracy and efficacy
70 5,757 5,3191 4,9409 4,6183
for all figures. However, the detailed analysis is presented in
80 6,4972 6,0289 5,6459 5,2772
Section 4.1.
90 7,1972 6,5017 6,1472 5,8919 The influence of mobility on system performance is evi-
100 7,8986 6,9288 6,4278 6,1502 dent in Figs. 5, 6, 7, 8, 9, and 10, as portrayed with significant
110 8,6176 7,3467 6,6772 6,2864 clarity. Hence, Figs. 5, 6, and 7 clearly show the effect of
120 9,4133 7,7384 6,9133 6,4174 mobility on QoS in the edge layer with various edge network
mobility rates. In addition, Figs. 8, 9, and 10 also indicate the
effect of mobility in the edge layer on the performance of the
throughput, and mean response time of each layer separately. fog layer with various edge network mobility rates. Figures 5,
In order to achieve more realistic QoS results for each exper- 6 and 7 represent M Q L E , M RTE and γ E results considering
iment, this paper extends its analysis to address mobility different mobility rates in the edge layer, m 1 = 0.1, 0.3, 0.5,
concerns in edge and fog networks. Therefore, a two-stage 0.7, and 0,9, with the external arrival rate of the edge layer
open network is incorporated, ensuring a comprehensive increasing by 10 tasks/sec. The M Q L E decreases as the m 1
examination of the subject matter. Please note that the main increases, as presented in Fig. 5. Mobile users move to the fog
focus is on providing operational space to optimise and layer due to mobility. Hence, M Q L E decreases since more
enhance overall system performance under mobile condi- users move to the fog layer to get a service without interrup-
tions. The operational space is obtained and presented in tion. Similarly, latency decreases as well since fewer tasks
order to obtain more accurate and realistic QoS output mea- will be staying in the system to get service due to mobility.
surements by analysing each network’s performance outputs Figure 6 indicates this situation clearly. High mobility to the
separately. The operational spaces are analysed considering fog layer (e.g. m 1 =0.7, 0.9) leads to better response time in
different m 1 and m 2 values. The external arrivals of the fog the edge layer, as shown in Fig. 6. On the other hand, when
and the cloud layers are not considered similar to the refer- m 1 =0.1 the system has the highest delay since more tasks
ence scenario presented in [24]. m 1 =m 2 =0.1 represents the are waiting in the system to be served, especially for highly
case of low mobility, whereas m 1 =m 2 =0.9 represents the high loaded systems. For instance, the response time is around 0.9

Fig. 11 M Q L F results with

various fog layer mobility rates
(m 2 ) when m 1 =0,7

123
Mobile Networks and Applications

seconds for m 1 =0.1 when λ E =90 tasks/sec however for the the edge network is alleviated, resulting in improved overall
high mobility case, e.g. m 1 =0.9, the response time is around system performance.
0.18 seconds. On the other hand, QoS and mobility analysis of the fog
Nonetheless, for the γ E results shown in Fig. 7, there is a network from edge to fog networks are presented in Figs. 8,
decrease in throughput as mobility increases. Because more 9, and 10. Figure 8 compares M Q L F results with different
requests are migrated from the system as a result of mobility, mobility rates in the edge subsystem m 1 = 0.1, 0.3, 0.5, 0.7,
throughput decreases; however, this improves edge network and 0,9 as the edge network’s external arrival rate increases
efficiency because these tasks are handled by the fog network from 10 to 120. The throughput and M RTF results are shown
in the proposed architecture. In other words, an interesting in Figs. 9 and 10, respectively. It should be noted that Figs.
observation is the reduction in throughput as mobility levels 8, 9, and 10 consider a scenario of low mobility (m 2 =0.1)
increase. This decline can be attributed to the migration of between the fog and cloud networks. The high mobility case
a larger number of tasks from the system due to mobility. is considered in Table 4 and Figs. 11, 12, and 13. This clarifi-
Although this decrease in throughput may initially appear cation is explicitly mentioned in all figures to ensure a better
detrimental, it actually contributes to the enhancement of understanding of the experimental setup.
edge network efficiency within the proposed architecture. Due to the migration from the edge to the fog network, the
By offloading these tasks to the fog network, the burden on number of tasks and throughput increase in the fog network as

Fig. 12 γ F results with various

fog layer mobility rates (m 2 )

