IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO.

16, AUGUST 15, 2021 12705

METO: Matching-Theory-Based Efficient Task

Offloading in IoT-Fog Interconnection Networks
Chittaranjan Swain , Graduate Student Member, IEEE, Manmath Narayan Sahoo , Member, IEEE,
Anurag Satpathy , Graduate Student Member, IEEE, Khan Muhammad , Member, IEEE,
Sambit Bakshi, Senior Member, IEEE, Joel J. P. C. Rodrigues , Fellow, IEEE,
and Victor Hugo C. de Albuquerque , Senior Member, IEEE

Abstract—Typical cloud systems are often prone to inher- and heterogeneous tasks. Extensive simulations across both envi-
ent wide area network (WAN) latency. To address this issue ronments confirm that the proposed algorithm outperforms the
fog computing is proposed that enables resource-constrained existing schemes with respect to improved energy consumption,
Internet-of-Things (IoT) devices, to execute deadline-sensitive completion time, and execution time. Moreover, METO also
tasks at the edge of the network. These devices can extend shows the reduced number of outages across baselines used for
their battery lifespan by intelligently offloading computations as comparison.
tasks to fog nodes (FNs) in their vicinity. However, finding an
optimal offloading plan in a densely connected IoT-fog network Index Terms—Criteria importance though inter criteria cor-
is proven to be N P -Hard. Hence, in this article, we propose relation (CRITIC), fog computing, Gale–Shapley algorithm,
a matching theory-based efficient task offloading strategy called matching, task offloading, technique for order of preference by
METO that aims to reduce the total system energy and number similarity to ideal solution (TOPSIS).
of outages (number of tasks exceeding the deadline) in an IoT-fog
interconnection network. As resource allocation involves multiple I. I NTRODUCTION
criteria, their weights are derived using criteria importance HE burgeon of Internet of Things (IoT) has led to the emer-
though inter criteria correlation (CRITIC). Furthermore, to rank
the alternatives we use the technique for order of preference by
T gence of a variety of applications catering to the demands
of autonomous vehicles, mobile crowdsensing [1], [2], smart and
similarity to ideal solution (TOPSIS). Based on this ranking,
we formulate the overall offloading problem as a one-to-many connected communities [3], [4], including smart home devices,
matching game and utilize the deferred acceptance algorithm e-healthcare, smart grids, smart buildings, smart phones [5],
(DAA) to produce a stable assignment. Simulation is performed etc. However, executing such applications on IoT devices in
in two different settings comprising offloading of homogeneous a densely connected modern network introduces the following
challenges. First, these emerging applications work on massive
data and are typically resource hungry [6]. Second, applications,
Manuscript received April 7, 2020; revised July 26, 2020; accepted
September 2, 2020. Date of publication September 21, 2020; date of current such as connected vehicles, online gaming, and augmented real-
version August 6, 2021. This work was supported in part by the Institute for ity (AR) have stringent requirements on the maximum tolerable
Information and Communication Technology Planning & Evaluation (IITP), latency [7], [8]. Third, these IoT devices not only suffer from
South Korea, funded by the Ministry of Science and ICT (MSIT), South Korea,
under Project 2019-0-00795 [Project Name: Development of Integrated Cross- limited computing and networking capabilities but also have
Model Data Processing Platform Supporting a Unified Analysis of Various restricted battery lifespan [9]. These limitations render appli-
Big Data Models]; in part by the FCT/MCTES through National Funds and cation execution on IoT devices an inevitable challenge. As an
when applicable co-funded EU funds under Project UIDB/50008/2020; and
in part by the Brazilian National Council for Scientific and Technological immediate remedy to the resource limitations of IoT devices,
Development CNPq under Grant 309335/2017-5. The work of Victor Hugo cloud computing can be considered as a viable solution. Cloud
C. de Albuquerque was supported by the Brazilian National Council computing augments the resource limitations by permitting IoT
for Research and Development (CNPq) under Grant 304315/2017-6 and
Grant 430274/2018-1. (Corresponding authors: Khan Muhammad; Manmath devices to offload computations to a remote cloud over a wire-
Narayan Sahoo.) less channel. However, intermittent channels, scarce spectrum
Chittaranjan Swain, Manmath Narayan Sahoo, Anurag Satpathy, and resources, and high propagation delays make cloud computing
Sambit Bakshi are with the Department of Computer Science and
Engineering, National Institute of Technology Rourkela, Rourkela 769008, an infeasible platform for applications with ultralow latency
India (e-mail: [email protected]; [email protected]; requirements [10]. To overcome this impediment, an emerging
[email protected]; [email protected]). paradigm called fog computing was introduced [11]. Fog com-
Khan Muhammad is with the Department of Software, Sejong University,
Seoul 143-747, South Korea (e-mail: [email protected]). puting provides computing, storage, and networking resources
Joel J. P. C. Rodrigues is with the Covilhã Delegation, Instituto de close to the end users. In contrast to cloud computing, fog com-
Telecomunicações, 6200-161 Covilhã, Portugal, and also with the Department puting provides additional benefits, such as mobility support,
of PPGEE, Federal University of Piauí, Teresina 4049-550, Brazil (e-mail:
[email protected]). location awareness, and improved Quality of Service (QoS) in
Victor Hugo C. de Albuquerque is with LAPISCO, Federal Institute of terms of latency incurred and energy consumed [12].
Education, Science and Technology of Ceará, Fortaleza 60811-905, Brazil, To make the most of fog computing, resource-constrained
and also with the ARMTEC Tecnologia em Robótica, Fortaleza 60811-341,
Brazil (e-mail: [email protected]). IoT devices often offload computation tasks that are executed
Digital Object Identifier 10.1109/JIOT.2020.3025631 on nearby fog nodes (FNs). Primarily these devices sense and
2327-4662 
c 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on January 01,2024 at 07:21:44 UTC from IEEE Xplore. Restrictions apply.
 12706 IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO. 16, AUGUST 15, 2021

