Journal of Systems Architecture 118 (2021) 102225

Contents lists available at ScienceDirect

Journal of Systems Architecture

journal homepage: www.elsevier.com/locate/sysarc

A Survey on Task Offloading in Multi-access Edge Computing

Akhirul Islam a ,∗, Arindam Debnath a , Manojit Ghose a , Suchetana Chakraborty b
a
Department of Computer Science and Engineering, IIIT Guwahati, India
b
Department of Computer Science and Engineering, IIT Jodhpur, India

ARTICLE INFO ABSTRACT

Keywords: With the advent of new technologies in both hardware and software, we are in the need of a new type of
Multi-access edge computing application that requires huge computation power and minimal delay. Applications such as face recognition,
Task offloading augmented reality, virtual reality, automated vehicles, industrial IoT, etc. belong to this category. Cloud
Mobile edge computing
computing technology is one of the candidates to satisfy the computation requirement of resource-intensive
Survey
applications running in UEs (User Equipment) as it has ample computational capacity, but the latency
requirement for these applications cannot be satisfied by the cloud due to the propagation delay between
UEs and the cloud. To solve the latency issues for the delay-sensitive applications a new network paradigm
has emerged recently known as Multi-Access Edge Computing (MEC) (also known as mobile edge computing)
in which computation can be done at the network edge of UE devices. To execute the resource-intensive tasks
of UEs in the MEC servers hosted in the network edge, a UE device has to offload some of the tasks to MEC
servers. Few survey papers talk about task offloading in MEC, but most of them do not have in-depth analysis
and classification exclusive to MEC task offloading. In this paper, we are providing a comprehensive survey
on the task offloading scheme for MEC proposed by many researchers. We will also discuss issues, challenges,
and future research direction in the area of task offloading to MEC servers.

1. Introduction MDs and save critical energy. However, cloud servers introduce a good
amount of delay and jitter due to long propagation delay between
In recent years mobile devices (MDs) have become an indispensable MDs and cloud servers because of their geographical location [5]. Task
part of human lives. Rapid growth in sensing and communication completion time can also be increased further if the MDs offload all
technologies has recently opened up the possibilities of numerous ap- their tasks along with the generated data to the cloud by congesting
plications towards realizing the true digital age [1]. With the increasing the backhaul network. Multi-Access Edge Computing (MEC) is a new
trend of smartphone usage, the usability of mobile devices (MD) is not
computing paradigm that has recently been introduced in practice to
limited to serve over-the-wire connectivity anymore. Smart applications
address these problems [6–8]. The basic principle of MEC is to bring
for gaming, social media interaction, business, infotainment, or even
the computation power of MCC to the edge of a network close to
the basic utilities heavily use the in-built sensors including camera
and microphone to offer customized services. On-device processing and the MDs [9]. The applications running on MDs can now offload the
storage of these huge amounts of multimodal sensory data are not compute-intensive tasks to the nearby MEC servers at low overhead,
feasible considering the limited resources of the mobile devices. Hence, which can significantly alleviate the problem of network congestion
resource-intensive mobile applications such as real-time online games, and improve the response time for the running applications [10–12].
virtual reality (VR) and augmented reality, image processing, voice, MEC servers are deployed at the edge of the network in cellular base
gesture or face recognition, or other similar user-adaptive services pose stations or small cells or both [13]. A wide range of mobile applications
critical design challenges in achieving the desired performance for nowadays tends to avail the MEC services for providing a better quality
resource-constrained MDs [2]. of experience (QoE) to the users. Emerging applications like distributed
Until now Mobile Cloud Computing (MCC) was considered to be content delivery and caching, web performance enhancements, context-
an efficient solution to this problem [3] as the cloud servers have
awareness, content optimization, and various smart city services in
huge computation powers and storage resources [4], so offloading
the domains of healthcare, transportation, energy, manufacturing, and
computation-intensive tasks to cloud servers reduces the burden of

∗ Corresponding author.
E-mail addresses: [email protected] (A. Islam), [email protected] (A. Debnath), [email protected] (M. Ghose), [email protected]
(S. Chakraborty).

https://doi.org/10.1016/j.sysarc.2021.102225
Received 17 April 2021; Received in revised form 18 June 2021; Accepted 20 June 2021
Available online 24 June 2021
1383-7621/© 2021 Elsevier B.V. All rights reserved.
A. Islam et al. Journal of Systems Architecture 118 (2021) 102225

industrial automation [1,14] heavily rely on MEC infrastructure for task within the MEC, and mobility management. They have classified the
offloading. task offloading mainly on two parameters i.e. execution delay and en-
The main objective of task offloading is to minimize the task ex- ergy consumption. Authors in [33], mainly focused on task offloading
ecution time of applications running at UEs and to save energy con- based on the multi-task learning approach. Authors in [34] have tried
sumption of UEs. The problem of efficient task offloading is multi-fold to classify the existing work on task offloading based on 1. gaming
as MEC servers are not as resourceful as the cloud servers, so task and cooperation between edge and the cloud and 2. heuristic based
offloading from UEs to the MEC servers has to be done in such a way computation offloading. The survey [35] presents an overview of the
that the resources of the MEC servers are utilized judiciously. The MEC task offloading technology with the collaboration of edges and clouds.
servers hosted on different BS may not be equipped with the same set Authors in [36] highlighted the importance of offloading modeling and
of services, and the computation load in the MEC servers may also explore future research trends in this area. The survey [37] reviews
vary dynamically with time. Optimal resource allocation for offloaded task offloading methods for edge computing environments into five
tasks, fair and stable load distribution among the MEC servers, and categories i.e. offloading destination, load balance of edge servers,
seamless coordination between the cloud and edge are only one facet device mobility, application partitioning, and partition granularity. The
of the problem to utilize the MEC services effectively [15]. The other authors in [38] have reviewed the computation offloading approaches
challenge remains in understanding the dimension of the designated based on game-theoretic approach for edge computing. A summary of
tasks to offload. Depending on the required QoE parameters for the the above survey papers is presented in Table 2. Most of the surveys
running applications, the tasks can also vary in terms of priorities, mentioned above mainly focused on the breadth of different topics on
progress, computing overhead, and other associated dependencies [16]. MEC, but these literature lacks in-depth analysis and classification of
The selection, scheduling, and placement of tasks consist of the other MEC task offloading. We are trying to make an in-depth classification
set of challenges associated with the problem of task offloading. Thus of task offloading and the architectural model used for the task of-
the problem of optimal task offloading boils down to selecting the floading which is different from the above surveys. We also discuss the
right task to offload to the right MEC server at the right time to meet associated design challenges, practical use cases, and potential research
the optimal performance of a running application while utilizing the directions in this field.
MEC resources efficiently. This is clearly a multi-objective optimization The rest of the paper is organized as follows. Section 2 presents a
problem and known to be NP-hard [17,18]. formal definition of the task-offloading problem in MEC and evaluation
Task offloading in the context of our discussion can be defined metrics for the task offloading. In Section 3 we describe different system
as uploading a complete module of an application that includes the models under three broad categories and Section 4 is about different
computation, required data, and other dependent libraries to a remote offloading strategies. Real-life use cases of MEC are briefed in Section 5,
server and obtaining the result of computation from the remote server. while technical challenges and future research direction are described
While discussing the task offloading, it is also important to define in Section 6 and finally, Section 7 concludes the whole paper. We have
scheduling as it is an essential part of task offloading. Scheduling is a summarized the definitions of the acronyms that will be frequently used
method that executes a set of tasks (or subtasks) on a set of computing in this paper in Table 1 for ease of reference.
resources to optimize a goal [19]. One of the primary objectives of
scheduling is to maximize the utilization of resources while minimizing 2. Task offloading
the total task execution time [20]. In our discussion, scheduling can
happen at UEs, the MEC server, or on the cloud in the case of collab- The principal objective of MEC lies in supporting the resource con-
orative MEC system models. A UE can either schedule a task for local strained mobile devices with extended cloud services at the edge of the
execution or offloads it to a MEC server. A MEC server schedules the network. This is particularly helpful for emerging delay-sensitive and
task for execution locally or forwards the task to the cloud or other high compute-intensive mobile applications as they can save critical
MEC server depending on different system models. A similar concept resources through offloading few tasks to the nearby MEC hosts. The
as task offloading is computation offloading which means sending only key objectives of task offloading are to improve QoE by reducing
the compute-intensive part of a module to a remote server and receiving computational delay, to extend the UE lifetime by reducing device
the result of computation from the remote server. energy consumption [11], and to maximize the revenue for the network
The state-of-the-art literature is quite rich in providing various operators and service providers.
solution strategies to address the stated problem of MEC task offload-
ing [21–23]. In the absence of standardization, the existing solutions 2.1. Task: Definition and model
vary widely in the assumption of architecture and system model, design
of task model, consideration of the objectives, and use of an offloading A task can be defined as a basic unit of work to be executed towards
mechanism to name a few. Over the last couple of years MEC has the accomplishment of an application service. When an application runs
been emerging as a key enabling technology towards the realization on mobile devices, it accomplishes a series of tasks, sometimes in a
of IoT and 5G visions [24–26]. MEC research lies at the intersection loop, to provide the desired service. From a computer programming
of mobile computing and wireless communications, which naturally perspective, a task is a logical unit of a program containing a bunch
opened up the possibilities of many research opportunities. In recent of instructions tightly or loosely coupled to implement specific func-
years, researchers from both academia and industry have significantly tionality. Depending on how the application program is developed, a
contributed to addressing various challenges related to MEC, including task may be an entire program or a set of one or more functions in a
system and network modeling, resource allocation, implementation, program. Since one user program may avail the services of other utility
and standardization. programs, a utility program may also be considered as a separate task
MEC network paradigm is still not a mature technology, and re- (or sub-tasks). For example, an operation like matrix multiplication,
search in this direction is also in the early stage. There are few survey derivation of a complex function, evaluating a complex mathematical
publications [1,16,27–31] available related to MEC which mainly fo- expression, searching, sorting, or executing a specific algorithm can be
cused on MEC taxonomy, future research directions and more specific treated as a task.
MEC attributes like communication, computation offloading, resource A task  , can be formally defined as a vector (𝐼, 𝑆, 𝐶), where 𝐼
allocation, security, virtualization, and applications. Though these arti- is the set of instructions, 𝑆 is the input size in terms of bits, and 𝐶
cles talk about computation offloading, they described the task offload- denotes the set of CPU cycles needed by the task. Each 𝐶𝑖 represents
ing very briefly. In the survey [32], the authors have focused mainly on the CPU cycle requirement for the corresponding instruction 𝐼𝑖 in a
decisions on computation offloading, allocation of computing resources task [39]. The order of instructions to be executed to accomplish a task

2
A. Islam et al. Journal of Systems Architecture 118 (2021) 102225

Table 1
Summary of important acronyms.
Acronym Definition Acronym Definition
MD Mobile Device MNO Mobile Network Operator
UE User Equipment ACCO Approximation Collaborative Computation
Offloading
MEC Multi-access Edge Computing SDN Software Defined Network
MCC Mobile Cloud Computing DRL Deep Reinforcement Learning
QoE Quality of Experience MRL Meta Reinforcement Learning
QoS Quality of Service MBS Macro Base Station
DAG Directed Acyclic Graph eNB Evolved Node B
BS Base Station HGOS Heuristic Greedy Offloading Scheme
SC Small Cell MIP Mixed Integer Program
KKT Karush–Kuhn–Tucker MINLP Mixed Integer Nonlinear Programming
IOSDTO Iterative Optimization of Service Deploying and BBM Branch and Bound Method
Task Offloading
RNIS Radio Network Information Service DPOA Dynamic Programming based Offloading Algorithm
DTOS Dynamic Task Offloading and Scheduling IPM Interior Point Method
MRLCO Meta Reinforce Learning Computation offloading MDP Markov Decision Processes
VPN Virtual Private Network BnB Branch and Bound
CPE Customer Premises Equipment VM Virtual Machine
V2I Vehicle-to-infrastructure RAN Radio Access Network
MOSP Modified Online Saddle Point CCC Centralized Cloud Computing