123
 Mobile Networks and Applications

m 1 increases. This is because more tasks migrate to the fog performance. The M Q L F results of the fog network are pre-
layer to get better service without encountering any prob- sented in Table 4 and Fig. 11 with different m 2 for m 1 =0,3
lems. This also leads to better performance on the edge and m 1 =0,7, respectively. The M Q L F increases with the
network. However, the considered fog network will be busy growth of the λ E as expected. The effect of the m 2 on system
due to the migration from the edge network. Hence, the fog performance, on the other hand, is clearly shown in Table 4
network has more tasks in the system, as shown in Fig. 8 as and Fig. 11. The higher mobility rate in fog computing low-
m 1 increases. In addition, the throughput of the fog network ers the tasks regardless of m 1 . Because of the low mobility
increases with mobility. However, as clearly seen in Fig. 9 the (e.g., m 2 =0.2), the fog network has more tasks in the sys-
throughput results remain constant after some point. Since tem. The M Q L F decreases accordingly as the m 2 increases
the service rate of the fog network could not be used to serve in both cases. However, M Q L F gets significantly affected
tasks after a certain point, the throughput remains constant when there is high mobility from the edge network, as shown
depending on the λ E . For instance, the throughput results are in Fig. 11. If the mobility rate is less in the fog network,
22.8 and 68.9 tasks/sec when λ E = 70 tasks/sec for m 1 =0.3 the system becomes fully loaded very fast due to the high
and m 1 =0.9, respectively. In addition, the M RTF increases mobility rate from the edge network. In other words, the high
significantly, especially with the high mobility rates from the mobility from the edge networks leads to more tasks in the fog
edge network. Figure 10 clearly depicts this situation. network with lower m 2 rates. For instance, M Q L F =14,98
The analysis is done in Table 4 and Figs. 11, 12 and and 2,46, respectively, for m 1 =0,3 and 0,7 when m 2 =0,2 and
13 focuses the effect of m 2 on fog and cloud network λ E =30 are considered.

Fig. 13 M Q L C results of the

cloud network with different m 2

123
Mobile Networks and Applications

The throughput results of the fog network, on the other The cloud network’s M Q L C results, on the other hand,
hand, are shown in Fig. 12a and b for m 1 =0,3 and m 1 =0,7, are also shown in Fig. 13. As stated before, the through-
respectively. The results also show the effect of mobility from put results of the fog network are the input of the cloud
both networks on system performance. The results shown in network in the proposed system to obtain realistic QoS mea-
Fig. 12 clearly demonstrate the importance of mobility anal- surements. Hence, the results obtained in Fig. 12a and b
ysis (e.g. m 1 and m 2 ) in such an architecture. Figure 12a are used as inputs in Fig. 13a and b, respectively. As the
shows the results of the relatively slow-mobile users from throughput of the fog network increases in Fig. 12a, the
the edge to the fog network, considering different mobil- M Q L C increases accordingly since the arrival rates of the
ity rates in the fog network. Although the fog system has a cloud network are dependent on the fog network’s through-
high mobility rate, it has successfully served all the tasks. put. However, the M Q L C decreases in Fig. 13b at some point
Thus, throughput increases even though the mobility rate since the arrival rate of the could network also decrease due
increases in the fog network. On the other hand, as the den- to mobility issues. The results show that the proposed model
sity of mobile users coming from the edge network increases, is capable of successfully obtaining QoS measurements for
as shown in Fig. 12b (e.g. m 1 =0,7), it is observed that the all networks while taking into account many factors, partic-
throughput of the fog network decreases after a certain point. ularly mobility issues. Hence, the operational space of the
This is due to the fact that users with high mobility rates in proposed networks is given in Figs. 14 and 15 to indicate the
both systems leave the fog system for neighbouring networks complete picture.
without authorising the system. Hence, the throughput of the Figures 14 and 15 show the operational areas of the users
fog network decreases. This is a very important and interest- in acquiring a resource at the edge and in fog computing,
ing finding in such an environment. These results indicated respectively, when λ=70. The mean queue length, through-
that mobility issues should be carefully considered while put, and mean response time results for both networks are
designing such an integrated environment to obtain the best represented in 3D. For instance, both networks are almost
QoS measurements. full (M Q L E =69,691 tasks/sec and M Q L F =51,70 tasks/sec)

Fig. 14 Operational space representation of edge network for all mobility probabilities when λ=70