collect data that is offloaded to the FNs for instant process- 4) To validate the effectiveness of METO we test its
ing [13], [14]. Offloading is primarily categorized into two performance in two different simulation settings, i.e.,
types, viz., 1) full and 2) partial. In case of full offloading with homogeneous and heterogeneous tasks. Thorough
a task is entirely executed at an FN, whereas a part of it is simulation analysis across both environments confirm
executed in partial offloading. In this article, we focus our that the proposed algorithm outperforms the existing
attention on full offloading. A fundamental challenge in such a schemes with respect to improved energy consumption,
scenario is to effectively offload computation tasks of different completion time, and execution time. Moreover, METO
IoT devices to FNs with heterogeneous resources [7]. The also realizes the reduced number of outages.
literature devoted to task offloading is primarily focused on The remainder of this article is organized as follows.
addressing two independent parameters. Considering latency Section II discusses the literature that we have reviewed. In
minimization, Chiti et al. [15] discussed a matching-theory- Section III, we discuss the system model in detail. Section IV
based framework in a full offloading scenario. On the other talks about the matching-theory-based solution approach.
hand, Yousefour et al. [16] discussed a collaborative full Performance analysis of METO is discussed in Section V and
offloading policy to reduce service delays in executing tasks conclusions are drawn in Section VI.
at FNs. Some other works that have concentrated on latency
reduction are discussed in [17]–[20]. As IoT devices and II. R ELATED W ORK
FNs are battery powered, energy consumption is a crucial In this section, we discuss the literature that we have
parameter that dictates the lifespan of such devices. In this reviewed on task offloading in IoT-fog interconnection
context, Haber et al. [6] presented a convex approximation environments. Primarily the researchers have independently
strategy (SCA) for edge cloud that aims to reduce the com- addressed latency and energy concerns. In the context
putational overheads and energy consumption. Alternatively, of latency minimization for full offloading scenarios,
Zhang et al. [9] discussed a fair offloading mechanism to Chiti et al. [15] discussed a matching game with externali-
optimize the energy consumption. Some recent works targeting ties considering the queuing delay at the FNs. On the other
energy minimization are discussed in [21]–[23]. hand, Yousefour et al. [16] focused on building a collabora-
The research on task offloading independently addresses tive offloading policy aimed at reducing the service delay of
latency minimization [15]–[20] and energy consumption [9], applications hosted at IoT devices. To boost the processing
[21], [22]. However, Zhang et al. [9] pointed out that there capacity and save communication bandwidth, Hu et al. [25]
exists a strong correlation between the parameters. Hence, it discussed a secure fog computing-based facial identification
is essential to jointly consider the latency and energy con- and resolution framework for an amalgamation of IoT and
sumption in the offloading process. Therefore, in this article, big data applications. In order to provide flexibility to users
we propose an efficient task offloading algorithm (METO) to in choosing the execution environment among local, device
jointly optimize latency and energy consumption in a densely to device (D2D), and cloud, Chen and Zhang [26] proposed
connected IoT-fog network. Furthermore, to make the imple- a hybrid task offloading strategy based on minimum weight
mentation pragmatic, we also aim to reduce the number of matching problem over three-layered graphs aimed at mini-
outages, i.e., number of tasks having completion time greater mizing the total execution cost. Some other works that have
than a given deadline. As we deal with different stakeholders focused on reducing the number of outages are discussed
with contrasting agendas, we use matching theory as a solu- in [27] and [28]. For partial offloading scenarios specific to
tion framework that elegantly captures the design objectives of mobile computing environments, Kan et al. [18] proposed a
different stakeholders using the concept of preferences [24]. heuristic that aims to jointly optimize the radio and computa-
Moreover, matching-theory-based solutions are computation- tional resources consumed in offloading. Considering resource
ally inexpensive and scalable. Motivated by this, we formulate allocation over radio channels, Le et al. [19] proposed a bisec-
the overall task offloading problem as a one-to-many matching tion search method based on time-division multiple access
game between the IoT devices and FNs. (TDMA) and frequency-division multiple access (FDMA)
In a nutshell the contributions of this article can be high- focused at minimizing the overall completion time of tasks.
lighted as follows. Keeping in view energy consumption, Haber et al. [6]
1) We propose a model called METO that aims to discussed a convex approximation method (SCA) for full
jointly optimize the latency and energy consumption offloading in mobile-edge computing (MEC) environment
for offloading deadline sensitive tasks in a IoT-fog directed at reducing the energy consumption of mobile devices.
interconnection network. Chen et al. [21] modeled an energy-aware resource alloca-
2) The overall problem is modelled as a one-to-many match- tion problem and solved it using a concave–convex procedure
ing game between a set of tasks and FNs. The preference (CCCP). On the other hand, Liu et al. [30] formulated the same
profile of the players is generated using a hybrid CRITIC as a convex optimization problem and aimed at reducing the
and TOPSIS approach. CRITIC is used to derive criteria system energy and latency incurred by users. Chiti et al. [29]
weights and TOPSIS is used to rank the alternatives. discussed a matching game in an integrated edge-fog com-
3) To evaluate the performance of METO, we perform puting system to minimize the latency and energy overheads.
extensive simulations and compare its performance with Some other relevant literature on energy minimization in partial
three different baselines: SMETO [22], a matching- offloading environments can be found in [9] and [22].
theory-based model inspired by Chiti et al. [15], and Table I summarizes the parameters used to compare dif-
a random allocation strategy. ferent offloading techniques. It can be observed from the
Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on January 01,2024 at 07:21:44 UTC from IEEE Xplore. Restrictions apply.
 SWAIN et al.: METO: MATCHING-THEORY-BASED EFFICIENT TASK OFFLOADING IN IoT-FOG INTERCONNECTION NETWORKS 12707

TABLE I
S UMMARY OF THE L ITERATURE ON TASK O FFLOADING
FN VRUs of same
Fog Network
computational speed

FN
FNC

FN

Uplink and Downlink

Channels

table that, majority of the literature that we have reviewed

have considered latency and energy independently. However, IoT Devices
...
Zhang et al. [9] pointed out that there is a strong correlation
between the above parameters which motivates us toward joint
allocation. The closest work in the literature that considers Fig. 1. IoT-fog interconnection network.
both parameters is discussed in Chiti et al. [29]. However, the
authors fail to consider the overall system energy and have FN fi is considered to be same. Since a VRU can host a task,
restricted themselves to reducing the energy consumption of Qi denotes the maximum number of tasks that can be executed
IoT devices. Therefore, in this article, we propose METO that in parallel.
is focused on joint optimization of latency and system energy As IoT devices do not possess any capabilities for local
in a heavily interconnected IoT-fog network. Furthermore, to execution, the tasks must be offloaded to suitable FNs. The
make the model real time we also focus on reducing the offloading comprises the following steps.
number of outages, i.e., the number of tasks exceeding the 1) Transmission of a task to an FN.
deadlines. The overall problem is formulated as a one-to-many 2) Execution at the FN.
matching game between FNs and IoT devices [29]. To the 3) Retrieval of the computed results.
best of our knowledge, we are the first to consider the joint The transmissions are carried out over a noisy wireless
optimization of latency and system energy. Next, we discuss channel. The overheads in the offloading process from a user’s
in detail the system model and solution approach. standpoint consist of 1) energy consumption and 2) over-
III. S YSTEM M ODEL all latency. On the other hand, the FNs aim at reducing
1) the number of outages and 2) energy consumption. The
We consider an IoT-fog interconnection network as depicted above-mentioned parameters are mathematically modeled in
in Fig. 1. It consists of a set of IoT devices and FNs, and the following sections.
a specific FN is designated as the FN controller (FNC).
The agents, i.e., FNs and IoT devices distributively generate A. Latency Computation
preferences based on local information and dispatch it to The overall latency incurred in successfully offloading and
the FNC over a secure channel [31]–[33]. After receiving executing a task comprises of the following.
the preferences, the FNC assigns tasks to FNs based on 1) Transmission Delay: It refers to the amount of time
a matching game, which is discussed later in this article. expended in transferring the input data from an IoT
The interconnection network comprises “n” FNs denoted by device to an FN over a wireless channel. This delay is
F = {f1 , f2 , . . . , fn } and a set of “m” IoT devices represented dependent on the task size and channel state information.
by D = {d1 , d2 , . . . , dm }. Assuming that each IoT device 2) Queuing Delay: It is defined as the time spent by a task
generates a single task, the set of tasks can be defined as at an FN’s queue prior to its execution. This work con-
T = {t1 , t2 , . . . , tm }, where tj corresponds to a task generated siders a parallel execution environment that guarantees
by the IoT device dj . The IoT devices have limited compu- exclusive computational resources, i.e., an FN fi cannot
tational resources and cannot perform local computations, be assigned more than its quota Qi at any point, hence
hence are dependent on nearby FNs for processing. this delay is set to 0.
An IoT device dj sends an offloading request in the form 3) Computation Delay: This delay refers to the time for
of a quadruple Ij , Oj , Cj , Dj . Here, Ij and Oj correspond to executing a task on an FN. This is dependent on the FN’s
the input and output size (bits) and, Cj and Dj denote the computational capabilities and the incoming task size.
computational demand (cycles) and deadline (s), of the task tj . 4) Receiving Delay: Finally, the receiving delay refers to
The computational capacity of an FN fi is logically partitioned the time required in transmitting the processed output
into homogeneous virtual resource units (VRUs), each having from an FN to the concerned IoT device. For some
capacity Ci [34]. For an FN fi , number of VRUs, also referred lightweight applications, such as face recognition the
to as its quota, is denoted by Qi . To add heterogeneity to the receiving delay is ignored [35], [36]. However, for
architecture it is considered that the number of VRUs of FNs smart city, smart grid, smart transportation systems, and
are different, i.e., for any two FNs fi and fi ∈ F, Qi = Qi . deadline-based applications the receiving delay cannot
However, the computational capabilities of each VRU at an be disregarded [5].