Table 2 Moreover, a task can be partitioned into multiple sub-tasks based on

Some references of existing surveys in task offloading.
the complexity of the corresponding application design and desired
Paper Brief summary performance characteristics. These sub-tasks may either execute inde-
[32] The authors in this survey have focused mainly on decisions on pendently or depend on other tasks based on the functionality and logic
computation offloading, allocation of computing resources within of control flow for the running application program.
the MEC, and mobility management. They have classified the task Those tasks which are well integrated, highly coupled, or relatively
offloading mainly on two parameters i.e. execution delay and
simple and cannot be partitioned into multiple dependent or inde-
energy consumption.
pendent tasks, that has to be executed as a whole either locally at
[33] The authors in this survey mainly focused on task offloading based
on the multi-task learning approach the mobile devices or offloaded to MEC servers called binary offload-
[34] Tried to classify the existing work on task offloading based on 1. ing [30]. Such a task can be represented by a three-field notation
gaming and cooperation between edge and the cloud and 2.  (𝐿, 𝐷, 𝑋). This commonly used notation contains the information of
heuristic the task input data size 𝐿 (in bits), the task completion deadline 𝐷 (in
[36] This survey highlights the importance of offloading modeling and seconds), and the computation intensity 𝑋 (in CPU cycles per bit).
explores future research trends in this area. They have classified Therefore, a computation task can either be executed locally at the
exiting offloading schemes into four categories 1. offloading from
ED to EC 2. offloading from EC to cloud computing 3. offloading
UE or be offloaded to a nearby MEC host for processing. The offloading
from one EC server to another (others) 4. hierarchical offloading decision is made and the offloading process is triggered by the running
[35] This survey presents an overview of the task offloading technology application.
with the collaboration of edges and clouds. The main parameters
they have considered while classifying the task offloading are 2.2. Evaluation metrics for local execution
response time, device energy, service provider cost, load balance
between edge and cloud, and device mobility
[37] This survey reviews task offloading methods for edge computing
Let us define 𝑐 𝐿 be the current computing capability of the mobile
environments into five categories 1.offloading destination 2. load device in CPU cycles per unit time, 𝑤𝑖 be the CPU cycles needed in total
balance of edge servers 3. device mobility 4. application to complete the task 𝑖 . Thus total execution time of the task, denoted
partitioning and 5. partition granularity by 𝛿𝑖 , can be expressed as
[38] The authors have reviewed the computation offloading approaches
𝑤
based on game-theoretic for edge computing 𝛿𝑖 = 𝐿𝑖 (1)
𝑐
Also, we can obtain the energy consumption for local task process-
ing as [40]:
is dependent on the system architecture and it is out of the scope of
this paper. Other attributes associated with a task might include the 𝜓𝑖 = 𝛼𝑤𝑖 (2)
priority, deadline, the list of associated applications, and the level of
where 𝛼 is the coefficient denoting the consumed energy per CPU
dependency. Generally, each task is assigned a unique task ID so that
cycle.
every task can be uniquely identified within a UE and mapped with the
corresponding application. The level of dependency for a task defines
its offloading capacity and a less dependent task can be seamlessly 2.3. Evaluation metrics for MEC-server-side execution
offloaded to the MEC server. However, it is important to be careful
while scheduling a task for offloading without violating any operational For task offloading to the MEC server, the application QoE is charac-
constraints for high dependency level tasks. The dependency can be terized by the primary objective of reducing the average task duration
either bound to the system architecture or the execution of other tasks time with the limited battery capacity. Let us denote the computing
and a task that is closely dependent on the platform should not be resource (CPU cycles per unit time) of the designated MEC servers to be
offloaded. If the dependency is in terms of the level of association 𝑐 𝑆 . The total task duration, denoted by 𝛥𝑖 for any task 𝑖 can be defined
with other tasks, then care must be taken in scheduling the task while as the total response-time for the task starting from its submission time
offloading. till its completion. Hence, 𝛥𝑖 is composed of the task dispatch time (𝑇𝑑 ),
Task-related dependency is quite common as the functional progress the propagation time for uploading the task to the MEC server (𝑇𝑢 ), the
of a task is often dependent on the completion of one or more tasks. task execution time (𝑇𝑒 ) and the propagation time for downloading the

3
A. Islam et al. Journal of Systems Architecture 118 (2021) 102225

task output to the UE (𝑇𝑟 ). We are assuming the value of 𝑇𝑑 is a very

small and constant for all tasks. 𝑇𝑟 can also be ignored, considering the
output data size for the task to be negligible compared to its input [40–
42]. Therefore, the total task duration for task 𝑖 can be obtained
as:
𝑆𝑖 𝑤
𝛥𝑖 = 𝑇𝑢 + 𝑇𝑒 = + 𝑖 (3)
𝐵𝑢𝑝 𝑐 𝑆

where 𝐵𝑢𝑝 is the up-link data rate and 𝑆𝑖 is the size of input data to
be uploaded as part of the task 𝑖 . We obtain the energy consumption
at an UE with transmission power 𝜌 for offloading the task 𝑖 to edge
server is as follows:
𝜌(𝑆𝑖 + 𝜔𝑖 )
𝛹𝑖 = (4)
𝐵𝑢𝑝

where 𝜔𝑖 is the extra data (metadata in bits) needed to carry the task
offloading data to a MEC server for the task 𝑖 . The size of the metadata
depends on different protocols used to transmit the task offloading data Fig. 1. Taxonomy of task offloading sub problems.
to MEC server.

2.4. Task offloading problem formulation

Based on the evaluation metrics defined in Sections 2.2 and 2.3, we

are going to formulate the task offloading problem in this section. Let
us define the integer decision variable, 𝑥𝑖 ∈ {0, 1}, which indicates task
𝑖 is processed locally (𝑥𝑖 = 1) or in MEC server (𝑥𝑖 = 0). Now, the
objective function is to minimize the total latency for executing 𝑛 such
tasks. Mathematically, the problem can be formulated as below (similar
to [36]).
∑
𝑛
𝑚𝑖𝑛𝑖𝑚𝑖𝑧𝑒 [𝑥𝑖 𝛿𝑖 + (1 − 𝑥𝑖 )𝛥𝑖 ] (5)
𝑖=1

subject to following constraints:

Fig. 2. Single BS with single MEC server.
∑
𝑛 ∑
𝑛
𝑥𝑖 𝜓𝑖 + [(1 − 𝑥𝑖 )𝛹𝑖 + 𝜀] ≤ E𝑚𝑎𝑥 (6)
𝑖=1 𝑖=1

The constraint (6) indicates the total energy consumption of a UE 2.5. Task offloading subproblems
should be less than or equal to its remaining battery capacity E𝑚𝑎𝑥 . Here
𝜀 signifies the minimum amount of energy required to receive the task Task offloading as a whole is a complex problem and it further can
be divided into different subproblems. Some of the subproblems are
output from the MEC server. The task duration minimization problem
exclusive to UE and MEC whereas some of them are common to both
(5) can be proved to be NP-hard [18].
UE and MEC. Though discussing these subproblems in details are out of
Task offloading problems can have many stakeholders based on the the scope of this literature, we have tried to categorize them as shown
MEC architecture like MDs, MEC servers, cloud, etc. Task offloading is in Fig. 1. A state-of-art review of existing task offloading work, we have
a challenging problem to solve for the following reason: described in Section 4 shown in Fig. 5. In Section 4, different authors
have tried to address one or more sub-problems provided in this section
1. It is a multi-objective optimization problem and known to be
shown in Fig. 1.
NP-hard which cannot be solved using traditional deterministic
algorithms in polynomial time.
3. MEC architecture classification
2. To solve this problem, look-ahead knowledge on the amount of
arriving tasks at the MDs is required. Although The European Telecommunication Standards Institute
3. While formulating the task offloading algorithm, we need to (ETSI) has proposed some generic reference architecture [43], there is
consider wireless channel conditions from the MDs to the MEC no concrete standardized architectural framework for MEC till now to
servers and computing resources across the entire network. the best of our knowledge. In the absence of any concrete standardized
4. The task offloading problem itself includes many subproblems architectural framework, diverse design practices can be observed
like task scheduling, resource allocation, task offloading decision across various MEC research communities in modeling the system
making, and so on, which are closely interrelated. architecture. The solution approaches proposed by different authors for
5. MEC servers are not homogeneous. Along with varying load, MEC task offloading are mainly defined by the system model assump-
the variety in hosted services across the MEC servers creates a tions and reference architecture followed. After analyzing most of the
problem in host selection. The service mapping is conventionally recent research papers on MEC task offloading, we have categorized the
a dynamic effort and may suffer from instability resulting in a MEC architecture into three broad classes as follows. We are assuming
sub-optimal solution. the MEC servers are hosted in the base stations only in the classification
of MEC architecture.

4
A. Islam et al. Journal of Systems Architecture 118 (2021) 102225

Fig. 3. Single BS with multiple MEC servers.

Fig. 5. Task Offloading Strategies (Section 4).

running on the target BS or (iii) If the task deadline cannot be

met with the support from MEC servers in the target BS.

4. Offloading strategies

Task offloading is the key problem in realizing the MEC-enabled

smart IoT applications. This is why it has got huge traction from both
academia as well as the industry in terms of research contributions
in the fields. The adopted architecture and system model widely vary
across the MEC research communities coming up with diverse task
offloading solutions. Hence, it is unfair to evaluate all these different
solutions under one umbrella. As discussed in Section 2, the task
offloading is an NP-hard problem, and we can only find an approximate
Fig. 4. Multiple BS with multiple MEC servers. solution to this problem. However, there will always be a trade-off
between the performance gain achieved through offloading the task and
offloading overhead incurred by the algorithm used.
The existing literature is widely diverse in terms of adopted archi-
1. Single MEC server in single BS/Small Cells: As shown in
tecture and system model assumptions, consideration of task model,
Fig. 2, there exists one MEC server connected to a base station
use of algorithmic paradigm and computational strategies, and the task
and MDs offload their tasks to the nearest BS based on the offloading problem formulation. We have broadly categorized different
offloading decisions made by the UEs. In this architecture, only solution strategies into three different perspectives, which can further
one MEC server serves the requests for all MDs connected to that be sub-divided into various classes of problem-solving approaches taken
BS and thus creating a high possibility of resource unavailability as follows.
at the MEC server. A detailed comparison of different works on MEC task offloading is
2. Multiple MEC server with single BS/Small Cells: As shown in depicted based on different categories and parameters in Table 3. We
Fig. 3, there exist multiple MEC servers connected to a single have tried to summarize the research works in the direction of MEC
BS. MDs offload the tasks based on their task offloading policy task offloading based on the different categories like system model,
to the nearest BS, and the tasks are distributed among all the computation offloading model, evaluation methodology, the method
MEC servers connected to that BS. There is less chance of server used to solve the problem, etc. We have also tried to highlight some lim-
overloading as compared to the first model described above. In itations of the different proposed solutions in different research works
this architecture, no more offloading is done to other BS if the in Table 3. We have categorized only those works (tasks offloading on
MEC) which are generic and not specific to a particular application
target BS is overloaded.
like vehicular technology [62–73], UAV [74–76], Industry4.0 [77], or
3. Distributed MEC servers with multiple BS/Small Cells: As
urban informatics [78] and others.
shown in Fig. 4, there will be multiple MEC servers connected
to a BS. MDs will offload the tasks based on their task offloading 4.1. Computational perspective
decisions to the nearest BS, and the tasks will be distributed
among all the MEC servers connected to that BS or route the A comparative analysis of different computational system models
tasks to other MEC servers running on other BSs. In this architec- with respect to the performance benefits is shown in Table 4. In this
ture a task can be routed to other BSs depending on the following section, we summarized different task offloading strategies with respect
conditions (i) All the MEC servers in the target BS are overloaded to different computational models. We have also presented this in
(ii) The required application is not available in any MEC server Table 5 with a detailed insight.