123
 Mobile Networks and Applications

Fig. 15 Operational space representation of fog network for all mobility probabilities when λ=70

when the users are highly mobile, e.g. m 1 = m 2 = 0, 9. Since The obtained results for M Q L E , M RTE , γ E and M Q L F ,
the users are constantly moving, both networks accept users M RTF , γ F are presented in Tables 5 and 6 for the scenarios
into the system or direct them to other systems. Therefore, considered in Figs. 5, 6, 7, 8, 9, and 10 with E=10,L E =70,
systems are always loaded due their high mobility. In addi- Q E =60, μ E =10, F=13, L F =52, Q F =39, μ F =5, respectively.
tion, the mean response time of the loaded systems is always The tables clearly display a comparison of the proposed ana-
high. This case is clearly seen in Figs. 14c and 15c. 2,30 lytical model and DES results.
and 2,54 seconds are the highest response times of the edge Furthermore, the disparities between the proposed analyt-
and fog networks for the highly mobile users when λ=70. ical model and DES are depicted in all the figures. The tables,
The throughput results are also shown in Figs. 14b and 15b namely Tables 5 and 6, exhibit the maximum discrepancies
to figure out the operational area. As a result, by analysing for mean queue length, throughput, and mean response time
the findings in Figs. 14 and 15, the decision to acquire a for both edge and fog networks. The maximum discrepancies
resource for users can be made effectively while taking into for M Q L E , M RTE and γ E are 1.45%, 1.96%, and 1,87%,
account both mobility probabilities. Therefore, the models respectively, in Table 5. On the other hand, the maximum
and analysis presented here suggest that the necessary design discrepancies are found to be less than 0,61%, 1,35%, and
and development can be applied to enhance the QoS of such 0,66% in Table 6 for M Q L F , M RTF and γ F , respectively. In
systems. addition, Table 7 depicts the comparison of M Q L C results
of cloud layer and the DES for C = 16, μC = 5, Q C =
∞, E = 10, L E = 70, Q E = 60, μ E = 10, F = 13, L F =
4.1 Addressing the accuracy of the analytical 52, Q F = 39, μ F = 5, m 1 = 0.7. Overall, the observed dis-
approach crepancies in all figures and tables fall comfortably within the
5% confidence interval of the DES, indicating a high degree
To assess the level of accuracy between the proposed solu- of agreement between the proposed analytical models and
tion approach and DES, a comparative analysis is conducted. the DES.

123
Mobile Networks and Applications

Table 5 Comparison of
M Q L E , m 1 =0.1 γ E ,m 1 =0.5 M RTE , m 1 =0.9
M Q L E , M RTE and γ E results
λE Analyt. DES D(%) Analty. DES D(%) Analty. DES D(%)
of edge layer and DES for E=10,
L E =70, Q E =60, μ E =10, F=13, 10 1 1,0085 0,8428 10 10,1328 1,3105 0,1 0,1001 0,0999
L F =52,Q F =39,μ F =5, m 2 =0.1,
D=Discrepancy 20 2,1 2,1185 0,8732 21 21,0552 0,2621 0,1 0,1011 1,088
30 3,2 3,2471 1,4505 31 31,5923 1,8748 0,1 0,1015 1,4778
40 4,4 4,4092 0,2086 42 42,101 0,2398 0,1 0,1018 1,7681
50 5,5 5,571 1,2744 52 52,6407 1,2171 0,1 0,1019 1,8645
60 7 7,0291 0,4139 63 63,1663 0,2632 0,1 0,102 1,9607
70 42,3 42,3012 0,0028 73 73,6946 0,9425 0,11 0,1101 0,0908
80 64 64,0618 0,0964 84 84,2061 0,2447 0,12 0,1203 0,2493
90 66,8 66,8121 0,0181 94 94,5889 0,6225 0,18 0,1812 0,6622
100 67,8 67,8284 0,04187 99 99,827 0,8284 0,42 0,4223 0,5446
110 68,3 68,3515 0,0753 100 100,0075 0,0074 0,61 0,6148 0,7807
120 68,6 68,6706 0,1028 100,01 100,0153 0,0052 0,65 0,6539 0,5964

