Computer Networks 195 (2021) 108177

Contents lists available at ScienceDirect

Computer Networks
journal homepage: www.elsevier.com/locate/comnet

Survey paper

Task offloading in Edge and Cloud Computing: A survey on mathematical,

artificial intelligence and control theory solutions
Firdose Saeik a , Marios Avgeris b , Dimitrios Spatharakis b , Nina Santi c , Dimitrios Dechouniotis b ,
John Violos a , Aris Leivadeas a ,∗, Nikolaos Athanasopoulos d , Nathalie Mitton c ,
Symeon Papavassiliou b
a
Department of Software and Information Technology Engineering, École de Technologie Supérieure, Montréal, Canada
b
Department of Electrical and Computer Engineering, National Technical University of Athens, Greece
c
INRIA, France
d
School of Electronics, Electrical Engineering and Computer Science, Queen’s University, Belfast, UK

ARTICLE INFO ABSTRACT

Keywords: Next generation communication networks are expected to accommodate a high number of new and resource-
Edge Computing voracious applications that can be offered to a large range of end users. Even though end devices are becoming
Task offloading more powerful, the available local resources cannot cope with the requirements of these applications. This
Resource allocation
has created a new challenge called task offloading, where computation intensive tasks need to be offloaded
Control theory
to more resource powerful remote devices. Naturally, the Cloud Computing is a well-tested infrastructure
Mathematical optimization
Artificial intelligence
that can facilitate the task offloading. However, Cloud Computing as a centralized and distant infrastructure
creates significant communication delays that cannot satisfy the requirements of the emerging delay-sensitive
applications. To this end, the concept of Edge Computing has been proposed, where the Cloud Computing
capabilities are repositioned closer to the end devices at the edge of the network. This paper provides a detailed
survey of how the Edge and/or Cloud can be combined together to facilitate the task offloading problem.
Particular emphasis is given on the mathematical, artificial intelligence and control theory optimization
approaches that can be used to satisfy the various objectives, constraints and dynamic conditions of this
end-to-end application execution approach. The survey concludes with identifying open challenges and future
directions of the problem at hand.

1. Introduction 5G is an exemplary wireless communication system that tries to

minimize the gap between the new emergent applications and their
Wireless communications have come a long way over the last high-performance requirements. Specifically, 5G promises the support
40 years allowing a plethora of new applications and services to of increased bandwidth and connection density, as well as low-latency
proliferate. This wireless growth has revolutionized the way humans communication, with the induction of the enhanced mobile broadband
and machines interact with each other and between them. Specifically, (eMMB), the massive machine-type communication (mMTC) and the
as wireless technologies evolve, the data rate, mobility, coverage and ultra-reliable low latency communication (uRLLC) services [2]. How-
spectral efficiency rapidly increase [1], permitting radical changes ever, even though the performance of the wireless access networks
on the grounds of our society and our personal communication. At continues to increase, allowing the support of new and more intelli-
the same time, with the advent of the Internet of Things (IoT) and
gent applications, the end devices cannot always cope with the strict
emergent applications such as Virtual Reality (VR) and driverless cars,
computational requirements of these resource voracious applications.
the demand for wireless communications with even higher-speeds
Inevitably, the answer to where we can find an increased avail-
and ubiquitous connectivity becomes a necessity that requires more
ability of computational resources, accompanied with the necessary
efficient wireless communication systems.

∗ Corresponding author.
E-mail addresses: firdose.saeik.1@ens.etsmtl.ca (F. Saeik), mavgeris@netmode.ntua.gr (M. Avgeris), dspatharakis@netmode.ntua.gr (D. Spatharakis),
nina.santi@inria.fr (N. Santi), ddechou@netmode.ntua.gr (D. Dechouniotis), violos@mail.ntua.gr (J. Violos), aris.leivadeas@etsmtl.ca (A. Leivadeas),
n.athanasopoulos@qub.ac.uk (N. Athanasopoulos), nathalie.mitton@inria.fr (N. Mitton), papavass@mail.ntua.gr (S. Papavassiliou).

https://doi.org/10.1016/j.comnet.2021.108177
Received 19 October 2020; Received in revised form 24 January 2021; Accepted 14 May 2021
Available online 18 May 2021
1389-1286/© 2021 Elsevier B.V. All rights reserved.
F. Saeik et al. Computer Networks 195 (2021) 108177

reliability to offer a seamless communication for the wireless applica- been recently proposed, emphasizing on the mathematical models,
tions, always lies around the Cloud. The Cloud is a well tested and optimization techniques, machine learning algorithms and control the-
used solution that can extend the resource capabilities of the end ory approaches. Section 5 presents the open challenges. Finally, we
devices with powerful data center topologies. Besides, Cloud is well conclude and provide suggestions for future work in Section 6.
equipped with the appropriate automation tools and platforms in order
to offer the necessary transparency to the end devices, while hiding the 2. Computing paradigms: Overview & use cases
complexity and the logistic details of this resource extension.
Hence, the practice of offloading computation intensive tasks of As already discussed, over the last two decades Cloud Computing
resource-intensive applications from the end devices to a centralized has been the dominant service delivery paradigm. However, modern
Cloud infrastructure, is a well explored solution [3][4]. Nonetheless, applications come with strict requirements which cannot be met via
as the focus of new applications turned towards high throughput and execution in remote Cloud resources (e.g., ultra low delay). Thus,
low latency communications, the Cloud started to expose its inherent the current trend of resource provisioning is to augment the edge
limitations. The long distance between the end devices and the Cloud of the network with computing capabilities. Towards this direction,
infrastructure, the use of a best-effort and unreliable intermittent trans- the emerging service model of Edge Computing promises to mitigate
port network, the cost of traversing the backhaul network and the the limitations of Cloud Computing. To clarify the ambiguity behind
increased security surface throughout this long communication path, the terminology and architectures used in the literature, this section
created the need for alternative solutions. provides the fundamental elements of the various modern computing
There is no question, that these substitute solutions should introduce infrastructures such as Cloud Computing, Mobile Cloud Computing,
a more distributed infrastructure that will enhance the local efficiency Mobile Edge Computing and Fog Computing. Furthermore, emerging
by bringing Cloud-like capabilities closer to the end devices, at the use cases concerning task offloading at the Edge and/or Cloud are
edge of the network. This is exactly how the term Edge Computing presented.
was coined. Even though, multiple flavors of the Edge Computing exist
2.1. Modern computing paradigms
(e.g., Fog Computing, Mobile Cloud Computing, Cloudlet, Mobile Edge
Computing), they all agree that additional and existing computational
2.1.1. Cloud computing
and networking resources at the edge of the network should be inserted
Cloud Computing has revolutionized the Internet and completely
and regrouped.
transformed the way that applications, software and resources are
This new infrastructural component that creates an additional re-
offered to the end users. According to NIST [20], Cloud Computing
source layer between the end devices and the Cloud, is able to re-
is defined as ‘‘a model for enabling convenient, on-demand network
duce the increased bandwidth consumption at the backhaul, transport
access to a shared pool of configurable computing resources (e.g., net-
and Cloud networks and also reduce the communication delay and
works, servers, storage, applications and services) that can be rapidly
support applications with real time requirements. In particular, end
provisioned and released with minimal management effort or service
devices are now capable of offloading their resource-intensive tasks
provider interaction". The Cloud paradigm brings unique benefits. In
to a nearby Edge device, thus minimizing the overall execution time
particular, with this model, computing resources are offered to the end
without adding excessive communication paths towards a distant Cloud
users on demand, in a self-service fashion, independent of the type of
infrastructure. This solution, called task offloading, allows enhancing
the device and the location of the user. Furthermore, the computational
the user’s experience by providing lower latency, better reliability and
and network resources available at the Cloud can be shared and dynam-
improved energy efficiency for battery-powered devices.
ically scaled. This is achieved by adopting virtualization as the enabling
Even though the notion of Edge Computing exists for almost a technology of the Cloud, allowing the resources to be allocated and
decade, the problem of task offloading has only recently started to released with minimal interaction, while users pay for the service they
be investigated. Nonetheless, it has gained a lot of attention from the consume according to its usage [21].
industry and the academic community, leading to the publication of Despite bringing numerous advantages, Cloud Computing poses
many scientific and research papers over the last couple of years. A some serious limitations. These limitations, although they exist since
great effort has also been made to classify and categorize the different the beginning of the Cloud, they did not surface until recently. The
types of task offloading by a number of surveys and tutorials. reason is that new communication technologies, new applications and
These surveys focused on multiple aspects such as architecture [5– services have increased the data volume generated and at the same
7], resource allocation [8,9], communication [8,10,11], caching [10], time also increased the demands for low latency communications.
mobility management [6,10,12], integration with wireless, IoT and 5G Hence, offering Cloud Computing resources in a centralized manner
technologies [5,8,13,14], decision on task offloading [6,11], applica- far away from the users, can create serious delay bottlenecks. This
tion partitioning [12,15], application models [8,12,16,17], application delay can be disastrous for mission critical applications, such as health
scenarios [5,8,10,15,18,19] and algorithms [11,12,17]. related applications, or time-critical applications like real-time control
In this survey paper, we also attempt to study the task offloading in manufacturing. Another disadvantage is that forwarding the traffic
problem, emphasizing, however, on novel algorithmic and control ap- from the end devices to the Cloud, usually involves traversing the
proaches. Thus, in contrast with recent surveys on task offloading, our core Internet. This can create three serious problems. Firstly, sending
contribution is twofold; firstly, we provide a comprehensive survey hundreds of TB of data from the devices to the Cloud can certainly
of task offloading within three subfields: (i) Optimization algorithms create traffic hotspots in the Internet infrastructure, which can further
(ii) Artificial Intelligence techniques and (iii) Control theory; secondly, affect the communication delay. Secondly, the existence of various
a categorization of the above techniques is provided based on their different networks and administrative domains between the Cloud and
objective function, the granularity level, the use of the Edge and/or the front-end devices, can create an unstable and intermittent network
Cloud infrastructures and the incorporation of mobility in the overall connectivity. Finally, the data, before being sent to the Cloud, proba-
solution, depending on the type of the edge devices. bly have to traverse a backhaul network (e.g., cellular and satellite).
This paper is organized as follows. Section 2 presents an overview This backhaul network may be costly and lossy, creating additional
of the various computing paradigms and relevant technologies evolved problems to this end-to-end communication. Of course, there may
in the last decade, along with some potential use cases for task of- be additional limitations, regarding for example the security aspects
floading of interactive applications. Following, Section 2.2.6 formally of the communication, since this end device-to-Cloud communication
defines the task offloading problem along with the challenges in- increases the surface of threat. However, in the particular survey we
volved. Section 4 covers different task offloading solutions that have emphasize only on the networking and data processing limitations.