5
A. Islam et al. Journal of Systems Architecture 118 (2021) 102225

Table 3
Task offloading detailed comparison.
Paper Problems Method Performance Evaluation Supported Computation System Limitation
addressed applied metrics methodology UEs offloading model
model
[44] Task placement, Nonlinear Task duration, Simulation Multi UEs Mixed Distributed Ignores the delay
Resource allocation programming, Energy consumption Offloading and energy
SDTO consumption of
sending request
from mobile device
to the SD-UDN
controller.
[45] Task placement, Greedy algorithm Network delay, Numerical Multi UEs Full Distributed No consideration of
Resource allocation energy consumption Offloading the neighboring
base stations which
have MEC servers.
[46] Task scheduling, Mixed Integer Task admission and Numerical Multi UEs Full Distributed As the no. of UEs
Resource allocation Programming execution time Offloading increases the
admission rate
decreases
[47] Resource Greedy heuristic Energy Numerical Single UE Partial Centralized Can only achieve a
allocation consumption Offloading near-optimal
solution.
[48] Task offloading, BBM, IPM Revenue of MNO Simulation Multi UEs Full Centralized No fault tolerance
Service deployment Offloading and scalability
[49] Joint task IPM, BBA Energy Numerical Multi UEs Mixed Centralized Not better than AOP
offloading, resource consumption, Task Offloading in terms of Average
allocation processing delay task processing
delay when task
complexity is
increased.
[50] Task scheduling Deep learning Latency Simulation Single UE Partial Distributed When operating at
Offloading large-scale, some UE
may drop out due
to broken
connection or
insufficient power.
[15] Collaborative Greedy, GT Energy Simulation Multi UEs Full Centralized OECCO is
computation consumption Offloading unworkable for a
offloading, task large no. of
placement computation tasks
and has a very high
computation
complexity.
[41] Task offloading and Greedy heuristic Average time and Numerical Multi UEs Full Distributed Complexity of task
resource allocation energy consumption Offloading offloading of the
proposed scheme is
high
[51] Resource allocation, Greedy heuristic Energy Simulation Multi UEs Partial Centralized In case of splitting
channel allocation, consumption Offloading tasks, they have not
power control mentioned how they
are splitting the
tasks.
[52] Task partition, Task Greedy Heuristic Task execution Numerical and Multi UEs Mixed Distributed For better
scheduling latency Simulation Offloading performance high
bandwidth and
large numbers of
MEC server needed
[53] Task scheduling Collaborative Latency Simulation Multi UEs Full Centralized Average task failure
offloading Offloading increases as the task
size increases for all
the cases.

(continued on next page)

4.1.1. Centralized model all requests are processed by a central server, the centralized computing
Centralized computing refers to a computing model involving a model has lower efficiency and a higher risk of failure, compared to
central server with high computing capability and sophisticated ap- the decentralized or distributed computing model. However, the cen-
plications. The central server connects to client computers that have tralized model provides higher security and reliability with respect to
very low processing capability, and hence when a task needs to be user data, as all the data is stored in one place and tasks are performed
executed, the clients simply send requests to the central server, which on a single site reducing the chances of multi-party exposure [84,85].
then performs most of the processing. The connection between the Task offloading strategies proposed in [51] and [48] are based on a
central server and clients can be either direct or over a network [83]. As centralized computation model. In [51], a greedy sequential selection

6
A. Islam et al. Journal of Systems Architecture 118 (2021) 102225

Table 3 (continued).
Paper Problems Method Performance Evaluation Supported Computation System Limitation
addressed applied metrics methodology UEs offloading model
model
[54] Dynamic task Distributed Total Simulation Multi UEs Full Centralized Mao’s Method [13]
placement and algorithm energy-efficiency Offloading and Dinh’s Method
Resource allocation cost at UEs [13] are slightly
better method when
the number of tasks
is small.
[55] Task partition, BnB, Iterative Latency Simulation Multi UEs Partial Distributed Non-linear
placement, and Heuristic Offloading applications with
Resource allocation complex internal
dependencies are
not considered.
[56] Task allocation and Greedy algorithm Total cost(energy Numerical Single UE Full Distributed Channel congestion
Computational consumption, Offloading may occur due to
frequency scaling execution delay) limited channel
resources and
higher the
communication
Cost.
[57] Resource Sub-optimal Energy Simulation Multi UEs Full Centralized Not considered the
allocation resource allocation consumption Offloading effect of
algorithm heterogeneous
computation loads
[58] Resource Sub-gradient System delay Numerical Multi UEs Partial Centralized Co-channel
allocation algorithm Offloading interference is not
considered.
[59] Task computation, Heuristic algorithm Total energy Simulation Multi UEs Full Distributed When the number
Resource allocation cost(UE, MEC Offloading of UEs grows the
system) interference among
UEs become severe
which leads to a
relatively high sum
energy cost.
[60] Task partition Approximation Energy Simulation Single UE Partial Distributed Codec energy model
consumption Offloading is not suitable for
less complex local
computations
[61] Task offloading and Reinforcement System cost(energy Simulation Multi UEs Partial Distributed As resources are
resource allocation learning and delay) Offloading allocated
dynamically based
on the task arrival,
it will be a
challenging task
when there is
uncertainty in task
arrival, and there is
no central
controller.

Table 4 the mobile network operator (MNO), while meeting the requirements
Analysis Table for different computational model. of resource-constrained mobile devices. According to their proposed
Centralized Decentralized Distributed model different services can run on the virtual machines of the MEC
Design complexity Low High High server. The main goal is to efficiently use MEC resources to maximize
Scalability Low High High the revenue of the MNO while assigning the offloaded task to one of
Horizontal scalability No Yes Yes
Vertical scalability Yes Yes Yes
the services associated with a virtual machine.
Fault tolerance Low High High
Single point of failure Yes Yes Yes 4.1.2. Decentralized model
Availability Low High High In decentralized systems, every server makes its own decisions. The
Design complexity Low High High
final behavior of the system is defined by the aggregated decisions
made through a synchronized cooperation among the participating
servers [86,87]. This type of system is highly available, has high fault
algorithm SMSEF (Select Maximum Saved Energy First) for task of- tolerance, and is not a single point of failure. Designing this kind of
floading is proposed. The main objective of this work is to maximize system is a big challenge as it involves associated challenges like replica
the energy saved for mobile devices. In the proposed solution a task management and consistency enforcement across all copies of data
can be divided into subtasks and the divided subtasks can be com- distributed over multiple servers. A decentralized dynamic computation
puted locally or remotely. The methodology followed in splitting the offloading framework based on the deep deterministic policy gradient
tasks as well as the feasibility analysis for performing such a splitting (DDPG) [88] algorithm has been built, which enables each user to
operation is not discussed in the paper though. In [48], the authors independently learn efficient offloading policies from only local ob-
have emphasized maximizing the total revenue from the perspective of servations for dynamic power allocation in a continuous domain [82].

7
A. Islam et al. Journal of Systems Architecture 118 (2021) 102225

Table 5
Task offloading comparison of papers based on computational model.
Computational Design objective Reference Proposed solution
model
Minimizing energy consumption for MDs [51] A greedy Select Maximum Saved Energy First (SMSEF)
Centralized
algorithm.
Maximizing the total revenue [48] The iterative optimization of service deploying and task
offloading (IOSDTO)
Minimizing the total delay in finishing the tasks [79] Deep reinforcement learning-based joint task offloading
and bandwidth allocation.
Minimizing total energy consumption while satisfying the [15] Cloud-MEC Optimal Collaborative computation Offloading
maximum execution latency constraints
Minimize the network delay [45] Distributed bandit optimization (DBO) algorithm
Distributed Minimal total latency [50] MRL(Meta Reinforce Learning) based method MRLCO
(Meta Reinforce Learning Computation offloading)
Minimizing latency [80] Deep reinforcement learning (DRL) based task offloading
in MEC
Minimize the energy consumption at the mobile devices [49] Improved branch and bound algorithm
Reduce latency [46] Dynamic Task Offloading and Scheduling problem (DTOS)
as a Mixed Integer Program (MIP) (DTOS-MIP)
Minimize the average task duration with the limited [44] Software defined Task offloading(SDTO)
battery capacity
Minimizing energy consumption for MDs [47] Heuristic greedy offloading scheme
Minimize the time consumption and reduce the number [53] Collaborative Offloading
of task failure
Minimize the processing and transmission delay and [61] Reinforcement learning based
reduce the power consumption for these two operations state–action-reward-state–action (RL-SARSA) algorithm
Maximization of offloading efficiency subject to task [81] The optimal mobile edge server selection strategy.
Decentralized
deadline constraints algorithm.
Minimize long-term average computation cost of MDs [82] Decentralized dynamic computation offloading.

While proposing the solution [82], authors have considered multi-input not addressed the problem of task distribution across the servers. One of
multi-output (MIMO) enabled multi-user MEC system with stochastic the solutions performing approximation collaborative computation of-
wireless channels. A peer-to-peer enabled decentralized edge server floading (ACCO) [15] proposed task distribution based on a centralized
management scheme for edge computing environment is proposed controller. However, the work does not discuss how the central con-
in [81] by the authors. The task offloading scheme based on this model troller distributes the tasks among the servers. In [44], the authors have
involves maximizing the offloading efficiency by selecting the optimal discussed the problem of task offloading for software-defined ultra-
MEC server, subject to meeting the task deadline constraint is proposed dense networks, which is based on software-defined network (SDN),
by the authors. and the SDN controller is the brain of the system and responsible for
task distribution among the servers. Though the SDN controller has the
4.1.3. Distributed model overall view of the network which takes the task distribution decisions,
Following the distributed computing paradigm, multiple MEC server the authors have not mentioned what are the statistics it collects and
entities communicate and coordinate with each other to appear as from which all entities for formulating the task distribution decision
a single coherent system to the end-users. There are two general model.
ways in which the distributed systems function: (1) each node works
toward a common goal and the end-user views the results as one 4.2. Decision-making perspective
cohesive unit. (2) each node has its end-user and the distributed
system facilitates sharing resources or communication services. Typ- In this section, we have provided a detailed classification of existing
ically in this paradigm, a controller distributes the tasks to different task offloading works based on decision-making entities like UEs, cloud,
servers based on various criteria like application type, latency deadline, SDN controller, etc. and we have also presented the same in Table 6.
types of resource requirement, and other parameters [89,90]. The
distributed cluster can reside in a single BS or multiple BSs based on 4.2.1. Cloud-assisted
the system model. A distributed system is highly available and can In this approach, all the task offloading decisions are made at the
scale both horizontally and vertically. Although there are many advan- cloud servers. The MEC servers send information about computation
tages of a distributed system, on the other hand, it has some design resources, network topology, different application types, and their cor-
challenges such as heterogeneity, transparency, fault tolerance and fail- responding QoS requirements to a centralized cloud server. Similarly,
ure management, scalability, concurrency, openness and extensibility, all the UE devices send various task attributes to the cloud. On the
migration and load balancing, and security [91]. other hand, the cloud server processes the information received from
Task offloading schemes proposed in [15,44–47,49,50,53,80] are MEC servers and UE devices, formulates the task offloading decision
based on the distributed computing paradigm. The system model used policy, and sends them to UEs. In [80], a deep reinforcement learning
in [15,45,49,50,80] consists of multiple computing servers in a single (DRL) based offloading framework is integrated into the MEC system
base station. A task offloaded to a base station must be computed on the where UEs or mobile devices sends the application DAGs (Direct Acyclic
servers belonging to that BS. But the system model used in [44,46,47] Graph) to a centralized cloud. A deep reinforcement learning (DRL)
has the provision of redirecting the tasks to the servers of another based task offloading decision model is formulated in the centralized
base station if the task requirements cannot be satisfied by the MEC cloud by training the collected data that includes application DAGs and
servers residing in the current base station. Most of these works have the current environment of the MEC system.