Table 6 Comparison of
M Q L F , m 1 =0.9 γ F ,m 1 =0.5 M RTF , m 1 =0.1
M Q L F , M RTF and γ F results
λE Analyt. DES D(%) Analty. DES D(%) Analty. DES D(%)
of fog layer and DES for E=10,
L E =70, Q E =60, μ E =10, F=13, 10 1,967 1,9791 0,6113 5,2 5,2022 0,0422 0,2007 0,2002 0,2497
L F =52, Q F =39, μ F =5,
m 2 =0.1, D=Discrepancy 20 3,966 3,9765 0,264 10,5 10,570 0,6622 0,1998 0,19712 1,3595
30 5,956 5,9408 0,2558 15,7 15,7213 0,1354 0,2000 0,2005 0,2493
40 8,022 8,0218 0,0024 21,0 21,0125 0,0594 0,2001 0,2002 0,0499
50 10,757 10,7454 0,1079 26,3 26,3315 0,1196 0,2001 0,2003 0,0998
60 18,3706 18,3806 0,0544 31,5 31,5863 0,2732 0,2000 0,20036 0,1796
70 48,1705 48,1605 0,0207 36,8 36,8124 0,0336 0,1999 0,19994 0,02
80 49,54 49,5098 0,0609 42,1 42,1242 0,0574 0,1999 0,19996 0,03
90 49,6288 49,6134 0,031 47,2 47,2994 0,2101 0,200 0,2006 0,2991
100 49,6876 49,6724 0,0306 49,9 49,903 0,006 0,2002 0,20029 0,0449
110 49,7343 49,7222 0,0243 50,0 50,0243 0,0485 0,2017 0,20176 0,0297
120 49,7710 49,7613 0,0194 50,001 50,022 0,0419 0,2030 0,203026 0,01289

Table 7 Comparison of M Q L C
M Q L C , m 1 = 0.7 and m 2 = 0.1 M Q L C , m 1 = 0.3 and m 2 = 0.9
results of cloud layer and DES
λE Analyt. DES D Analyt. DES D
for C = 16, μC = 5, Q C =
∞, E = 10, L E = 70, Q E = 10 1,50 1,5056 0,3719 0,8318 0,8217 1,2291
60, μ E = 10, F = 13, L F =
52, Q F = 39, μ F = 5, 20 3,009 3,0097 0,0232 1,6448 1,6446 0,0121
D=Discrepancy 30 4,514 4,5153 0,0287 2,4789 2,4673 0,4701
40 6,0201 6,0227 0,0431 3,2963 3,2868 0,2890
50 7,5408 7,5312 0,12747 4,0998 4,1096 0,2384
60 9,0815 9,066 0,1709 4,9343 4,9337 0,0121
70 10,7028 10,7015 0,0121 5,7385 5,7391 0,0104
80 12,686 12,6524 0,2655 6,0005 6,0009 0,0066
90 13,4825 13,4528 0,2207 6,0027 6,0029 0,0033
100 13,4198 13,4891 0,5137 6,0043 6,0035 0,0133
110 13,576 13,494 0,6076 6.0103 6.0111 0,0133
120 13,5362 13,5241 0,0894 6,0318 6,0322 0,0066