2
F. Saeik et al. Computer Networks 195 (2021) 108177

2.1.2. Mobile cloud computing 2.1.4. Edge computing

Apparently, Cloud Computing can be used for offloading tasks from Fog Computing has managed to bypass many of the limitations of
mobile devices to a more powerful infrastructure. This approach cre- Cloud Computing, increasing the performance of IoT and mobile ap-
ated the notion of Mobile Cloud Computing (MCC). A Mobile Cloud plications in terms of task offloading. However, stringent requirements
is defined as a mobile device that can execute a resource-intensive such as ultra-low latency, user experience, stability and high reliability
application on a distant high-performance compute server or compute have created the need for even higher localized information near end
cluster and support thin client user interactions with the application users. Thus, Edge Computing is another similar concept, that can be
over the Internet [22–25]. MCC can be thus described as the inte- defined as a network layer encompassing end devices and their users, in
gration of mobile devices with Cloud Computing technology. It offers order to provide local computing capabilities on sensors, smart meters
computing, storage, services and applications over the Internet and the or other network-accessible devices. Following the same mentality
typical advantages found in a Cloud Computing environment such as where Cloud can be found in a distant location far away from the end
cost reduction and resource flexibility. In addition, MCC can potentially user and Fog can be found closer to the end user, Edge Computing
save energy for mobile users by offloading high-energy consuming has also been associated with the term Mist Computing. As the name
applications to the Cloud [21]. suggests, Mist Computing covers the computational and communication
However, such an approach still carries the typical Cloud limitations capacity available on the same level with the end devices. According
presented above. Thus, the concept of MCC can be modified to offer to NIST [37], Mist Computing is defined as a lightweight and primitive
the necessary Cloud resources closer to the mobile devices. This new type of Fog Computing which resides at the very edge of the network,
flavor of MCC is called Cloudlet [26] and it allows the mobile devices bringing the layer of Fog Computing closer to the smart end devices.
to offload their workload to a local ‘‘mini cloud", consisting of multiple Mist Computing uses microcomputers and microcontrollers to feed into
multi-core hardware equipment directly connected to an Access Point nodes of Fog Computing and theoretically into centralized (Cloud)
(AP) or Base Station (BS). Therefore, Cloudlet can be seen as a trusted, Computing.
resource-rich computer or computer cluster, which is connected to the In light of this, Mobile Edge Computing (MEC) was developed as
Internet and is available for use from mobile devices in proximity. a key technology to assist wireless networks with Cloud Computing-
Due to the sheer proximity of Cloudlet, sharp interactive response for like capabilities to provide low-latency and context-aware services
immersive applications that magnify human cognition is much easier directly from the network Edge [38–48]. Mobile Edge Computing,
to attain. Instead of depending on a remote server, a mobile user lately renamed as Multi-Access Edge Computing, was initiated under
instantiates a ‘‘Cloudlet" on the local network and uses it via a wireless the umbrella of the European Telecommunication Standards Institute
LAN. These proposed Cloudlets can be placed in common areas such as (ETSI) [47]. A key objective of the ETSI initiative is to standardize the
railway stations, restaurants and coffee shops, so that mobile devices APIs between the mobile Edge platform and the application service sce-
could connect to them and act as a thin client. This opposes to the narios (augmented reality (AR), mixed reality (MR), intelligent video
use of a centralized Cloud server that would raise issues of latency and acceleration and Internet of Things gateway) and promote innovation
bandwidth. in an open environment [38]. ETSI’s reference architecture is largely
based on the concept of Network Function Virtualization (NFV), where
2.1.3. Fog computing MEC applications can be offered as Virtualized Network Functions
Fog Computing is another approach for expanding the Cloud Com- (VNFs).
puting concept to the edge of the network, thus enabling a new range
of apps and services [27,28]. Fog Computing was the first industry 2.1.5. Computing paradigms comparison
initiative to explicitly define an architecture for applying utility Cloud Almost all of the paradigms discussed above have as a common
at the edge of the network, and was standardized by the OpenFog ground that they are offering remote computational and communi-
consortium [29]. Specifically, the term Fog Computing was coined cation capabilities to the end devices. Furthermore, except from the
by Cisco in 2012 and is defined as ‘‘the process of extending Cloud Cloud, the rest of the paradigms are able to offer these capabilities
Computing capabilities at the edge of the network. Fog incorporates at the edge of the network, as close to the end devices as possible.
computing, storage and network resources close to the IoT layer to Nonetheless, there are some differences between them.
facilitate the data processing’’. [27,30]. From the previous definition First of all, in terms of available resources, as we move farther from
it becomes evident that Fog Computing was introduced in order to the end devices the available resources increase in quantity, with the
facilitate the monitoring, control and analysis of IoT devices in real Cloud having practically infinite capacity. Since a public Cloud may
time, removing the long communication delay between the IoT devices have dozens of data centers, each equipped with hundreds of servers,
and the central analytics application servers in a remote Cloud. around the world, there is no actual problem of resource depletion. In
Hence, Fog Computing expands Cloud Computing by installing lo- contrast, in Fog, MCC, and MEC infrastructures, resource availability is
calized computing facilities at the user’s premises, delivering Cloud mostly limited due to the fact that they are comprised of micro-data
data to mobile users with fast local connections. The aim is to process centers with few servers of lower capabilities than the ones that we
in part workload and services locally on Fog devices (such as hardened usually find in the Cloud. On top of that, in edge infrastructures, we
routers, switches, IP video cameras), rather than being transmitted also find heterogeneous hardware resources with even lower resource
to the Cloud [28]. As such, Fog Computing introduces an interme- availability such as wireless routers and gateways, street cabinets and
diate infrastructure layer between mobile users and Cloud, in order Raspberry Pi’s.
to support low-latency and high-speed services. Moreover, Fog Com- Secondly, delay can be another factor of comparison between the
puting can support and promote applications that do not suit the different infrastructures. As mentioned before, Cloud Computing is not
Cloud [31], such as (i) applications involving very low and consistent always a feasible solution for providing low latency communication. To
latency, (ii) geographically distributed systems such as pipeline control this end, the available infrastructure at the edge of the network is the
and sensor networks (iii) mobile applications like smart connected most favorable option to reduce the communication delay. Nonetheless,
vehicles and (iv) large-scale adaptive control systems, such as smart since there are various levels of Edge at the WAN, LAN, or access net-
energy delivery and smart traffic lights. As such, in the literature, work, different levels of delays can be produced, regarding where the
the typical applications usually combined with the Fog Computing computing resources are located. Obviously, going at the level of the
are mostly IoT related [32–34], cache networks [35] and immersive access network, i.e., the extreme Edge or Mist, the delay is minimized
media services (AR/VR, a 360-degree video and free-viewpoint video) since we do not have to account propagation and transmission delays
applications [36]. involved in traversing the LAN, WAN or a backhaul network. However,

3
F. Saeik et al. Computer Networks 195 (2021) 108177

in this case, another factor rises; that of the energy consumption. The 2.2.2. Autonomous vehicles
available hardware at the Mist is usually battery supplied, imposing the Similar to the immersive media services, autonomous vehicles is
double burden of both limited resources and limited lifetime. another type of application that task offloading can be utilized. The key
Thirdly, some of the paradigms were conceived under the scope of objectives here are to reduce the latency and the transmission cost and
providing computational resources to specific applications. For exam- increase the efficiency of traffic management. Use case-oriented ser-
ple, Fog Computing was introduced to facilitate some of the top IoT vices concerning autonomous driving include: Highway Pilot, Parking
application domains and vertical markets, such as energy, industry, Pilot, Fully Automated Vehicle and Vehicle on Demand [62].
transportation, agriculture, and healthcare [29]. On the other hand, Edge Computing is considered as the key technology in connected
MCC is mostly associated with providing remote computational re- vehicles, adding computation capabilities and geo-distributed services
sources to mobile applications, while MEC introduces the necessary to BSs and Edge devices distributed on the roadside. The idea is to
flexibility to host multiple applications in the areas of video analyt- analyze data from proximate vehicles and roadside sensors and broad-
ics, location services, IoT, augmented reality, optimized local content cast messages to drivers at a very low latency [63]. For example, in
distribution and data caching among others [49]. Particular emphasis an intelligent transportation system, low-level devices can be used for
should be placed on the uniqueness of the MEC. In particular, even the decision-making processes of the transportation [64]. Specifically,
though MEC is oriented to cellular Radio Access Networks (RAN), it can
the decision-making tasks can be distributed to Edge devices instead of
be practically applied to any kind of access network. Furthermore, the
sending all the data to a centralized server. Moreover, task offloading
way that MEC has been defined and standardized by ETSI, promotes
can enable real-time traffic management [65].
an open environment where third-party developers, application and
service providers can all participate together towards expediting the
2.2.3. Robotics
introduction of new applications targeting to respond to emerging user
Very complex robotic applications have been emerging during the
requirements.
last decade, related among others to autonomous mobile agents, manip-
Finally, the decision of which paradigm to follow, usually includes
the requirements of security and confidentiality. Certainly, Cloud Com- ulators and collaborative tasks. Efficient, safe and autonomous robot
puting as a popular and successful technology, has many safeguards and operation in manufacturing, health care, learning and exploration,
tools to provide a certain level of security and confidentiality. Nonethe- requires running computation and memory intensive algorithms related
less, several security threats still exist making Cloud-based security an to image processing, planning, localization, mapping and autonomous
active open-challenge. Additionally, sending data to the Cloud over the learning. Consequently, during the last few years, task offloading is
Internet can be susceptible to attacks. In contrast, by employing an gaining attention that has lead to the new paradigms of Cloud, Edge
Edge infrastructure, the necessary security and confidentiality can be and Fog robotics [66–70].
attained since the data of the end devices usually stay within the local Specifically, many offloading opportunities emerge in planning and
network. SLAM (simultaneous localization and mapping) algorithms for robotic
From the above, the pros and cons of each computing paradigm manipulators [71,72], mobile robots [73–77] and learning in gen-
can be extracted. Even though the MCC, Fog, and MEC can overcome eral [78,79], among others. It is worth noting that there are already
certain limitations of the Cloud, they usually cannot be offered as a available commercial products that allow task offloading in robotic
standalone solution. In other words, the notion of Edge Computing in applications [80–82].
general, did not emerge to replace the Cloud but rather to complement
it. Thus, it is very important to create collaborative solutions (possibly 2.2.4. Video streaming
utilizing more than two computing paradigms) that will enable a In general, the video streaming use cases fall under the content
smooth continuum from the end device to the Cloud, with the goal to delivery network (CDN) [83] category. The key objective of CDN
satisfy the stringent requirements of novel and future applications. networks is to reduce the cost and the number of bits transmitted over
2.2. Use cases the network, by maintaining an adequate QoE [84]. The mechanisms
to reduce the overall cost and traffic while providing a high QoE in
Following the above definitions of the computing paradigms and applications ranging from simple video streaming to HTTP, to Adap-
the respective infrastructures, in this part of the survey we refer to tive BitRate (ABR) and 360-degree video applications, can be further
some typical applications that leverage task offloading at the available improved by applying task offloading techniques.
resources at the edge of the network in order to increase their per- Offloading can be implemented on Cloud-based solutions, where
formance. These real-world applications can range from simple data appropriate resource allocation techniques can be used to increase
processing to immersive multimedia applications. Following, we briefly user satisfaction [85] or deployment costs of the CDN networks [84].
describe the role of task offloading in the particular set of applications.
Nonetheless, task offloading at the Edge can supplement the achieved
performance. For example, multi-user mobile media delivery can be
2.2.1. Immersive applications
enhanced by enabling the gateways (i.e., BSs) to perform appropriate
Current developments in computer vision have made possible the
scheduling strategies [86]. An Edge infrastructure can also be used to
launching of mixed reality applications, such as VR and AR, that can
facilitate the caching and transcoding mechanisms in a distributed fash-
offer immersive experience even in wireless environments. At the same
time, the development of increasingly advanced mobile devices such ion [87]. Regarding latency, data compression tasks can be offloaded
as smart glasses, can help us identify objects, superimpose contextual at the Edge [88], removing the burden of local compression models
knowledge on our field of vision and create a three-dimensional view of and reducing at the same time the application response time [89].
the surrounding environment. As these devices become smarter, more Task offloading can be partially implemented by differentiating flows
and more sensor data in our environment can be aggregated, pro- based on their quality and performing the video compression at the
cessed and served, requiring however high bandwidth and low response Edge, only for the high-quality video flows (e.g., 360-degree video
time communications. Hence, task offloading can be an advantageous streaming) [90].
solution for this type of applications.
Specifically, both Cloud and Edge-based task offloading mechanisms 2.2.5. IoT
can be used in AR/VR applications [36,50–61]. The objectives include The impact of task offloading can be maximized in the context of IoT
reducing the energy consumption in mobile devices, increasing the applications. The reason is that the IoT devices are usually constrained
speed of computation intensive operations, reducing the average CPU in terms of available resources and battery capacity. Inevitably, only
load to overcome computation intensive tasks and improving the user’s small and non resource demanding tasks can be executed locally. Task
Quality of Experience (QoE). offloading in IoT usually focuses on reducing task execution time,