8
A. Islam et al. Journal of Systems Architecture 118 (2021) 102225

Table 6
Task offloading comparison of papers based on decision model.
Decision model Design objective Reference Proposed solution
Cloud-assisted Minimizing latency [80] Deep reinforcement learning (DRL) based task offloading
in MEC
Minimize the network delay [45] Distributed bandit optimization (DBO) algorithm
Minimizing energy consumption for MDs [47] Heuristic greedy offloading scheme
UE driven
Minimizing total energy consumption while satisfying the [15] Cloud-MEC Optimal Collaborative computation Offloading
maximum execution latency constraints
Minimizing energy consumption for MDs [51] A greedy Select Maximum Saved Energy First (SMSEF)
algorithm.
Minimize the energy consumption at the mobile devices [49] Improved branch and bound algorithm
Reduce latency [46] Dynamic Task Offloading and Scheduling problem (DTOS)
SDN based
as a Mixed Integer Program (MIP) (DTOS-MIP)
Minimize the average task duration with the limited [44] Software defined Task offloading(SDTO)
battery capacity
Minimal total latency [50] MRL(Meta Reinforce Learning) based method MRLCO
Collaborative
(Meta Reinforce Learning Computation offloading)
Minimizing the latency [92] A joint iterative algorithm based on the Lagrangian dual
decomposition
Maximizing the total revenue [48] The iterative optimization of service deploying and task
offloading (IOSDTO)

4.2.2. UE driven (eNB) if the current one is unable to satisfy the task requirement [46].
In this case, UE itself formulates the task offloading decision policy However, the main challenge lies in selecting a suitable MEC server
and offloads the task accordingly. As per this strategy, UE takes re- without knowing if the current server can satisfy the request or a
sponsibility for making the decision of offloading the tasks. It may use neighbor eNB has a server that can serve the request. The server load
different algorithm design paradigms like greedy technique, dynamic varies dynamically, and hence the server selection for task offloading
programming, integer programming, and so on to formulate the deci- is a time-critical decision. The current literature has not addressed that
sion model. Since UE is taking the offloading decision, it has to expend problem yet. The arbitrary selection of the MEC server running on
some of its computational resources for this purpose which may result another BS to redirect the offloaded task might often lead to the ping
in added energy consumption. Task offloading strategy in [15,45,47, pong effect. In the work described in [44], the decision of forwarding
49,51] are UE driven where each UE decides whether to offload the a task to another base station is taken by the SDN controller. Though
tasks to MEC server or cloud or compute locally. The task offloading the SDN framework has many advantages like network flexibility,
decision in [45,47,49,51] are based on maximizing the energy savings easy troubleshooting, it has its limitations when it comes to security,
at UEs or minimizing the energy consumption at the mobile devices. scalability, high availability, reliability, and elasticity issues [96]. The
The limitations of these works come with not considering the CPU current literature is still in its nascent stage in addressing all these
cycles needed for computing the energy requirements for the tasks issues towards solving the MEC offloading problem through SDN-based
under consideration. The energy-efficient task offloading algorithms solutions.
in [47,49] also considers the delay constraint of a task while formu-
lating the offloading decisions. One of the main objectives of the task 4.2.4. Collaborative
offloading algorithm [45] is to minimize the network delay along with A collaboration among UEs, MEC servers, and the mobile cloud
maximizing the energy savings at the UE side. An ultra-dense network can be classified into three categories based on the participation in
is considered in [47] where MEC servers are attached to a macro base the task offloading decision-making processes like 1. combination of
station (MBS) and small cell (SC) BS. In this architecture [47], the MEC UEs, MEC servers, and the mobile cloud 2. MEC servers of different
servers are attached to both MBS and SCs which adds more complexity service providers. 3. MEC servers of different networks with the same
in selecting the MEC servers in the task offloading decisions. service provider. UEs decide which base station to offload the task
or to compute the task locally. A controller in the BS will decide
4.2.3. SDN based which server to redirect the request based on the requirement of a
Software-defined networking (SDN) technology is an approach to task. It may also be possible that the controller can select a server
network management that enables dynamic, programmatically efficient in the neighbors’ BS depending on the MEC system model used. The
network configuration to improve network performance and monitor- system model adopted in [48,50,53] is collaborative involving both
ing, making it more like cloud computing than traditional network UEs as well as MEC servers. In [53], authors have proposed two
management [96]. SDN attempts to centralize the network intelligence types of collaboration: one is the collaboration between mobile users
in one network component by disassociating the forwarding process and MEC server for reducing the number of task failures and another
of network packets (data plane) from the routing process (control one is the collaboration between MEC server with cloud for reducing
plane). The control plane consists of one or more controllers, which latency as well as executing a large number of computation tasks. The
are considered to be the brain of the SDN network where the whole MRL(Meta Reinforce Learning) based method MRLCO (Meta Reinforce
intelligence is incorporated [97]. Learning Computation offloading) algorithm has two parts of which
In this approach, the task offloading decision is modeled at the SDN one is executed in UEs and the other in MEC servers. In this proposed
controller based on different information collected from UEs and MEC method, the UEs collaborate with MEC and cloud for decision-making
servers from different BS. The decision model is then broadcast to UEs on task offloading. A joint problem of a cooperative computation task
periodically by the SDN controllers. The task offloading mechanism for offloading scheme proposed in [92] where the authors have considered
MEC in [44,46] is based on SDN architecture. The task offloading model vertical cooperation among UE devices, MEC server nodes, and cloud
in [44] is based on an ultra-dense network on top of SDN architecture. server nodes, and the horizontal computation cooperation between
A task can be forwarded to a MEC server on another base station edge nodes.

9
A. Islam et al. Journal of Systems Architecture 118 (2021) 102225

Table 7
Task offloading comparison of papers based on algorithmic paradigm.
Algorithmic Design objective Reference Proposed approach
paradigm
Minimizing energy consumption for MDs [51] A greedy Select Maximum Saved Energy First (SMSEF)
Greedy heuristics algorithm
Minimizing energy consumption for MDs [47] Heuristic greedy Offloading scheme
Minimizing total energy consumption while satisfying the [15] The Algorithm approximation collaborative computation
maximum execution latency constraints offloading solution (ACCO)
Minimizing total energy consumption while satisfying the [93] A heuristic swap matching based algorithm
maximum execution latency constraints
Minimize the task completion time [94] A heuristic algorithm
Maximizing the total revenue [48] Convex optimization and branch and bound are used to
Integer solve the MIP problem
programming
Reduce latency [46] Dynamic Task Offloading and Scheduling (DTOS)
Minimize the energy consumption at the mobile devices [49] Improved branch and bound algorithm
Minimize the average task duration with the limited [44] MINLP problem converted to a convex problem and used
battery capacity the Karush–Kuhn–Tucker (KKT) condition to obtain an
optimal solution
Minimal total latency [50] Dynamic Task Offloading and Scheduling problem (DTOS)
Machine learning
as a Mixed Integer Program (MIP) (DTOS-MIP)
Minimizing latency [80] Deep reinforcement learning (DRL) based task offloading
in MEC
Minimizing the total delay in finishing the tasks [79] Deep reinforcement learning-based joint task offloading
and bandwidth allocation.
Minimize the processing and transmission delay and [61] Reinforcement learning based
reduce the power consumption for these two operations state–action-reward-state–action (RL-SARSA) algorithm.
Maximizing the total revenue [48] Iterative optimization of service deploying and task
Branch and bound
offloading (IOSDTO)
Minimize the energy consumption at the mobile devices [49] Improved branch and bound algorithm
Minimizing latency and device energy consumption [95] Reformulation-Linearization Technique (RLT) based
solution

Convex Minimize the average task duration with the limited [44] MINLP problem converted to a convex problem and used
Optimization Battery capacity the Karush–Kuhn–Tucker (KKT) condition to obtain an
optimal solution

4.3. Algorithmic paradigm (i) total CPU cycles needed for a task (ii) network delay from UEs to
MEC server and (iii) time per CPU cycle for both UEs and MEC servers.
In this section, we have classified existing works on task offloading To get all those information and process them, the UEs have to use the
under the heading algorithmic paradigm. A summary of this section is computation power which in turn consumes energy. It is also important
illustrated in Table 7. for UEs to be agnostic about the MEC server information. Authors
in [93], formulated the task offloading problem as an optimization
4.3.1. Greedy heuristics problem and proposed a heuristic swap matching-based algorithm to
As already discussed in Section 2, task offloading is an optimization solve this problem. Authors in [94] proposed a heuristic algorithm for
problem and it is NP-hard, we can use a greedy heuristic to solve it. task offloading to MEC server that aims to minimize the task completion
A greedy heuristic algorithm generally tends to find a near-optimal time. This algorithm assigns priority to each UE device according to
solution without costing too much computation time or memory space. their local computation frequency and input data size. The device with
This algorithm does not need any complex derivatives or careful choice higher priority will have precedence to occupy channels and offload
of initial values for building a solution space. [98]. The task-offloading the task to MEC server attached to a BS.
algorithms proposed in [15,47,51] are based on greedy approaches.
SMSEF (Select Maximum Saved Energy First) is a greedy sequential 4.3.2. Integer programming
selection algorithm [51] that chooses the most energy-saving task at The objective of integer programming algorithms is to minimize
or maximize a function subject to equality, inequality, and integer
each iteration. The heuristic greedy offloading scheme (HGOS) [47]
constraints [99]. Integer constraints restrict some or all of the variables
first calculates the task processing time based on local computation,
in the optimization problem to take on only integer values. When there
computation at the MEC server connected to the MBS, and computation
are integer constraints on only some of the variables, the problem is
at the MEC server connected to the SC. In the second phase, it is decided
called a mixed-integer program (MIP). An integer linear program in the
where to offload the tasks based on the processing time. The main
canonical form is expressed as below [100].
challenge in HGOS is to calculate the value of the processing time for a
task concerning local, MEC, or cloud platform. The Algorithm approxi- maximize 𝑐 𝑇 𝑥
mation collaborative computation offloading solution (ACCO) [15] also
subject to 𝐴𝑥 ≤ 𝑏, 𝑥 ≥ 0, 𝑥 ∈ 𝑍 𝑛 (7)
follows a greedy approach with the lowest energy consumption by the
MDs as the evaluation parameter. The computational complexity of where c, b are vectors and A is a matrix with all entries to be integers.
ACCO is 𝑂(𝑛2 ), where 𝑛 is the number of computation tasks. For greedy The task offloading optimization problem can be formulated into
algorithms to apply, we need prior knowledge about the optimization a mixed-integer program with constraints like remaining energy of
parameters and for the task offloading scenario these parameters are the UE devices, computation abilities of UEs and MEC servers, other