123
 Mobile Networks and Applications

5 Conclusions and future works Funding Not Applicable

Availability of data and materials Not Applicable

In this paper, a stochastic performance model is presented
together with the solution approach that obtains QoS output
Declarations
measurements considering mobility issues in an integrated
cloud-fog-edge computing environment. A two-stage open
queuing model and a M/M/c queueing model are consid- Conflict of interest/Competing interests The authors declare no con-
ered for the edge-fog and cloud networks, respectively. The flict of interest
throughput of the fog system is taken as the input of the
cloud subsystem. Hence, these dynamic input/output strate-
gies and mobility analysis give an opportunity to get more
realistic QoS results. Furthermore, using the proposed model,
the operational space for such environments can be obtained. References
The experimental results obtained from the proposed model
1. Qi Q, Tao F (2019) A smart manufacturing service system based
were compared and contrasted with the discrete event simu-
on edge computing, fog computing, and cloud computing. IEEE
lation for validation. The mean queue length, throughput, and Access 7:86769–86777. https://doi.org/10.1109/ACCESS.2019.
mean response time results were evaluated for QoS output 2923610
measurements. The numerical results showed that important 2. Wang H, Liu T, Kim B, Lin C-W, Shiraishi S, Xie J, Han Z
(2020) Architectural design alternatives based on cloud/edge/fog
QoS output measurements, characteristic features, and oper-
computing for connected vehicles. IEEE Commun Surv Tutorials
ational space of such an environment can be obtained before 22(4):2349–2377. https://doi.org/10.1109/COMST.2020.3020854
designing and deploying such an architecture with mobility 3. Abbasi M, Yaghoobikia M, Rafiee M, Jolfaei A, Khosravi MR
constraints. Furthermore, the proposed analytical model and (2020) Efficient resource management and workload allocation in
fog-cloud computing paradigm in iot using learning classifier sys-
solution approaches are effective and can be easily applied
tems. Comput Commun 153:217–228. https://doi.org/10.1016/j.
to any similar system. In future work, we will consider the comcom.2020.02.017
breakdown and repair behaviour of the CPUs in each network 4. Yan J, Huang Y, Gupta A, Gupta A, Liu C, Li J, Cheng L (2022)
with mobility issues. And try more realistic, intelligent, eas- Energy aware systems for real-time job scheduling in cloud data
centers: A deep reinforcement learning approach. Comput Electri-
ily extensible modelling strategies.
cal Eng 99:107688. https://doi.org/10.1016/j.compeleceng.2022.
107688
5. Rahman MA, Tari Z, Zhu D, Piccialli F, Wang X (2020) Ieee access
Appendix A special section editorial: Cloud-fog-edge computing in cyber-
physical-socia systems (cpss). IEEE Access 8:222859–222864.
https://doi.org/10.1109/ACCESS.2020.3042584
r C   r n−C
∞
 ∞ 6. Rapuzzi R, Repetto M (2018) Building situational awareness
rn
= for network threats in fog/edge computing: Emerging paradigms
C n−C C! C! C beyond the security perimeter model. Future Generation Comput
n=C n=C
∞ 
Syst 85:235–249. https://doi.org/10.1016/j.future.2018.04.007
rC  r m 7. Rathi S, Nagpal R, Mehrotra D, Srivastava G (2022) A metric
= focused performance assessment of fog computing environments:
C! C
m=0 A critical review. Comput Electric Eng 103:108350. https://doi.
rC 1 org/10.1016/j.compeleceng.2022.108350
= (A1) 8. Murtaza F, Akhunzada A, ul Islam S, Boudjadar J, Buyya R, (2020)
C! 1 − r Qos-aware service provisioning in fog computing. J Netw Comput
C Appl 165:102674. https://doi.org/10.1016/j.jnca.2020.102674
∞
 ∞
 9. Jha DN, Alwasel K, Alshoshan A, Huang X, Naha RK, Bat-
rn
QC = (n − C)Pn = (n − C) P0 tula SK, Garg S, Puthal D, James P, Zomaya A, Dustdar S,
C n−C C! Ranjan R (2020) Iotsim-edge: A simulation framework for mod-
n=C+1 n=C+1
∞ ∞
eling the behavior of internet of things and edge computing
r C P0  r C P0  environments. Software: Practice and Experience 50(6):844–
= (n − C)ρ n−C = mρ m 867. https://arxiv.org/abs/https://onlinelibrary.wiley.com/doi/pdf/
C! C!
n=C+1 m=1 10.1002/spe.2787. https://doi.org/10.1002/spe.2787
∞ ∞
r ρ P0   10. Cavalieri d’Oro E, Colombo S, Gribaudo M, Iacono M, Manca
C r C ρ P0 d
= mρ m−1 = ρm D, Piaz-zolla P (2019) Modeling and evaluating a complex edge
C! C! dρ computing based systems: An emergency management support sys-
m=1 m=1
  tem case study. Int Things 6:100054. https://doi.org/10.1016/j.iot.
r C ρ P0
d 1 r ρ P0
C
2019.100054
= −1 =
C! dρ 1 − ρ C!(1 − ρ)2 11. Ferrández-Pastor F-J, Mora H, Jimeno-Morenilla A, Volckaert B
(2018) Deploy-ment of iot edge and fog computing technologies
rnρ
= P0 (A2) to develop smart building services. Sustainability 10(11). https://
C!(1 − ρ)2 doi.org/10.3390/su10113832