4
F. Saeik et al. Computer Networks 195 (2021) 108177

response time and energy consumption. IoT use cases that can ben- 3.1. Granularity levels of task offloading
efit from task offloading can refer to health, agriculture, smart city,
industry and energy related applications among others. Task offloading aims at optimizing the offloading of computation
IoT, from the very beginning, has been largely based on Cloud- intensive tasks from the end user device to a remote site, under various
centric approaches in order to offload the tasks of data processing and computational, communication and mobility constraints. The process
of task offloading, as shown in Fig. 2, consists of (i) various hardware
analysis of massive data produced from millions of IoT devices [91].
components, such as end user devices and Edge/Cloud devices, (ii) mul-
However, the long delays added from the Cloud, combined with the
tiple computing processes, including task splitting and computational
introduction of new mission critical IoT applications, has pushed the
processing either locally or remotely and (iii) networking components
academic and industrial community to take advantage of the Edge con- for transferring data between the hardware components involved.
cept. Hence, local IoT Clouds have emerged, with the goal to maximize In more detail, as Fig. 2 illustrates, a mobile device can execute
the number of offloaded tasks that can be executed in close proximity an application comprising of multiple tasks. The end device, through a
to the IoT devices [92] and to maximize the battery lifetime of the de- task splitting process, decides which of these tasks should be executed
vices [93,94]. However, the scalability issues of the IoT market which locally and which ones should be offloaded to the Edge or Cloud
is currently consisted of dozens of billions of devices often necessitates infrastructures. This decision is based on a plethora of factors that are
a Cloud–Edge collaboration during the task offloading [95–97]. presented in the following sections, including the QoS requirements and
battery lifetime of the device, among others. Following, the tasks that
are to be executed remotely are transferred through the wireless access
2.2.6. Physical disaster management network to the gateway and from there to a remote physical machine
In case of disaster management, the process of task offloading (either at the Edge or Cloud), where they are executed following an
is crucial since it affects the efficiency of the rescue operations. In appropriate computational approach (e.g., creating a VM or container).
addition, the network can be unstable and simply offloading tasks to At the same time, the tasks that remain on the device are executed
the Cloud could be difficult and require too much time. So, optimal locally using the available computational resources of the end device.
offloading strategies to local services, rather than remote Clouds, would The last step is to combine the results of both local and remote executed
allow for precious time saving and preservation of battery of mobile tasks to provide the final output of the application. Based on this
phones, sensors and autonomous agents in the field. process, we describe the different types of task offloading according to
the task splitting decision taken, i.e., the granularity level, as follows:
Unmanned Aerial Vehicles (UAVs), which possess great mobility
and versatility, are at the core of disaster management scenario by
3.1.1. Partial offloading at the edge
providing situational awareness and computing resources. But, as they In this type of offloading, part of the computing tasks is processed
are battery-powered, they cannot undertake the full computation of locally, while the rest is offloaded to the Edge. Partial offloading is
all the involved data and need to offload tasks to near Edge Comput- typically the most effective, since it can benefit from both local and
ing servers. This challenge is addressed in different papers [98–100]. remote resources. Nonetheless, another level of complexity is added
In general, task offloading in the context of disaster remains little since it needs to be decided and scheduled which tasks should be
explored [101,102]. offloaded while taking into account the possible energy and resource
constraints of the end device.

3. Task offloading & challenges 3.1.2. Full offloading at the edge

In this case, all of the computing tasks are offloaded and processed
In the previous sections, we have provided a short description of at the Edge. Full offloading is usually translated into a simple resource
the task offloading, its infrastructural components, and the importance allocation problem, where tasks can be executed on virtual machines
of this solution for new and emerging scenarios. In this part of the or containers at the Edge. Energy-savings at the end devices can be
maximized, however we need to take into account other sources of
survey, a more detailed definition of task offloading is provided, while
energy dissipation such as the transmission power of the device. Finally,
also, the typical objectives, the performance evaluation metrics, and the
a precise network path from the device to the Edge site, where the
challenges encountered during task offloading are presented.
tasks are offloaded, has to be set up carefully, so as to comply with
Generally, task offloading can be defined as the transfer of resource- the possible QoE/QoS constraints.
intensive computational tasks to an external, resource-rich platform
such as the ones used in Cloud, Edge or Fog Computing. Offloading 3.1.3. Partial/full offloading at the edge and cloud
the whole or part of the set of tasks to another processor or server, During this type of offloading, a collaboration between the Edge and
can be used to accelerate resource-intensive and latency-sensitive ap- Cloud infrastructures is established in order to execute the offloaded
plications [65,90,103]. Task offloading is a complex process and can tasks. This type of collaboration can be proved advantageous in large-
be affected by a number of different factors [24]. In particular, this scale scenarios where the available Edge resources are not enough
process involves application partitioning, offloading decision making to host all of the tasks offloaded from the end users. Herein, the
and distributed task execution [4,15,104]. A typical infrastructure in- main challenge is the two-level task offloading decision. If a partial
volved in an offloading scenario is illustrated in Fig. 1. From this, offloading mechanism is followed, the first level of decision lies on
the local device, in order to decide the set of tasks that needs to be
it becomes evident that the network’s Edge infrastructure creates an
offloaded at a remote location. In other words, the local device has to
additional resource layer between the end devices and the external
decide which tasks can be executed locally in the device and which
platform. This layer is capable of reducing bandwidth consumption
ones should be offloaded either at the Edge or the Cloud. On the tasks
on the backhaul, transport and Cloud networks, thus reducing any
decided to be offloaded, a second-level of decision will be performed. In
communication delays, supporting applications with real-time require- particular, the Edge will perform a second task partitioning, regardless
ments, improving the energy efficiency and consequently increasing the of the type of offloading (i.e., partial or full) to determine the subset of
lifetime of battery-powered devices. At this point, we should note that, tasks to be executed at the Edge and the subset of tasks to be executed
for the rest of the paper, we refer to the term Edge Computing as the at the Cloud. In the latter case, particular attention should be paid on
whole set of resources that can be found at the edge of the network, the transport network that facilitates the interconnection of the two
including the Fog and Edge nodes. infrastructures and the delay constraints that it may impose.

5
F. Saeik et al. Computer Networks 195 (2021) 108177

Fig. 1. Infrastructure components during task offloading.

Fig. 2. Task offloading process.