10
A. Islam et al. Journal of Systems Architecture 118 (2021) 102225

resources needed for computation of a task etc. Task offloading prob- past experiences and keeps refining the model and which helps UEs to
lem in [46,48] is formulated as Mixed Integer Program (MIP) while offload tasks more intelligently and efficiently.
in [44,49] the problem is formulated as Mixed-Integer Nonlinear Pro- A machine learning-based task offloading scheme is proposed in [50,
gramming (MINLP). A non convex mixed-integer optimization problem 80]. Deep reinforcement learning (DRL) [61,79,80] and MRL (Meta
is formulated to solve task offloading in [101]. An improvement of a Reinforce Learning) [50] based approaches for learning the best task
branch and bound algorithm is proposed to address the MINLP in [49] offloading decisions also exists in the literature. In [50], the authors
instead of transforming the MINLP to a relaxed optimization problem as have proposed an idea to model the dynamic computation offloading
a convex one, which can be solved by the interior point method (IPM). process as multiple MDP (Markov Decision Processes)s, where the
Although the IPM can find the optimal solution for a relaxed problem, learning of offloading policies is decomposed into two parts: effectively
it may not be practical from the implementation point of view in the learning a meta policy among different MDPs, and fast learning a
MEC network. The MINLP problem in [44] is divided into two sub- specific policy for each MDP based on the meta policy. They have
problems as (1) computing resource allocation problem and (2) task converted the offloading decision process as a sequence prediction
placement problem. For resource allocation, the problem is proved to process and designed a custom seq2seq neural network to represent the
be a convex problem and authors have used the Karush–Kuhn–Tucker offloading policy. The MRLCO (Meta Reinforce Learning Computation
(KKT) condition to obtain an optimal solution. Based on the solution of offloading) aims to leverage the computation resources from both the
the above sub-problem, they adopted the task placement algorithm to UE and the MEC host for the training process. The proposed model
obtain the solution for this sub-problem. The MIP problem in [48] is in [80] is divided into three layers: application, MEC, and cloud.
solved by decomposing the problem into two dependent sub-problems The cloud layer gathers application DAGs from mobile devices and
as (1) task offloading problem and (2) service deployment problem. For conducts DRL training based on both the gathered DAGs and the
task offloading it becomes a convex optimization problem when the environment. The offloading problem is then formulated as a MDP,
decision of service deployment is given. On the other hand, when the and a new S2S (Sequence-to-Sequence) neural network is proposed to
offloading strategy is fixed, the problem turns into a 0–1 programming model the offloading policy. The S2S neural network is trained via the
problem. The authors have solved the mentioned sub-problem using the Proximal Policy Optimization algorithm [105], which has high training
branch and bound method (BBM) and the interior point method (IPM) performance and stability.
respectively and named it iterative optimization of service deploying Though there are many advantages, deep neural networks are often
and task offloading (IOSDTO). Dynamic Task Offloading and Schedul- much harder to train than shallow neural networks. To use machine
ing (DTOS) [46] problem is solved by decomposing it into a master learning in task offloading, it is important to design a system properly
problem(MP) which solves the task assignment and the application keeping the following points in mind (i) whether the machine learning
resource allocation problems, and multiple subproblems (SP) to resolve model to formulate in the central cloud or the MEC server. (ii) how the
the task scheduling problem. machine learning model gathers the relevant information (iii) how the
formulated task offloading model is pushed to UEs.
Perhaps the simplex method is effective for solving linear pro-
grams, but there is no single technique for solving integer programs.
4.3.4. Branch and bound
Instead, several procedures have been developed, and the performance
Branch and bound (BnB) is a general programming paradigm used
of any particular technique appears to be highly problem-dependent.
to solve hard optimization problems. The complexity of these problems
Methods to date can be classified broadly into three approaches as
is generally in exponential order. Branching is the process of creating
follows: (i) enumeration techniques, including the branch-and-bound
sub-problems, and bounding is to ignore partial solutions that are not
procedure; (ii) cutting-plane techniques; and (iii) group-theoretic tech-
better than the current best solution. BnB methodology can be a good
niques [102]. Sometimes we may need to use the composite meth-
candidate to solve many optimization problems easily for which direct
ods (more than one method combined) to solve the integer prob-
methods of a solution either do not exist or insufficient [106]. Task
lem. Though we can solve the task offloading problems using integer
offloading is an optimization problem with many constraints and it can
programs, it is difficult to model and can be hard to solve.
be formulated into an integer or mixed-integer program. As discussed
in Section 2, it is an NP-hard problem, we can relax some variables
4.3.3. Machine learning to solve this problem using the BnB algorithm. The task offloading
Machine learning(ML) is a sub-field of artificial intelligence (AI), problem in [48] is formulated as MIP while in [49] as MINLP. The
which contains all the approaches that allow machines to learn from IOSDTO algorithm for task offloading sub-problem discussed in 4.3.2 is
data without being explicitly programmed [103]. Machine learning is solved using the BnB methodology. An improvement of a BnB algorithm
closely related to computational statistics, which focuses on making is also proposed to address the MINLP problem in [48]. In [49], the
predictions by using computational resources [104]. By leveraging authors proposed the task offloading problem as a MINLP optimization
complex mathematical and statistical tools, ML renders machines ca- problem, which is NP-hard. They have introduced a relaxing solution
pable of performing independently intellectual tasks that have been that converts binary decision variables to real values and proposed a
traditionally solved by human beings. A model for task offloading BnB-based solution. The authors in [95] have proposed a QoE-driven
decisions can be formulated by training the different data set from computation offloading algorithm based on the BnB technique, where
UEs and MEC servers. Since the formulation of a decision model using they considered latency time and energy consumption of UEs as the
machine learning is computation-intensive and needs huge resources, main attributes to measure QoE. The main goal of this paper [95] is
it is better to run machine learning algorithms in the far cloud which to enhance the quality of service experienced by users by minimiz-
does not have resource constraints. The UEs periodically send the ing the weighted sum of the total latency delay and device energy
information of its computation power, task-related information like consumption.
CPU cycle requirements, latency deadline, etc. Similarly MEC servers
also periodically send the information of their resource capacity, type 4.3.5. Dynamic programming
of applications supported to the cloud. The machine-learning algorithm Dynamic programming is an algorithmic design strategy that trans-
running in the cloud uses this information from UEs and MEC servers forms a complex problem into a sequence of simpler problems to
and formulates the task offloading decisions. The decision model for- solve an optimization problem. Its efficiency has been proved in many
mulated by the deep learning algorithm a subfield of ML in the cloud is areas of wireless communication for solving optimization problems
sent back to UEs, which in turn takes the offloading decision based on within an acceptable bound on computation time, such as optimal
that model. The good thing about machine learning is, it learns from path searching in data packets routing [107], resource scheduling and

11
A. Islam et al. Journal of Systems Architecture 118 (2021) 102225

allocation [108], power management [109,110], and so on. Dynamic routed to the core network [117], which will consume a great
Programming based Offloading Algorithm (DPOA) [107] is an adap- amount of network bandwidth due to the nature of the video
tive offloading algorithm that can determine the offloading policy stream. If we can perform these analysis tasks close to the
dynamically according to a changing wireless environment. By apply- source of the video stream, it will help in getting rid of network
ing a dynamic programming technique to an existing task offloading congestion caused by the video stream uploading and the latency
algorithm (if suitable), we may reduce the time complexity of that of task transmission between the source device and distant
algorithm. By reducing the time complexity of the task offloading cloud [118]. A MEC-based video analysis system is a suitable
algorithm, we can save the energy of mobile devices and reduce the choice for this kind of application. A real-life example could be
task completion time. It can also be used to cope up with dynamic envi- a factory surveillance system. In this system, a large number of
ronments like varying channel bandwidth between mobile devices and cameras can stream the surveillance videos to the edge server
base stations or small cell stations. However, dynamic programming- directly. The edge servers pre-process and analyze the streaming
based solutions suffer from a few drawbacks (1) It needs a significant video and alert the administrator for any emergency in real-time
amount of memory to store the calculated result of every sub-problem and send the filtered video frames to the cloud servers. Cloud
without ensuring if the stored value will be utilized or not. As a result, servers process and analyze the filtered video frames from the
it uses more energy and resources of mobile devices. (2) There is no edge servers. The edge servers in this case help in alerting the
general method to solve a DP (dynamic programming) problem and administrator in real-time and reduces the traffic in the backhaul
the method to solve a DP problem varies from problem to problem. network [119].
3. Distributed Content Delivery and Caching [14]: One of the most
4.3.6. Convex optimization popular services of mobile networks is content, and especially
A convex optimization problem is a problem consisting of mini- video delivery. Thousands of video content items are uploaded
mizing a convex function over a convex set. Mathematically a convex daily to content provider’s networks and stored in large volumes
problem can be represented as below in the content provider’s centralized databases. These contents
are then transcoded from source format to final delivery format
min 𝑓 (𝑥)
before distributed to multiple streaming servers. Despite best
such that 𝑥 ∈ S distribution efforts from content providers, many users may
where S is a convex set and 𝑓 (𝑥) is a convex function over S. A set S experience service interruptions because of buffering problems
is convex if for all members x, y ∈ S and all 𝜃 ∈ [0, 1], we have that and jitter due to lack of content in the proximity of end-users.
𝜃𝑥 + (1 − 𝜃)𝑦 ∈ S. Convex optimization is a subfield of optimization A smart distributed video caching at the edge using the MEC
problem and a convex optimization can be solved in polynomial-time framework [120] by extending Content Delivery Network (CDN)
algorithms [111], though the optimization problem is in general a NP- services can enhance the quality of experience (QoE) of the
hard [112,113]. The computing resource allocation problem [44] is users. This solution can also help in reducing backhaul and core
proved as a convex problem and they used the Karush–Kuhn–Tucker network usage and congestion. The streaming service providers
(KKT) condition to obtain an optimal solution. like Netflix,1 Amazon Prime,2 etc. can cache the popular content
of a particular area to the edge servers of that area. This caching
5. Use cases will make the content streaming faster by reducing the latency
which will directly improve the user experience.
As MEC server hosts at the edge of the network, which becomes suit- 4. Application-Awareness and Content Optimization: A MEC server
able for numerous modern applications. We will discuss some typical resides at the edge of the server which is closer to the radio
use cases of the MEC network paradigm as below. network and the end-user, it can provide information about the
application performance. A MEC platform can be used to provide
1. IoT Applications: IoT devices have less computation power and application-aware performance optimization in near real-time by
need to prolong battery lives, so it is important to offload altering the encoding of streaming content inside the edge [121].
the computation-intensive tasks for remote processing and re- A video streaming can be optimized at the MEC server based
trieve the results once the processing completes [114]. It is on network resource availability, the screen size of MDs, etc.
also difficult for some IoT devices to act without the aid of which helps in increasing the efficiency of network resource
an external entity, so these IoT applications have to obtain usages and the quality of experiences of end users [14]. Another
distributed information for computation [16]. Although all these example of application-awareness and content optimization is to
requirements are supported by a distant centralized cloud that use social networks as valuable sensors of human activity [122].
has high-performance computation capabilities and can collect Information collected from social networks can be useful for
distributed information. One of the main challenges of central- optimizing the resource allocation for specific applications based
ized cloud computing is to fulfill the latency deadline for a on indicated popularity in a particular region.
section of IoT applications which can be solved by offloading 5. Connected Vehicles: The connected vehicle technology can en-
computation intensive tasks to a MEC server [115]. A MEC large safety, reduce traffic congestion, sense vehicle behaviors,
server can also act as an IoT gateway to aggregate the mes- and provide opportunities for diverse value-added services such
sages from IoT devices nearby. In addition, the MEC server as the car finder and parking location [123,124]. However,
can pre-process messages and send only meaningful messages this technology is still in a nascent stage and one of the main
to the cloud. It will help in reducing the congestion in the challenges is to satisfy the shortcoming is the latency require-
backhaul network and significantly reduces the communication ment with the existing connected car clouds, which contributes
and processing overhead [27,116]. to an end-to-end latency between 100 ms to 1s [16]. A MEC
2. Video Stream Analysis: Different applications come under video platform can be used as a key enabling technology for connected
stream analysis such as vehicular license plate recognition, vehicles by adding computation and geo-distributed services to
face recognition, and home security surveillance, etc. Generally, roadside BSs [16]. The services deployed in the MEC server
video analysis algorithms have high computation complexity,
so it is preferable to offload the analysis tasks away from
video-capturing devices [16]. If these processing tasks are to 1
https://www.netflix.com/.
2
be handled in the central cloud, the video stream should be https://www.primevideo.com/.