123
Mobile Networks and Applications

12. Qiu M, Kung S-Y, Gai K (2020) Intelligent security and opti- tive node scaling approach. Comput Electrical Eng 60:30–44.
mization in edge/-fog computing. Future Generation Comput Syst https://doi.org/10.1016/j.compeleceng.2017.04.018
107:1140–1142. https://doi.org/10.1016/j.future.2019.06.002 31. Mas L, Vilaplana J, Mateo J (2022) A queuing theory model for fog
13. Petrovic N, Tosic M (2020) Smada-fog: Semantic model driven computing. J Supercomput 78(8):11138–11155. https://doi.org/10.
approach to deployment and adaptivity in fog computing. Simul 1007/s11227-022-04328-3
Model Pract Theory 101:102033. https://doi.org/10.1016/j.simpat. 32. Tadakamalla U, Menascé D (2018) Fogqn: An analytic model for
2019.102033 fog/cloud computing. In: 2018 IEEE/ACM International Confer-
14. Mutlag AA, Abd Ghani MK, Arunkumar N, Mohammed MA, ence on Utility and Cloud Computing Companion (UCC Com-
Mohd O (2019) Enabling technologies for fog computing in health- panion), pp. 307–313. https://doi.org/10.1109/UCC-Companion.
care iot systems. Future Generation Comput Syst 90:62–78. https:// 2018.00073
doi.org/10.1016/j.future.2018.07.049 33. El Kafhali S, Salah K, Ben Alla S (2018) Performance evalua-
15. Haverkort BR, Marie R, Rubino G, Trivedi KS (2001) Performa- tion of iot-fog-cloud deployment for healthcare services. In: 2018
bility Modelling: Techniques and Tools. Wiley, England 4th International Conference on Cloud Computing Technologies
16. Lera I, Guerrero C, Juiz C (2019) Yafs: A simulator for iot scenarios and Applications (Cloudtech), pp. 1–6. https://doi.org/10.1109/
in fog computing. IEEE Access 7:91745–91758 CloudTech.2018.8713355
17. Mechalikh C, Taktak H, Moussa F (2019) Pureedgesim: A simu- 34. Chakka R (1998) Spectral expansion solution for some finite capac-
lation toolkit for performance evaluation of cloud, fog, and pure ity queues. Annals Operations Res 79:27–44
edge computing environments. In: 2019 International Conference 35. Ever E (2007) Performability modelling of homogenous and het-
on High Performance Computing and Simulation (HPCS), pp. 700– erogeneous multiserver systems with breakdowns and repairs. PhD
707. https://doi.org/10.1109/HPCS48598.2019.9188059 thesis, Middlesex University
18. Puliafito C, Gonçalves DM, Lopes MM, Martins LL, Madeira E, 36. Aldag M, Kirsal Y, Ülker S (2022) An analytical modelling and
Mingozzi E, Rana O, Bittencourt LF (2020) Mobfogsim: Simula- qos evaluation of fault-tolerant load balancer and web servers in
tion of mobility and migration for fog computing. Simul Model fog computing. J Supercomput 78(10):12136–12158. https://doi.
Pract and Theory 101:102062 org/10.1007/s11227-022-04345-2
19. Mohan N, Kangasharju J (2016) Edge-fog cloud: A distributed 37. Firouzi F, Farahani B, Marinšek A (2022) The convergence and
cloud for internet of things computations. In: 2016 Cloudification interplay of edge, fog, and cloud in the ai-driven internet of
of the Internet of Things (CIoT), pp. 1–6. https://doi.org/10.1109/ things (iot). Inf Syst 107:101840. https://doi.org/10.1016/j.is.2021.
CIOT.2016.7872914 101840
20. Gupta H, Vahid Dastjerdi A, Ghosh SK, Buyya R (2017) ifogsim: 38. Ma X, Xu H, Gao H, Bian M, Hussain W (2023) Real-time virtual
A toolkit for modeling and simulation of resource management machine scheduling in industry iot network: A reinforcement learn-
techniques in the internet of things, edge and fog computing envi- ing method. IEEE Trans Industrial Informatics 19(2):2129–2139.
ronments. Softw: Pract Exp 47(9):1275–1296 https://doi.org/10.1109/TII.2022.3211622
21. Gupta H, Vahid Dastjerdi A, Ghosh SK, Buyya R (2017) ifogsim: 39. Gao H, Huang W, Liu T, Yin Y, Li Y (2022) Ppo2: Location
A toolkit for modeling and simulation of resource management privacy-oriented task offloading to edge computing using rein-
techniques in the internet of things, edge and fog computing envi- forcement learning for intelligent autonomous transport systems.
