Download as pdf or txt
Download as pdf or txt
You are on page 1of 23

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/350770133

Mobility-aware computational offloading in mobile edge networks: a survey

Article  in  Cluster Computing · December 2021


DOI: 10.1007/s10586-021-03268-6

CITATIONS READS

19 803

8 authors, including:

Sardar Khaliq uz Zaman Ali Imran Jehangiri


COMSATS University Islamabad Hazara University
11 PUBLICATIONS   78 CITATIONS    20 PUBLICATIONS   129 CITATIONS   

SEE PROFILE SEE PROFILE

Tahir Maqsood Zulfiqar Ahmad


COMSATS University Islamabad Hazara University
25 PUBLICATIONS   191 CITATIONS    17 PUBLICATIONS   73 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Post Crises Development View project

Energy Efficient and Sustainable Cloud Data Centers View project

All content following this page was uploaded by Junaid Shuja on 03 May 2021.

The user has requested enhancement of the downloaded file.


Cluster Computing
https://doi.org/10.1007/s10586-021-03268-6 (0123456789().,-volV)(0123456789().
,- volV)

Mobility-aware computational offloading in mobile edge networks:


a survey
Sardar Khaliq uz Zaman1 • Ali Imran Jehangiri1 • Tahir Maqsood2 • Zulfiqar Ahmad1 • Arif Iqbal Umar1 •

Junaid Shuja2,4 • Eisa Alanazi3 • Waleed Alasmary4

Received: 29 September 2020 / Revised: 5 February 2021 / Accepted: 14 March 2021


 The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2021

Abstract
Technological evolution of mobile devices, such as smart phones, laptops, wearable and other handheld devices have come
up with the emergence of different user applications in learning, social networking, entertainment, and community
computing domains. Many of such applications are fully or partially offloaded to the nearby server capable with high
computing and storage resources. The delivery of task offloading results to the users is a challenge in those networks where
the frequency of user mobility is high, leading to increased latency, higher energy consumption and inefficient resource
utilization. In this paper, we survey the existing studies which optimize the task offloading in edge networks with mobility
management. We formulate taxonomy of the research domain for classification of research works. We compare the listed
state-of-the-art research works based on the components identified from taxonomy. Moreover, we debate future research
directions for mobility, security, and scalability aware MEC offloading.

Keywords Mobile Edge Computing (MEC)  Mobility aware  Computational offloading

1 Introduction to potential battery drainage in smart/mobile devices [1].


Other than limited battery power, memory, processing, and
With the advent of modern computing cores and memory storage constraints also hinder the successful execution at
architectures, today‘s mobile devices have become smart mobile devices. The computationally intensive tasks can be
but still limited to computing and residual battery capacity. offloaded to the proximate cloud and edge servers equipped
There are many computationally intensive applications with sufficient computing resources [3].
such as 3D modeling, augmented/virtual reality (AR/VR), The concept of computation offloading or cyber-forag-
online games, ultra-HD image and video processing, arti- ing was first introduced in Mobile Cloud Computing
ficial intelligence and Internet of things (IoT) based (MCC) [4, 5]. In computational offloading, mobile devices
applications which are resource hungry and generate offload their computationally intensive tasks to the central
tremendous amount of data [1] [2]. Such applications may cloud data centers. The mobile devices face a major chal-
induce heavy amount of computational workload, leading lenge while contending for computational offloading. The
time and resources saved by mobile devices is sometimes
offset by latency involved in discovering appropriate cloud
& Junaid Shuja
skhaleeq@cuiatd.edu.pk server and migrate the workload. Several real-time mobile
applications, such as online gaming, audio/video confer-
1
Department of Information Technology, Hazara University encing, and financial transactions require high quality of
Mansehra, Mansehra, Pakistan service (QoS) and low latency. Failing to provide desired
2
Department of Computer Science, COMSATS University QoS adversely affects the user experience [3].
Islamabad (CUI), Abbottabad Campus, Islamabad, Pakistan To reduce network delays and improve user experience
3
Department of Computer Science, Umm Al-Qura University, compared to MCC, researchers proposed the idea of
Makkah, Saudi Arabia cloudlets. A cloudlet can be referred as small scale data
4
Department of Computer Engineering, Umm Al-Qura center that is placed near to the users [6, 7]. A few fixed
University, Makkah, Saudi Arabia

123
Cluster Computing

servers are connected through the wireless network in a Various surveys have been produced which cover the dif-
micro cellular region to perform the computation and ferent aspects of Mobile edge computing. Those aspects
offloading [8]. To overcome the shortcomings of both the include but not limited to, architectures and frameworks
MCC and cloudlet, the European Telecommunications [9], task/code offloading [4], technology integration [17],
Standards Institute (ETSI) presented the idea of Mobile machine learning models to optimize the offloading and
Edge Computing (MEC). MEC brings the computation and caching decisions[16], and device to device communica-
storage resources close to the edge of mobile network to tion [18]. However, we could not find an adequate number
execute the highly resource demanding applications [9]. of studies that purely address the mobility aware offloading
The MEC paradigm enhances the computing services and in mobile edge networks. In this article, we focus on
reduce the latency in application access [10]. Large Inter- mobility aware MEC solutions.
net service providers, like Microsoft, Google, Facebook, The contributions of this paper are:
and Netflix are deploying their MEC services to relieve the
• We survey of the recent research efforts in Mobile Edge
mobile users and delegate their workflows to the native
Computing with respect to mobility management.
edge servers. Hence, the most of the tasks are performed at
• We formulate the taxonomy based on key features,
edge servers resulting in reduced latency and power con-
parameters, models, and constraints.
sumption than the traditional cloud systems [11, 12]. MEC
• We compare the state-of-the-art MEC offloading solu-
offers proximate service provisioning to end users, to
tions based on taxonomized parameters and features.
reduce the access latencies compared to distant cloud ser-
• We highlight and identify some open research chal-
vices [13]. Figure 1 illustrates the generic MEC
lenges and future directions relating to mobility aware
architecture.
MEC offloading.
Handling user mobility is a critical challenge in MEC
[14]. The users are equipped with wireless multi-access The paper is structured as; Section II explains the sig-
technologies, such as 5G and Wi-Fi. Multi-access wireless nificance of computation offloading in MEC and provides a
technologies provide users with freedom to move between glimpse of user mobility in offloading decisions supported
access points and base stations leading to variable internet by a use case. The use case elucidates the impact of
connectivity. As a result, mobility management in MEC mobility affecting the delivery of offloading results to the
becomes a challenge due to intermittent connectivity. Such mobile user. The detailed taxonomy of task offloading
scenarios can result in frequent disconnections, and ele- techniques for Mobile Edge Computing is formulated in
vated noise levels. However, not all MEC solutions tackle Section III. Section IV presents analysis of state-of-the-art
this challenge and assume the user to be static [15, 16]. on MEC computation offloading with focus on mobility-
awareness. In section V, we compare the state-of-the-art
based on the parameters of the taxonomy. Section VI
Data Center
presents the research issues and challenges pertaining to
mobility aware MEC. Section VII provides the concluding
Tier-3 remarks for the article.
Upload
d
Download
2 Background: MEC and mobility
management
Tier-2 Cellular BS

Modern wearable and low power IoT devices cannot per-


sBSs
form the full execution of compute intensive applications.
Tier-1 The mobile users frequently download and install the
applications from desired software repositories, such as
Wifi Routers Google play, Samsung store and Apple store [18]. Those
applications include but not limited to games, text editors,
End
nd Users photo/video editors, and face and speech recognition. In
computation offloading, such applications are partitioned
ANETs
ANE
N TTss
into small tasks, some of which are executed at the nearby
MEC servers thus preserving the energy and speeding up
the computation process.
On the other hand, the mobile user certainly consumes
Fig. 1 A 3-tier heterogeneous MEC system some amount of energy, for the transmission of offloaded

123
Cluster Computing

tasks to the MEC server, and subsequently for the receiving previously associated server. Once the mobile user moves
of computation results from the server. Therefore, a major out of the previous edge server, the offloaded task will be
challenge in computation offloading is to take the decision, failed, thus extend the response time of mobile applica-
whether to offload the task or not. In case of offloading tions. Such offloading failures result in energy and com-
decision, another question arises about the application putational resources wastage [11]. Consequently, before
selection and amount of workflow needed to offload [9]. A the task offloading the MEC network should analyze the
decision made on computation offloading could possibly user‘s moving patterns and predict its next location. The
result in: task then may be offloaded to a suitable server at predicted
location.
• Local execution: refers to the computation of user
application or a set of tasks at the mobile device locally
(see Fig. 1). The data is not offloaded to the MEC
3 Task offloading in MEC: TAXONOMY
server. This will occur when a mobile user fails to find
the appropriate MEC server due unavailability or
Various studies have been performed on the optimization
required resources, overloading, or link failure.
problems relating to task offloading in mobile edge com-
• Full offloading: The entire user application is computed
puting. Those studies have been classified below, based on
and processed by the MEC server.
specific optimization objectives, constraints, algorithms,
• Partial offloading: Certain portion of a user submitted
mobility models, and support networks.
application workflow comprising a single task or
multiple tasks is executed locally whereas rest of the
application tasks is offloaded to MEC servers.
4 Objectives
The importance of task delegation/offloading or
rescheduling has great importance, to reduce the energy The mobile edge computing has various benefits which can
consumption and latency, with optimum resource utiliza- be categorized based on different performance objectives.
tion [19–21]. In 5G cellular networks, traditional high Some of the key objectives are described below.
compute-intensive applications (such as online HD games,
augmentation reality, IoT applications, and 3D modeling, 4.1 Energy efficiency
etc.) cannot be performed in the device itself [3, 22].
The task offloading is an emerging research area in The mobile devices are limited in energy and storage due
mobile edge computing (Fig. 2). The task offloading to their small design. Nevertheless, the users expect more
request generally interrupts the user connection with the from them, in terms of application usage, like remote
server. More elaborately, a mobile user submits the task surveillance, community gaming, and e-health [23]. Those
delegation request to the edge server and then moves away applications are essentially energy hungry and providing
disconnecting him from the network. Further, the suc- power to the mobile devices round the clock is a critical
cessful delivery of the execution results to the user job. However, with efficient computation at the user end
becomes a challenge and leads to increased latency and the energy can be saved to adequate levels. For energy
energy consumption [22]. Further, it becomes more crucial saving the power-hungry tasks may be offloaded from
when the frequency of user mobility is high i-e., the user mobile devices to the edge server. The experiments show
rapidly changes its location and switches between different that substantial power saving can be achieved by compu-
small cell base stations (SBSs). The motivation behind this tation offloading. For instance, eyeDentify is a multimedia
survey is to explore such schemes that consider the user application that can complete up to 44 times more com-
mobility as a key factor during the offloading decisions. putation load by exploiting the task offloading [24]. It also
We extend this discussion with the help of an illustrative enhances the battery life of edge devices to execute dif-
scenario (see Figs. 3, 4). ferent AR applications by 30–50% [25].
We envision a scenario in which the mobile users in Several computation offloading algorithms target to
different areas are connected to their corresponding 5G reduce the energy consumption offset with the latency
base stations (roadside/roof top units). The base stations are involved during offloaded application [1, 9]. Usually,
equipped with networking, computation, and storage computational offloading techniques first calculate the
capabilities. Such a unified view creates various task expected delay and energy consumption for the target
offloading scenarios. The mobile user submits some com- application and then offload applications tasks to MEC.
putationally intensive task to the nearby server. However, Same approach is used by a three-stage task offloading
due to rapid mobility the user makes a handoff to the other algorithm for efficient energy consumption is proposed in
base station right after the submission of workflow to the