12
A. Islam et al. Journal of Systems Architecture 118 (2021) 102225

can receive and analyze messages from proximate vehicles and There is no standardization about the allocation of resources
roadside sensors and can propagate the hazard warnings and for different applications running inside the MEC servers to the
latency-sensitive messages rapidly which allows the drivers to best of our knowledge. It is challenging to devise a technique to
react immediately and make it possible for autonomous driving. allocate resources so that the resource allocation is efficient and
The tech giant Google had already announced its plan to develop at the same time it satisfies the resource requirements of a task.
self-driving vehicles back in 2009 [125]. • Task Scheduling: As task scheduling is an integral part of the task
6. Industrial Internet-of-Things (Industry 4.0): The fourth industrial offloading framework, so an efficient task scheduling algorithm is
revolution defines the next generation of manufacturing tech- needed. Once a task is offloaded to MEC servers, it is important
nology where information technologies are utilized in almost to schedule the task for one of the application processes. As the
all operations to provide self optimization and automation for nature task scheduling algorithm is NP-hard, it is a huge challenge
production, referred to as smart factory [126]. In a smart fac- to formulate an efficient task scheduling algorithm.
tory environment, all the components of the production system • The decision of task offloading to which MEC server or base
and auxiliary machinery, named industrial Internet-of-Things station is a challenging task as i. the nearest base station may
(IIoT) devices, are equipped with internet connection [127]. not host the MEC servers, ii. it may not have enough resources
Based on the information provided by the IIoT devices, the to serve the request, or iii. the requested application may not
available. As most of the tasks offloaded to MEC servers are delay-
smart manufacturing applications return management and con-
sensitive, we need to decide concisely which MEC server or base
trol commands to the IIoT devices and these devices act on the
station to offload the tasks.
command provided by the smart manufacturing applications.
To achieve the optimal result, the smart factory rigorously re-
6.2. Future research direction
quires ultra-low latency and a reliable working environment to
acquire the precise control commands from the smart manu- As research on task offloading to MEC is still in evolving stage, there
facturing applications for operations. Due to the low latency are ample opportunities for research in this area.
requirement of IIoT, the MEC platform can be a great fit to host
smart-manufacturing applications. 1. Though there is a good progress in the usage of machine learning
(ML) [50] for decision making in task offloading to MEC, it is a
6. Discussion quite challenging task to formulate a machine learning model
which takes into consideration of following attributes a. data
6.1. Technical challenges rate between MDs and the BS which hosts MEC servers, b.
application type of the task to be offloaded, c. computational
Recent years have witnessed substantial research efforts on task capabilities and the energy status of the MDs, d. resource avail-
ability of the nearest BS which hosts the MEC servers, e. channel
offloading for MEC as surveyed in the preceding sections. However,
quality information between MDs and the BS, f. mobility of the
there exists ample scopes for further research in various emerging
MD devices etc. We can proceed in the direction of modeling a
directions centered around MEC task offloading problems that are
machine learning algorithm that considers the above attributes.
still largely uncharted. In this section, we are discussing technical
2. If we consider a MEC system model of single BS with multiple
challenges towards solving the MEC task offloading problem.
MEC servers attached to it and there is no cooperation among
• UEs and MEC server connection management: If a UE offloads the BSs. A decision on task load balancing or distribution among
a task to a MEC server attached to the nearest BS/SCs and MEC servers depends on many factors like types of the appli-
before receiving the task computation result, that UE connects cation running in different servers, resource utilization of these
to a different MEC server of different BS/SCs. It is a key design servers, application type of tasks and their different attributes,
issue to manage the connection between UEs and MEC servers to etc. In [44], authors have discussed task distribution among MEC
make sure UEs receive the task computation result in the scenario servers based on a centralized decision-maker SDN controller,
mentioned above. but as per our research, there is no distributed algorithm to
• Estimation of computation time of a task: A decision on task distribute tasks among the servers inside a BS in an efficient way.
offloading depends on several factors like computation time of 3. There is little research on task offloading for a distributed MEC
requirements of a task, channel bandwidth, the proximity of server system model that includes multiple BSs. A BS can for-
ward a task to another BS if it cannot satisfy the requirement of
MEC server, and many others. For task offloading decision to
a task offloaded to it by a UE. A BS can forward a task to another
make the UEs have to calculate the approximate computation
BS for different reasons such as application needed by the UE is
time requirements of the task. The offloading decision-making
not available, resource constraints in the MEC servers, etc. We
algorithm will be more efficient if a UE can calculate the most
cannot just blindly forward a task to neighbor BS, we have to
accurate computation time requirement of the task.
devise an efficient algorithm to forward the task to the nearest
• Distributed MEC servers with multiple base stations or small cell
base station which has the required application and sufficient
base stations: Authors in [44] have proposed a distributed model, resources to serve the UEs. This is one of the open areas we can
but the MEC servers are managed by a central SDN controller. focus on.
So it is not a purely distributed model as the SDN controller is a 4. An efficient task placement mechanism in the MEC is also impor-
centralized one. If the central SDN controller fails, the MEC server tant to satisfy the task offloading criteria like delay, computa-
topology will be in a bad state as UEs will not receive any infor- tional resource need, and application type constraints. We need
mation from the SDN controller which is used for making task to focus on this area to formulate an efficient task placement
offloading decisions. This means there is no fault tolerance for mechanism that satisfies all the task offloading constraints.
the SDN controller. It is challenging to create a truly distributed 5. After reviewing the literature on task offloading on MEC, we can
system of MEC servers hosted across multiple BSs. firmly claim that impressive progress is happening in this field
• Resource allocation for a task in MEC servers: It is important corresponding to distributed and centralized MEC systems. But
to allocate computation and storage resources to an application we could not find much research progress in the area of the de-
running on the MEC server to handle the task offloaded by UEs. centralized MEC systems model. Task offloading in decentralized
Resources for an application can be allocated in various ways MEC systems model could be one of the research areas we can
by initiating VMs, Docker containers, or any other technologies. focus on in the future.

13
A. Islam et al. Journal of Systems Architecture 118 (2021) 102225

7. Conclusion [18] Y. Pochet, L.A. Wolsey, Production Planning By Mixed Integer Programming,
Springer Science & Business Media, 2006.
[19] P. Brucker, Scheduling Algorithms, fifth ed., Springer, 2010, ISBN: 3642089070,
The advancement of new types of delay-sensitive applications like
9783642089077.
face recognition augmented/virtual reality etc. needs a significant [20] P. Salot, A survey of various scheduling algorithm in cloud computing
amount of computation power. Since both the requirements are not environment, Int. J. Res. Eng. Technol. 2 (2) (2013) 131–135.
able to be addressed by either local computation or cloud computing, [21] X. Lin, Y. Wang, Q. Xie, M. Pedram, Task scheduling with dynamic voltage
as a result a new network paradigm MEC is gaining popularity. One and frequency scaling for energy minimization in the mobile cloud computing
environment, IEEE Trans. Serv. Comput. 8 (2) (2014) 175–186.
of the critical parts of the MEC is the efficient task offloading scheme. [22] T.Q. Dinh, J. Tang, Q.D. La, T.Q. Quek, Offloading in mobile edge computing:
Though research in this area is still in the early stage, we have tried Task allocation and computational frequency scaling, IEEE Trans. Commun. 65
our best to provide a comprehensive survey of research progress in (8) (2017) 3571–3584.
this field. First, we have divided the different task offloading models [23] H. Wu, W.J. Knottenbelt, K. Wolter, An efficient application partitioning
algorithm in mobile environments, IEEE Trans. Parallel Distrib. Syst. 30 (7)
into three broad categories and conducted an overall literature review
(2019) 1464–1480.
on recent research efforts on task offloading for MEC. We further [24] Y.C. Hu, M. Patel, D. Sabella, N. Sprecher, V. Young, Mobile edge computing—A
discussed several challenges regarding task offloading and provided key technology towards 5G, ETSI White Pap. 11 (11) (2015) 1–16.
some potential future research direction in this area. The extensive [25] W. Shi, S. Dustdar, The promise of edge computing, Computer 49 (5) (2016)
overview and research outlook on task offloading in MEC provided 78–81.
[26] O. Salman, I. Elhajj, A. Kayssi, A. Chehab, Edge computing enabling the Internet
in this survey hopefully can serve as handy references and valuable of Things, in: 2015 IEEE 2nd World Forum on Internet of Things, WF-IoT, IEEE,
guidelines for further in-depth investigations of this area. 2015, pp. 603–608.
[27] H. Tanaka, M. Yoshida, K. Mori, N. Takahashi, Multi-access edge computing: A
Declaration of competing interest survey, J. Inf. Process. 26 (2018) 87–97.
[28] N. Abbas, Y. Zhang, A. Taherkordi, T. Skeie, Mobile edge computing: A survey,
IEEE Internet Things J. 5 (1) (2017) 450–465.
The authors declare that they have no known competing finan- [29] Q.-V. Pham, F. Fang, V.N. Ha, M.J. Piran, M. Le, L.B. Le, W.-J. Hwang, Z.
cial interests or personal relationships that could have appeared to Ding, A survey of multi-access edge computing in 5G and beyond: Funda-
influence the work reported in this paper. mentals, technology integration, and state-of-the-art, IEEE Access 8 (2020)
116974–117017.
[30] Y. Mao, C. You, J. Zhang, K. Huang, K.B. Letaief, Mobile edge computing:
References Survey and research outlook, 2017, arXiv preprint arXiv:1701.01090.
[31] S. Wang, X. Zhang, Y. Zhang, L. Wang, J. Yang, W. Wang, A survey on mobile
[1] P. Porambage, J. Okwuibe, M. Liyanage, M. Ylianttila, T. Taleb, Survey on edge networks: Convergence of computing, caching and communications, Ieee
multi-access edge computing for internet of things realization, IEEE Commun. Access 5 (2017) 6757–6779.
Surv. Tutor. 20 (4) (2018) 2961–2991. [32] P. Mach, Z. Becvar, Mobile edge computing: A survey on architecture and
[2] J. Lee, J.-W. Kim, J. Lee, Mobile personal multi-access edge computing computation offloading, IEEE Commun. Surv. Tutor. 19 (3) (2017) 1628–1656.
architecture composed of individual user devices, Appl. Sci. 10 (13) (2020) [33] B. Yang, X. Cao, J. Bassey, X. Li, L. Qian, Computation offloading in multi-access
4643. edge computing: A multi-task learning approach, IEEE Trans. Mob. Comput.
[3] Z. Sanaei, S. Abolfazli, A. Gani, R. Buyya, Heterogeneity in mobile cloud (2020).
computing: taxonomy and open challenges, IEEE Commun. Surv. Tutor. 16 (1) [34] C. Jiang, X. Cheng, H. Gao, X. Zhou, J. Wan, Toward computation offloading
(2013) 369–392. in edge computing: A survey, IEEE Access 7 (2019) 131543–131558.
[4] B. Liu, X. Xu, L. Qi, Q. Ni, W. Dou, Task scheduling with precedence and [35] B. Wang, C. Wang, W. Huang, Y. Song, X. Qin, A survey and taxonomy on task
placement constraints for resource utilization improvement in multi-user MEC offloading for edge-cloud computing, IEEE Access 8 (2020) 186080–186101.
environment, J. Syst. Archit. 114 (2021) 101970. [36] H. Lin, S. Zeadally, Z. Chen, H. Labiod, L. Wang, A survey on computation
[5] P. Ranaweera, A.D. Jurcut, M. Liyanage, Realizing multi-access edge computing offloading modeling for edge computing, J. Netw. Comput. Appl. (2020)
feasibility: Security perspective, in: 2019 IEEE Conference on Standards for 102781.
Communications and Networking, CSCN, IEEE, 2019, pp. 1–7. [37] J. Wang, J. Pan, F. Esposito, P. Calyam, Z. Yang, P. Mohapatra, Edge cloud
[6] Y.C. Hu, M. Patel, D. Sabella, N. Sprecher, V. Young, Mobile edge computing—A offloading algorithms: Issues, methods, and perspectives, ACM Comput. Surv.
key technology towards 5G, ETSI White Pap. 11 (11) (2015) 1–16. 52 (1) (2019) 1–23.
[7] X. Xu, B. Shen, X. Yin, M.R. Khosravi, H. Wu, L. Qi, S. Wan, Edge server [38] A. Shakarami, A. Shahidinejad, M. Ghobaei-Arani, A review on the computation
quantification and placement for offloading social media services in industrial offloading approaches in mobile edge computing: A g ame-theoretic perspective,
cognitive IoV, IEEE Trans. Ind. Inf. (2020). Softw. - Pract. Exp. 50 (9) (2020) 1719–1759.
[8] H. Dai, X. Zeng, Z. Yu, T. Wang, A scheduling algorithm for autonomous driving [39] S. Thananjeyan, C.A. Chan, E. Wong, A. Nirmalathas, Mobility-aware energy
tasks on mobile edge computing servers, J. Syst. Archit. 94 (2019) 14–23. optimization in hosts selection for computation offloading in multi-access edge
[9] X. Deng, J. Li, E. Liu, H. Zhang, Task allocation algorithm and optimization computing, IEEE Open J. Commun. Soc. 1 (2020) 1056–1065.
model on edge collaboration, J. Syst. Archit. 110 (2020) 101778. [40] X. Chen, Decentralized computation offloading game for mobile cloud
[10] C. Dou, S. Zhang, H. Wang, L. Sun, Y. Huang, W. Yue, Adhd fmri short-time computing, IEEE Trans. Parallel Distrib. Syst. 26 (4) (2014) 974–983.
analysis method for edge computing based on multi-instance learning, J. Syst. [41] T.X. Tran, D. Pompili, Joint task offloading and resource allocation for multi-
Archit. 111 (2020) 101834. server mobile-edge computing networks, IEEE Trans. Veh. Technol. 68 (1)
[11] P. Cong, J. Zhou, L. Li, K. Cao, T. Wei, K. Li, A survey of hierarchical energy (2018) 856–868.
optimization for mobile edge computing: A perspective from end devices to the [42] X. Lyu, H. Tian, C. Sengul, P. Zhang, Multiuser joint task offloading and
cloud, ACM Comput. Surv. 53 (2) (2020) 1–44. resource optimization in proximate clouds, IEEE Trans. Veh. Technol. 66 (4)
[12] C. Wu, Q. Peng, Y. Xia, Y. Ma, W. Zheng, H. Xie, S. Pang, F. Li, X. Fu, X. (2016) 3435–3447.
Li, et al., Online user allocation in mobile edge computing environments: A [43] I. ETSI, Multi-access edge computing (MEC); framework and reference
decentralized reactive approach, J. Syst. Archit. 113 (2021) 101904. architecture, Tech. rep., 2016, 2019.
[13] J.-W. Ryu, Q.-V. Pham, H.N. Luan, W.-J. Hwang, J.-D. Kim, J.-T. Lee, [44] M. Chen, Y. Hao, Task offloading for mobile edge computing in software defined
Multi-access edge computing empowered heterogeneous networks: A novel ultra-dense network, IEEE J. Sel. Areas Commun. 36 (3) (2018) 587–597.
architecture and potential works, Symmetry 11 (7) (2019) 842. [45] Z. Sun, M.R. Nakhai, An online learning algorithm for distributed task
[14] T. Taleb, K. Samdanis, B. Mada, H. Flinck, S. Dutta, D. Sabella, On multi-access offloading in multi-access edge computing, IEEE Trans. Signal Process. (2020).
edge computing: A survey of the emerging 5G network edge cloud architecture [46] H.A. Alameddine, S. Sharafeddine, S. Sebbah, S. Ayoubi, C. Assi, Dynamic task
and orchestration, IEEE Commun. Surv. Tutor. 19 (3) (2017) 1657–1681. offloading and scheduling for low-latency IoT services in multi-access edge
[15] H. Guo, J. Liu, Collaborative computation offloading for multiaccess edge computing, IEEE J. Sel. Areas Commun. 37 (3) (2019) 668–682.
computing over fiber–wireless networks, IEEE Trans. Veh. Technol. 67 (5) [47] H. Guo, J. Liu, J. Zhang, Computation offloading for multi-access mobile edge
(2018) 4514–4526. computing in ultra-dense networks, IEEE Commun. Mag. 56 (8) (2018) 14–19.
[16] Y. Mao, C. You, J. Zhang, K. Huang, K.B. Letaief, A survey on mobile edge [48] W. Ni, H. Tian, X. Lyu, S. Fan, Service-dependent task offloading for multiuser
computing: The communication perspective, IEEE Commun. Surv. Tutor. 19 mobile edge computing system, Electron. Lett. 55 (15) (2019) 839–841.
(4) (2017) 2322–2358. [49] T.T. Vu, N. Van Huynh, D.T. Hoang, D.N. Nguyen, E. Dutkiewicz, Offloading
[17] Q.-V. Pham, L.B. Le, S.-H. Chung, W.-J. Hwang, Mobile edge computing with energy efficiency with delay constraint for cooperative mobile edge computing
wireless backhaul: Joint task offloading and resource allocation, IEEE Access 7 networks, in: 2018 IEEE Global Communications Conference, GLOBECOM, IEEE,
(2019) 16444–16459. 2018, pp. 1–6.