3.2. Mobility of end devices mobility adds another level of splitting decision, as it needs to be
decided at which Edge site should the tasks be offloaded while the
End device mobility is one of the most critical components when user is on the move (or not). Even though mobility is considered
it comes to task offloading decision. End devices can either be con- a challenge, it can generate a number of opportunities for the task
sidered as static or mobile for the time window which spans between offloading. First of all, it can initiate a load balancing technique to
initiating and finishing the offloading of their tasks. In the latter case, allow the system to provide the necessary services in distributed Edge

6
F. Saeik et al. Computer Networks 195 (2021) 108177

site scenarios [45,105,106]. Secondly, complementing mobility with immediately and return to the mobile user. On the other hand, when
appropriate prediction solutions can enhance the system’s capacity, by the task requires significantly longer time, then the task could be split
finding the potential next associated BS/AP of the user [87,107,108]. into sub-tasks and be transferred to neighboring BSs along the user’s
This can be even more beneficial in a dense scenario, where the system trajectory [6,8,131].
can analyze the active users and their mobility patterns and allocate Most of the mobility tackling solutions integrate mechanisms to
the resources in an online manner to existing and newly requested obtain the device’s current and future positions, as well as tune the of-
services. Moreover, mobility can benefit from handover mechanisms floading infrastructure accordingly, to achieve the objectives described
that can enable service migrations between BSs and their Edge servers. in the following subsections. Using the type of the mobility tackling
However, as the requirements of zero millisecond handover are studied mechanism as a criterion, mobility solutions can be categorized in the
by the 5G community, mobility with prediction mechanisms is starting following classes:
to gain attention, in order to predict beforehand where the tasks should Proactive — Behavior Related: Nowadays, services running on
be offloaded. This behavior can be decisive for the overall performance mobile devices, e.g., Google Location Services [132], constantly track
in dense Edge deployments with multiple mobile users [109]. Based on and log the historic mobility behavior of the user; the proliferation
the above, we firstly describe the different types of end devices accord- of smartphone devices has made trajectory pattern crowdsourcing a
ing to their mobility level. Then, we present some mechanisms from reality. What is more, distributing intelligence at the Edge has allowed
the literature on how the Edge infrastructures adapt to the potential for logging the times an end user device connects and disconnects to a
mobility of the offloading devices. smart access point, thus extracting users’ periodic movement patterns.
Based on this data, mobility information can be estimated and lever-
3.2.1. Static (low-mobility) aged towards predicting the users’ position at any given moment [133,
In this situation, the devices are considered static or relatively 134]. Specifically, this mobility information can be extracted in a
static (low levels of mobility) during the task offloading procedure. probabilistic way, by utilizing Mobility Markov Chains (MMC) to model
Apparently, this type of device mobility is considered trivial when it the historic behavior of a user as a discrete stochastic process; in this
comes to studying the task offloading problem, as it does not impose way, the prediction of a user moving to a specific location depends on
any type of dynamicity to the network conditions. In certain cases, their previous visited locations and the probability distribution of the
purely non-mobile devices like stationary IoT sensors are engaged in transitions between them [135]. Complementary to the regularity in
task offloading at the Edge, [110,111]. Other works apply this assump- the users’ mobility patterns, a Markov model can be trained to estimate
tion on mobile devices, to reduce the complexity of their proposed the expected network quality and the expected staying time under the
offloading solutions; for example, the authors in [112] assume that coverage of each Edge server [136]. Then, one way to leverage the
the statistics of the utilized wireless links remain unchanged during extracted information is in favor of bringing the Edge Computing and
the processing of the users’ tasks, reflecting a relatively static or low- storage resources closer to the user; this can be achieved by proactively
mobility scenario. In a similar manner, in [113] and [114], the time to installing the services that the users will consume in the Edge servers
transfer the task response from a cell to another is excluded from the located in the positions that they will most probably visit, thus reducing
total execution time of an offloaded mobile task, by assuming that the the network delay during task offloading [137–139]. The probability
end device stays in the same cell during the task offloading process. density function of the sojourn time, i.e., the time a user is expected to
spend within the coverage area of an Edge site, can also be exploited
3.2.2. Mobile (high-mobility) towards predicting the user’s next location and seamlessly migrating
On the other hand, when the devices involved in task offloading are the service to be used for the task offloading appropriately [140]. For
highly mobile, i.e., their movement has a direct effect on the network example, MAGA [141] is a mobility-aware mechanism for partial task
conditions considered for task offloading and the respective resource offloading that falls in this category. Frequent mobility patterns are
allocation, the problem becomes significantly more complex. Several inferred by a tailor matching subsequence method and then a genetic
studies consider mobility at the Edge [65,115–118]. Specifically, the algorithm is used for the offloading decision.
focus is placed on the user contacts (inter-contacts) in which the user Proactive — Trajectory Related: Another way of proactively deal-
can offload the task, based on the mobility pattern and an opportunistic ing with mobility at the Edge is by exploiting the user’s ongoing
computing decision [119–123]. The opportunistic computing is taking movement characteristics, i.e., trajectory, duration and speed. By ap-
advantage of the contact patterns regulated by the mobility of the plying these characteristics on specific translational motion models, one
devices (e.g., which Edge site the node is visiting and what type of can predict the location and the time of the next Edge server han-
interactions occur on daily basis), in order to determine the amount of dover [142]. Apart from utilizing motion models, periodically receiving
computation to be offloaded to other devices. Opportunistic computing timestamped geolocation updates from a moving user, enables produc-
takes also advantage of the contact patterns regulated by the mobility of ing real-time travel information for route segments which can be used
these devices, to determine the amount of computation to be offloaded for trajectory estimation [143]. In a cooperative Edge infrastructure
to other devices. Furthermore, in a mobility scenario, users may either scenario, taking advantage of the mobility information can guide Edge
transfer their tasks to remote servers or peer devices, possibly through servers to route the collected offloaded tasks to adjacent servers at
the gateways or even via the Edge servers [124–127]. For instance, mo- the next location on the user’s moving direction. In this way, when
bility can influence the decision on which BS and Edge server to select users arrive at the coverage area of the next Edge site, they receive
and when to perform the handover [124]. When trying to minimize the product of their completed offloaded task, with the minimum addi-
the execution delay, user mobility information needs to be combined tional delay [144]. As an example, a two-step offloading mechanism for
with the task characteristics and resource availability, in order to make smart touristic services [145,146] is based on estimating the location
the best task scheduling decision [125]. This mobility information is and density of users. Every mobile device takes the initial offloading
usually captured by trajectory prediction models [128,129] that can decision based on a dead reckoning technique and measurements of its
actually uncover motion patterns of the users in real-time scenarios. WiFi signal strength. Secondly, at the Edge side, a Kalman filter is used
However, in most of the studied scenarios [6,8,105,130], the tasks to predict the number of users and a controller is responsible for the
are usually offloaded to BSs that are in close proximity to the user’s final offloading decision and the allocation of resources to VMs.
position and have sufficient resources to satisfy the time execution Reactive: The evolution of the Edge–Cloud continuum and the
requirements. For example, when the tasks are lightweight and can growing adoption it receives, has recently enabled network infrastruc-
be executed within a satisfactory time period, without migrating to a tures to quickly and efficiently adapt to the rapidly changing user
neighbor Edge site, then the execution of the task should be processed environment, in real time. When it comes to task offloading on the

7
F. Saeik et al. Computer Networks 195 (2021) 108177

move, Edge servers can utilize a central agent, located at the Cloud, to modulation and coding scheme, together with other radio parameters,
form a mobility-aware offloading infrastructure that tracks the users’ needs to be taken into account, as they contribute in significant energy
position and optimally routes the task and its response through the and consequently battery consumption [41,165]. This type of energy
closest server to the users’ locations [147]. In a similar manner, the consumption can increase, especially when network conditions are not
whole virtual server can migrate to the topologically closest Edge server favorable. Secondly, by reviewing the full offloading from a complete,
to the user, reactively, every time a relocation is detected [148]. For network-wide view, one can easily understand that the problem is
instance, utilizing IP tracking, remote caching and the Software-Defined simply pushed to the Edge and/or Cloud infrastructures. Thus, energy
Networking (SDN) paradigm, can set the ground for efficient and timely consumption minimization needs to be pursued at all layers of this
task offloading as well; an SDN controller is able to track the user’s end-to-end communication model [170,171].
network location, i.e., the Edge server in proximity and quickly react To evaluate this objective, a number of different metrics can be
to changes in it by rerouting the offloaded task’s response [149]. used; the most common is the average power consumption measured
by aggregating the power consumption on the hardware equipment
3.3. Task offloading objectives used [159,172]. Alternatively, energy consumption can be used, ex-
pressed as the power consumption over time. Normally a minimiza-
When solving the task offloading problem, a number of differ- tion of power consumption leads to energy consumption minimization
ent objectives may be applied, as the different stakeholders and ac- as well [107,173–175]. Furthermore, as the end devices are usually
tors (e.g., Cloud providers, Edge providers, Mobile Users and Service battery powered, the energy savings can be expressed as battery sav-
Providers) target a variety of goals. An objective function helps to ings [170,176]. Finally, another way to provide the necessary en-
formally and mathematically formulate these goals and guide the of- vironmental and economic sustainability is through minimizing the
floading solution. Objectives of the task offloading problem can be electricity cost [177]. Electricity cost depends on location and time.
categorized as follows: Hence, appropriate allocation of offloaded tasks potentially reduces
the electricity cost, cutting down on operational costs while providing
3.3.1. Delay benefits to the environment [157,167].
Task execution delay minimization is one of the main objectives
during the task offloading problem [42,65,115,150–153]. Regardless
3.3.3. Bandwidth/spectrum
of the type of task offloading, the overall goal is narrowed down to
The available bandwidth at the access network and how it can
reducing the total task execution delay. This delay, can be broken down
be shared by multiple users in order to offload the tasks, is also
into a number of different delay contributors. The first source of delay
a significant constraint. However, due to the great influence it has
is the task execution delay, coming from the task that can be either
on the task offloading performance, it can also be considered as an
executed locally at the device, at the Edge or the Cloud. In the case
objective [166,176]. Due to the limited available bandwidth, especially
of offloading the task at a remote Edge or Cloud site, we need to
in IoT networks and dense cellular networks, the careful allocation of
take into consideration the transmission and propagation delay at the
spectrum becomes of utmost importance.
various layers of the infrastructure (access, Edge, transport and Cloud
The objective of spectrum allocation is often associated with the
networks), in both directions (i.e., sending the task and receiving the
transmission rate and power level of each end user [176], as well as
response). On top of that, processing and queuing delay at the vari-
the duration of the transmission of each device [151], in order to op-
ous processing and forwarding devices should be taken into account.
timally share the available bandwidth. Thus, when trying to optimally
Finally, an additional delay contributor can be the time to optimally
deploy the available spectrum, an efficient metric is to evaluate the
partition the task delay, during the task offloading decision [154]. The
spectrum utilization in accordance with the number of offloaded tasks,
delay objective can be expressed as either the minimization of the aver-
power transmission and bandwidth consumption [64,152,173,174]. In
age delay of each task [155] or the total delay of all the involved tasks
view of the dynamic wireless conditions, the optimal scheduling of
of a mobile application. This objective is directly proportional to the
the bandwidth needs to follow the time-varying channel state and be
available resources [115,150] and the network conditions [151,154].
also associated with the arrival rate of the tasks [166]. Throughput
The total execution delay can also be used to assess the impact of
task offloading to the QoS achieved. Thus, according to the type of the is another typical metric applied to evaluate the overall spectrum
application used and the part of the infrastructure under consideration, utilization, since it reveals how timely and efficiently the task can be
the task execution delay can be associated with: (i) the response time offloaded at the remote infrastructures [170,178].
(i.e., the time duration from when a user requests a service until the
service actually initiates [65,118,156]), (ii) the delay variation, in order 3.3.4. Load balancing
to reveal how robust the task offloading solution can be, both over time How to carefully allocate and schedule the available physical hard-
and over dynamic traffic profiles, while also estimating the number of ware resources is another objective to consider during task offload-
SLA (Service Level Agreement) violations noticed [107,157,158] and ing [12,17,42]. Specifically, there is a high interest in path optimiza-
(iii) the network delay, including the four delay contributors (i.e., prop- tion, efficient resource usage and load balancing when solving the task
agation, transmission, processing and queuing delay) at the different offloading problem. The goal is to provide the necessary scalability by
parts of the infrastructure [65,107,118,152,155,156,158–160]. increasing the resource availability, increasing the number of offloaded
tasks concurrently deployed at the Edge and/or Cloud [17,179], max-
3.3.2. Energy imizing the resource sharing and fairness among multiple users [42]
The second most common objective during task offloading is the and facilitating the offloading of future tasks.
minimization of the energy consumption [161–165]. This energy con- This objective can be translated into minimizing the overall resource
sumption typically refers to the end devices [116,166–169]. The reason usage (e.g., minimizing the average percentage of the computational
is that mobile and IoT devices are usually battery powered, thus a major and communication resource utilization), minimizing the maximum
concern is how to maximize the lifetime of the battery by reducing resource utilization of the infrastructure or minimizing the variance
the device’s energy consumption. Inevitably, it is reasonable to assume of resource utilization [95]. Load balancing can be applied either at
that, by following a full offloading approach, the maximum energy sav- each layer of the infrastructure or between the different layers. For
ings can be pursued. However, a number of other energy contributors example, the appropriate distribution of traffic between an Edge and a
need to be taken into account, even when a full offloading approach Cloud infrastructure, can be considered as an alternative load balancing
is followed. First of all, during the offloading, the transmission power, objective [171,180].