123
Cluster Computing

Fig. 2 Viable results of


computation offloading decision

Fig. 3 Task offloading scenarios


with user mobility and handoff

[26]. We discuss some more techniques on energy aware Mach and Becvar [28] propose a distributed cloud-aware
task offloading ahead with joint optimization objectives. power control algorithm (CaPC) for delay sensitive appli-
cations. CaPC iteratively finds the suitable time when
4.2 Latency power control is triggered. Fog Computing also offers low
latency but at the cost of capacity limitations, however the
Delay is one of the most used performance metrics that joint operations of cloud and fog computing systems can
significantly affects the user experience in MEC. The user decrease the service delay and guarantee to remove the
applications in ultra-dense 5G networks have lower latency capacity limitations [27].
requirements, i.e., a round trip time of 1 ms, that is 10
times less than the 10 ms as of 4G [27]. The real-time 4.3 Caching
applications are highly delay bounded while offloading the
tasks to cloud servers. Therefore, increasing the density of Task caching is an emerging idea in which the offloading
sBSs with computing abilities is a reasonable approach. task is cached at the edge server to avoid the repeated

123
Cluster Computing

Fig. 4 Taxonomy of MEC Mobile Edge


offloading [16] Computing

Support
Objectives Constraints Mobility Models Algorithms Networks

Mobility Software Machine Edge


Energy Efficiency aware Defined Network Learning Network

Context Nature
Latency Stochastic FOG
aware Inspired

Caching Safety aware Random Walk Cloudlet

Scalability User Movement


Joint Optimization
aware Tracking

offloading of the same task and enhance the user experi- allocation reduces the power consumption offset by the
ence [29]. This helps to minimize the delay and save the delay constraints. A genetic algorithm based multi-objec-
energy and computation time [30]. Several studies have tive workflow-based computation offloading framework
been done on caching to save the resources. However, the concurrently considers energy consumption and task
major concern is the storage capacity of the Edge servers. deadline constraints while the task offloading to edge ser-
Several caching policies and algorithms have been pro- vers in [36], but the user mobility and the topology changes
posed by the researchers in mobile edge computing. The were not considered.
Least Frequently Used (LFU), and the Least Recently Used
(LRU) [31] are the conventional caching schemes widely 4.4.2 Caching and offloading
used in MEC. These schemes are simple but cannot per-
form well in heterogeneous task and server sizes. Further, To perform the computational offloading the storage and
they also ignore the task size and latency involved in computing both capabilities of an MEC server are con-
fetching the tasks. sidered. However, for caching only, considering the storage
capacity is enough. Hence, the joint optimization of both
4.4 Joint optimization caching and offloading objectives can also be exploited to
enhance the MEC system efficiency [29]. In [37] the
Several studies have been proposed to consider the task authors proposed a light weight scheme that reduces the
offloading as joint optimization problem in mobile edge latency and energy consumption by optimizing the com-
computing. We discuss them in the two subsections, putation, caching, and offloading policies.
(a) Energy and Delay, (b) Caching and Offloading.
4.5 Constraints
4.4.1 Energy efficiency and reduced latency
4.5.1 Mobility aware
A multi-stage sequential game model is proposed to meet
the energy and delay requirements at the same time [32]. A User mobility is tightly coupled with the, network con-
heuristic offloading decision algorithm (HODA), jointly nection between edge node and the mobile device. A
optimizes the delay and resource utilization [33]. An mobile user will get disconnected after the delegation of
adaptive sequential task offloading (ASTO) technique certain task or workflow to the nearby edge server. While,
proposed in [34] facilitates the mobile devices to take the successful delivery of the execution results to the user
offloading decisions sequentially on the basis of channel becomes a challenge and endeavors to increased latency
interference and offered computing resources. This could and energy consumption, since the user has changed its
achieve good results in reducing delay and energy con- next location [22]. In addition, that becomes more critical
sumption. An energy-efficient computation offloading when the frequency of user mobility is higher. Here a key
(EECO) mechanisms is presented in [35] which incorporate challenge is to mitigate effects of mobility and reviving the
the multi-access characteristics of the next generation 5G offloading priority function. Further, to assure the quality
networks. A joint optimization of task offloading and sBSs of service for the user, with time critical services and

123
Cluster Computing

higher mobility, MEC systems can reserve some dedicated probability (SOP) function has been used to assess the
resources and offer reliable computing service for such security level of uplink channel during task offloading.
users. Section 4 depicts the detailed analysis on such Similar algorithm has been proposed in [45] that does not
mobility-aware techniques. consider the CSI of malicious nodes and work under
realistic scenarios. The intrusion detection rate of this
4.6 Context aware technique is higher than NOMA. However, being an
exhaustive search-based technique, its complexity is higher
The distinguishing feature of the MEC system is its ability and cannot be considered as an affordable solution in MEC
to leverage the proximity of edge servers to the consumer paradigms where the latency is off paramount importance.
end. This feature can be exploited to track the users‘ real-
time information like; behaviors, moving paths, locations, 4.7.2 Reliability
and operating environments. Therefore, context-aware
services can be easily provided to the users [38]. One of the In terms of service provisioning the reliability and avail-
commonly used context-aware services is the recommen- ability of resources is also an important perspective to be
dation system, which is broadly used in online recom- guaranteed, which ultimately will enhance the user quality
mendation of food, location and venues [39, 40]. In of experience and service level agreements. Therefore,
addition, AR applications can also be exploited in video working on some proactive defense measures to ensure the
launches of new cars in an auto show. The viewers can be network and data integrity would be an exciting research
suggested the location of cars based on their interest. area. A security mechanism termed as HoneyBot is pro-
posed in [46] to cope with the Denial of Service (DoS)
4.7 Safety aware attacks. However, its accuracy and speed, is not up to the
mark. In layered architectures like MEC, the fault tolerance
Some of the interesting research directions in MEC include plays vital role in successful execution of user submitted
safety measures such as security, reliability, and privacy. tasks. Hence, the protection and restoration mechanisms
MEC systems are usually more resource constrained than across different layers of MEC would improve the avail-
centralized cloud systems and are more prone to the ability of computing and caching resources.
security attacks. We discuss the aforesaid issues as follows.
4.7.3 Privacy
4.7.1 Security
Majority of the reported works related to privacy issues are
Security is a critical challenge in edge paradigm ecosystem based on centralized mobile networks such as conventional
due to several factors discussed ahead. The edge networks clouds and mobile clouds. A very few works are available
involve different enabling technologies like; wireless net- that address the privacy aspects in MEC [47], specifically
works, distributed systems and virtualization platforms the task offloading. The mobile devices frequently offload
[41]. Several security protocols are proposed for fully their tasks to the nearby server with much focus on latency
connected mesh networks, however their applicability in minimization and energy savings. Nevertheless, the serving
mobile edge networks is questionable due to the intermit- node may not be a trusted one and might track the user
tent connectivities in MEC. Though, some studies cover locations and other contextual information that compro-
the secure communication aspects in Mobile cloud net- mises the user privacy. The aforesaid issue can be resolved
works [42]. However, such solutions do not well suit for if the task is offloaded to a distant server, enabling the
mobile edge networks, because the working surroundings intermediate servers to ensure the user privacy, but that
of edge devices may face several new threats different from will increase the latency of the system.
MCC [27]. To achieve a balance among minimizing the latency and
More elaborately, the task offloading in MEC is quite maintaining the privacy, a pioneer technique is proposed in
susceptible to be snooped or intruded because the wireless [47] termed as Post-Decision State (PDS) learning frame-
medium is naturally shared and open. To handle this issue a work. The PDS model uses the averaging operations to
physical layer security instead of traditional encryption is determine an objective function inspired by Markovian
well suited. The authors in [43] proposed a secrecy channel chain. The averaging techniques work well in noise less
encoding scheme to avoid the eavesdropping attacks. They channels [48], which is an unrealistic condition in wireless
assumed the intruder‘s Channel State Information (CSI) to mobile networks. Further, PDS is a deep learning-based
be known. The authors in [44] developed a nonorthogonal framework that induces higher time complexity and tends
multiple access (NOMA) based mechanism that detects the to increase the latency.
malicious nodes in MEC network. The secrecy outage

123
Cluster Computing

Moreover, the data privacy concerns become more [56]. More elaborately, the authors in [57] devised a
critical in D2D communication. At this level, there is no mechanism in which the mobility is managed by reallo-
third-party trusted entity that guarantee the user privacy cating the already offloaded tasks in the mobile edge net-
and integrity [41]. The users need to develop a trust rela- work. The task migration leads to reduced delay and
tionship for secure and private communication. Here, the distance between mobile user and the sBSs. A Software-
opportunity to develop microlevel third party solutions for Defined Network (SDN) based framework is proposed in
trust building would be an interesting research area. [56]. The authors considered the physical locations of the
mobile user within the cell radius and the edge server, but
4.8 Scalability aware the trajectory of the mobile user is not taken into
consideration.
Edge computing reshapes the entire technological land-
scape of IoT devices. It addresses the shortcomings of the 4.9.2 Stochastic models
resource-limited devices in terms of execution time and
energy consumption. Though, edge computing is a rapidly A Markov model is built on the basis of recorded user
adaptive technology, but accommodating exponentially mobility patterns on different networks [58, 59]. This
growing users to offer round the clock services in such trained mobility model is helpful to predict the future
distributed network is a challenging task [9]. When mul- network conditions, which is further useful in dynamic task
tiple users request to offload their tasks on the same link, it assignment decisions. Another approach for mobility
creates congestion in the edge network. awareness could be the allocation of computing resources
A virtual machine (VM) migration technique is pre- in edge networks on the basis of number of active users
sented for computation offloading, which highlights the [60].
effective utilization of edge resources and manages the
load over the edger server [49]. The VM migrations sim- 4.9.3 Random walk models
plifies the synchronization and task scheduling in the
computation offloading process but involves the complex- One of the simplest mobility models is Random Way Point
ity of context gathering and remote execution control [50]. (RWP) model. In this model a mobile node can move
A multi-user and multi-tasking computation offloading freely in any direction (zig-zag) from the base station [61].
model is presented in [51]. The model uses edge computing Further, RWP model is mathematically analyzable and
architecture under strict latency, energy, and security easy to implement. However, this model frequently suffers
constraints. The model enhances the resource utilization of to a non-uniform stationary mobile node distribution
edge servers and reduced the latency. leading to complicate the handover analysis or predictive
A distributed computation offloading strategy is pro- decisions [62]. A RWP based robust framework is pro-
posed using densely populated edge computing architec- posed in [63], which considers the user location during task
ture [52]. In this scheme, IoT devices and edge servers are execution or offloading. Another RWP based strategy for
deployed on large scale, and a multi-objective optimization the performance evaluation of offloading systems is
function is developed to achieve effective resource allo- exploited as the mobility model of mobile hosts in [64].
cation with low latency. However, computation load is not The authors examine the unreachability, mean time to
balanced across the edge servers. Similar approach is fol- recover (MTTR), and total execution time of workflow
lowed in [53] to reduce execution time, and the energy (makespan). This model is incapable to minimize the tasks
consumption of heterogeneous IoT devices. execution time and energy consumption. Another RWP
A task aggregation and prioritization, technique is pro- based interesting approach is to handle computational
posed in [54] to handle the scalability problem. In [55], a offloading based on user mobility patterns, i.e., to assume
distributed task execution model for personalized health the movement as random function. With this method, the
care is introduced to reduce energy and latency. However, data rate of a channel can be estimated by signal strength,
the applicability of this model in public edge networks is which is further used to estimate the amount of time
not explored yet. required to offload a particular task [65]. Self-similar Least
Action Walk (SLAW) is a mobility model which predicts
4.9 Mobility models the users’ next expected location on the basis of walk
behavior [18]. SLAW is particularly effective in predicting
4.9.1 Software defined networking (SDN) models the mobility patterns of device-to-device users having the
common task offloading specifications within a same
In MEC the workflows are offloaded on the basis of system geographical location. The afore said models are widely
resource utilization and geographic locations of the servers used in delay tolerant networks, however their deployment