14
A. Islam et al. Journal of Systems Architecture 118 (2021) 102225

[50] J. Wang, J. Hu, G. Min, A.Y. Zomaya, N. Georgalas, Fast adaptive task [78] K. Zhang, Y. Zhu, S. Leng, Y. He, S. Maharjan, Y. Zhang, Deep learning
offloading in edge computing based on meta reinforcement learning, IEEE empowered task offloading for mobile edge computing in urban informatics,
Trans. Parallel Distrib. Syst. 32 (1) (2020) 242–253. IEEE Internet Things J. 6 (5) (2019) 7635–7647.
[51] F. Wei, S. Chen, W. Zou, A greedy algorithm for task offloading in mobile edge [79] L. Huang, X. Feng, C. Zhang, L. Qian, Y. Wu, Deep reinforcement learning-
computing system, China Commun. 15 (11) (2018) 149–157. based joint task offloading and bandwidth allocation for multi-user mobile edge
[52] T. Yang, R. Chai, L. Zhang, Latency optimization-based joint task offloading computing, Digit. Commun. Netw. 5 (1) (2019) 10–17.
and scheduling for multi-user MEC system, in: 2020 29th Wireless and Optical [80] J. Wang, J. Hu, G. Min, W. Zhan, Q. Ni, N. Georgalas, Computation offloading
Communications Conference, WOCC, IEEE, 2020, pp. 1–6. in multi-access edge computing using a deep sequential model based on
[53] M.D. Hossain, L.N. Huynh, T. Sultana, T.D. Nguyen, J.H. Park, C.S. Hong, E.- reinforcement learning, IEEE Commun. Mag. 57 (5) (2019) 64–69.
N. Huh, Collaborative task offloading for overloaded mobile edge computing [81] W. Tang, X. Zhao, W. Rafique, L. Qi, W. Dou, Q. Ni, An offloading method
in small-cell networks, in: 2020 International Conference on Information using decentralized P2P-enabled mobile edge servers in edge computing, J. Syst.
Networking, ICOIN, IEEE, 2020, pp. 717–722. Archit. 94 (2019) 1–13.
[54] H. Yu, Q. Wang, S. Guo, Energy-efficient task offloading and resource schedul- [82] Z. Chen, X. Wang, Decentralized computation offloading for multi-user mobile
ing for mobile edge computing, in: 2018 IEEE International Conference on edge computing: A deep reinforcement learning approach, EURASIP J. Wireless
Networking, Architecture and Storage, NAS, IEEE, 2018, pp. 1–4. Commun. Networking 2020 (1) (2020) 1–21.
[55] Z. Ning, P. Dong, X. Kong, F. Xia, A cooperative partial computation offloading [83] J.L. King, Centralized versus decentralized computing: Organizational con-
scheme for mobile edge computing enabled Internet of Things, IEEE Internet siderations and management options, ACM Comput. Surv. 15 (4) (1983)
Things J. 6 (3) (2018) 4804–4814. 319–349.
[56] T.Q. Dinh, J. Tang, Q.D. La, T.Q. Quek, Offloading in mobile edge computing:
[84] A. Baneriee, M.S. Chishti, A. Rahman, R. Chanazain, Centralized framework for
Task allocation and computational frequency scaling, IEEE Trans. Commun. 65
workload distribution in fog computing, in: 2018 3rd International Conference
(8) (2017) 3571–3584.
for Convergence in Technology, I2CT, IEEE, 2018, pp. 1–5.
[57] C. You, K. Huang, H. Chae, B.-H. Kim, Energy-efficient resource allocation
[85] A. Bozorgchenani, D. Tarchi, G.E. Corazza, Centralized and distributed architec-
for mobile-edge computation offloading, IEEE Trans. Wireless Commun. 16 (3)
tures for energy and delay efficient fog network-based edge computing services,
(2016) 1397–1411.
IEEE Trans. Green Commun. Netw. 3 (1) (2018) 250–263.
[58] J. Ren, G. Yu, Y. Cai, Y. He, Latency optimization for resource allocation
[86] A.J. Ferrer, J.M. Marquès, J. Jorba, Towards the decentralised cloud: Survey
in mobile-edge computation offloading, IEEE Trans. Wireless Commun. 17 (8)
on approaches and challenges for mobile, ad hoc, and edge computing, ACM
(2018) 5506–5519.
Comput. Surv. 51 (6) (2019) 1–36.
[59] L. Yang, H. Zhang, M. Li, J. Guo, H. Ji, Mobile edge computing empowered
[87] B. Pon, Blockchain will usher in the era of decentralised computing, LSE Bus.
energy efficient task offloading in 5G, IEEE Trans. Veh. Technol. 67 (7) (2018)
Rev. (2016).
6398–6409.
[60] S. Chouhan, Energy optimal partial computation offloading framework for mo- [88] T.P. Lillicrap, J.J. Hunt, A. Pritzel, N. Heess, T. Erez, Y. Tassa, D. Silver, D.
bile devices in multi-access edge computing, in: 2019 International Conference Wierstra, Continuous control with deep reinforcement learning, 2015, arXiv
on Software, Telecommunications and Computer Networks, SoftCOM, IEEE, preprint arXiv:1509.02971.
2019, pp. 1–6. [89] H. Attiya, J. Welch, Distributed Computing: Fundamentals, Simulations, and
[61] T. Alfakih, M.M. Hassan, A. Gumaei, C. Savaglio, G. Fortino, Task offloading and Advanced Topics, Vol. 19, John Wiley & Sons, 2004.
resource allocation for mobile edge computing by deep reinforcement learning [90] J. Waldo, G. Wyant, A. Wollrath, S. Kendall, A note on distributed computing,
based on SARSA, IEEE Access 8 (2020) 54074–54084. in: International Workshop on Mobile Object Systems, Springer, 1996, pp.
[62] C. Yang, Y. Liu, X. Chen, W. Zhong, S. Xie, Efficient mobility-aware task 49–64.
offloading for vehicular edge computing networks, IEEE Access 7 (2019) [91] K. Nadiminti, M.D. De Assunçao, R. Buyya, Distributed systems and recent
26652–26664. innovations: Challenges and benefits, InfoNet Mag. 16 (3) (2006) 1–5.
[63] G. Qiao, S. Leng, K. Zhang, Y. He, Collaborative task offloading in vehicular [92] Z. Kuang, Z. Ma, Z. Li, X. Deng, Cooperative computation offloading and
edge multi-access networks, IEEE Commun. Mag. 56 (8) (2018) 48–54. resource allocation for delay minimization in mobile edge computing, J. Syst.
[64] J. Zhang, H. Guo, J. Liu, Y. Zhang, Task offloading in vehicular edge computing Archit. (2021) 102167.
networks: A load-balancing solution, IEEE Trans. Veh. Technol. 69 (2) (2019) [93] B. Gu, Z. Zhou, S. Mumtaz, V. Frascolla, A.K. Bashir, Context-aware task
2092–2104. offloading for multi-access edge computing: Matching with externalities, in:
[65] Y. Sun, X. Guo, J. Song, S. Zhou, Z. Jiang, X. Liu, Z. Niu, Adaptive learning- 2018 IEEE Global Communications Conference, GLOBECOM, IEEE, 2018, pp.
based task offloading for vehicular edge computing systems, IEEE Trans. Veh. 1–6.
Technol. 68 (4) (2019) 3061–3074. [94] T.-Y. Kan, Y. Chiang, H.-Y. Wei, Task offloading and resource allocation
[66] P. Liu, J. Li, Z. Sun, Matching-based task offloading for vehicular edge in mobile-edge computing system, in: 2018 27th Wireless and Optical
computing, IEEE Access 7 (2019) 27628–27640. Communication Conference, WOCC, IEEE, 2018, pp. 1–4.
[67] H. Guo, J. Zhang, J. Liu, FiWi-enhanced vehicular edge computing networks: [95] J. Luo, X. Deng, H. Zhang, H. Qi, QoE-driven computation offloading for edge
Collaborative task offloading, IEEE Veh. Technol. Mag. 14 (1) (2018) 45–53. computing, J. Syst. Archit. 97 (2019) 34–39.
[68] B. Gu, Z. Zhou, Task offloading in vehicular mobile edge computing: A [96] K. Benzekki, A. El Fergougui, A. Elbelrhiti Elalaoui, Software-defined
matching-theoretic framework, IEEE Veh. Technol. Mag. 14 (3) (2019) 100–106. networking (SDN): a survey, Secur. Commun. Netw. 9 (18) (2016) 5803–5833.
[69] H. Guo, J. Liu, J. Ren, Y. Zhang, Intelligent task offloading in vehicular edge [97] E. Haleplidis, K. Pentikousis, S. Denazis, J.H. Salim, D. Meyer, O. Koufopavlou,
computing networks, IEEE Wirel. Commun. 27 (4) (2020) 126–132. Software-defined networking (SDN): Layers and architecture terminology, in:
[70] Y. Wang, P. Lang, D. Tian, J. Zhou, X. Duan, Y. Cao, D. Zhao, A game- RFC 7426, IRTF, 2015.
based computation offloading method in vehicular multiaccess edge computing [98] Z.W. Geem, J.H. Kim, G.V. Loganathan, A new heuristic optimization algorithm:
networks, IEEE Internet Things J. 7 (6) (2020) 4987–4996. harmony search, Simulation 76 (2) (2001) 60–68.
[71] S. Raza, W. Liu, M. Ahmed, M.R. Anwar, M.A. Mirza, Q. Sun, S. Wang, An
[99] A. Schrijver, Theory of Linear and Integer Programming, John Wiley & Sons,
efficient task offloading scheme in vehicular edge computing, J. Cloud Comput.
1998.
9 (2020) 1–14.
[100] C.H. Papadimitriou, K. Steiglitz, Combinatorial Optimization: Algorithms and
[72] H. Peng, Q. Ye, X.S. Shen, SDN-based resource management for autonomous
Complexity, Courier Corporation, 1998.
vehicular networks: A multi-access edge computing approach, IEEE Wirel.
[101] Z. Kuang, L. Li, J. Gao, L. Zhao, A. Liu, Partial offloading scheduling and power
Commun. 26 (4) (2019) 156–162.
allocation for mobile edge computing systems, IEEE Internet Things J. 6 (4)
[73] X. Wang, T. Wei, L. Kong, L. He, F. Wu, G. Chen, ECASS: Edge computing based
(2019) 6774–6785.
auxiliary sensing system for self-driving vehicles, J. Syst. Archit. 97 (2019)
258–268. [102] L.A. Wolsey, Integer Programming, Vol. 52, John Wiley & Sons, 1998.
[74] J. Chen, S. Chen, S. Luo, Q. Wang, B. Cao, X. Li, An intelligent task offloading [103] D. Jakhar, I. Kaur, Artificial intelligence, machine learning and deep learning:
algorithm (iTOA) for UAV edge computing network, Digit. Commun. Netw. 6 definitions and differences, Clin. Exp. Dermatol. 45 (1) (2020) 131–132.
(4) (2020) 433–443. [104] P. Ongsulee, Artificial intelligence, machine learning and deep learning, in:
[75] J. Xiong, H. Guo, J. Liu, Task offloading in UAV-aided edge computing: Bit 2017 15th International Conference on ICT and Knowledge Engineering,
allocation and trajectory optimization, IEEE Commun. Lett. 23 (3) (2019) ICT&KE, IEEE, 2017, pp. 1–6.
538–541. [105] J. Schulman, F. Wolski, P. Dhariwal, A. Radford, O. Klimov, Proximal policy
[76] Q. Wang, Z. Li, K. Nai, Y. Chen, M. Wen, Dynamic resource allocation for optimization algorithms, 2017, arXiv preprint arXiv:1707.06347.
jointing vehicle-edge deep neural network inference, J. Syst. Archit. 117 (2021) [106] E.L. Lawler, D.E. Wood, Branch-and-bound methods: A survey, Oper. Res. 14
102133. (4) (1966) 699–719.
[77] Z. Cao, P. Zhou, R. Li, S. Huang, D. Wu, Multiagent deep reinforcement learning [107] Y. Liu, M.J. Lee, An effective dynamic programming offloading algorithm in
for joint multichannel access and task offloading of mobile-edge computing in mobile cloud computing system, in: 2014 IEEE Wireless Communications and
industry 4.0, IEEE Internet Things J. 7 (7) (2020) 6201–6213. Networking Conference, WCNC, IEEE, 2014, pp. 1868–1873.