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on January 01,2024 at 07:21:44 UTC from IEEE Xplore. Restrictions apply.
 12708 IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO. 16, AUGUST 15, 2021

To calculate the transmission and receiving delays over a total amount of energy Ei expended by fi for executing a task
wireless channel, it is essential to determine the uplink and tj , considering the aforementioned operations, is calculated as
downlink rates between IoT devices and FNs. In this work,    
it is considered that each device dj has an active uplink and cmp
Ei = pi ∗ Tj,i
tran
+ Tj,i
recv
+ Tj,i ∗ pj . (8)
downlink channel with an FN fi having bandwidth Bj,i and Bi,j ,
respectively, [6]. The uplink rate between dj and fi is denoted
by Uj,i and is calculated as per C. Problem Formulation
  The overall objective of METO is to construct an offload-
pj hj,i
Uj,i = Bj,i log2 1 + . (1) ing plan that efficiently assigns tasks to FNs such that the
w0
total system energy expended and overall latency incurred, are
In the above equation, pj denotes the transmission power of reduced simultaneously. Furthermore, METO tries to ensure
dj , hj,i is the uplink channel gain, and w0 represents the noise that the offloading suffers a minimum number of outages. Let
power. For an input size Ij and a given uplink rate Uj,i , the xj,i be a binary indicator variable, that follows definition as
transmission time Tj,i tran can be calculated as per
in (9), signifies if a task tj is assigned to fi
Ij
tran
Tj,i = . (2) 
Uj,i 1 : if tj is assigned to fi
xj,i = (9)
As mentioned in [6], the downlink rate Di,j for an FN fi with 0 : otherwise.
a downlink channel gain hi,j and transmission power pi is
expressed as The total system energy consumed, denoted as Esys , in the
  offloading process comprises energy consumed by all FNs and
pi hi,j IoT devices is expressed as
Di,j = Bi,j log2 1 + . (3)
w0

m 
n
 
m
Similar to (2), for a given output size Oj and downlink rate Di,j , Esys = xj,i ∗ Ei + Ej . (10)
the receiving delay Tj,ircv is presented as per
j=1 i=1 j=1

Oj
rcv
Tj,i = . (4) Since the tasks are with deadline constraints, the offloading
Di,j mechanism should ensure minimum number of outages. Let O
cmp be the set of tasks suffering outages that is defined formally as
The computation delay of tj at fi , denoted by Tj,i , is men-
tioned in (5). Here, Ci refers to the computational capacity of
a VRU at an FN fi O = tj ∈ T | xj,i ∗ Tj,i
tot
> Dj ; i ∈ [1, n] . (11)
cmp Cj
Tj,i = . (5) The overall objective of METO is expressed in (12)
Ci
tot
 sys
From (2), (4), and (5) the total time (completion time) Ti,j min E and min(|O|) (12)
for offloading and executing a task tj at fi is calculated in 
m
accordance with the following equation: s.t. C1 : xj,i ≤ Qi (13)
cmp j=1
tot
Tj,i = Tj,i
tran
+ Tj,i
rcv
+ Tj,i . (6)
n
C2 : xj,i = 1 (14)
B. Energy Computation i=1
During offloading, both IoT devices and FNs consume 
n

energy. The energy consumption models for both are discussed C3 : |T| ≤ Qi . (15)
i=1
in sequel. Note that the transmission power of IoT devices and
FNs varies with its distance from each other. The task offloading problem is subject to the constraints
1) Energy Consumption of IoT Devices: The energy con- expressed in (13)–(15). Constraint C1 ensures that an FN fi ∈ F
sumption Ej of an IoT device dj for offloading a task to fi is cannot host tasks more than its quota Qi . Constraint C2 ensures
dependent on transmission and receiving delays [29], and is that a task is assigned to one FN only. Constraint C3 confirms
calculated as that the aggregate number of tasks in the system is not more
 
Ej = pj ∗ Tj,i
tran
+ Tj,i
rcv
. (7) than the total quota of all the FNs taken together. The problem
expressed as (12)–(15) is a combinatorial optimization
2) Energy Consumption of Fog Nodes: Energy consump- problem and is proven to be N P-Hard [7]. In fact, for the
tion of the FNs reasons to the following three operations: increasing number of IoT devices and FNs it is almost infea-
1) receiving a task from an IoT device; 2) executing a task; sible to obtain an optimal solution in polynomial time. Taking
and 3) transmitting the obtained results back to the IoT device. the above factors into consideration the offloading problem is
As reported in [29], the reception time at the FN is same as formulated as a one-to-many matching game that generates a
the transmission time of an IoT device and vice versa. The suboptimal offloading plan in the polynomial time frame.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on January 01,2024 at 07:21:44 UTC from IEEE Xplore. Restrictions apply.
 SWAIN et al.: METO: MATCHING-THEORY-BASED EFFICIENT TASK OFFLOADING IN IoT-FOG INTERCONNECTION NETWORKS 12709

IV. T HE P ROPOSED S OLUTION A PPROACH Algorithm 1: Criteria Weight Determination Using

Matching theory has recently emerged as an apt frame- CRITIC
work to perform association between two distinct sets taking Input: Decision matrices of all agents a ∈ T ∪ F
individual preference of each agent into consideration [34]. Result: Weight Matrix W
The preference profile of each agent portrays the satisfaction 1 for each agent a ∈ T ∪ F do