123
Cluster Computing

and efficacy in delay-sensitive networks such as next vehicles is designed to develop an integrated resource
generation 5G networks or edge cloudlets is not known management system. Similar method is also used for
[66]. energy efficiency in mobile devices while delegating the
tasks to the nearby edge servers [75].
4.9.4 User movement tracking (UMT) based We observed that most of the machine learning models
proposed to predict user mobility use Deep Neural Net-
Further exploration of aforesaid studies highlighted some works (DNN) or Reinforcement Learning (RL). Using3
limitations, such as they considered the user allocation DNN or RL models involve higher time complexity due to
problem as a static global optimization problem and two reasons [76, 77]. First, state space and complexity
focused to determine the optimal or near optimal solutions. increase with the increase in number of heterogenous cells
Those approaches can be unproductive while dealing with and number of users present in the cells. Second, the DNNs
the systems operated with real-time connectivity, because and RL models are essentially designed for linear mapping
of high time-complexity. In addition, the existing studies, and systems with several input parameters and multiple
e.g., [67–69] reflected the distance between edge users and features. On the other hand, the input data required for a
MEC servers as the primary objective and inclined to learning model, i.e., user mobility traces are non-linear in
ignore the users‘ motion information. Therefore, when the nature. Further, in cellular networks, the number of input
users are highly mobile, the schedules produced for task re- parameters and features are limited and handcrafted.
allocation can have bad user-perceived performance.
Consequently, such mobility related information, e.g., user 4.9.6 Nature-inspired/evolutionary
motion information and speed, also need to be utilized for
the decision making of allocation at real-time. The task assignment is critical problem for real-time delay
In further works related to mobility, besides geographic sensitive applications. Hence, the workflow of the IoT
location of the fixed edge servers, the availability of mobile devices must be evenly distributed among the edge nodes,
resources is determined on the basis of a Head-light Model and the performance parameters like delay, bandwidth,
(HLM) [70]. The model captures a zone alongside the execution time of the edge nodes, and existing system load
direction of movement, analogous to the headlight of a should also be considered aside. Maintaining the tradeoffs
moving vehicle. Moreover, the HLM also considers the between the aforesaid performance parameters is an NP-
mean encounter time between a mobile and a cloudlet hard problem [78]. In such conditions, it would be quite
network, which is statistically calculated from the user’s challenging to use the traditional task assignment approa-
historic movement logs. Another statistical mobility model ches. However, the evolutionary and nature inspired algo-
advocates that the inter-contact time between any two users rithms such as Genetic algorithm (GA), particle swarm
is complied with an exponential distribution [71]. optimization (PSO), and Ant Colony Optimization (ACO)
can work very well in such multi-objective scenarios [79].
Algorithms For optimal task assignment and offloading to
Additionally, the ACO algorithm is used for load bal-
the edge servers, the placement algorithms play a key role.
ancing in edge clouds with reduced response time of the
Various sort of algorithms such as learning based, evolu-
tasks [80]. The ACO algorithm is used for task scheduling
tionary/nature-inspired and heuristics based have been in
in cloudlets [81], and for deadline-aware task scheduling in
practice.
fog networks [82]. In [78] a metaheuristic algorithm using
ACO is proposed for task offloading. The algorithm bal-
4.9.5 Machine learning
ances the workload across the fog nodes and optimizes the
resource utilization with defined QoS constraints.
Machine learning is an artificial intelligence-based art that
analyses systems and obtain the knowledge form data.
4.10 Support networks
Several learning methods have been proposed recently for
task offloading in MEC. An energy efficient online learn-
The mobile edge networks have come up new computing
ing-based task offloading technique is proposed in [72]. For
and storage paradigms. Different types of support networks
better resource utilization while offloading the tasks, an
have been proposed by the industry and academia. In this
intelligent machine learning-based technique with metric
section, we present some state-of-the-art architectures
prediction feature is proposed in [73]. Deep neural net-
including MEC, Fog Computing, Cloudlet and Edge
works are also famous in optimization problems, and sev-
Caching.
eral lightweight frameworks have been introduced for
modern consumer-based networks such as MEC. In [74], a
deep reinforcement learning method for interconnected

123
Cluster Computing

4.10.1 Edge A dynamic mobility-aware offloading algorithm for


edge computing (DMPO) is proposed in [67]. In DMPO
Edge network is an ETSI launched standard to enable the offloading path is minimized after movement of user
hosting of different consumer-oriented applications on equipment. The size of the user data is also predicted, and
mobile devices. The ETSI has defined the specifications, mobility path of the users is predicted for a short term. The
requirements, scenarios, architectures, and APIs for the algorithm works good if base stations are evenly deployed
Edge network. Edge servers are placed at the proximity of and reduces the energy consumption and latency. But it
base stations. They can handle the user requests or forward does not provide optimum solution when base stations are
them to the backhaul data centers if needed [27]. unevenly deployed. Also considers the mobility in one
dimension.
4.10.2 FOG Energy-aware mobility management (EMM) framework
for larger mobile edge networks is proposed in [68]. In
Like cloud computing FOG computing is a layered archi- EMM the information about the incoming tasks can be
tecture designed for IoT applications. A FOG network classified into two classes. The first is user-side information
consists of several IoT devices in proximity which can (user location, available base stations and size of the input
interact and collaborate with each other on a massive scale. data or tasks) and second is the BS-side information
FOG system forms a cluster of IoT devices to exploit their (channel frequency, gain and co-channel interference).
storage and computation power and provides a better There are two deployment scenarios; in the first scenario
control regarding the communication and networking. the user knows both states, mentioned above, i.e., it has
global state information (GSI). In the second scenario the
4.10.3 Cloudlet user only has the user state information; therefore, it would
have to determine the base station side information. The
Cloudlet is a 3-tier extended cloud architecture, in which authors further assume that a serving base station is not
the mobile user or IoT devices are connected with a small changed for one task, which means that the user knows
nearby cloud datacenter located at the immediate Wi-Fi about GSI and it may avoid the unnecessary handoffs. The
access point or small cell base stations [27]. The cloudlets user will remain associated with the same base station until
are self-managed, energy efficient and easy to deploy. its task completion at least. Those deployment scenarios
Cloudlets are quite exciting to use for real time and delay increase the energy consumption. To tackle this, the
sensitive applications and can collaborate with the 5G authors use the Lyapunov optimization formula to balance
cellular networks. the energy consumption and latency involved.
To improve the user reallocation process, a three stage
mobility aware model is proposed in MobMig [11]. In first
5 Mobility-aware offloading in mobile step the overloaded MEC servers are identified which have
environments: state of the art no remaining capacity even though there are more users in
the queue waiting for the resources. For all overloaded
As mentioned earlier, this survey mainly focusses the servers, the users which can be reassigned to the other
mobility-aware task offloading at the mobile edge net- underloaded servers within the same coverage area are
works. Therefore, we discuss stat-of-the-art mobility-aware identified. In the last step the users needed to migrate are
schemes used for task offloading across different mobile shifted to the idle or underloaded servers, and new users are
environments, such as MEC, FOG, Cloudlets, and vehic- accepted in the system. In [83] Mobility-aware Workflow
ular edge networks (VEN). In addition, we highlight the Offloading and Scheduling Strategy (MAWOSS) is pro-
characteristics or features of the mobility-aware schemes posed. The algorithm first identifies the task in a ready
used in different edge-centric paradigms. queue needed to be executed. Then the distance between
the end device and different edge servers is calculated and
5.1 MEC the server with fastest transfer speed is selected. The
algorithm minimizes the energy consumption and response
Mostly existing schemes on task offloading consider the time but does not handle the mobility fair enough because
user location to be fixed instead of dynamic mobility. This it is designed for single user environment and not well
can be quite impractical in the real scenarios. For optimal suited for the multiuser edge networks.
performance, with minimal latency the mobility of the end To handle the user mobility a time slot-based Lyapunov-
devices must be considered during offloading decisions. In optimization inspired Mobility-Aware Dynamic Service
this section we elaborate the state-of-the-art techniques to Placement (MDSP) algorithm is proposed that enables the
handle the user mobility in Mobile Edge Computing. mobile users to send service requests to their local MEC

123
Cluster Computing

server [84]. An SDN controller collects the requests-in- proposed in [90]. This framework uses Dijkstra‘s algorithm
formation and chooses the best MEC node to server a with machine learning to predicts the user next location and
corresponding user. The SDN controller checks the dif- the amount of energy required for task offloading and user
ferent parameters in the decision process such as, compu- migration.
tation capacity and computing delay. However, the
limitation of the system is that it considers the service 5.2 FOG
delay but does not consider the queuing delay of a task.
Queuing delay is a key consideration in time-varying FOG computing is an extended form of traditional cloud
user environment where the task offloading is not a one- computing. FOG provides the network and computing
time operation [85]. Some mobility aware techniques per- resources in a distrusted way close to the end users.
form the base station allocation to the end users based on However, the FOG has limited-service coverage; hence the
channel conditions. In [86] a Heuristic Mobility Aware mobile users might get out of radio coverage resulting in
Offloading Algorithm (HMAOA) is presented to optimize more handovers. Higher handovers have adverse effects
the task offloading in MEC environments for high speed like increased service cost, higher energy consumption and
moving users. The next expected base station of the user is increased latency. Therefore, considering the user mobility
predicted by considering not only the computational while migrating the task to the FOG Computing Node
resources but also the channel fading, noise, interference, (FCN) has significant importance.
and distance from the end user to the base station. The We discuss some of the works related to mobility in
algorithm promises to reduce the latency and energy con- FOG networks. A mixed Integer Nonlinear programming
sumption, but its mobility prediction feature produces the problem (MINLP) inspired scheme using Gini Coefficient
short-term results. Due to this limitation the tasks that need based FCNs selection algorithm (GCFSA) is presented in
to be offloaded or cached again cannot be guaranteed to [91]. The algorithm assumes that the end user has sub-
have the optimal services. mitted a workflow to the FCN and tends to move away
Further works on mobility awareness include Mobility- without collecting the execution results. Their mobility
Aware Deep Reinforcement Learning (M-DRL) [87] model focuses on the current state of the active mobile
technique for optimal service provisioning in MEC. To be users to predict the sojourn time which ultimately reduces
more specific the authors developed a glimpse mobility the probability of migrations. The model also reduces the
model to predict a series of next expected locations. The offloading failures. Another scheme that focuses to mini-
model uses an independent training algorithm which con- mize the offloading failures called mobility-aware and
siders the user current location as 2-dimensional vector genetic algorithm- (MAGA)is proposed in [92]. MAGA
having horizontal and vertical distance of mobile user from supports multisite offloading scenario, offering finest
the base station and mathematically determines the base offloading decisions to enhance the offloading success
station with least reported delay. Similar methodology is frequency and minimize the power usage of the mobile
adopted in [88] called as Mobility-aware Edge Service devices while satisfying the task completion time require-
Placement method (MAESP). This is a multi-objective ments in FOG environments. However, computational cost
method assisted with machine learning technique of algorithm is high and is well-suited for single-job
(CMABDO) which considers user mobility information scenarios.
(time, location) as context of the user to predict the next An offloading framework proposed in [93] gathers
possible location. The authors in [89] worked out to Global Positioning System (GPS) traces of mobile user and
develop a mobility prediction technique based on social predict the user‘s next position. This information is useful
generative adversarial networks (SGAN). for Wireless Access Point (WAP) selection during
Traditional mobility prediction techniques only consider offloading. The aforesaid WAP selection technique also
a specific user‘s next expected location just based on its considers the WAP location, network bandwidth and
historical mobility patterns. However, the above-mentioned latency constraints. The authors in [94] proposed a mobility
model not only focuses the user movement records but also aware collaborative framework for FOG networks (Mobi-
considers the moving activities of other users in the IoST). In this framework, if the user changes its location
vicinity, to enhance the prediction accuracy. The authors and gets disconnected from the base station, the base sta-
aim to reduce the response time and offloading failures tion intimates the central cloud about this and transfers the
count of the system. Migration energy and offloading user code and data to the cloud. The cloud node then
energy are also the important factors to be considered while analyses the user mobility patterns and predicts the user‘s
offloading a group of tasks. An energy-cum-mobility aware new location with the help of Markovian model. The main
offloading framework (MAEO) that forecasts the power modules of the aforesaid framework are i) modeling of user
required, during user migration and for task offloading is patterns, ii) user‘s next location prediction, and iii) results