8
F. Saeik et al. Computer Networks 195 (2021) 108177

3.3.5. Deployment cost 173,189–196]. Energy and latency optimization can also be combined
The task offloading problem can be often seen as a resource allo- with an optimal spectrum allocation through appropriate power and
cation problem where appropriate resources at the Edge and/or Cloud channel allocation [162,167,168,174]. Regarding load balancing, it can
need to be reserved according to a deployment utility cost, in order be jointly optimized with the delay, since both objectives are tightly
to execute the offloaded tasks in a virtualized environment. This de- correlated [42].
ployment cost can be modeled in various ways, each having a different
interpretation. For example, the deployment cost can be defined as the 3.4. Challenges of task offloading
aggregation of computational and communication resources that a set
of tasks needs in order to be provisioned. This is typical when mobile Achieving the aforementioned objectives of task offloading entails
or IoT applications needs to be complemented with specific network a series of challenges. In this section, we identify these challenges and
functions (e.g., security, compression, QoS) that is otherwise difficult classify them into two main categories.
for a user to achieve locally in his device [180–182].
In general, this cost can be expressed as the monetary cost for 3.4.1. Network dynamics challenges
computing processing (e.g., $/CPU hour), memory (e.g., $/GB) and net- Dynamic Network Conditions: The mobile and IoT networks are
work bandwidth (e.g., $/GB/day), induced for using network resources. characterized as quickly varying access networks that create dynamic
These costs can be associated with the Cloud and Edge providers network and traffic conditions. This is a significant challenge that adds
and how and to whom they lease their infrastructure. Furthermore, an extra level of complexity during the task offloading problem, since
since this cost usually follows a ‘‘pay-as-you-go" model, the number it is very difficult to pre-specify the behavior of the network. Aspects
of physical resources used for the total number of tasks offloaded can of noise, interference, fading and signal reflection can significantly
reflect the deployment cost [152,157]. impact the wireless communications, aggressively altering the overall
throughput and delay of the wireless transmission. This necessitates an
3.3.6. Model accuracy analysis and prediction of the network conditions, in order to accu-
All the objectives described so far, are some typical objectives that rately estimate when a task offloading decision positively affects the
can be used regardless of the optimization solution/strategy followed. performance. Besides this, prediction can be combined with a resource
However, when it comes to Artificial Intelligence (AI) and Machine allocation mechanism at the Edge and Cloud, since the amount of
Learning (ML) -based approaches, an additional objective can be the resources required for the task execution is directly proportional to the
model accuracy for the prediction of the behavior of the applications amount of traffic (i.e., the request rate) that will actually end up at the
(e.g., mobility [183] and network related features [184]). Usually the Edge or Cloud.
AI/ML objectives are used as secondary or complementing objectives of Dynamic User Behavior: Another level of impediment and dynam-
the ones presented above [185]. In addition to that, regarding critical ics, during task offloading, is added by the random behavior of the
applications that build or use ML models during the task offloading, mobile users and how they employ their mobile applications. These
particular emphasis should be placed on the training time, inference behavioral aspects are very difficult to foresee and quantify, creating
time and the cost of the required computational power. as result arbitrary user-based traffic profiles. A categorization of the
Model accuracy can be optimized by using the appropriate ML mobile applications based on the users’ preferences, the transmission
metrics. For example, regression metrics [186] can express how close patterns, the spatial and temporal correlation of the user generated
the predictions of a model are compared to the actual values, by using traffic, as well as other traffic related characteristics, can be of utmost
the R-squared, Root Mean Squared Error (RMSE) and Mean Absolute importance for the subsequent resource scheduling and allocation.
Error (MAE) metrics. In clustering models, the evaluation metrics [187] Accordingly, machine learning and data analytic techniques should be
measure the cohesion and separation of groups of observations. An applied to estimate the users’ behavior and the rate of task generation.
example of such an approach is the Sum of Squared Error (SSE) metric, Edge/Cloud Dynamics: Although the Edge and Cloud layers can
which aggregates the distance of each observation from its nearest work together in harmony, they still have their own dynamics. Cloud
cluster. When using classification ML techniques [188], typical metrics sites are centralized while Edge sites are distributed having only a
used to evaluate the accuracy are the precision (i.e., the percentage local view of the network.This leads to different dynamics between
of relevant observations among the retrieved observations), Recall these two layers. Specifically, contrary to Cloud, Edge has a spatial
(i.e., the percentage of the total amount of relevant observations that dynamic exactly because of its location awareness. Additionally, end
were actually retrieved) and F-Measure (i.e., the harmonic mean of devices can dynamically re-purpose and re-associate themselves to dif-
precision and recall). ferent applications by offloading different type of tasks or simply new
devices could appear or disappear. This inevitably creates an additional
3.3.7. Multi-objective dynamic factor for the Edge. Obviously, the initial task offloading and
The typical objectives presented so far are usually conflicting, mak- allocation decision over the Edge could be performed in an optimal way
ing task offloading a very challenging problem. For example, aiming to by the Cloud, since it possesses this centralized system view; however,
minimize the latency can lead to higher energy consumption on the end the latter may not react timely to local dynamics. Therefore, Edge
device, by deciding to execute the tasks locally and vice versa. When servers should meet the burden to locally decide to move services’ tasks
adopting load balancing objectives, offloaded tasks can be distributed along time. Thus, a new challenge arises in order to (i) address these
among different Edge sites or between the Edge and the Cloud, in dynamics, (ii) create a consistent view of the tasks to be executed that
order to reduce the total delay and energy consumption. Similarly, Cloud and Edge should share, and (iii) place the different services at
when minimizing the deployment cost, offloaded tasks may be gathered the right places in due time.
in one single Edge site; this results in an uneven resource utilization
that can also create significant congestion in the infrastructure and 3.4.2. Resource allocation challenges
thus higher communication delays. Finally, when trying to optimize Task offloading is strongly affected by the resource allocation mech-
the spectrum allocation independently of the available communication anisms that decide how and where the offloaded tasks will be executed
and computational resources, it can result in poor offloading decisions in a remote platform. Thus, the task offloading and resource alloca-
in terms of delay, deployment cost and load balancing. Hence, multi- tion decisions are coupled and should be addressed jointly. The main
objective solutions can be used in order to explore the trade-off between challenges this issue creates follow.
the various objectives. The most common multi-objective approaches Partitioning Decision: The decision of which task to offload is the
consider jointly minimizing the delay and energy consumption [117, first and most significant challenge to address, as it comprises the core

9
F. Saeik et al. Computer Networks 195 (2021) 108177

Table 1 by executing the tasks and effectuating actuation in minimal time. How-
Comparison of Mathematical Optimization (MO), Artificial Intelligence (AI), and Control
ever, at the same time, a meticulous design of the task management
Theory (CT) approaches.
control modules is required at the Edge. The placement of controllers
MO AI CT
and their mapping to the sites that they will serve, the decision of which
Stability ✓
task is going to be offloaded at each site and how the load is going to
Low complexity ✓
Optimality ✓ ✓ be distributed between the sites, are some of the challenges that fall in
Online training ✓ this category.
Reachability ✓ ✓
Real-time decision ✓ ✓ ✓
4. A taxonomy of task offloading approaches