15
A. Islam et al. Journal of Systems Architecture 118 (2021) 102225

[108] H.A. Lagar-Cavilla, N. Tolia, E. De Lara, M. Satyanarayanan, D. O’Hallaron, [127] M. Wollschlaeger, T. Sauter, J. Jasperneite, The future of industrial communi-
Interactive resource-intensive applications made easy, in: ACM/IFIP/USENIX cation: Automation networks in the era of the internet of things and industry
International Conference on Distributed Systems Platforms and Open Distributed 4.0, IEEE Ind. Electron. Mag. 11 (1) (2017) 17–27.
Processing, Springer, 2007, pp. 143–163.
[109] D. Huang, P. Wang, D. Niyato, A dynamic offloading algorithm for mobile Akhirul Islam received the Bachelor’s (B.E.) degree in
computing, IEEE Trans. Wireless Commun. 11 (6) (2012) 1991–1995. Computer Science and Engineering from Dibrugarh Uni-
[110] N. Takahashi, H. Tanaka, R. Kawamura, Analysis of process assignment in multi- versity, India in 2009 and currently pursuing Ph.D. from
tier mobile cloud computing and application to edge accelerated web browsing, Indian Institute of Information Technology, Guwahati, India.
in: 2015 3rd IEEE International Conference on Mobile Cloud Computing, He also has 11+ years of experience in software industry
Services, and Engineering, IEEE, 2015, pp. 233–234. and working as a Staff Software Dev Engineer in FireEye,
[111] Y. Nesterov, A. Nemirovskii, Interior-Point Polynomial Algorithms in Convex Bangalore, India R&D center. His present research focus is
Programming, SIAM, 1994. on Multi-access Edge Computing(MEC), specially in MEC
[112] K.G. Murty, S.N. Kabadi, Some NP-Complete Problems in Quadratic and task offloading.
Nonlinear Programming, Tech. rep., 1985.
[113] S. Sahni, Computationally related problems, SIAM J. Comput. 3 (4) (1974)
262–279.
[114] J. Yang, Z. Lu, J. Wu, Smart-toy-edge-computing-oriented data exchange based Arindam Debnath received the Bachelor’s (B.Tech.) degree
on blockchain, J. Syst. Archit. 87 (2018) 36–48. in Information Technology from Gauhati University, India
[115] Y. Ai, M. Peng, K. Zhang, Edge computing technologies for Internet of Things: in 2009 and currently pursuing Master’s degree (M.Tech.)
a primer, Digit. Commun. Netw. 4 (2) (2018) 77–86. from Indian Institute of Information Technology, Guwahati,
[116] D. Sabella, A. Reznik, R. Frazao, Multi-Access Edge Computing in Action, CRC India.
Press, 2019.
[117] A. Anjum, T. Abdullah, M. Tariq, Y. Baltaci, N. Antonopoulos, Video stream
analysis in clouds: An object detection and classification framework for high
performance video analytics, IEEE Trans. Cloud Comput. (2016).
[118] B. Varghese, N. Wang, S. Barbhuiya, P. Kilpatrick, D.S. Nikolopoulos, Challenges
and opportunities in edge computing, in: 2016 IEEE International Conference
on Smart Cloud, SmartCloud, IEEE, 2016, pp. 20–26.
Manojit Ghose is working as an Assistant Professor in
[119] J. Wang, J. Pan, F. Esposito, Elastic urban video surveillance system using edge
the Department of Computer Science and Engineering at
computing, in: Proceedings of the Workshop on Smart Internet of Things, 2017,
IIIT Guwahati. Prior to joining IIIT Guwahati, he worked
pp. 1–6.
as an Assistant Professor at Dibrugarh University. He has
[120] S. Wang, M. Binstin, A smart transmission optimization mechanism for sports
completed his B.E. from Dibrugarh University, M.Tech. and
events based on edge computing, Internet Technol. Lett. 4 (1) (2021) e174.
Ph.D. from IIT Guwahati in Computer Science and Engi-
[121] M.A. Hoque, M. Siekkinen, J.K. Nurminen, Energy efficient multimedia stream-
neering. His research area broadly includes multiprocessor
ing to mobile devices—A survey, IEEE Commun. Surv. Tutor. 16 (1) (2012)
scheduling, cloud computing, and computer architecture. He
579–597.
is a member of IEEE and serves as TPC member for several
[122] P. Garcia Lopez, A. Montresor, D. Epema, A. Datta, T. Higashino, A. Iamnitchi,
conferences of repute.
M. Barcellos, P. Felber, E. Riviere, Edge-Centric Computing: Vision and
Challenges, ACM New York, NY, USA, 2015.
[123] P. Papadimitratos, A. De La Fortelle, K. Evenssen, R. Brignolo, S. Cosenza,
Suchetana Chakraborty is currently working as an As-
Vehicular communication systems: Enabling technologies, applications, and
sistant Professor at the department of Computer Science
future outlook on intelligent transportation, IEEE Commun. Mag. 47 (11) (2009)
and Engineering, Indian Institute of Technology Jodhpur,
84–95.
India. Prior she was a faculty member at the department
[124] N. Lu, N. Cheng, N. Zhang, X. Shen, J.W. Mark, Connected vehicles: Solutions
of CSE, IIIT Guwahati, Assam, India. Her research interests
and challenges, IEEE Internet Things J. 1 (4) (2014) 289–299.
are broadly in the areas of distributed systems, ubiquitous
[125] L. Lin, X. Liao, H. Jin, P. Li, Computation offloading toward edge computing,
computing, wireless sensor and ad hoc networks, edge com-
Proc. IEEE 107 (8) (2019) 1584–1607.
puting and Internet of Vehicles. She received her M.Tech.
[126] N.-N. Dao, Y. Lee, S. Cho, E. Kim, K.-S. Chung, C. Keum, Multi-tier multi-
and Ph.D. degree from Department of CSE, IIT Guwahati in
access edge computing: The role for the fourth industrial revolution, in:
the year 2011 and 2014, respectively. She is a member of
2017 International Conference on Information and Communication Technology
IEEE, IEEE ComSoc, ACM and IMOBILE, the Indian chapter
Convergence, ICTC, IEEE, 2017, pp. 1280–1282.
of ACM SIGMOBILE. She can be contacted at suchetana@
iitj.ac.in.

16