123
Cluster Computing

delivery. The user location prediction is based on, day of upcoming set of tasks but does not consider the other
the week, time of the day, structure of user surroundings performance parameters such as, energy consumption, load
(roads/buildings). When user is re-connected to the new balancing and task execution time. An offline heuristic
base station, the results are subsequently delivered to the algorithm for service placement to multiple cloudlets is
mobile device. The model improves the energy consump- proposed in [99] which dynamically considers the distances
tion and reduces the delay. between the user and cloudlets. A stochastic method for the
Apart from probabilistic approaches the deterministic offloading of task in real time environment is presented in
approaches are also in practice, but their accuracy is [59]. Like the previous model this method also considers
questionable. However, these deterministic methods can the distances between the user and the cloudlets for suc-
achieve good results by exploiting the real time user cessful task migration.
mobility information. In this regard, Dynamically trading-
off fog and edge resources for performance and mobility- 5.4 Vehicular edge networks
aware IoT services (URMILA) is a novel method proposed
in [95]. URMILA determines the mobile user next location With the advent of high-speed internet technologies and
holistically by observing the motion path in real time. This IoT, the connected vehicles have prevailed over the tradi-
feature is achieved by creating a session between the user tional vehicular Adhoc networks. The modern multimedia
and FCN using RESTful API. User movement path, start and delay tolerant applications like autonomous driving
and end locations are determined by using Open Street demand the stringent quality of service (QoS) in the
Maps and Google Maps APIs. The optimization parameters vehicular networks employed with benchmarked commu-
considered are energy consumption and response time, but nication technologies and efficient decision systems.
the framework is only suitable for single unified applica- Meeting such requirements is a challenge and strongly
tions/tasks and does not consider the microservice based constrained with high computational power, dynamic data
applications which are multithreaded in nature. The authors forwarding techniques, minimum congestion, and maxi-
in [96] used Bayesian network based prediction service mum response time with minimum service/execution
called MoBaaS [97] to proposed the Enhanced Follow-me- delay. In such densely populated network, the high-speed
Cloud (EFMEC). The user mobility is first predicted with private/public vehicles frequently and rapidly change their
MoBaaS then applies a multiple attribute decision algo- locations and associations with the roadside units (RSU)
rithm for service placement. Nevertheless, this technique resulting in higher deadline misses, energy and resource
only works when the historical record of user mobility wastage.
patterns is available. Different studies have been proposed to handle the
mobility of users in VEN. In [100] the authors presented
5.3 Cloudlets Mobility-aware Computation Offloading in MEC based
Vehicular Wireless Networks (MACVWA) considering the
In next generation 5G network the number of mobile users random mobility of the vehicles and handovers across the
and IoT devices is a multi-billion figure with diverse net- RSUs. Each RSU holds an edge server to assist the vehicle
work resources. With the increasing demands of users, the in task computation. The RSUs are interpointed with each
concept of service provisioning is orchestrated from cen- other through a support network, usually but not limited to
tralized cloud-datacenters to distributed clouds associated the metropolitan cellular network. The vehicular device
with the corresponding edge servers or mini data centers, offloads the task and gets back its computation results.
which enhances the fault tolerance capability, reduces the During the transmission if handover occurs, the core net-
latency of the user applications. work transfers the executed content from old RSU com-
Like the edge servers, the concept of task offloading is municates to the new one. The handover is determined
applied in cloudlets. Different studies have been proposed based on real-time signal to noise ratio (SNR). Velocity
to optimally offload the group of tasks to cloudlets. Some variations are considered to be discrete values with the help
of them focus on user mobility as well. The researchers in of Boundless Simulation Area mobility model (BSA)
[98] proposed a Virtualized Network Framework (VNF) [101]. The vehicular movements are consequently consid-
which enables the mobile users to freely move among ered as Markovian chain.
cloudlets with strict joint delay and mobility requirements. Soft-VAN: Mobility-Aware Task Offloading in Soft-
The system checks the mobility profile of the user based on ware-Defined Vehicular Network is another scheme [102]
its previous locations traces. Using different mining and to optimize the computational offloading and network
machine learning techniques the next location of the user is latency in vehicular networks. This SDN based
approximated for task offloading. The algorithm jointly scheme works in two phases, node selection and task
optimizes the delay and service orchestration required for computation. An integer linear program is formulated in

123
Cluster Computing

the first phase to select the finest number of RSUs. In leased to the vehicle requesting for task offloading. The
contrast, non-linear optimization problem is formulated in roadside units periodically broadcast the resource blocks of
the second phase to minimize latency in task computation available spectrum. The mobile devices in a VEN deter-
considering the other system constraints as well. Due to its mine their bandwidth requirements based on their task
NP-hard nature the computation problem is divided into sizes and required computational power and time. The
two sub-problems and solved by evolutionary algorithms. spectrum is then allocated to those devices through
First subproblem refers to the selection of optimal edge OFDMA technology.
server if the offloading is required. In second subproblem A vehicular cloud is consisted of moving vehicles in an
the path required for downloading the completed task is SMDP-Based Resource Allocation in Vehicular Cloud
selected. The SDN controller predicts the best RSU and Computing Systems [108]. Each vehicle is assumed to have
associated edge server for vehicular user. Like many earlier equal computing and storage capacity. A central controller
stated studies this scheme also used Markov model for collects the tasks needed to be offloaded, and delegates
prediction activity while energy consumption and latency them to the other vehicular nodes federated in cloud which
has been improved. reduces the delay, saves energy, and efficiently utilizes the
A joint optimization framework known as Mobility- resources. However, this is a sort of homogenous system
Aware Edge Caching and Computing (MAECC) is pro- with single point of failure involved.
posed in [103]. Caching and offloading is optimized using
multi-time scale Deep Q-Network (DQN) algorithm. The
model is specifically designed for limited storage and 6 Comparisons
computational resources at the user end and RSUs. The
model utilizes the RSUs and other vehicles on the road for In the previous sections, we provided the detailed literature
offloading or caching. review of task offloading techniques with respect to dif-
The authors in [104] proposed AVE: Autonomous ferent optimization objectives, constraints, models, algo-
vehicular edge computing framework with ACO-based rithms, and support networks. In this section we first
scheduling approach for VEN. Global positioning system compare the mobility models and then the generalized
(GPS) is exploited to track the vehicle‘s locations. Based offloading techniques used in mobile environments.
on GPS results moving vehicles are grouped into different
clusters inspired by ACO. The Intragroup task offloading is 6.1 Comparison of mobility models
then performed to optimize the task computation process
and reduce the latency. Similar approach is followed in For task offloading in MEC some models use SDN based
[105] and [106] to cooperatively use the vehicle resources techniques to handle the mobility and consider the user‘s
to enhance the computation and communication capacities geographic location or recorded user mobility patterns.
of the moving vehicles. SND based models do not consider the mobile user‘s tra-
A spectrum leasing methodology is proposed for task jectory in real time. Then, some predictive models such as
offloading in VENs [107]. The radio spectrum resources of Markov chain, exist that consider the real time network
cellular network and RSUs of an edge network are jointly performance and MEC server utilization. Markovian

Table 1 Features/parameters present in different mobility models


Mobility models Features
Geographic Recorded mobility Path Server MTTR Makespan User
location patterns prediction load Mobility

1 SDN based framework 4 9 9 9 9 9 9


[56]
2 Stochastic Models [59] 9 4 9 9 9 9 9
3 RWP based [62, 63] 9 9 4 9 9 9 4
4 RWP based [64] 9 9 9 9 4 4 4
5 RWP based [65, 109] 9 9 9 9 9 9 4
6 SLAW [66] 9 9 4 9 9 9 4
7 Head-light Model (HLM) 9 9 4 9 9 9 4
[70]

123
Cluster Computing

Table 2 Features/parameters in offloading schemes in MEC


Offloading work areas in MEC Features/parameters
Energy Latency Load Mobility Channel State Execution
Consumption Balancing Prediction Information Time

1 Linear Search Algorithm [110] 9 4 9 9 9 4


2 Distributed cloud-aware power control (CaPC) 4 9 4 9 9 9
algorithm [28]
3 Multi-stage sequential game model [32] 4 4 9 9 9 9
4 Heuristic offloading decision algorithm 9 4 4 9 9 9
(HODA) [33]
5 Adaptive sequential task offloading (ASTO) 4 4 9 9 9 9
[34]
6 Energy-efficient computation offloading 4 9 4 9 9 9
(EECO) [35]
7 Genetic Algorithm based multi-objective 4 9 9 9 9 9
work-flow algorithm [36]
8 Minimized latency and energy consumption 4 4 9 9 9 4
[1, 9, 110, 111]
9 A 3-stage Task offloading [26]framework 9 9 9 9 4 4
10 Mixed-integer non-linear-programing problem 4 4 9 9 9 9
[112]

models are helpful to predict the future network conditions optimize the task offloading. Table 1 summarizes the above
and dynamic task assignment decisions but the complexity discussion.
of solving the problem and accuracy of cost modeling is
adversely proportional to each other. Performing compu- 6.2 Comparison of generalized offloading
tation offloading based on task completion time, is another techniques
approach which exploits the genetic algorithms to evolve
the optimum solution. Random Way point and Leavy Walk The searching-based algorithms produce better offloading
models are also used for user location prediction, by con- decisions but the at the cost of higher computational
sidering the mobility as a random function. However, the complexity. Game-theory based heuristic reduces the
aforesaid models lack to offer an optimal solution. RWP energy and delay requirements at the same time. Multistage
offers better mobility prediction but increased task execu- task offloading framework classifies the mobile devices and
tion time and energy consumption. SLAW is suitable for determine task priority then offload it to suitable sBSs. This
homogenous task offloading specifications for local users, scheme also considers Interference, collisions and number
but its efficacy is low in delay sensitive networks. of contenders is also considered. Some schemes incorpo-
In addition, the models based on user walk patterns rate the multi-access characteristics of the next generation
sometimes result in non-uniform stationary mobile node 5G networks for energy efficiency. Delay and resource
distribution, increased energy consumption and execution utilization can jointly be optimized by some techniques. A
time. Then we have some heuristic based models which joint optimization of task offloading and sBSs allocation
generate good schedules for task offloading but ignore the reduces the power consumption offset by the delay con-
user motion information and have higher time complexity straints. Genetic algorithm based multi-objective work-
so cannot be used in real time systems. In the Headlight flow-based computation offloading framework
model there is no global view of the network entities and concurrently considers energy consumption and task
resources, and it forecast the user locations based on sta- deadline constraints while the task offloading to edge ser-
tistical analysis of historical mobility record. Such tech- vers, but the user mobility and the topology changes were
niques cannot well predict the user location. Analysis also not considered. Table 2 summarizes the previous
showed that most of the above-mentioned works regarding discussion.
mobility-aware models are relatively simple, and straight-
movement based. Hence, it is needed to develop new
schemes which consider the dynamic mobility patterns to