This section lists the most prominent task offloading solutions that
of the task offloading problem. As shown in Fig. 2, when new tasks have been proposed in the literature. These algorithmic solutions can
are generated by an application, an intelligent mechanism is required be divided into three categories: (i) Mathematical Optimization algo-
in order to decide whether the task should be executed locally or to rithms; (ii) Artificial Intelligence techniques and (iii) Control Theory-
be offloaded to a remote infrastructure. This partition decision of the based approaches.
tasks is associated with the task execution delay, the transmission delay
and the energy consumption. A poor partitioning decision may result 4.1. Methodology comparison
in performance bottlenecks regarding the execution of the application.
Thus, a compromise between when and which tasks should be offloaded Before providing the most common approaches for each category,
to the Cloud/Edge has to be explored, taking also into consideration it is worth investigating the advantages and disadvantages of each
any possible transmission costs in terms of energy, delay and money. category and the level of their efficiency. Accordingly, we present the
Resource Availability: The performance of an application is closely main characteristics of each category, while Table 1 summarizes the
dependent on the resources available at the end user/Cloud continuum. main features supported.
In general, as we move towards the core of the infrastructure, the Mathematical optimization is the most common solution category
available resources increase in amount, paying the price, however, of applied in resource allocation and scheduling networking problems.
a higher application delay. Hence, sharing these resources is a crucial The reason lies in the fact that, traditionally, these types of problems
challenge which needs an efficient resource allocation and management can be mathematically formulated and solved, using a great vari-
mechanism that will be able to guarantee the performance require-
ety of existing solutions. Usually, the main goal for this category of
ments. Thus, alongside the offloading decision, the resource allocation
algorithms is to find the optimal solution among a set of possible solu-
mechanism should fulfill various functional and non-functional require-
tions. For example, in the context of task offloading the mathematical
ments. The primary goal of resource scheduling is the respect of the QoS
optimization approaches will have to appropriately model the input
requirements of the application. Additionally, the resource allocation
(e.g., Edge/Cloud infrastructure, end users, available resources, task
should guarantee important properties, such as stability, reachability,
distribution size and duration) and, according to a certain objective,
safety and robustness against internal uncertainties and external dis-
decide when and where the tasks should be optimally offloaded.
turbances. In terms of functional requirements, the resource allocation
The optimality can be achieved through exhaustive search optimiza-
strategy should be implemented with commercial or open-source re-
tion solvers in the expense of a high complexity and execution time.
source orchestration tools, that enable scalability, interoperability and
Nonetheless, mathematical optimization approaches can reduce their
the transparent development of the applications over heterogeneous
time complexity when relaxing any hard constraints of the input and
hardware and software technologies.
altering their final goal into finding a sub-optimal but fast and real-
Performance Modeling: Measuring the performance of a task of-
time solution. Even though these types of solutions can fit and be
floading solution is an additional challenge. The task offloading prob-
used in real scenarios, they sometimes suffer from their static nature
lem can be modeled as a system where the energy and/or delay are
the typical output variables and the available computing resources and inability to adapt and model the dynamic challenges inherited
(e.g., CPU, memory), incoming requests and network bandwidth are from the problem at hand. Under such circumstances, the algorithms
the input variables. In most of the current studies, the proposed per- should be re-executed and re-customized every time a change happens
formance models are single-input/single-output, empirically derived or in time and/or space (e.g., dynamical arrivals of end users, mobility
fixed. Although this assumption is realistic, the processing time of an and equipment failures).
offloaded task depends on many time-varying parameters, which are AI task offloading mechanisms have also seen great progress nowa-
usually not easily measured. On the other hand, multi-input/multi- days. Data driven models, learning from batch or online data, provide
output models are more accurate, but the identification process is usu- real-time task offloading decisions and elastic resource allocation. De-
ally strenuous. Specifically, the offloading decision performs adequately cisions are made by generalizing historical data, recognizing in an
only for specific operating conditions, being unable to guarantee sta- automatic way the prevailing data patterns and evaluating the possible
bility under fluctuating workloads and heterogeneous communication destination states of the main actors in the Cloud/Edge environment.
infrastructures, such as in IoT. Hence, this system model should be The state of Edge infrastructures includes the status of computing nodes
adapted in order to include the performance metrics, expressed as state in terms of resource utilization, the number of application requests and
variables, that can be regulated by the control parameters (i.e., the the user requirements contracted as SLAs.
input variables). This framework will be capable of capturing struc- Contemporary AI and ML task offloading is characterized by flexible
tural changes interpreted as discrete jumps in the dynamics, e.g., user adaptation and automatic learning. ML models have the advantages
mobility, changes in network conditions and addition/removal of Edge that they are not explicitly designed by human experts and they are
servers. self-trained based on the available data. In addition, they can handle
Task Management: As stated before, offloading tasks to the Cloud multi-dimensional and multi-variety data in a unified way and they are
follows a centralized approach in which the Cloud infrastructure serves capable of identifying hidden patterns. The main weakness of AI-based
the whole set of tasks coming from the access network layer. In con- models lies in the case of significant inconsistency between training and
trast, at the Edge layer, the infrastructure is usually distributed in testing data properties, which may lead to performance degradation.
multiple geographically dispersed Edge sites. Obviously, this is one of This means that the data should be selected and gathered with diligent
the core advantages of Edge Computing, as it creates a local efficiency attention to detail and special emphasis should be given to the data

10
F. Saeik et al. Computer Networks 195 (2021) 108177

preparation tasks of synchronization, transformation, and normaliza- Similarly, these types of problems allow the formation of multi-object-
tion. Lastly, we should take into consideration that the large amount ive functions. For instance, a system cost representing the weighted
and high frequency of data make the training model a computational cost of delay and energy consumption among all available users, can
heavy and resource-intensive process. be expressed as a non-linear objective function in a mixed Cloud–
System Theory provides various models for describing the opera- Edge task offloading environment [198]. In case only the energy exists
tion of a process. Additionally, Control Theory provides many formal in the objective functions, typically the latency requirements can be
methodologies to analyze and control the performance of the pro-
imposed as constraints along with other various conditions (e.g., power
cess. Both of them have been widely used for industrial processes
consumption levels, channel states and resource heterogeneity) [199].
while, during the last decades, they have been introduced to com-
puter networks. System theory provides black- or gray-box training
algorithms, namely system identification, to compute multiple-input-
4.2.2. Heuristics
multiple-output (MIMO) models that capture the system dynamics of
Heuristic approaches can introduce fast but sub-optimal solutions.
continuous or discrete systems. However, system identification must
be performed offline and the computed model may have low accuracy. The main advantage of heuristics is that they are simple algorithms
The control-based task offloading solutions enable real-time decisions devised to address the problem at hand, with low execution time. In
against the dynamic network and workload conditions. Additionally, contrast with MIP algorithms, they do not require specialized optimiza-
apart from reaching an optimal operating point, the control-based tion tools to be solved and can rather be expressed as pseudo-code, eas-
offloading solutions can guarantee various system properties, such as ily implementable in any programming language. To this end, heuristic
stability and reachability. The guarantee of stability means that the solutions are very popular to be applied in the task offloading problem.
system will reach specific operating conditions and will remain on them These types of solutions can range from optimizing the offloading
against any disturbance. The reachability property means that, given decision of the user while minimizing the overall cost of energy, compu-
the current system state, we can compute all possible destination states. tation and delay, by applying appropriate relaxation and randomization
Although, the complexity of a controller increases with the complexity techniques [194], to optimizing the resource allocation at the Edge, by
of the system model (linear or not) and the properties to be guaranteed, considering a Cross-Entropy optimization approach [162]. Heuristics
the design is an offline process and the real-time application of control
can also be used to optimize non quantifiable parameters, such as QoE
law is simple.
in a Cloud–Edge collaboration [197], by satisfying the various compu-
tational and bandwidth restrictions, applying appropriate fairness and
4.2. Mathematical optimization algorithms
popularity techniques [85].
The task offloading problem is usually formulated as a mathematical The efficiency of the heuristic becomes much more evident when
problem, which tries to find an optimal or near-optimal solution. the task offloading problem is modeled as a non-linear constrained
The problem can be formulated by defining the objective function as optimization problem, or when the experimentation covers large-scale
described in Section 3.3 and the optimization strategy used. These offloading scenarios [162]. In this case, greedy heuristics can be used
strategies may include Mixed Integer Programming, heuristics, meta- to estimate the exact solution [42,85,200].
heuristics and game theory approaches, among others. Following, we
present the main optimization strategies found in the literature, while
a summary of them along with the objective of the study, algorithm 4.2.3. Game theory
developed and type of offloading is listed in Table 2. Furthermore, Lately, there has been also a use of game theoretic approaches to
Fig. 3 illustrates the key components of the existing mathematical deal with resource allocation problems. Through game theory, the task
optimization approaches, shedding light into the well and less explored offloading problem can be introduced as a resource allocation game.
proposed solutions.
For example, the problem of the partial task offloading in a multi-user,
Edge Computing infrastructure and a multi-channel wireless interfer-
4.2.1. Mixed integer programming
ence environment, can be formulated as an offloading game [201].
Mixed Integer Programming (MIP) formulations provide a flexible
and mathematically precise way of formulating many real-world prob- This game tries to maximize the spectrum efficiency during offloading
lems. Specifically, integer programming is a commonly used technique by allocating the proper channel to each user/player. The specific
for resource allocation and scheduling in wired and wireless networks. approach can be complemented with a second matching game that will
The two main problem types that MIP addresses are: (i) network syn- aim to maximize the efficiency of resource allocation at the Edge, by
thesis and (ii) resource assignment problems [208]. MIP optimization appropriately selecting the right Edge servers [196]. A multi-step/slot
approaches facilitate also the introduction of a multi-objective function game theoretic approach can be followed in order to find the optimal
optimizing more than one goals under various offloading constraints state, expressed as the Nash Equilibrium. Specifically, in each step
(e.g., delay, energy and load balancing). Hence it can be often used as the end user/player can make a decision on whether to offload their
an optimization strategy during the task offloading problem. Usually, tasks in order to reach a potentially optimal offloading. A similar
MIP provides a linear objective function (MILP), where at least one of slotted approach can be followed by treating each user as a player
the variables takes integer/binary values. Even though these types of with the goal to optimize the CPU-cycle frequency and offloading
problems can provide the optimal solution, they can be very complex
decision, in order to maximize the energy efficiency. Game theory
or even computationally intractable for large scale experimentation
has also been used in an Edge–Cloud interplay, where players can be
scenarios. However, they can often be used as benchmark approaches
considered as the corresponding infrastructures [202]. In particular, by
during the performance evaluation. For example, in the context of
formulating a Stackelberg game, a leader player is assigned the goal to
task offloading, they can be used to minimize the weighted amount
of mobile energy consumption in a multi-user system, under latency maximize a utility function expressed in order to obtain the optimal
constraints [192]. Regarding delay, the objective function of the MIP revenue for the Edge and Cloud providers, while satisfying the delay
can include both the transmission and processing delay, especially for requirements. A different objective can be considered in an Edge–Cloud
IoT mission-critical applications in an Edge–Cloud collaboration [197]. collaboration, where the two infrastructures comprise the players of
In case the objective is non-linear, a Mixed Integer non-linear the problem and try to minimize the overall energy consumption under
(MINLP) or quadratic (MIQP) programming formulation is modeled. delay constraints [203], by using a potential game [209].