levels in the matching procedure. To address the issues of fair- 2 Normalize the decision matrix of an agent ‘a’ as per
ness of allocation, the notion of stability is also defined [24]. Eq. (16).
Although it might be enticing to formulate task offload- 3 Evaluate the standard deviation σk for each criterion
ing as an optimization problem, the optimization solutions, in the normalized decision matrix.
however, suffer from the following drawbacks: 1) they are 4 Calculate each criterion quantity of information QIk
incapable of considering objectives of different stakeholders based on Eq. (17).
and 2) are computationally expensive, hence are not scal- 5 Determine each criterion weight wk calculated as per
able. Motivated by this, a matching-theory-based efficient task Eq. (18) and form the criteria weight vector Wa for
offloading model called METO is proposed. Moreover, as agent ‘a’.
an FN can execute multiple such offloaded tasks simultane- 6 Add Wa as next row in W.
ously, it is appropriate to model the overall problem as a
one-to-many game. As multiple criteria are considered in the
decision-making process, a hybridized CRITIC and TOPSIS-
based techniques are used to generate the preferences of the B. CRITIC-Based Criteria Weight Determination
agents. TOPSIS is preferred in comparison with other con- The overall working of CRITIC follows Algorithm 1. We
temporary techniques, such as analytical hierarchy process only highlight important steps used in the process. The detailed
(AHP), preference ranking organization method for enrich- documentation of CRITIC can be found in [38]. Every agent
ment evaluation (PROMETHE), and simple additive weights a ∈ T ∪ F having an index l constructs a normalized decision
(SAWs) due to its higher accuracy [37]. However, TOPSIS matrix B∗l . For example considering a task tj , each entry b∗i,k ∈
requires the criteria weights, as input, to rank the alternatives. B∗j is normalized as per (16). Here, bworst
k and bbest
k represent
For this, a CRITIC-based weight determination strategy is th
the worst and best entries along the k criterion in Bj [39]
used. The overall working of METO consists of the following
four steps: 1) decision matrix formulation; 2) CRITIC-based bi,k − bk worst
b∗i,k = ∀i ∈ [1, n] ∀k ∈ [1, c]. (16)
criteria weight determination; 3) ranking based on TOPSIS; bk best − bk worst
and 4) efficient task offloading using a matching game. The
Each agent a ∈ T ∪ F then constructs a criteria correlation
detailed discussion on these steps is provided subsequently.
matrix which is a symmetric matrix [38]. The correlation
matrix Sj of dimension c × c, has entries sk,k indicating
A. Decision Matrix Formulation the linear correlation coefficient value of criterion k with k .
Depending on the value of sk,k the relationship between the
A decision matrix is constructed for every agent a ∈ T ∪ F.
criterion k and k can be termed as either strong, moderate or
It comprises entries corresponding to each criterion across dif-
weak. The quantity of information contained in kth criterion
ferent alternatives. For instance each task tj ∈ T constructs a
denoted by QIk is determined as per (17), where σk is the
n × c decision matrix denoted by Bj . Here, c refers to the total
standard deviation of kth criterion column in the normalized
number of evaluation criteria. An entry bi,k ∈ Bj corresponds
decision matrix
to the value of ith FN against criterion k. A decision matrix
in this context is portrayed as follows: 
c

QIk = σk ∗ 1 − sk,k ∀k ∈ [1, c]. (17)
⎡ ⎤ k =1
b1,1 b1,2 ··· b1,c
⎢b2,1 b2,2 ··· b2,c ⎥ The weight of the kth criterion is then derived as
⎢ ⎥
Bj = ⎢ . .. .. .. ⎥ .
⎣ .. . . . ⎦ QI
wk = c k . (18)
bn,1 bn,2 ··· bn,c n×c k=1 QIk

Thus for each agent “a,” a criteria weight vector Wa is

In a similar way, an m × c decision matrix Bi for each agent
obtained. Finally, a weight matrix W for all agents is con-
fi ∈ F can be constructed. In this work, to calculate the prefer-
structed considering individual criteria weight vectors. Using
ence of task tj , two criteria are considered, viz., latency of the
these weights alternatives are ranked based on TOPSIS.
task and energy consumption of the IoT device, as mentioned
in (6) and (7), respectively. Similarly, the preference of an FN
fi is also calculated based on two criteria, viz., energy con- C. Ranking Based on TOPSIS
sumption of the FN to execute a task as given in (8) and the An agent a ∈ T∪F ranks the alternatives using TOPSIS fol-
deadline of the task. After constructing the decision matrices, lowing, Algorithm 2. For details on TOPSIS the readers may
a CRITIC-based criteria weight determination followed by the refer to [38], [40], and [41]. In continuation with the previous
ranking of alternatives using TOPSIS is performed. example, the decision matrix Bj of task tj is normalized to

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on January 01,2024 at 07:21:44 UTC from IEEE Xplore. Restrictions apply.
 12710 IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO. 16, AUGUST 15, 2021

Algorithm 2: TOPSIS-Based Rank Calculation fi that allows each agent tj ∈ T to indicate preference over
Input: Decision matrices of all agents a ∈ T ∪ F all agents fi ∈ F in the opposite set, and vice versa.
Result: Performance Scores P(.) Definition 2: The matching game is based on a mapping
1 for each agent a ∈ T ∪ F do function λ : T ∪ F → 2T∪F such that
   
2 Normalize the decision matrix using Eq. (19). λ tj ⊂ F and λ tj  = 1 (24a)
3 Calculate the weighted normalized decision matrix
λ(fi ) ⊆ T and |λ(fi )| ≤ Qi (24b)
based on Eq. (20). 
4 Evaluate the positive and negative ideal solutions for fi ∈ λ tj ⇔ tj ∈ λ(fi ). (24c)
each criterion k as minimum and maximum value of Condition (24a) enforces that each task is assigned to
kth column weighted normalized matrix. exactly one FN. An FN can only accept tasks until its quota
5 Determine the distance of each alternative from the is met. This is captured in (24b). Condition (24c) states that
positive and negative ideal solutions based on Eq. a task tj is matched to an FN fi iff fi is matched to tj .
(21) and (22). Definition 3: A matching function λ is individually rational
6 Compute the performance score for each alternative iff there does not exist an agent a ∈ T∪F such that φ a λ(a),
following Eq. (23), and rank them in decreasing i.e., agent a prefers to remain unmatched in comparison to its
order. current match λ(a).
This implies that the partner obtained in the matching pro-
cess should not be unacceptable, i.e., the agent should not
prefer to remain unmatched than its currently matched partner.
obtain Bj . Each entry ri,k
 of B is derived from
j Definition 4: A matching function λ is said to be blocked
 bi,k by a pair of agents (tj , fi ) iff tj ∈
/ λ(fi ), fi tj λ(tj ) and tj fi
ri,k =  ∀i ∈ [1, n] ∀k ∈ [1, c]. (19) tj , tj ∈ λ(fi ) and Qi = 0.
n 2
i=1 bi,k Definition 5: A matching function λ is said to be stable iff
it is individually rational, and is not blocked by any pair of
Next considering Bj as input, the weighted normalized deci-
agents.
sion matrix B#j , having entries ri,k
# , is formulated as
The preference profiles of all agents are complete, strict, and
  transitive. A complete preference is one where an agent a of a
#
ri,k = wk ∗ ri,k ∀i ∈ [1, n] ∀k ∈ [1, c]. (20)
set ranks all the agents of the other set. Strictness ensures that
Positive ideal solution (PSk+ ) and negative ideal solution (PSk− ) an agent a of a set is not indifferent between any two agents
from each criterion of B#j are computed as minimum and maxi- of the other set and has a concrete preference between them,
mum value of kth criterion, respectively. Let di+ and di− denote implying the absence of ties. On the other hand, transitivity
the distance of alternative i from positive and negative ideal implies that if an agent a of a set has preferences of the type
solutions, respectively, which are derived as per b a c and c a d over elements b, c, d of the other set, then
 we can safely conclude that b a d. Task tj assigns preferences
 c  2 ∀fi ∈ F based on performance score Pi . Therefore

di+ =  # − PS+
ri,k ∀i[1, n] (21)
k fi tj f i ⇐⇒ Pi > Pi .
k=1
 Likewise fi assigns preference ∀tj ∈ T such that
 c  2

di = 
− # − PS−
ri,k k ∀i[1, n]. (22) tj fi t j , ⇐⇒ Pj > Pj .
k=1
The overall working of the matching algorithm is depicted
The performance score of alternative i, denoted by Pi , is cal- in Algorithm 3. Note that METO is based on the classical
culated using the following equation. The alternatives are then deferred acceptance algorithm (DAA) [44]. The input to the
ranked in decreasing order of their performance scores algorithm is the set of agents T and F, and the preferences
of each agent a, denoted by P(a). The algorithm generates as
di− output a stable matching through λ. Initially all the tasks and
Pi = . (23)
di− + di+ FNs are set to be free and available for allocation. Steps 2–4
involve sending proposals for any unassigned task tj to its
most preferred FN fi∗ to which it has not proposed yet. For
D. Efficient Task Offloading Using Matching Game every FN that receives at least one proposal form any task tj
Taking as inputs the preferences obtained using TOPSIS, performs the following operations:
we formulate a one-to-many matching game between the set 1) if its quota is not full, it performs immediate assignment
of tasks T and set of FNs F. The matching algorithm runs of tj and updates its quota (steps 6 and 7);
at the FNC. The agents distributively generate and dispatch 2) if the FN has exhausted its quota then two cases may
their preferences to the FNC via a secure channel [42], [43]. arise:
Formally, the matching game is expressed as per Definition 1. a) FN finds a replacement task tj such that tj fi tj ,
Definition 1: Let T and F be two sets of agents. A matching where tj is the least preferred task in λ(fi )
game defined over (T, F) has two preference relations tj and (steps 9–11);