123
Cluster Computing

6.3 Comparison of mobility-aware offloading from the end user. In MDSP an SDN controller using
techniques in mobile environments Lyapunov optimization technique reduces the computation
and service delay. It does optimize the queuing delay. For
This section summarizes the aforesaid mobility-aware high-speed edge users HMAOA technique is presented that
offloading techniques used in different mobile environ- predicts locations with the help of channel state informa-
ments. We see that majority of the said works focus to tion. It reduces the latency, but mobility prediction is not
optimize energy, latency, and execution time. We extend for long term.
our discussion to these three performance parameters. A deep reinforcement learning technique M-DRL using
glimpse mobility model predicts a series of next possible
6.3.1 Energy efficiency location. Due to its two-dimensional foreseeing feature, the
distance between user and edge server is well optimized
DMPO minimizes the offloading path after user movement. resulting in least reported delay. Another similar scheme is
It predicts the user data size and next location for very short MAESP which is context-aware delay minimization
term. It does not outperform when the base stations are not framework enabled with mobility prediction feature.
equidistant. EMM categorizes the information in two Another interesting approach is SGAN. To enhance the
classes. First contains the user task size, its location and no prediction accuracy, it not only considers the moving pat-
of base stations. The second class contains channel state terns of a certain user but also the mobility behavior of the
information. The user handoffs are energy consumption is other users residing in the vicinity. With such efficient
minimized based on those classes. MAWOSS first make a location prediction, it optimizes the involved delays as
ready queue of tasks then based on distance between user well.
and offloading site, select the suitable server for task MAEO is a Dijkstra algorithm-based technique that
offloading. It does not work well in multiuser system. groups the users with same mobility and computation
In heuristics-based techniques the HMAOA outperforms constraints in homogenous groups. It creates the shortest
for high-speed moving users. It predicts the user next paths to resolve the task interdependencies and optimizes
location by foreseeing both computational and spectral the latency by considering migration and offloading cost.
resources. Migration energy and offloading energy are GCFSA is an algorithm which operates by considering the
considered for optimization in MAEO using machine post mobility effects. It also reduces the offloading failures
learning algorithm. GCFSA provides solution for those which ultimately reduce the delays.
users who move away just after submitting the offloading
data. It predicts a sojourn time to minimize the probability 6.3.3 Execution time
of migrations and reduces the offloading failures. MAGA is
another effort to reduce the offloading failures and energy The execution time of tasks is one of the important per-
consumption but its complexity is higher. ENDA predicts formance parameters to be considered while taking
the user next location using GPS and Open MP. Its scope is offloading decisions in the mobile environments. It
limited to small area. Mobi-IoST is a collaborative becomes more critical in vehicular edge networks, since
framework to handle mobility in different mobile envi- the self-driving cars cannot afford the delays caused due to
ronments. A central cloud periodically monitors the mobile greater execution time of the offloaded tasks. Apart from
nodes and facilitates them in task offloading from one edge mobility prediction the task size can also be predicted as
server to the other one. Its uses Markov model for pre- performed in DMPO. Although, the prediction is for short
diction along with energy saving. Similarly, URMILA also term even then the ultimate objectives can be achieved.
observes the users but in real time. It also reduces the The only limitation of DMPO is it assumes the distribution
energy consumption but cannot be used in complex and of base stations to be even, which is somewhat unrealistic.
multithread dependent applications. MACVWA uses the EMM focus on energy saving. However, for its operation it
SNR values and velocity variations to predict the user required to classify the incoming set task based on their
mobility and energy saving. Soft-VAN uses SDN with execution time. The aforesaid data centric nature of the
Markov chains to optimize the energy consumption. algorithm enables it to optimize the execution times of
task.
6.3.2 Latency/delay minimization Similar approach is adopted in MAWOSS which works
only in single user environment. HMAOA is designed for
MAWOSS is a mobility aware task offloading technique high-speed moving devices. While taking offloading deci-
that reduces the latency of the offloaded task. The fastest sions it considers the task execution time along other
and nearest edge server is selected considering the distance parameters. However, due to its week mobility prediction
capability, the tasks needed to offload or cached again and

123
Cluster Computing

Table 3 Comparison of mobility-aware task offloading techniques in mobile environments


Offloading schemes in Features/parameters
MEC
Energy Latency Load Execution Channel state Distance between user Offloading
Efficiency Reduction Balancing Time information & offloading site failures
reduction

1 MEC MobMig [11] 4 9 4 9 9 9 9


2 DMPO [113] 4 9 9 4 9 9 9
3 EMM [114] 4 9 4 4 4 9 9
4 MAWOSS 4 4 9 4 9 9 9
[83]
5 MDSP [84] 9 4 9 9 4 9 9
6 HMAOA [86] 4 4 9 4 4 4 9
7 (M-DRL) [87] 9 4 9 9 9 4 9
8 MAESP [88] 9 4 9 9 9 4 9
9 SGAN based 9 4 9 9 9 9 4
[89]
10 MAEO [90] 4 4 9 9 9 9 9
11 FOG GCFSA [91] 4 4 9 9 9 9 4
12 MAGA [92] 4 9 9 4 9 9 4
13 ENDA [93] 4 9 4 9 9 9 9
14 Mobi-IoST 4 4 9 9 9 4 9
[94]
15 URMILA [95] 4 4 4 9 9 4 9
16 EFMEC [97] 4 4 9 9 9 9 9
17 VEN MACVWA 4 4 9 4 4 4 9
[100]
18 Soft-VAN 4 4 9 9 9 9 9
[102]
19 MAECC 9 9 9 9 9 9 4
[103]
20 AVE [104] 9 4 9 9 9 4 9
21 Spectrum 9 4 9 9 9 4 4
Leasing
[107]
22 SMDP [108] 4 4 4 9 9 9 9

again cannot be executed. MAGA is an evolutionary and delay. Since MEC is a developing technology, there is
algorithm that well optimizes the energy usage, latency, still enough room for improvement and research, which
and execution time of task, but at the cost of higher com- needs attention of both industry and academia. In the
putational complexity. Table 3 summarizes the previous subsequent sections, first we discuss some open research
discussion. challenges and opportunities at a glance, related to scala-
bility (Sect. 6.1), and security (Sect. 6.2). Conclusively, we
present in detail the mobility related research challenges
7 Future research challenges and opportunities as a crux of the survey in (Sect. 6.3).
and opportunities
7.1 Scalability
In previous sections, we showed that MEC has got plenty
of attention because of efficient resource and service pro- Scalability is one of the critical components of mobile edge
visioning close to the user end. MEC enables offloading of networks. Various studies are discussed earlier that propose
real time, resource hungry and multimedia applications to the scalable offloading solutions. They include different
the proximate servers while optimizing the energy costs resource provisioning and task allocation schemes in the

123
Cluster Computing

MEC network. Further, some techniques exploit the edge intrusions effectively [115], hence the exploration of new
server capacity, to balance the workload across the net- methods to mitigate the intrusions is also worth doing.
work. Some techniques use task aggregation and prioriti- Reliable computation is a vital requirement to enhance the
zation mechanism. Some techniques focus to offload tasks uptime of network resources. The edge networks are dis-
by foreseeing the channel conditions. Some studies pro- tributed in nature and can easily be targeted by DoS
posed mutual coordination-based solutions for scalable attacks. Thus, working on some proactive defense mea-
offloading decisions. Moreover, existing studies mainly sures to ensure the network and data integrity would be an
focus on the offloading, caching and mobility management exciting research area. Additionally, in layered architec-
to optimize the network performance using centralized tures like MEC, the fault tolerance plays vital role in
SDN technology. One central SDN controller cannot successful execution of user submitted tasks [116]. Hence,
effectively scale down the network resources and service the protection and restoration mechanisms across different
provisioning and can suffer single of failure. The distri- layers of MEC would improve the availability of comput-
bution of SDN controller to lower level controllers can ing and caching resources. Ensuring the user data privacy
balance the network load and provide better control of the in MEC, specifically in D2D communication it is hard to
system in large scale networks [27]. Thus, developing an offer trusted services at the network edges.
edge cloud-level service orchestrator to induce flexible
manageability can also improve the network scalability. 7.3 Mobility awareness in mobile environments
We could not find notable works related to scalable system
design in MEC. Further, some novel technique that forms Various mobility aware models to handle the task
clusters of the edge devices (an IoT layer) that classifies offloading in MEC have been discussed earlier. Those
and group the similar tasks. The cluster can be formed models consider different performance parameters to
based on the tasks‘ computational requirement, priority, optimize the system efficiency. Some models use SDN
and delay requirements, which is a worthful research area based techniques to handle the mobility and consider the
to be explored. We could not find a study in this direction device location and perform averaging functions to predict
that uses the clustering mechanism to scale up the edge the user locations. Some other predictive models consider
network. the real time network performance and MEC server uti-
lization. Random Way point and Leavy Walk models are
7.2 Security, privacy and reliability also used for user location prediction. Nevertheless, the
aforesaid models lack to offer an optimal solution as they
MEC systems are usually more resource constrained than are unable to predict locations based on realistic user tra-
centralized cloud systems and are more prone to the jectories. The stochastic models increase the time com-
security attacks. Security is a critical challenge in edge plexity. The random walk models escalate the power
paradigm ecosystem due to several factors discussed ear- consumption and execution time. We also noted that the
lier. The edge networks involve different enabling tech- said models are relatively simple. Therefore, exploring new
nologies like; wireless networks, distributed systems and schemes which consider the dynamic and realistic mobility
virtualization platforms [41]. Several security protocols are patterns to optimize the prediction accuracy and enhance
proposed for fully connected mesh networks, however their the task offloading process would be an appealing research
applicability in mobile edge networks is questionable due direction.
to the intermittent connectivities in MEC. Though, some Existing techniques explored yet handle computation
studies cover the secure communication aspects in Mobile offloading and real-mobility management separately.
cloud networks [42]. However, such solutions do not well Hence, the impact of mobility on some state-of-the-art
suit for mobile edge networks, because the working sur- computation offloading techniques is evident to be ana-
roundings of edge devices may face several new threats lyzed. Without considering the real-time mobility factors,
different from MCC [27]. More elaborately, the security such as channel conditions, MEC servers’ utilization, and
mechanism should be unaware of the intruder‘s channel network traffic load, computation offloading techniques
state information (CSI). To the best of our knowledge no produce sub-optimal task assignment. Such sub-optimal
notable study offers secure task offloading without know- task placement has adverse effects on QoS and energy
ing the user CSI. consumption [117]. There is an opportunity to develop a
Further, we noticed that the technical limitations of edge unified approach that jointly considers computation
and IoT devices lack the hardware protection as compared offloading and mobility management.
to commodity servers in centralized systems. Another Further, the suitable mobility models for accurately
important aspect of security in MEC is the secure authen- modeling user movement in real-time MEC scenarios are
tication. The traditional solutions do not handle the also a challenging task. The user mobility traces in mobile