11
F. Saeik et al. Computer Networks 195 (2021) 108177

Table 2
Taxonomy of mathematical optimization task offloading algorithms.
Reference Optimization objective Algorithms Mobility Offloading
Delay Energy Bandwidth/ Load Deployment MIP Heuristic Game Contract Local Static Mobile Partial Full Edge–
spectrum balancing cost theory theory search Cloud
[41] ✓ ✓ ✓ ✓ ✓
[42] ✓ ✓ ✓ ✓ ✓
[65] ✓ ✓ ✓ ✓ ✓
[85] ✓ ✓ ✓ ✓ ✓ ✓
[89] ✓ ✓ ✓ ✓ ✓
[116] ✓ ✓ ✓ ✓
[117] ✓ ✓ ✓ ✓ ✓ ✓ ✓
[121] ✓ ✓ ✓ ✓ ✓ ✓
[150] ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
[151] ✓ ✓ ✓ ✓ ✓
[152] ✓ ✓ ✓ ✓ ✓ ✓ ✓
[154] ✓ ✓ ✓ ✓ ✓ ✓
[155] ✓ ✓ ✓ ✓ ✓
[161] ✓ ✓ ✓ ✓ ✓
[162] ✓ ✓ ✓ ✓ ✓ ✓ ✓
[163] ✓ ✓ ✓ ✓ ✓
[164] ✓ ✓ ✓ ✓
[165] ✓ ✓ ✓ ✓ ✓ ✓ ✓
[166] ✓ ✓ ✓ ✓ ✓
[167] ✓ ✓ ✓ ✓ ✓
[168] ✓ ✓ ✓ ✓ ✓ ✓
[169] ✓ ✓ ✓ ✓ ✓ ✓ ✓
[170] ✓ ✓ ✓ ✓ ✓
[173] ✓ ✓ ✓ ✓ ✓ ✓
[174] ✓ ✓ ✓ ✓ ✓ ✓
[176] ✓ ✓ ✓ ✓ ✓
[189] ✓ ✓ ✓ ✓ ✓
[191] ✓ ✓ ✓ ✓ ✓ ✓
[192] ✓ ✓ ✓ ✓ ✓
[193] ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
[194] ✓ ✓ ✓ ✓ ✓
[195] ✓ ✓ ✓ ✓ ✓
[196] ✓ ✓ ✓ ✓ ✓ ✓
[197] ✓ ✓ ✓ ✓ ✓ ✓
[198] ✓ ✓ ✓ ✓ ✓ ✓
[199] ✓ ✓ ✓ ✓ ✓
[200] ✓ ✓ ✓ ✓ ✓
[201] ✓ ✓ ✓ ✓
[202] ✓ ✓ ✓ ✓
[203] ✓ ✓ ✓ ✓ ✓
[204] ✓ ✓ ✓ ✓ ✓
[205] ✓ ✓ ✓ ✓ ✓
[206] ✓ ✓ ✓ ✓ ✓ ✓
[207] ✓ ✓ ✓ ✓ ✓

4.2.4. Contract theory to macro base station (MBS) offloading scenario, described in [206].
Naturally, task offloading introduces conflicts between the partici- In the case of opportunistically offloading part of the cellular traffic to
pating parties; for example, on the one hand, users and devices seek to coexisting networks, towards alleviating the overload problems caused
maximize energy and spectrum efficiency while on the other, small cells by traffic demands, a contract theory-based incentive mechanism can
and Edge servers try to minimize consumption of their own resources, motivate users to leverage their delay tolerance in exchange for service
like battery capacity and computing power. Conflicts like these might cost [207].
cause reluctance by third parties to participate, which could subse-
quently raise barriers to the development of attractive traffic offloading
4.2.5. Local search
solutions. Contract theory is an approach originating from real world
Local search algorithms adopt mechanisms of perturbation to ex-
economics, that dictates the design of contracts to achieve cooperation
plore neighbor solutions in the search space, that allow to gradually
between the conflicting sides. In a broad sense, contract theory studies
converge to local optimum solutions. Due to the problem-agnostic
the design of formal and informal agreements that motivate agents
nature of the local search algorithms, they can be used as a compo-
with conflicting interests to take mutually beneficial actions. In the
wireless networks domain, the agents include the BS, service provider nent of heuristics and meta-heuristics in order to provide solutions
and the spectrum owner, as well as the small cells, smart devices and very close to optimality [211]. For instance, a one-dimensional local
users [210]. The late boom of contract theory applications in task search algorithm can be used for a partial task offloading solution,
offloading has managed to deal with many early challenges of the with the goal to minimize the average execution delay, expressed as
field; for instance, combining contract theory with game theory and a Markov chain process, under the energy constraints imposed by the
a monetary rewards system, can eliminate the influence of information device [155]. When both energy and latency constitute the objective
asymmetry in a user–Edge server relationship [204]. Similar incentives of the partial task offloading problem, a univariate search technique
can be utilized when dealing with small-cell base stations (SBSs) and can be used [189]. This type of search allows to transform a non-
heterogeneous ultra-dense networks (HetUDNs) [205]. When the goal is convex problem into a convex one, by finding a local optimum solution.
to optimize bandwidth allocation in data offloading, dynamic program- An iterative local search can further reduce the gap with an optimal
ming concepts can integrate with contract formulation, as in the UAV solution, where multiple iterations of the local search can result in

12
F. Saeik et al. Computer Networks 195 (2021) 108177

Fig. 3. Summary of mathematical optimization task offloading algorithms.

a better resource allocation of computing and channel resources in a offloading problem, by learning from data and tasks distributed across
partial offloading scenario [174]. When the goal is to optimize the the Edge infrastructure, AI can enable a smart, real-time, and dynamic
energy sustainability of the end-user devices, by selecting the proper resource management framework [11,214–216]. On another perspec-
Edge resources and at the same time to minimize the execution time of tive, AI techniques can also be applied to avoid costly data offloading,
the allocation, a simple bi-section search algorithm can be used [151]. by enabling data estimation or prediction, like in dual prediction
approaches [217].
4.3. AI-based optimization algorithms Similar to traditional mathematical optimization solutions, when us-
ing AI, a problem can be formulated by defining the objective function
In the above section, we provided traditional mathematical and and the algorithmic strategy to be followed. Following, we present the
algorithmic approaches to derive the optimal or near-optimal solution, major AI techniques used in the literature to address the task offloading
in the context of task offloading. However, these approaches may suffer problem. A summary of the related work with the objective of the
from the following issues: (i) Most of the solutions investigated so far study, the algorithm developed and type of offloading, is listed in
fail to take into consideration the dynamic network conditions. Since Table 3, while Fig. 4 illustrates the distribution of the key components
this is a random variable, it is difficult to estimate and reflect this of the existing AI optimization approaches.
behavior during the allocation of the Edge and Cloud resources and
during the task partitioning decision; (ii) The traditional approaches are 4.3.1. Machine learning
rather opportunistic, addressing the challenges of the task offloading As a sub-category of AI, Machine Learning (ML) gives devices or
in a short-term scale. However, in this manner, we cannot capture computer systems the capability to learn useful patterns and behaviors
the long-term time and space varying conditions in all the layers of from historic data and make decisions about new ones. The models are
this end-to-end communication model. In other words, the solutions built without explicit programming; in the case of ML parameterized
presented above lack the ‘‘intelligence" to better adapt holistically to models, such as Linear Discriminant Analysis, Logistic Regression and
the inherent challenges of the problem at hand. This prepares the Naive Bayes, the models are built by tuning a fixed number of parame-
ground for using artificial intelligence techniques for the task offloading ters of a predefined mapping function. In the case of non-parametric
problem. models, such as the RBF-kernel Support Vector Machines, Decision
Artificial Intelligence (AI) techniques include multi-disciplinary Trees and K-Nearest Neighbor, the models use a flexible number of
techniques from machine learning, consensus-based and constraint- parameters with no prior knowledge about the data distribution and
based algorithms and they have been widely used in different computer mapping function. In both cases, a mapping function is a function
systems and network scenarios [8,185,200,212,213]. AI techniques are that maps the independent data variables to the dependent variables,
becoming successful alternatives also for solving optimization problems i.e., the variables the model predicts.
that include the mathematical formulation of uncertain, stochastic and ML models can be divided in supervised, unsupervised and re-
dynamic information, thus making them excellent candidates for the inforcement learning ones, based on the available training data. A
task offloading problem. Furthermore, AI can potentially reduce the prominent ML subfield is Deep Learning, which involves Artificial
complexity by enabling recursive feedback-based learning and local Neural Networks (ANN) with multiple layers of representation. ML
interactions and thus faster speed in seeking sub-optimal solutions models have been used successfully to overcome the challenges of task
than traditional techniques [8,185,212]. For example, during the task offloading and resource allocation, as described below.

13
F. Saeik et al. Computer Networks 195 (2021) 108177

Table 3
Taxonomy of AI-based task offloading algorithms.
Reference Optimization objective Algorithms Mobility Granularity
Delay Energy Bandwidth/ Load Deployment Model Machine Population Constraint Static Mobile Partial Full Edge–
spectrum balancing cost accuracy learning Cloud
[218] ✓ ✓ ✓ ✓ ✓ ✓
[219] ✓ ✓ ✓ ✓ ✓
[220] ✓ ✓ ✓ ✓ ✓ ✓
[221] ✓ ✓ ✓ ✓ ✓ ✓
[222] ✓ ✓ ✓ ✓ ✓ ✓ ✓
[223] ✓ ✓ ✓ ✓ ✓
[224] ✓ ✓ ✓ ✓
[225] ✓ ✓ ✓ ✓
[226] ✓ ✓ ✓ ✓
[227] ✓ ✓ ✓ ✓ ✓
[228] ✓ ✓ ✓ ✓
[229] ✓ ✓ ✓ ✓
[230] ✓ ✓ ✓ ✓ ✓ ✓
[231] ✓ ✓ ✓ ✓ ✓ ✓
[232] ✓ ✓ ✓ ✓ ✓
[233] ✓ ✓ ✓ ✓
[234] ✓ ✓ ✓ ✓ ✓ ✓
[235] ✓ ✓ ✓ ✓
[236] ✓ ✓ ✓ ✓ ✓ ✓
[237] ✓ ✓ ✓ ✓
[238] ✓ ✓ ✓ ✓
[239] ✓ ✓ ✓ ✓ ✓ ✓
[240] ✓ ✓ ✓ ✓ ✓
[241] ✓ ✓ ✓ ✓ ✓
[242] ✓ ✓ ✓ ✓ ✓
[243] ✓ ✓ ✓ ✓ ✓ ✓
[244] ✓ ✓ ✓ ✓ ✓ ✓
[245] ✓ ✓ ✓ ✓ ✓
[246] ✓ ✓ ✓ ✓ ✓ ✓
[247] ✓ ✓ ✓ ✓
[248] ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
[249] ✓ ✓ ✓ ✓ ✓
[250] ✓ ✓ ✓ ✓
[251] ✓ ✓ ✓ ✓ ✓
[252] ✓ ✓ ✓ ✓
[253] ✓ ✓ ✓ ✓ ✓
[254] ✓ ✓ ✓ ✓ ✓
[255] ✓ ✓ ✓ ✓ ✓
[256] ✓ ✓ ✓ ✓
[257] ✓ ✓ ✓ ✓ ✓
[258] ✓ ✓ ✓ ✓
[259] ✓ ✓ ✓ ✓ ✓ ✓ ✓
[260] ✓ ✓ ✓ ✓ ✓ ✓
[261] ✓ ✓ ✓ ✓ ✓
[262] ✓ ✓ ✓ ✓ ✓ ✓