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on January 01,2024 at 07:21:44 UTC from IEEE Xplore. Restrictions apply.
 SWAIN et al.: METO: MATCHING-THEORY-BASED EFFICIENT TASK OFFLOADING IN IoT-FOG INTERCONNECTION NETWORKS 12711

TABLE II
Algorithm 3: Proposed Task Offloading Algorithm K EY PARAMETER VALUE [6], [9], [22], [46]
(METO)
Input: T, F, P(a) ∀a ∈ T ∪ F
umber
Result: λ:T ∪ F → 2T∪F
1 Initialize all tj ∈ T and fi ∈ F as free.
2 while ∃ tj such that tj is free, and P(tj )  = ∅ do
3 fi∗ = highest ranked FN in P(tj ) to which tj has not
proposed yet.
4 Send proposal to fi∗ .
5 if Qi∗ > 0 then
6 Assign tj to fi∗ .
7 Update quota as Qi∗ = Qi∗ − 1
8 else
9 Find a task tj in λ(fi∗ ) such that tj fi∗ tj
10 if ∃tj then
11 Remove and free tj from λ(fi∗ ) and add tj to
λ(fi∗ ).
12 else
13 Reject tj .

b) FN does not find a replacement task and rejects

task tj (step 13). Fig. 2. System energy versus the number of tasks.

E. Time Complexity Analysis assumed to have an active uplink and downlink channel of
The overall time complexity of METO is dependent on 10 MHz with each FN. Considering PCS-1900 GSM band,
three aspects: 1) weight determination using CRITIC, which the free space path loss in dB between an IoT device dj and
incurs O(n2 ) time [40]; 2) ranking based on TOPSIS that FN fi is calculated as PLj,i = 38.02 + 20log(dj,i ), where dj,i is
experiences a time complexity of O(n2 ) [40]; and 3) resource the distance between dj and fi [26]. The channel gain is then
allocation via one-to-many matching game that incurs O(m∗n) calculated as hj,i = 10−PLj,i/10 . The communication channel
time. Therefore the overall time complexity of METO can be is assumed to be noisy with noise power w0 = 10−10 W.
expressed as O(n2 ) + O(n2 ) + O(m ∗ n), which is polynomial. To evaluate the proficiency of METO, simulations are carried
out in two different offloading scenarios, viz., 1) heteroge-
V. P ERFORMANCE E VALUATION neous: where no two tasks have same parameter values and
2) homogeneous: where tasks have same parameter values.
To establish the superiority of METO, the performance is
compared with three baseline algorithms. The details of the
baseline algorithms are presented in Section V-B. All simula- B. Baselines Algorithms
tions have been carried out using the iFogSim simulator [45]. To demonstrate the effectiveness of METO, its
performance is compared with three baseline algorithms, viz.,
A. Environmental Setup 1) Zu et al. [22] (referred to as SMETO); 2) Chiti et al. [15]
(referred to as ME); and 3) a random allocation strategy
For the purpose of simulation, the key parameter values are
(referred to as RANDOM). Both SMETO and ME formulate
set in accordance with Table II. The fog network comprises
the task offloading problem as a matching game. The former
five heterogeneous FNs that are statically deployed across a
aims to reduce the energy consumption of IoT devices
200 m × 200 m square area. The computational capabili-
whereas the latter is focused on reducing the overall latency
ties of FNs are expressed in the form of VRUs generated
in a densely connected IoT-fog environment. For simulation
following the uniform distribution range of U[50, 500]. The
the caching delay in ME is ignored and the preferences from
computational rate (cycles/s) and computational power (W) are
IoT devices and FNs are set as per minimum completion
generated in the range of U[6, 10] GHz and U[0.35, 0.55] W.
and deadline, respectively. The RANDOM allocation strategy
The number of IoT devices varies in the range of 250–1000
randomly assigns tasks to FNs that adhere to the constraints
at an interval of 250. Each IoT device is assumed to gener-
expressed in (13)–(15).
ate a task that is offloaded to one of the FNs in the network.
The task related parameters, such as input size, output size,
computational demand, and deadline are generated uniformly C. Simulation Results
in the range U[300, 600] kb, U[10, 20] kb, U[210, 480] mil- The performance of METO is evaluated from two different
lion cycles, and U[30, 60] s respectively. Each IoT device is perspectives: 1) the system performance based on its ability

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on January 01,2024 at 07:21:44 UTC from IEEE Xplore. Restrictions apply.
 12712 IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO. 16, AUGUST 15, 2021

Fig. 3. Mean completion time versus the number of tasks. Fig. 5. System energy versus the number of tasks.

Fig. 4. Mean execution time versus the number of tasks. Fig. 6. Mean completion time versus the number of tasks.

to achieve a tradeoff between energy and latency and 2) the

performance considering other metrics, such as resource uti-
lization and outages.
1) System Performance: Fig. 2 depicts the total energy
consumption for offloading [250, 500, 750, 1000] number of
heterogeneous tasks. It can be observed from Fig. 2 that
METO outperforms all the baseline algorithms. In SMETO,
the preferences of IoT devices for FNs are generated based
on their energy consumption in transmitting the tasks and
receiving the results, i.e., nearer the FN higher is its prefer-
ence. The FNs are also energy conscious and prefer tasks that Fig. 7. Mean execution time versus the number of tasks.
consume less energy for execution, i.e., tasks having smaller
computational demands are preferred. In two sided matching completion, the FNs also consider the deadline of the tasks
markets the DAA favors the proposers and is not optimal from while generating preferences. The minimum completion time
receivers’ perspective [47]. In SMETO, as IoT devices are is mainly attributed to faster execution of tasks which reflects
the proposers, the solution tends to favor the preferences of lower average execution time as observed in Fig. 4. Preference
IoT devices by assigning tasks to nearer FNs. This results calculation in SMETO is solely dependent on the energy con-
in reducing the energy consumption of the IoT devices. On sumption. This implies the algorithm may map a task to an
the contrary, the FNs are not matched with their preferred FN that consumes less energy, however, with less compute
smaller tasks, thereby consuming higher energy. The prefer- capability. This adversely affects the average execution and
ences of IoT devices and FNs in ME are calculated based on completion time of the tasks. Though preference generation
minimum completion time and deadline of tasks, respectively. of ME is based on latency and deadline, the queuing delay in
ME suffers from higher energy consumption in comparison their approach forces the tasks to look for an alternate FN with
to both SMETO and METO because it does not consider it lower computational capabilities which elevates both execution
as a parameter while setting preferences from either ends. and completion time in comparison with METO. For homo-
Similarly, Fig. 5 indicates the energy consumption for homo- geneous tasks, similar nature of comparison about completion
geneous tasks. The behavior of the techniques is similar to and execution time can be inferred from Figs. 6 and 7.
that of a heterogeneous scenario. In RANDOM, the tasks are mapped to the FNs based on
Figs. 3 and 4 show the average completion and execu- randomly generated preferences. Since no evaluation criteria,
tion time of heterogeneous tasks. METO realizes improved such as energy consumption, latency, or deadline are not taken
performance with respect to these two parameters, in compar- into account by either category of agents, it falls behind all
ison to SMETO and RANDOM. This is because it considers other approaches used for comparison.
latency, in addition to energy consumption, while generat- 2) Other Performance Metrics: To further strengthen the
ing the preferences of IoT devices. To supplement faster claim of improved performance, METO is evaluated with three