ronments. Softw: Pract Exp 47(9):1275–1296 IEEE Trans Intell Transportation Syst 1–14. https://doi.org/10.
22. Kirsal Y (2018) Analytical modelling and optimization analysis of 1109/TITS.2022.3169421
large-scale communication systems and networks with repairmen 40. Xu Y, Lin J, Gao H, Li R, Jiang Z, Yin Y, Wu Y (2022)
policy. Computing 100(5):503–527 Machine learning-driven apps recommendation for energy opti-
23. Kirsal Y, Ever K, Y, Mapp GE, Raza M, (2021) 3d analytical mization in green communication and networking for connected
modelling and iterative solution for high performance computing and autonomous vehicles. IEEE Trans Green Commun Netw
clusters. IEEE Trans Cloud Comput 1–1. https://doi.org/10.1109/ 6(3):543–1552. https://doi.org/10.1109/TGCN.2022.3165262
TCC.2021.3055119 41. Gao Y, Guan H, Qi Z, Song T, Huan F, Liu L (2014) Service level
24. Fantacci R, Picano B (2020) Performance analysis of a delay con- agreement based energy-efficient resource management in cloud
strained data offloading scheme in an integrated cloud-fog-edge data centers. Comput Electrical Eng 40(5):1621–1633. https://doi.
computing system. IEEE Trans Vehic Technol 69(10):12004– org/10.1016/j.compeleceng.2013.11.001
12014. https://doi.org/10.1109/TVT.2020.3008926 42. Leontiou N, Dechouniotis D, Denazis S, Papavassiliou S (2018)
25. Vilaplana J, Solsona F, Teixidó I, Mateo J, Abella F, Rius J (2015) A hierarchical control framework of load balancing and resource
A queuing model considering resources sharing for cloud service allocation of cloud computing services. Comput Electrical Eng
performance. J Supercomput 71(11):4042–4055. https://doi.org/ 67:235–251. https://doi.org/10.1016/j.compeleceng.2018.03.035
10.1007/s11227-015-1503-z 43. Outamazirt A, Barkaoui K, Aïssani D (2018) Maximizing profit in
26. Battula SK, O’Reilly MM, Garg S, Montgomery J (2021) A generic cloud computing using m/g/c/k queuing model. In: 2018 Interna-
stochastic model for resource availability in fog computing envi- tional Symposium on Programming and Systems (ISPS), pp. 1–6.
ronments. IEEE Trans Parallel Distributed Syst 32(4):960–974. https://doi.org/10.1109/ISPS.2018.8379008
https://doi.org/10.1109/TPDS.2020.3037247 44. Adhikari S, Hutaihit M, Chakraborty M, Mahmood S, Durakovic
27. Banks J, Carson II JS, Automation B, Nelson BL, Nicol DM (2004) B, Pal S, Akila D, Obaid A (2021) Analysis of average waiting
Discrete-event system simulation fourth edition time and server utilization factor using queueing theory in cloud
28. Jafarnejad Ghomi E, Rahmani AM, Qader NN (2019) Apply- computing environment. Int J Nonlinear Anal Appl 12(Special
ing queue theory for modeling of cloud computing: A systematic Issue):1259–1267
review. Concurrency Comput: Pract Exp 31(17):51–86 45. Li W, Karthik C, Rajalakshmi M (2022) Big data visualization for
29. Liu X, Li S, Tong W (2015) A queuing model considering in-situ data exploration for sportsperson. Comput Electrical Eng
resources sharing for cloud service performance.J Supercomput 99:107829. https://doi.org/10.1016/j.compeleceng.2022.107829
71(11):4042–4055. https://doi.org/10.1007/s11227-015-1503-z 46. Ghosh S, Mukherjee A, Ghosh SK, Buyya R (2020) Mobi-iost:
30. Suresh S, Sakthivel S (2017) A novel performance constrained Mobility-aware cloud-fog-edge-iot collaborative framework for
power management framework for cloud computing using an adap- time-critical applications. IEEE Trans Netw Sci Eng 7(4):2271–
2285. https://doi.org/10.1109/TNSE.2019.2941754

123
 Mobile Networks and Applications

Publisher’s Note Springer Nature remains neutral with regard to juris-

dictional claims in published maps and institutional affiliations.

Springer Nature or its licensor (e.g. a society or other partner) holds

exclusive rights to this article under a publishing agreement with the
author(s) or other rightsholder(s); author self-archiving of the accepted
manuscript version of this article is solely governed by the terms of such
publishing agreement and applicable law.

123