123
Cluster Computing

edge networks can be expressed as non-linear function due 8 Conclusion


to random movements of mobile user [118]. Such linear
data can further leverage the machine learning based pat- In this article, we presented a detailed survey of mobility
tern recognition techniques to improve the computation aware MEC offloading techniques. The mobility of UE can
offloading in MEC. We explored that majority works lead to several challenges during the process of MEC
related to mobility prediction are based on reinforcement offload such as intermittent connectivity, security, and
learning (RL) due to its high accuracy [77]. However, RL reliability. As a result, machine learning, and nature
based models involve higher time complexity due to its inspired algorithms have been proposed to predict and
online learning procedures and can increase the latency of manage user mobility for efficient MEC offloading
the MEC system [77]. As mentioned earlier, the user frameworks. We presented the taxonomy of the state-of-
mobility traces are considered as historical records of the the-art MEC offloading frameworks in terms of objectives,
mobile users at different geographic location [119]. Those constraints, mobility models, algorithms, and support net-
records also contain the user speed, direction, and distance works. We listed a detailed literature review and compar-
from the base station. For such hand-crafted features and ison of fog, cloudlet, vehicular networks and MEC based
nonlinear location information, using deep neural networks mobility-aware offloading frameworks. At last, we listed
(DNN), fuzzy logic, genetic algorithms and RL based research challenges in the form of security, reliability, and
algorithms is not a practical approach in context of latency scalability for mobility aware MEC offloading frameworks.
and energy consumption. Therefore, developing some
novel light weight techniques with optimal latency to Acknowledgement The authors would like to thank the Deanship of
Scientific Research at Umm Al-Qura University for supporting this
predict the user location is a very interesting and needful work by Grant Code: 19-COM-1-01-0016.
research opportunity. For instance, on offline learning
model that uses time series analysis with simple artificial
neural network (ANN) technique can be used to predict the References
user locations [120].
In addition, the current communication protocols and 1. Wang, Z., Zhao, Z., Min, G., Huang, X., Ni, Q., Wang, R.: User
mobility management schemes do not address the high- mobility aware task/assignment for mobile edge computing.
Futur. Gener. Comput. Syst. 85, 1–8 (2018)
speed user, which can also be an interesting area to be 2. Sardar Khaliquz Zaman, Tahir Maqsood, Mazhar Ali, Kashif
figured out. In existing mobility management schemes, Bilal, Sajjad A. Madani, and A. U. R. Khan, A load balanced
machine learning techniques are applied only between the task scheduling heuristic for large-scale computing systems,
mobile node and edge server to predict the next possible Computer Systems Science and Engineering, vol. 34, pp. 1–12,
2019 .
location. However, in D2D communication/offloading the 3. Sheng, J., Hu, J., Teng, X., Wang, B., Pan, X.: Computation
repercussions of user mobility and offloading site selection offloading strategy in mobile edge computing. Information 10,
would be different and challenging than conventional 191 (2019)
offloading scenarios. Hence developing D2D mobility 4. J. Shuja, A. Gani, M. H. ur Rehman, E. Ahmed, S. A. Madani,
M. K. Khan, et al., Towards native code offloading based MCC
models with intelligence would be an interesting research frameworks for multimedia applications: a survey, Journal of
direction. Network and Computer Applications, vol. 75, pp. 335–354,
Moreover, we also noted that there are few studies 2016.
focusing on mobility in cloudlets. Even those studies have 5. Yu, W., Liang, F., He, X., Hatcher, W.G., Lu, C., Lin, J., et al.:
A survey on the edge computing for the Internet of Things. IEEE
ignored some critical issues. For instance, till now the Access 6, 6900–6919 (2017)
studies related to mobility aware offloading mostly handle 6. Ziming, Z., Fang, L., Zhiping, C., Nong, X.: Edge computing:
the offloading between single computing server and mul- platforms, applications and challenges. Journal of Computer
tiple mobile users. However, in cloudlets the edge resour- Research and Development 55, 327–337 (2018)
7. Sun, X., Ansari, N.: Latency aware workload offloading in the
ces form a federated cloud where single mobile users cloudlet network. IEEE Commun. Lett. 21, 1481–1484 (2017)
offloading request may be distributed among more than one 8. Yusong, S., Hui, S., Jie, C., Quan, Z., Wei, L.: Edge computing:
edge server, delegated by cloud controller. Handling the a new computing model in the age of internet of things. Com-
mobility here is a challenging task. Further, in such sce- puter Research Development 54, 907–924 (2017)
9. Mach, P., Becvar, Z.: Mobile edge computing: a survey on
nario the communication cost/delay optimization between architecture and computation offloading. IEEE Communica-
edge nodes to share the execution results is highly ques- tions: Surveys Tutorials 19, 1628–1656 (2017)
tionable. Like MEC and FOG, the reinforcement learning 10. Shuja, J., Gani, A., Naveed, A., Ahmed, E., Hsu, C.-H.: Case of
techniques with predictive capabilities can be implied here ARM emulation optimization for offloading mechanisms in
mobile cloud computing, Future Gener. Comput. Syst. 76,
to smartly handle the initial offloading process. 407–417 (2017)
11. Peng, Q., Xia, Y., Feng, Z., Lee, J., Wu, C., Luo, X., et al.:
Mobility-aware and migration-enabled online edge user

123
Cluster Computing

allocation in mobile edge computing. In: 2019 IEEE Interna- 29. Nur, F.N., Islam, S., Moon, N.N., Karim, A., Azam, S., Shan-
tional Conference on Web Services (ICWS), pp. 91–98, 2019 mugam, B.: Priority-based offloading and caching in mobile
12. Uzaman, S.K., Shuja, J., Maqsood, T., Rehman, F., Mustafa, S.: edge cloud. J. Commun. Softw. Syst. 15, 193–201 (2019)
A systems overview of commercial data centers: initial energy 30. Hao, Y., Chen, M., Hu, L., Hossain, M.S., Ghoneim, A.: Energy
and cost analysis. Int. J. Inf. Technol. Web Eng. 14, 42–65 efficient task caching and offloading for mobile edge computing.
(2019) IEEE Access 6, 11365–11373 (2018)
13. Ahmed, E., Ahmed, A., Yaqoob, I., Shuja, J., Gani, A., Imran, 31. Ioannou, A., Weber, S.: A survey of caching policies and for-
M., et al.: Bringing computation closer toward the user network: warding mechanisms in information-centric networking. IEEE
is edge computing the solution? IEEE Commun. Mag. 55, Commun. Surv. Tutor. 18, 2847–2886 (2016)
138–144 (2017) 32. Yang, J., Jiang, B., Lv, Z., Choo, K.-K.R.: A task scheduling
14. Saleem, M., Saleem, Y., Hayat, M.F.: Stochastic QoE-aware algorithm considering game theory designed for energy man-
optimization of multisource multimedia content delivery for agement in cloud computing. Futur. Gener. Comput. Syst. 105,
mobile cloud. Clust. Comput. 23, 1381–1396 (2020) 985 (2017)
15. Pham, Q.-V., Fang, F., Ha, V.N., Le, M., Ding, Z., Le, L.B., 33. Lyu, X., Tian, H., Sengul, C., Zhang, P.: Multiuser joint task
et al.: A survey of multi-access edge computing in 5G and offloading and resource optimization in proximate clouds. IEEE
beyond: fundamentals, technology integration, and state-of-the- Trans. Veh. Technol. 66, 3435–3447 (2016)
art, arXiv preprint arXiv:1906.08452, 2019 34. Deng, M., Tian, H., Lyu, X.: Adaptive sequential offloading
16. Shuja, J., Bilal, K., Alanazi, E., Alasmary, W., Alashaikh, A.: game for multi-cell mobile edge computing. In: 23rd Interna-
Applying machine learning techniques for caching in next- tional Conference on Telecommunications (ICT), 2016, pp. 1–5
generation edge networks: a comprehensive survey. J. Netw. 35. Yang, Y., Ma, Y., Xiang, W., Gu, X., Zhao, H.: Joint opti-
Comput. Appl. 181, 103005 (2021) mization of energy consumption and packet scheduling for
17. Pham, Q.-V., Fang, F., Ha, V.N., Piran, M.J., Le, M., Le, L.B., mobile edge computing in cyber-physical networks. IEEE
et al.: A survey of multi-access edge computing in 5G and Access 6, 15576–15586 (2018)
beyond: fundamentals, technology integration, and state-of-the- 36. Xu, X., Fu, S., Yuan, Y., Luo, Y., Qi, L., Lin, W., et al.: Mul-
art. IEEE Access 8, 116974–117017 (2020) tiobjective computation offloading for workflow management in
18. Waqas, M., Niu, Y., Li, Y., Ahmed, M., Jin, D., Chen, S., et al.: cloudlet-based mobile cloud using NSGA-II. Comput. Intell. 35,
Mobility-aware device-to-device communications: principles, 476–495 (2019)
practice and challenges. IEEE Commun. Surv. Tutor. 22, 1863 37. Cui, Y., He W., Ni, C., Guo, C., Liu, Z.: Energy-efficient
(2019) resource allocation for cache-assisted mobile edge computing.
19. Zaman, S.K., Tahir Maqsood, M.A., Bilal, K.: A load balanced In: 2017 IEEE 42nd Conference on Local Computer Networks
task scheduling heuristic for large-scale computing systems. (LCN), 2017, pp. 640–648
Comput. Syst. Sci. Eng. 34, 4 (2019) 38. Nunna, S., Kousaridas, A., Ibrahim, M., Dillinger, M.,
20. Sardar Khaliq uz Zaman, A.U.R.K., Malik, S.U.R., Khan, A.N., Thuemmler, C., Feussner, H., et al.: Enabling real-time context-
Maqsood, T., Madani, S.A.: Formal verification and perfor- aware collaboration through 5G and mobile edge computing. In:
mance evaluation of task scheduling heuristics for makespan 2015 12th International Conference on Information Technology-
optimization and workflow distribution in large-scale computing New Generations, 2015, pp. 601–605
systems. Comput. Syst. Sci. Eng. 32, 227 (2017) 39. Rehman, F., Khalid, O., Bilal, K., Madani, S.A.: Diet-Right: a
21. Al-Habob, A.A., Dobre, O.A., Armada, A.G., Muhaidat, S.: smart food recommendation system. KSII Trans. Internet Inf.
Task scheduling for mobile edge computing using genetic Syst. 11, 2910 (2017)
algorithm and conflict graphs. IEEE Trans. Veh. Technol. 40. Rehman, F., Khalid, O., Madani, S.A.: A comparative study of
(2020). https://doi.org/10.1109/TVT.2020.2995146 location-based recommendation systems. Knowledge Eng.
22. Khan, W.Z., Ahmed, E., Hakak, S., Yaqoob, I., Ahmed, A.: Review 32, e7 (2017)
Edge computing: a survey. Futur. Gener. Comput. Syst. 97, 41. R. Roman, J. Lopez, and M. Mambo, Mobile edge computing,
219–235 (2019) fog et al.: A survey and analysis of security threats and chal-
23. Mao, Y., You, C., Zhang, J., Huang, K., Letaief, K.B.: Mobile lenges, Future Gener. Comput. Syst., vol. 78, pp. 680–698,
edge computing: survey and research outlook, arXiv preprint 2018.
arXiv:1701.01090, 2017 42. Vemulapalli, C., Madria, S.K., Linderman, M.: Security
24. Kemp, R., Palmer, N., Kielmann, T., Seinstra, F., Drost, N., frameworks in mobile cloud computing. In: Handbook of
Maassen, J., et al.: Eyedentify: multimedia cyber foraging from Computer Networks and Cyber Security, Springer, Berlin,
a smartphone. In: 2009 11th IEEE International Symposium on pp. 1–41 (2020)
Multimedia, 2009, pp. 392–399 43. Zheng, T.-X., Wang, H.-M., Deng, H.: Improving anti-eaves-
25. Shi, B., Yang, J., Huang, Z., Hui, P.: Offloading guidelines for dropping ability without eavesdropper’s CSI: a practical secure
augmented reality applications on wearable devices. In: Pro- transmission design perspective. IEEE Wireless Communica-
ceedings of the 23rd ACM International Conference on Multi- tions Letters 7, 946–949 (2018)
media, 2015, pp. 1271–1274 44. Wu, W., Zhou, F., Hu, R.Q., Wang, B.: Energy-efficient
26. Chen, X., Jiao, L., Li, W., Fu, X.: Efficient multi-user compu- resource allocation for secure NOMA-enabled mobile edge
tation offloading for mobile-edge cloud computing. IEEE/ACM computing networks. IEEE Trans. Commun. 68, 493–505
Trans. Netw. 24, 2795–2808 (2015) (2020)
27. Wang, S., Zhang, X., Zhang, Y., Wang, L., Yang, J., Wang, W.: 45. Wu, W., Wang, X., Zhou, F., Wong, K.-K., Li, C., Wang, B.:
A survey on mobile edge networks: convergence of computing, Resource allocation for enhancing offloading security in
caching and communications. IEEE Access 5, 6757–6779 NOMA-enabled MEC networks. IEEE Syst. J. (2020). https://
(2017) doi.org/10.1109/JSYST.2020.3009723
28. Mach, P., Becvar, Z.: Cloud-aware power control for real-time 46. Mtibaa, A., Harras, K., Alnuweiri, H.: Friend or foe? Detecting
application offloading in mobile edge computing. Trans. Emerg. and isolating malicious nodes in mobile edge computing plat-
Telecommun. Technol. 27, 648–661 (2016) forms. In: 2015 IEEE 7th International Conference on Cloud