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on January 01,2024 at 07:21:44 UTC from IEEE Xplore. Restrictions apply.
 SWAIN et al.: METO: MATCHING-THEORY-BASED EFFICIENT TASK OFFLOADING IN IoT-FOG INTERCONNECTION NETWORKS 12713

Fig. 8. Number of outages versus the number of tasks. Fig. 11. Number of outages versus the number of tasks.

Fig. 12. System energy versus distance.

Fig. 9. System energy versus distance.

Fig. 13. Resource utilization versus the number of tasks.

Fig. 10. Resource utilization versus the number of tasks.

generating preference in the considered offloading techniques
in this article.
additional performance metrics. Figs. 8 and 11 report the total Thus, the obtained simulation results confirm the efficiency
number outages, i.e., number of tasks overshooting the dead- of the proposed approach over the state-of-the-art approaches
line. Both SMETO and ME suffer from higher number of considering different evaluation metrics.
outages. The former suffers as it ends up assigning tasks onto
slower FNs while the latter, on account of queuing delay, VI. C ONCLUSION
forces tasks to change their preference to less capable FNs In this article, a framework called METO is proposed that
subsequently leading to higher outages. On the other hand, aims to reduce the total energy consumption and overall
the proposed technique is focused on completion time from latency incurred considering a full offloading scenario in a
the IoT perspective and deadline from the FN’s standpoint densely connected IoT-fog network. The offloading problem
resulting in reduced outages. is proven to be N P-Hard and the complexity of finding a
Figs. 9 and 12 display the effect of varying distance of feasible solution increases with increasing problem size. To
IoT devices from FNs on energy consumption, considering obtain a suboptimal solution in polynomial time, the offload-
250 devices. It can be observed that with increasing distance, ing problem is formulated as a one-to-many matching game
energy consumption increases as higher energy is spent toward between IoT devices and FNs. As METO considers multiple
transmission and reception. criteria in decision making, a hybrid CRITIC and TOPSIS-
Figs. 10 and 13 show the average resource utilization of based technique is employed to generate the preferences of
FNs considering homogeneous and heterogeneous environ- both stakeholders. Furthermore, the classical DAA is used
ments. The algorithms show indifferent utilization levels for to find a stable assignment where no agent has incentive to
different scenarios considered for simulation. The resource uti- deviate from its current allocation. To validate the efficiency
lization of the FNs is directly dependent on the number of of METO, the performance is efficient against three different
VRUs used and is independent of any of the criteria used for baselines considering homogeneous and heterogeneous tasks.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on January 01,2024 at 07:21:44 UTC from IEEE Xplore. Restrictions apply.
 12714 IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO. 16, AUGUST 15, 2021

Thorough simulation analysis across both environments con- [21] X. Chen, Y. Cai, Q. Shi, M. Zhao, and G. Yu, “Energy-efficient resource
firm that the proposed algorithm outperforms the existing allocation for latency-sensitive mobile edge computing,” in Proc. IEEE
88th Veh. Technol. Conf. (VTC-Fall), 2018, pp. 1–5.
schemes with respect to improved energy consumption, com- [22] Y. Zu, F. Shen, F. Yan, L. Shen, F. Qin, and R. Yang, “SMETO: Stable
pletion time, and execution time. As METO also reduces the matching for energy-minimized task offloading in cloud-fog networks,”
number of outages, it is suitable for offloading real-time tasks. in Proc. IEEE 90th Veh. Technol. Conf. (VTC-Fall), 2019, pp. 1–5.
[23] Z. Zhou, M. Dong, K. Ota, G. Wang, and L. T. Yang, “Energy-efficient
Thus, METO stands as a de-facto choice for offloading tasks in resource allocation for D2D communications underlaying cloud-RAN-
the IoT-fog interconnection network. As an immediate future based LTE-A networks,” IEEE Internet Things J., vol. 3, no. 3,
direction, we would like to model the offloading problem in pp. 428–438, Nov. 2015.
[24] H. Xu and B. Li, “ANCHOR: A versatile and efficient framework for
the partial offloading environment with dynamic preferences. resource management in the cloud,” IEEE Trans. Parallel Distrib. Syst.,
vol. 24, no. 6, pp. 1066–1076, Oct. 2012.
[25] P. Hu, H. Ning, T. Qiu, H. Song, Y. Wang, and X. Yao, “Security and
R EFERENCES privacy preservation scheme of face identification and resolution frame-
[1] S. Curtis, B. Zafar, A. Gutub, and D. Manocha, “Right of way,” Vis. work using fog computing in Internet of Things,” IEEE Internet Things
Comput., vol. 29, no. 12, pp. 1277–1292, 2013. J., vol. 4, no. 5, pp. 1143–1155, Oct. 2017.
[2] S. Kim et al., “Velocity-based modeling of physical interactions in dense [26] X. Chen and J. Zhang, “When D2D meets cloud: Hybrid mobile task
crowds,” Vis. Comput., vol. 31, no. 5, pp. 541–555, 2015. offloadings in fog computing,” in Proc. IEEE Int. Conf. Commun. (ICC),
[3] Y. Sun, H. Song, A. J. Jara, and R. Bie, “Internet of Things and big data 2017, pp. 1–6.
analytics for smart and connected communities,” IEEE Access, vol. 4, [27] M. Mukherjee, M. Guo, J. Lloret, R. Iqbal, and Q. Zhang, “Deadline-
pp. 766–773, 2016. aware fair scheduling for offloaded tasks in fog computing with inter-
[4] N. Farooqi, A. Gutub, and M. O. Khozium, “Smart community chal- fog dependency,” IEEE Commun. Lett., vol. 24, no. 2, pp. 307–311,
lenges: Enabling IoT/M2M technology case study,” Life Sci. J., vol. 16, Feb. 2020.
no. 7, pp. 11–17, 2019. [28] F. Chiti, R. Fantacci, F. Paganelli, and B. Picano, “Virtual functions
[5] M. Adhikari, M. Mukherjee, and S. N. Srirama, “DPTO: A deadline and placement with time constraints in fog computing: A matching the-
priority-aware task offloading in fog computing framework leveraging ory perspective,” IEEE Trans. Netw. Service Manag., vol. 16, no. 3,
multi-level feedback queueing,” IEEE Internet Things J., vol. 7, no. 7, pp. 980–989, May 2019.
pp. 5773–5782, Jul. 2020. [29] F. Chiti, R. Fantacci, and B. Picano, “A matching game for tasks
[6] E. E. Haber, T. M. Nguyen, D. Ebrahimi, and C. Assi, “Computational offloading in integrated edge-fog computing systems,” Trans. Emerg.
cost and energy efficient task offloading in hierarchical edge-clouds,” in Telecommun. Technol., vol. 31, no. 2, 2020, Art. no. e3718.
Proc. IEEE 29th Annu. Int. Symp. Pers. Indoor Mobile Radio Commun. [30] Y. Liu, F. R. Yu, X. Li, H. Ji, and V. C. M. Leung, “Hybrid computa-
(PIMRC), 2018, pp. 1–6. tion offloading in fog and cloud networks with non-orthogonal multiple
[7] Z. Liu, X. Yang, Y. Yang, K. Wang, and G. Mao, “DATS: Dispersive access,” in Proc. IEEE Conf. Comput. Commun. Workshops (INFOCOM
stable task scheduling in heterogeneous fog networks,” IEEE Internet WKSHPS), 2018, pp. 154–159.
Things J., vol. 6, no. 2, pp. 3423–3436, Dec. 2018. [31] N. AlAssaf, B. AlKazemi, and A. Gutub, “Applicable light-weight cryp-
[8] D. Zhang, Y. Ma, Y. Zhang, S. Lin, X. S. Hu, and D. Wang, “A real-time tography to secure medical data in IoT systems,” J. Res. Eng. Appl. Sci.,
and non-cooperative task allocation framework for social sensing appli- vol. 2, no. 2, pp. 50–58, 2003.
cations in edge computing systems,” in Proc. IEEE Real-Time Embedded [32] N. Alassaf, A. Gutub, S. A. Parah, and M. Al Ghamdi, “Enhancing
Technol. Appl. Symp. (RTAS), 2018, pp. 316–326. speed of SIMON: A light-weight-cryptographic algorithm for IoT appli-
[9] G. Zhang, F. Shen, Y. Yang, H. Qian, and W. Yao, “Fair task offloading cations,” Multimedia Tools Appl., vol. 78, no. 23, pp. 32633–32657,
among fog nodes in fog computing networks,” in Proc. IEEE Int. Conf. 2019.
Commun. (ICC), 2018, pp. 1–6. [33] A. Gutub and F. Al-Shaarani, “Efficient implementation of multi-image
[10] R. K. Naha et al., “Fog computing: Survey of trends, architec- secret hiding based on LSB and DWT steganography comparisons,”
tures, requirements, and research directions,” IEEE Access, vol. 6, Arab. J. Sci. Eng., vol. 45, pp. 2631–2644, Feb. 2020.
pp. 47980–48009, 2018. [34] B. Gu, Z. Zhou, S. Mumtaz, V. Frascolla, and A. K. Bashir, “Context-
[11] F. Bonomi, R. Milito, J. Zhu, and S. Addepalli, “Fog computing and its aware task offloading for multi-access edge computing: Matching with
role in the Internet of Things,” in Proc. 1st Ed. MCC Workshop Mobile externalities,” in Proc. IEEE Global Commun. Conf. (GLOBECOM),
Cloud Comput., 2012, pp. 13–16. 2018, pp. 1–6.
[12] F. Haider, D. Zhang, M. St-Hilaire, and C. Makaya, “On the planning [35] X. Lyu, H. Tian, C. Sengul, and P. Zhang, “Multiuser joint task offload-
and design problem of fog computing networks,” IEEE Trans. Cloud ing and resource optimization in proximate clouds,” IEEE Trans. Veh.
Comput., early access, Oct. 8, 2018, doi: 10.1109/TCC.2018.2874484. Technol., vol. 66, no. 4, pp. 3435–3447, Jun. 2017.
[13] S. A. Aly, T. A. AlGhamdi, M. Salim, H. H. Amin, and A. A. Gutub, [36] X. Chen, “Decentralized computation offloading game for mobile
“Information gathering schemes for collaborative sensor devices,” cloud computing,” IEEE Trans. Parallel Distrib. Syst., vol. 26, no. 4,
Procedia Comput. Sci., vol. 32, pp. 1141–1146, Dec. 2014. pp. 974–983, Apr. 2015.
[14] S. A. Aly, T. A. Alghamdi, M. Salim, and A. A. Gutub, “Data dissem- [37] R. K. Dewi, B. T. Hanggara, and A. Pinandito, “A comparison between
ination and collection algorithms for collaborative sensor devices using AHP and hybrid AHP for mobile based culinary recommendation
dynamic cluster heads,” Trends Appl. Sci. Res., vol. 8, no. 2, p. 55, 2013. system,” Int. J. Interact. Mobile Technol., vol. 12, no. 1, pp. 133–140,
[15] F. Chiti, R. Fantacci, and B. Picano, “A matching theory framework 2018.
for tasks offloading in fog computing for IoT systems,” IEEE Internet [38] C. J. Slebi-Acevedo, P. Pascual-Muñoz, P. Lastra-González, and
Things J., vol. 5, no. 6, pp. 5089–5096, Sep. 2018. D. Castro-Fresno, “Multi-response optimization of porous asphalt mix-
[16] A. Yousefpour, G. Ishigaki, R. Gour, and J. P. Jue, “On reducing IoT tures reinforced with aramid and polyolefin fibers employing the
service delay via fog offloading,” IEEE Internet Things J., vol. 5, no. 2, CRITIC-TOPSIS based on taguchi methodology,” Materials, vol. 12,
pp. 998–1010, Jan. 2018. no. 22, p. 3789, 2019.
[17] A. Abouaomar, A. Kobbane, and S. Cherkaoui, “Matching-game [39] D. Diakoulaki, G. Mavrotas, and L. Papayannakis, “Determining objec-
for user-fog assignment,” in Proc. IEEE Global Commun. Conf. tive weights in multiple criteria problems: The critic method,” Comput.
(GLOBECOM), 2018, pp. 1–6. Oper. Res., vol. 22, no. 7, pp. 763–770, 1995.
[18] T. Kan, Y. Chiang, and H. Wei, “Task offloading and resource allocation [40] R. Hamdani and R. Wardoyo, “The complexity calculation for group
in mobile-edge computing system,” in Proc. 27th Wireless Opt. Commun. decision making using TOPSIS algorithm,” AIP Conf., vol. 1755, no. 1,
Conf. (WOCC), 2018, pp. 1–4. 2016, Art. no. 070007.
[19] H. Q. Le, H. Al-Shatri, and A. Klein, “Efficient resource allocation in [41] M. Mukherjee, L. Shu, and D. Wang, “Survey of fog computing:
mobile-edge computation offloading: Completion time minimization,” in Fundamental, network applications, and research challenges,” IEEE
Proc. IEEE Int. Symp. Inf. Theory (ISIT), 2017, pp. 2513–2517. Commun. Surveys Tuts., vol. 20, no. 3, pp. 1826–1857, 3rd Quart., 2018.
[20] H. Bao and Y. Liu, “A two-sided matching approach for distributed edge [42] N. Alassaf and A. Gutub, “Simulating light-weight-cryptography imple-
computation offloading,” in Proc. IEEE/CIC Int. Conf. Commun. China mentation for IoT healthcare data security applications,” Int. J. e-Health
(ICCC), 2019, pp. 535–540. Med. Commun., vol. 10, no. 4, pp. 1–15, 2019.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on January 01,2024 at 07:21:44 UTC from IEEE Xplore. Restrictions apply.
 SWAIN et al.: METO: MATCHING-THEORY-BASED EFFICIENT TASK OFFLOADING IN IoT-FOG INTERCONNECTION NETWORKS 12715