123
Cluster Computing

Computing Technology and Science (CloudCom), 2015, 65. Bhattacharya, A., De, P.: A survey of adaptation techniques in
pp. 42–49 computation offloading. J. Netw. Comput. Appl. 78, 97–115
47. He, X., Jin, R., Dai, H.: Deep PDS-learning for privacy-aware (2017)
offloading in MEC-enabled IoT. IEEE Internet Things J. 6, 66. Shahzamal, M., Pervez, M., Zaman, M., Hossain, W., Networks,
4547–4555 (2018) M.: Mobility models for delay tolerant network: a survey 6,
48. Gazori, P., Rahbari, D., Nickray, M.: Saving time and cost on 121–134 (2014)
the scheduling of fog-based IoT applications using deep rein- 67. Lai, P., He, Q., Abdelrazek, M., Chen, F., Hosking, J., Grundy,
forcement learning approach. Futur. Gener. Comput. Syst. 110, J., et al.: Optimal edge user allocation in edge computing with
1098–1115 (2020) variable sized vector bin packing. In: International Conference
49. Alam, M.G.R., Hassan, M.M., Uddin, M.Z., Almogren, A., on Service-Oriented Computing, 2018, pp. 230–245
Fortino, G.: Autonomic computation offloading in mobile edge 68. Yao, H., Bai, C., Xiong, M., Zeng, D., Fu, Z.J.: Heterogeneous
for IoT applications. Futur. Gener. Comput. Syst. 90, 149–157 cloudlet deployment and user-cloudlet association toward cost
(2019) effective fog computing. Concurr. Comput. 29, e3975 (2017)
50. Zhou, P., Finley, B., Li, X., Tarkoma, S., Kangasharju, J., 69. Deng, S., Huang, L., Hu, D., Zhao, J.L., Wu, Z.: Mobility-en-
Ammar, M., et al.: 5G MEC computation handoff for mobile abled service selection for composite services. IEEE Trans.
augmented reality, arXiv preprint arXiv:2101.00256, 2021 Serv. Comput. 9, 394–407 (2014)
51. Tamilselvan, L.: Client aware scalable cloudlet to augment edge 70. Qi, Q., Liao, J., Wang, J., Li, Q., Cao, Y.: Software defined
computing with mobile cloud migration service, iJIM, vol. 14, resource orchestration system for multitask application in
p. 165, 2020. heterogeneous mobile cloud computing. Mob. Inf. Syst. 2016,
52. Sonbol, K., Özkasap, Ö., Al-Oqily, I., Aloqaily, M.: EdgeKV: 1–17 (2016)
decentralized, scalable, and consistent storage for the edge. 71. Li, Y., Wang, W.: Can mobile cloudlets support mobile appli-
J. Parallel Distrib. Comput. 144, 28–40 (2020) cations? In: IEEE INFOCOM 2014-IEEE Conference on Com-
53. Chen, X.: Decentralized computation offloading game for puter Communications, 2014, pp. 1060–1068
mobile cloud computing. IEEE Trans. Parallel Distrib. Syst. 26, 72. Huang, M., Liu, W., Wang, T., Liu, A., Zhang, S.: A cloud-MEC
974–983 (2014) collaborative task offloading scheme with service orchestration.
54. Kabir, M.T., Masouros, C.: A scalable energy vs latency trade- IEEE Internet Things J. 7, 5792 (2019)
off in full-duplex mobile edge computing systems. IEEE Trans. 73. Liu, Y., Zeng, Z., Liu, X., Zhu, X., Bhuiyan, M.Z.A.: A novel
Commun. 67, 5848–5861 (2019) load balancing and low response delay framework for edge-
55. Abdellatif, A.A., Mohamed, A., Chiasserini, C.F., Tlili, M., cloud network based on SDN. IEEE Internet Things J. 7, 5922
Erbad, A.: Edge computing for smart health: context-aware (2019)
approaches, opportunities, and challenges. IEEE Network 33, 74. Zahoor, S., Javaid, N., Khan, A., Ruqia, B., Muhammad, F.J.,
196–203 (2019) Zahid, M.: A cloud-fog-based smart grid model for efficient
56. Chamola, V., Tham, C.-K., Chalapathi, G.S.: Latency aware resource utilization. In: 2018 14th International Wireless Com-
mobile task assignment and load balancing for edge cloudlets. munications & Mobile Computing Conference (IWCMC), 2018,
In: 2017 IEEE International Conference on Pervasive Comput- pp. 1154–1160
ing and Communications Workshops (PerCom Workshops), 75. Zahoor, S., Javaid, S., Javaid, N., Ashraf, M., Ishmanov, F.,
2017, pp. 587–592 Afzal, M.K.: Cloud–fog-based smart grid model for efficient
57. Alam, M.G.R., Tun, Y.K., Hong, C.S.: Multi-agent and rein- resource management. Sustainability 10, 2079 (2018)
forcement learning based code offloading in mobile fog. In: 76. Cao, B., Zhang, L., Li, Y., Feng, D., Cao, W.: Intelligent
International Conference on Information Networking (ICOIN), offloading in multi-access edge computing: a state-of-the-art
2016, pp. 285–290 review and framework. IEEE Commun. Mag. 57, 56–62 (2019)
58. Lee, K., Shin, I.: User mobility-aware decision making for 77. Shakarami, A., Ghobaei-Arani, M., Shahidinejad, A.: A survey
mobile computation offloading. In: IEEE 1st International on the computation offloading approaches in mobile edge
Conference on Cyber-Physical Systems, Networks, and Appli- computing: a machine learning-based perspective. Comput.
cations (CPSNA), 2013, pp. 116–119 Netw. 182, 107496 (2020)
59. Wang, S., Urgaonkar, R., Zafer, M., He, T., Chan, K., Leung, 78. Hussein, M.K., Mousa, M.H.: Efficient task offloading for IoT-
K.K.: Dynamic service migration in mobile edge computing based applications in fog computing using ant colony opti-
based on Markov decision process. IEEE/ACM Trans. Network. mization. IEEE Access 8, 37191–37201 (2020)
27, 1272–1288 (2019) 79. Peng, Z., Lin, J., Cui, D., Li, Q., He, J.: A multi-objective trade-
60. Bittencourt, L.F., Diaz-Montes, J., Buyya, R., Rana, O.F., off framework for cloud resource scheduling based on the Deep
Parashar, M.: Mobility-aware application scheduling in fog Q-network algorithm. Clust. Comput. 23, 2753–2767 (2020)
computing. IEEE Cloud Comput. 4, 26–35 (2017) 80. Hussein, M.K., Mousa, M.H., Alqarni, M.A.: A placement
61. Chiang, M., Balasubramanian, B., Bonomi, F.: Fog for 5G and architecture for a container as a service (CaaS) in a cloud
IoT, vol. 288: Wiley, New York (2017) environment. J. Cloud Comput. 8, 7 (2019)
62. Shahzamal, M., Pervez, M., Zaman, M., Hossain, M.: Mobility 81. Boveiri, H.R., Khayami, R., Elhoseny, M., Gunasekaran, M.: An
models for delay tolerant network: a survey. Int. J. Wirel. Mob. efficient Swarm-Intelligence approach for task scheduling in
Netw. 6, 121–134 (2014) cloud-based internet of things applications. J. Ambient. Intell.
63. Deng, S., Huang, L., Taheri, J., Zomaya, A.Y.: Computation Humaniz. Comput. 10, 3469–3479 (2019)
offloading for service workflow in mobile cloud computing. 82. Fan, J., Wei, X., Wang, T., Lan, T., Subramaniam, S.: Deadline-
IEEE Trans. Parallel Distrib. Syst. 26, 3317–3329 (2015) aware task scheduling in a tiered IoT infrastructure. In: GLO-
64. Ou, S., Wu, Y., Yang, K., Zhou, B.: Performance analysis of BECOM 2017–2017 IEEE Global Communications Conference,
fault-tolerant offloading systems for pervasive services in 2017, pp. 1–7
mobile wireless environments. In: 2008 IEEE International 83. Xu, J., Li, X., Liu, X., Zhang, C., Fan, L., Gong, L., et al.:
Conference on Communications, 2008, pp. 1856–1860 Mobility-aware workflow offloading and scheduling strategy for
mobile edge computing. In: International Conference on