[43] M. Alotaibi, D. Al-Hendi, B. Alroithy, M. AlGhamdi, and A. Gutub, Sambit Bakshi (Senior Member, IEEE) received the
“Secure mobile computing authentication utilizing hash, cryptography Ph.D. degree in computer science from the National
and steganography combination,” J. Inf. Security Cybercrimes Res., Institute of Technology, Rourkela, India, in 2015.
vol. 2, no. 1, pp. 1–13, 2019. He is currently with the Centre for Computer
[44] A. E. Roth, “Deferred acceptance algorithms: History, theory, prac- Vision and Pattern Recognition, National Institute
tice, and open questions,” Int. J. Game Theory, vol. 36, nos. 3–4, of Technology Rourkela, where he also serves as
pp. 537–569, 2008. an Assistant Professor with the Department of
[45] H. Gupta, A. V. Dastjerdi, S. K. Ghosh, and R. Buyya, “iFogSim: A Computer Science and Engineering.
toolkit for modeling and simulation of resource management techniques Dr. Bakshi received the Prestigious Innovative
in the Internet of Things, edge and fog computing environments,” Softw. Student Projects Award in 2011 from Indian
Pract. Exp., vol. 47, no. 9, pp. 1275–1296, 2017. National Academy of Engineering for his master’s
[46] S. Bi and Y. J. Zhang, “Computation rate maximization for wireless pow- thesis. He serves as an Associate Editor for the International Journal of
ered mobile-edge computing with binary computation offloading,” IEEE Biometrics in 2013. He is a Technical Committee Member of the IEEE
Trans. Wireless Commun., vol. 17, no. 6, pp. 4177–4190, Apr. 2018. Computer Society Technical Committee on Pattern Analysis and Machine
[47] D. Gale and L. S. Shapley, “College admissions and the stability of Intelligence.
marriage,” Amer. Math. Monthly, vol. 69, no. 1, pp. 9–15, 1962.

Chittaranjan Swain (Graduate Student Member, Joel J. P. C. Rodrigues (Fellow, IEEE) received
IEEE) received the M.Tech. degree in computer the five-year B.Sc. degree (licentiate) in computer
science from Vellore Institute of Technology, science and engineering from the University of
Vellore, India, in 2015. Coimbra, Coimbra, Portugal, in 1995, the M.Sc.
He is currently a Research Scholar with the degree and the Ph.D. degree in computer science
Department of Computer Science and Engineering, and engineering from the University of Beira Interior
National Institute of Technology Rourkela, (UBI), Covilhã, Portugal, in 2002 and 2006, respec-
Rourkela, India. His research interests include fog tively, and the Habilitation degree in computer sci-
computing, cloud computing, and IoT. ence and engineering from the University of Haute
Alsace, Mulhouse, France, in 2014.
He received the Academic Title of Aggregated
Manmath Narayan Sahoo (Member, IEEE) Professor in computer science and engineering from UBI in 2015. He is
received the Ph.D. degree in computer science a Professor with the Federal University of Piauí, Teresina, Brazil; a Senior
from the National Institute of Technology Rourkela, Researcher with the Instituto de Telecomunicações, Covilhã; and a collab-
Rourkela, India, in 2014. orator of the Post-Graduation Program on Teleinformatics Engineering with
He is an Assistant Professor with the Department the Federal University of Ceará (UFC), Fortaleza, Brazil. He has authored or
of Computer Science and Engineering, National coauthored over 900 papers in refereed international journals and conferences,
Institute of Technology Rourkela. His research three books, two patents, and one ITU-T Recommendation.
interests include mobile ad hoc networks, fault tol- Prof. Rodrigues had been awarded several Outstanding Leadership and
erance, sensor networks, and cloud computing. Outstanding Service Awards by IEEE Communications Society and several
Dr. Sahoo has served as a reviewer, a guest edi- best papers awards. He is the Leader of the Next Generation Networks
tor, the track chair, and the program chair in many and Applications Research Group (CNPq), an IEEE Distinguished Lecturer,
reputed journals and conferences. a Member Representative of the IEEE Communications Society on the
IEEE Biometrics Council, and the President of the scientific council at
Anurag Satpathy (Graduate Student Member, ParkUrbis—Covilhã Science and Technology Park. He was Director for
IEEE) received the B.Tech. degree in information Conference Development—IEEE ComSoc Board of Governors, Technical
technology from the International Institute Activities Committee Chair of the IEEE ComSoc Latin America Region
of Information Technology Bhubaneswar, Board, the Past-Chair of the IEEE ComSoc Technical Committee on eHealth,
Bhubaneswar, India, in 2014, and the M.Tech. the Past-Chair of the IEEE ComSoc Technical Committee on Communications
degree in computer science from Birla Institute of Software, a Steering Committee member of the IEEE Life Sciences Technical
Technology Mesra, Ranchi, India, in 2017. Community and Publications co-Chair. He is the Editor-in-Chief of the
He is currently a Research Scholar with the International Journal of E-Health and Medical Communications and edi-
Department of Computer Science and Engineering, torial board member of several high-reputed journals. He has been the
National Institute of Technology Rourkela, general chair and TPC Chair of many international conferences, includ-
Rourkela, India. His current research area includes ing IEEE ICC, IEEE GLOBECOM, IEEE HEALTHCOM, and IEEE
cloud computing, Internet of Things, and distributed systems. LatinCom. He is a member of the Internet Society, and a Senior Member
of ACM.
Khan Muhammad (Member, IEEE) received the
Ph.D. degree in digital contents from Sejong
University, Seoul, South Korea, in 2018. Victor Hugo C. de Albuquerque (Senior Member,
He is currently working as an Assistant Professor IEEE) received the graduation degree in mecha-
with the Department of Software and a Lead tronics engineering from the Federal Center of
Researcher with Intelligent Media Laboratory, Technological Education of Ceará (CEFETCE),
Sejong University. His research interests include Fortaleza, Brazil, in 2006, the M.Sc. degree
intelligent video surveillance (fire/smoke scene anal- in teleinformatics engineering from the Federal
ysis, transportation systems, and disaster manage- University of Ceará, Fortaleza, in 2008, and the
ment), medical image analysis, (brain MRI, diagnos- Ph.D. degree in mechanical engineering from the
tic hysteroscopy, and wireless capsule endoscopy), Federal University of Pará, Belém, Brazil, in
information security (steganography, encryption, watermarking, and image 2010.
hashing), video summarization, multimedia, computer vision, IoT, and smart He is a Professor and a Senior Researcher with
cities. He has filed/published over seven patents and 100 papers in peer- the LAPISCO/IFCE and ARMTEC Tecnologia em Robótica, Brazil. He
reviewed journals and conferences in the above areas. is a Full Professor and a Senior Researcher with the University of
Dr. Muhammad is serving as a reviewer for over 70 well-reputed jour- Fortaleza, UNIFOR, Fortaleza, and the Data Science Director with the
nals and conferences, from IEEE, ACM, Springer, Elsevier, Wily, SAGE, and Superintendency for Research and Public Safety Strategy of Ceará State
Hindawi publishers. He acted as a TPC member and the session chair at more (SUPESP/CE), Fortaleza. He is currently a Professor of the Graduate
than 10 conferences in related areas. He is the Editorial Board Member of Program in Applied Informatics of UNIFOR and the Leader of the Industrial
the Journal of Artificial Intelligence and Systems and a Review Editor for Informatics, Electronics and Health Research Group (CNPq). He is a spe-
the Section Mathematics of Computation and Data Science and Frontiers in cialist, mainly, in IoT, machine/deep learning, pattern recognition, and
Applied Mathematics and Statistics. robotic.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on January 01,2024 at 07:21:44 UTC from IEEE Xplore. Restrictions apply.