123
Cluster Computing

Algorithms and Architectures for Parallel Processing, 2019, 101. Gupta, A.K., Sadawarti, H., Verma, A.K.: Performance analysis
pp. 184–199 of MANET routing protocols in different mobility models. Int.
84. Ouyang, T., Zhou, Z., Chen, X.: Follow me at the edge: J. Inf. Technol. Comput. Sci. 5, 73–82 (2013)
mobility-aware dynamic service placement for mobile edge 102. Misra, S., Bera, S.: Soft-VAN: Mobility-aware task offloading in
computing. IEEE J. Sel. Areas Commun. 36, 2333–2345 (2018) software-defined vehicular network. IEEE Trans. Veh. Technol.
85. Keshavarznejad, M., Rezvani, M.H., Adabi, S.: Delay-aware 69, 2071–2078 (2019)
optimization of energy consumption for task offloading in fog 103. Hu, R.Q.: Mobility-aware edge caching and computing in
environments using metaheuristic algorithms. Clust. Comput. vehicle networks: a deep reinforcement learning. IEEE Trans.
(2021). https://doi.org/10.1007/s10586-020-03230-y Veh. Technol. 67, 10190–10203 (2018)
86. Zhan, W., Luo, C., Min, G., Wang, C., Zhu, Q., Duan, H.: Mobility- 104. Feng, J., Liu, Z., Wu, C., Ji, Y.: AVE: autonomous vehicular
aware multi-user offloading optimization for mobile edge com- edge computing framework with ACO-based scheduling. IEEE
puting. IEEE Trans. Veh. Technol. 69, 3341–3356 (2020) Trans. Veh. Technol. 66, 10660–10675 (2017)
87. Wu, C.-L., Chiu, T.-C., Wang, C.-Y., Pang, A.-C.: Mobility- 105. Huang, C.-M., Chen, Y.-F., Xu, S., Zhou, H.: The vehicular
aware deep reinforcement learning with glimpse mobility pre- social network (VSN)-based sharing of downloaded geo data
diction in edge computing. In: ICC 2020–2020 IEEE Interna- using the credit-based clustering scheme. IEEE Access 6,
tional Conference on Communications (ICC), 2020, pp. 1–7 58254–58271 (2018)
88. Zhao, X., Shi, Y., Chen, S.: MAESP: mobility aware edge ser- 106. Hou, X., Li, Y., Chen, M., Wu, D., Jin, D., Chen, S.: Vehicular
vice placement in mobile edge networks. Comput. Netw. (2020). fog computing: a viewpoint of vehicles as the infrastructures.
https://doi.org/10.1016/j.comnet.2020.107435 IEEE Trans. Veh. Technol. 65, 3860–3873 (2016)
89. Wu, C., Peng, Q., Xia, Y., Lee, J.: Mobility-aware tasks 107. Yang, C., Liu, Y., Chen, X., Zhong, W., Xie, S.: Efficient
offloading in mobile edge computing environment. In: 2019 mobility-aware task offloading for vehicular edge computing
Seventh International Symposium on Computing and Network- networks. IEEE Access 7, 26652–26664 (2019)
ing (CANDAR), 2019, pp. 204–210 108. Zheng, K., Meng, H., Chatzimisios, P., Lei, L., Shen, X.: An
90. Thananjeyan, S., Chan, C.A., Wong, E., Nirmalathas, A.: SMDP-based resource allocation in vehicular cloud computing
Mobility-aware energy optimization in hosts selection for systems. IEEE Trans. Industr. Electron. 62, 7920–7928 (2015)
computation offloading in multi-access edge computing. IEEE 109. Deng, S., Huang, L., Taheri, J., Zomaya, P., Systems, D.:
Open J. Commun. Soc. 1, 1056–1065 (2020) Computation offloading for service workflow in mobile. Cloud
91. Wang, D., Liu, Z., Wang, X., Lan, Y.: Mobility-aware task Comput. 26, 3317–3329 (2014)
offloading and migration schemes in fog computing networks. 110. Liu, J., Mao, Y., Zhang, J., Letaief, K.B.: Delay-optimal com-
IEEE Access 7, 43356–43368 (2019) putation task scheduling for mobile-edge computing systems. In:
92. Shi, Y., Chen, S., Xu, X.: MAGA: a mobility-aware computa- IEEE International Symposium on Information Theory (ISIT),
tion offloading decision for distributed mobile cloud computing. 2016, pp. 1451–1455
IEEE Internet Things J. 5, 164–174 (2017) 111. Zhang, K., Mao, Y., Leng, S., Zhao, Q., Li, L., Peng, X., et al.:
93. Li, J., Bu, K., Liu, X., Xiao, B.: Enda: embracing network Energy-efficient offloading for mobile edge computing in 5G
inconsistency for dynamic application offloading in mobile heterogeneous networks. IEEE Access 4, 5896–5907 (2016)
cloud computing. In: Proceedings of the Second ACM SIG- 112. Tran, T.X., Pompili, D.: Joint task offloading and resource
COMM Workshop on Mobile Cloud Computing, 2013, allocation for multi-server mobile-edge computing networks.
pp. 39–44 IEEE Trans. Veh. Technol. 68, 856–868 (2018)
94. Ghosh, S., Mukherjee, A., Ghosh, S.K., Buyya, R.: Mobi-IoST: 113. Yu, F., Chen, H., Xu, J.: DMPO: Dynamic mobility-aware
mobility-aware cloud-fog-edge-iot collaborative framework for partial offloading in mobile edge computing. Futur. Gener.
time-critical applications. IEEE Trans. Netw. Sci. Eng. 7, Comput. Syst. 89, 722–735 (2018)
2271–2285 (2019) 114. Sun, Y., Zhou, S., Xu, J.: EMM: Energy-aware mobility man-
95. Shekhar, S., Chhokra, A., Sun, H., Gokhale, A., Dubey, A., agement for mobile edge computing in ultra dense networks.
Koutsoukos, X., et al.: URMILA: dynamically trading-off fog IEEE J. Sel. Areas Commun. 35, 2637–2646 (2017)
and edge resources for performance and mobility-aware IoT 115. Dwivedi, S., Vardhan, M., Tripathi, S.: Building an efficient
services. J. Syst. Architect. 107, 101710 (2020) intrusion detection system using grasshopper optimization
96. Sousa, B., Zhao, Z., Karimzadeh, M., Palma, D., Fonseca, V., algorithm for anomaly detection. Clust. Comput. 1–20, 2021.
Simoes, P., et al.: Enabling a mobility prediction-aware follow- 116. Mishra, S.K., Manjula, R.: A meta-heuristic based multi
me cloud model. In: 2016 IEEE 41st Conference on Local objective optimization for load distribution in cloud data center
Computer Networks (LCN), 2016, pp. 486–494 under varying workloads. Clust. Comput. 23, 3079–3093 (2020)
97. Karimzadeh, M., Zhao, Z., Hendriks, L., Schmidt, R.D.O., la 117. Elashri, S., Azim, A.: Energy-efficient offloading of real-time
Fleur, S., van den Berg, H., et al.: Mobility and bandwidth tasks using cloud computing. Clust. Comput. 23, 1–16 (2020)
prediction as a service in virtualized LTE systems. In: 2015 118. Duong, T.M., Kwon, S.: Vertical handover analysis for ran-
IEEE 4th International Conference on Cloud Networking domly deployed small cells in heterogeneous networks. IEEE
(CloudNet), 2015, pp. 132–138 Trans. Wireless Commun. 19, 2282–2292 (2020)
98. Ma, Y., Liang, W., Guo, S.: Mobility-aware delay-sensitive 119. Puliafito, C., Gonçalves, D.M., Lopes, M.M., Martins, L.L.,
service provisioning for mobile edge computing. In: IEEE Madeira, E., Mingozzi, E., et al.: MobFogSim: simulation of
INFOCOM 2019-IEEE Conference on Computer Communica- mobility and migration for fog computing. Simul. Model. Pract.
tions Workshops (INFOCOM WKSHPS), 2019, pp. 270–276 Theory 101, 102062 (2020)
99. Bahreini, T., Grosu, D.: Efficient placement of multi-component 120. Ruan, L., Bai, Y., Li, S., He, S., Xiao, L.: Workload time series
applications in edge computing systems. In: Proceedings of the prediction in storage systems: a deep learning based approach.
Second ACM/IEEE Symposium on Edge Computing, 2017, Clust. Comput. (2021). https://doi.org/10.1007/s10586-020-
pp. 1–11 03214-y
100. Hoang, V.H., Ho, T.M., Le, L.B.: Mobility-aware computation
offloading in MEC-based vehicular wireless networks. IEEE
Commun. Lett. 24, 466–469 (2019)

123
Cluster Computing

Publisher’s Note Springer Nature remains neutral with regard to Zulfiqar Ahmad is Ph.D. Scholar
jurisdictional claims in published maps and institutional affiliations. in Department of Information
Technology Hazara University
Mansehra, Pakistan. He has
Sardar Khaliq uz Zaman is a received his M.Sc. (Computer
Ph.D. Scholar at Department of Science) degree from Hazara
Information Technology Hazara University Mansehra, Pakistan
University Mansehra, Pakistan in 2011 and M.S. (C.S.) degree
since 2018. He has done MS from COMSATS Institute of
(Computer Science) in 2014 Information Technology,
from Department of Computer Abbottabad, KPK, Pakistan in
Science, COMSATS University 2016. His research interest is
Islamabad (CUI) Abbottabad Cloud Computing, High Perfor-
Campus, Pakistan He joined the mance Computing, and Scien-
same Institute as Lecturer in tific Workflows Execution.
2015. His areas of interest/pub-
lications include: Mobile Edge Arif Iqbal Umar did his Ph.D.
Networks, Large-scale Comput- from China in 2010, M.S. in
ing systems, Social Informatics, 2006, M.Sc. in 1991 and B.Sc.
Cloud Computing and Data Centers Systems. With 6 years of research in 1987. He is Professor of
and teaching experience, he has published different articles in inter- Computer Science in the
national journals and conferences. He is an active member of Scalable Department of Information
Processing and Analytics Research in Communications (SPARC) Lab Technology, Hazara University
at CUI, Abbottabad Campus. Pakistan. He has teaching and
administration experience of 24
Ali Imran Jehangiri is a Lecturer years. He has published 55
in the Department of Informa- research papers in different
tion Technology, Hazara conferences and journals of
University, Mansehra, Pakistan. international repute. His
He received Ph.D. degree in research interest includes Data
Computer Science from the Mining, Machine Learning,
Georg-August-University Goet- Databases, Network Security and Secure Communications in IoT.
tingen, Germany in 2015. He
gained industrial experience Junaid Shuja completed his
with Service Computing work- B.S. in Computer and Informa-
ing as research assistant at tion Science from PIEAS,
GWDG. He is involved in Islamabad, in 2009 and M.S. in
research activities dealing with Computer Science from CIIT
parallel, Grid computing, Cloud Abbottabad, in 2012. He com-
computing and Big data. He is pleted his Ph.D. from University
author of several publications in international journals, and of Malaya, Malaysia in 2017.
conferences. His Ph.D. thesis focused on
execution of SIMD instructions
Tahir Maqsood Tahir Maqsood on heterogeneous mobile and
received the M.Sc. degree in cloud platforms. He is an
computer networks from Assistant Professor at CUI,
Northumbria University, U.K., Abbottabad Campus, Pakistan.
in 2007, and the Ph.D. degree in His research interests include
computer science from COM- application of machine learning techniques in edge computing, energy
SATS University Islamabad, efficient cloud data centers, and mobile cloud computing. He has
Pakistan, in 2017. He is cur- published research in more than 30 International journals and con-
rently an Assistant Professor ferences. He serves as Associate Editor at IEEE Access.
with COMSATS University
Islamabad, Abbottabad Campus,
Pakistan. His research interests
include resource allocation,
multi/manycore systems, reli-
able systems, Internet of Things,
and edge computing systems.

123
Cluster Computing

Eisa Alanazi received his B.Sc. CSAIL Laboratory, MIT, from 2016 to 2017. He subsequently joined
degree in Information Systems the College of Computer and Information Systems, Umm Al-Qura
from King Saud University, in University, as an Assistant Professor of computer engineering. His
2007, and the MSc and PhD mobility impact on the IEEE 802.11p article is among the most cited
degree from the University of Ad Hoc Networks journal articles list. He published articles in pres-
Regina, Canada in 2011 and tigious conferences and journals, such as the IEEE TRANSACTIONS
2017 respectively. He is cur- ON VEHICULAR TECHNOLOGY and the IEEE COMMUNICA-
rently an Assistant Professor at TIONS SURVEYS AND TUTORIALS. His current research interests
the Department of Computer include mobile computing, ubiquitous sensing, and intelligent trans-
Science, College of Computers portation systems.
and Information Systems in
Umm Al-Qura University in
Saudi Arabia. His research
interests include preference
learning and reasoning.

Waleed Alasmary received the


B.Sc. degree (Hons.) in Com-
puter Engineering from Umm
Al-Qura University, Saudi Ara-
bia, in 2005, the M.A.Sc. degree
in electrical and computer
engineering from the University
of Waterloo, Canada, in 2010,
and the Ph.D. degree in Elec-
trical and Computer Engineer-
ing from the University of
Toronto, Toronto, Canada, in
2015. During his Ph.D. degree,
he was a Visiting Research
Scholar with Network Research
Laboratory, UCLA, in 2014. He was a Fulbright Visiting Scholar with

123

View publication stats

You might also like