Supervised ML Models: Supervised ML models include classifi- Regression [226] has also been used to predict over-loaded and under-
cation and regression models. In classification, the model predicts loaded nodes, in order to facilitate a live migration process of tasks.
classes while in regression the model estimates continuous values. The Gaussian Process Regression [227] has been proposed to predict the
offloading decision can be formulated with a multiclass classification future workload of the tasks, allowing the deployment of new, delay
method and the resource allocation with a regression model [218]. sensitive applications and reducing energy consumption, blocking of
Classification and Regression Trees (CART) [219] have been used to requests and latency. The dynamic characteristics of applications and
select the fittest Edge device for offloading, while minimizing time and the complex Edge/Cloud Computing environment, have been modeled
energy by taking into consideration parameters such as the authenti- with the Support Vector Regressor [228] and the K-Nearest Neigh-
cation, confidentiality, integrity, availability, capacity, speed and cost. bor Regressor [229] for future load prediction and energy efficient
utilization of Edge servers respectively.
Classifiers such as JRIP and J48 [220] have been used for context-
Unsupervised ML Models: Clustering models discover groups of
sensitive offloading in a Mobile Cloud Computing environment using
objects that are similar, close and dense or share some common prop-
a robust profiling system. Logistic regression [221,222] has been used
erties. Clustering models are differentiated from Classification models
to calculate the load of each Edge node and enhance a dynamic re-
in that they do not require annotated data for training. Regarding task
source allocation strategy. A different classification approach classifies
offloading, clustering approaches have been used to group resources
Edge applications into classes of services [223], based on their QoS based on the distance between Edge nodes, wireless and computational
requirements, and maps them to Edge and Cloud resources. The Apriori resources [230,231] in order to minimize the response delay. In the
algorithm [224] has also been used to generate rules for every task, in same fashion, Edge sites can be grouped for different task resource
order to select the Edge node that offers the minimum completion time. demands [232] and Edge servers can be grouped using an analysis
Linear Regression [225] has been used to predict the total process- of the allocated computing resources [233]. Unmanned aerial vehicles
ing duration of each task on each candidate Edge node, in order to are also clustered to enable efficient multi-modal and multi-task of-
offload entire tasks to one Edge node instead of a local execution. Linear floading [234] and IoT users according to their priorities [235]. The

14
F. Saeik et al. Computer Networks 195 (2021) 108177

Fig. 4. Summary of AI-based task offloading algorithms.

dependencies between tasks can be represented by a graph and, by decisions, bandwidth allocation can be generated with a distributed
following a fuzzy clustering [236], makespan (i.e., the time difference deep learning-based offloading algorithm for Edge networks [241].
between the start and finish of tasks), monetary and energy costs can be Furthermore, the challenge of energy efficient task offloading with
minimized. The K-means clustering method [237] can provide efficient Deep Learning has been studied in the context of the internet of vehi-
task scheduling, thus increasing the utilization of the Edge devices, cles [242], users’ equipment [243] and specifically for delay-sensitive
based on the type of resource requirements in terms of CPU, I/O and and computation intensive tasks in Edge Computing networks [244].
communication. Lastly, a policy-based clustering approach [238] can Deep Reinforcement Learning: Generally, reinforcement Learning
provide energy efficient task offloading solutions, by organizing the models take actions in an environment in order to maximize a cu-
interactions among the Edge nodes. mulative reward or minimize expected loss. Reinforcement Learning
Deep Learning: Deep Learning has gained popularity in multiple is usually combined with Deep Learning in order to generalize with
decision and scheduling problems because of highly accurate outcomes, previously unseen data in terms of environment, states and actions. In
especially when large amounts of data are available. In Edge and the Edge Computing context, a Deep Reinforcement Learning model has
Cloud Computing, large amounts of data are being collected by re- been implemented to make the binary offloading decision on whether
source monitoring tools, application logging mechanisms and network the offloading will take place partially locally or fully remotely to an
sniffers [263]. A modular deep learning model can integrate different Edge server [245]. Deep Q-Networks [246,247] have been proposed
sources of data, manipulate the data observations with hierarchical to automatically infer the offloading decisions, in order to optimize
layers of representations and extract generalized knowledge that goes the system performance. Furthermore, Deep Q-Networks have been
beyond the historical observations. enhanced to capture the sequence of data with long short-term memory
Deep Learning can provide timely and accurate task offloading layers [248], for mobile tasks in a large-scale heterogeneous Edge
decisions, based on the resource usage of processing Edge nodes, the environment and Gated Recurrent Unit layers [249] in multi-Edge
workload and the QoS constraints defined in SLA [264]. A deep learn- networks.
ing model works as a function approximator that takes as input the
current infrastructure and workload status and outputs the appropriate 4.3.2. Population-based methods
processing nodes where each tasks will be offloaded. Further outputs Population-Based methods include a wide range of nature-inspired
of the deep learning models include the decisions for vertical or hori- algorithms and provide close to optimal solutions in combinatorial
zontal scaling up and VM migration, in order to guarantee the smooth problems, following a metaheuristic approach. Two main subcategories
operation of tasks execution in a dynamic and quick-change computing of Population-Based methods are the Swarm Intelligence and the Evo-
environment. Specifically, Deep Learning models have been imple- lutionary Algorithms. Both of them have been proposed in order to
mented to minimize the computation and task offloading overhead in provide efficient solutions in task offloading challenges.
varying network conditions and limited computation resources [239]. Swarm Intelligence (Consensus-based): The Swarm Intelligence
A Deep Learning model that also addresses the challenges of speed, properties make it a practical design model for algorithms that solve
power and security, while satisfying the quality of services, has been increasingly complex problems. In general, swarm algorithms strive
proposed in order to determine the combination of different devices and to allow the entire system to converge into a global consensus state,
dynamic tasks [240]. In addition, close to the optimal joint offloading while retaining the ability to perform assigned individual tasks in the

15
F. Saeik et al. Computer Networks 195 (2021) 108177

Fig. 5. Summary of control theory-based task offloading algorithms.

swarm. Ant Colony Optimization (ACO) and Particle Swarm Optimiza- is related to the artificial intelligence operations research and aims
tion (PSO) are the most common Swarm Intelligence algorithms. ACO to find feasible solutions by using methodologies such as constraint
can be applied for efficient task scheduling due to its strong global propagation, local search, backtracking and various heuristics. Specifi-
search ability [250] and improve the response time of IoT applications cally, task properties, user mobility and network constraints have been
by distributing effectively the tasks over the Edge nodes [251]. On jointly formulated as a CSP [258], in order to reduce the task execution
the other hand, PSO can be used to minimize both the transmission delay in Mobile Edge Computing infrastructures. A CSP formulation
and the processing delay, as a means to minimize the total end-to-end has also been used in combination with Min-conflicts scheduling algo-
delay during a partial task offloading at the Edge [252]. In this case, rithm [259], for achieving the necessary load balancing of the Edge
PSO can also incorporate other task offloading key mechanisms, such resources and minimizing energy consumption. A more demanding
as VM migration and transmission power management, to minimize offloading use case is the distributed processing of data streams. In
service delay as efficiently as possible, to provide high QoS for different this case, special emphasis is given to minimizing end-to-end latency
application profiles and to remain computationally feasible. Last but through the appropriate placement of the stream operators, either
not least, PSO has been used to jointly minimize energy consumption on Cloud nodes or Edge devices. A CSP optimization framework has
and completion time for high-quality solutions [253,254]. also been proposed in order to minimize this latency and satisfy the
Evolutionary Algorithms: Evolutionary Algorithms are based on constraints of power, bandwidth and CPU utilization [260].
One prominent approach to address a CSP comes from the Con-
natural selection principles, such as reproduction, mutation, recom-
straint Programming (CP). CP [266] is a programming paradigm used
bination and selection. They perform a lot of iterations on a set of
in solving complex problems, where instead of defining a sequence of
candidate solutions, aiming for the closer to optimal solutions to sur-
steps for the program to execute in order to obtain the result, one
vive as much as possible, while the unfit solutions tend to be discarded.
defines the relationships between variables in the form of constraints
Evolutionary Algorithms [255] have been used in the deployment of
that must be met. Afterwards, by following the steps of branching
Edge nodes and the offloading strategies. A subclass of Evolutionary
and exploration, CP finds feasible solutions to the problem. CP has
Algorithms is the Genetic Algorithms (GA), which is characterized by
been proposed for a generic and easy-to-upgrade placement service
the crossover principle in the reproduction of candidate solutions. GAs for Fog Computing with short resolution times and quality solutions.
have been used for sequential task offloading and proactive fault tol- Specifically, using Choco [261], a many times awarded constraint
erance [256]. Hybrid models, which combine GA with PSO, have also solver, we can estimate close to optimal solutions in terms of network
been proposed [257] and they achieve close to optimal task offloading infrastructure, applications graphs and metrics like the usage of storage,
of IoT applications, while minimizing the total makespan and energy network and energy resources. In addition, CP has been combined with
consumption. an event-based finite state model [262], in order to optimize mobile
battery life and guarantee QoS and cost minimization simultaneously.
4.3.3. Constraint satisfaction methods
The task offloading problem has also been re-defined as a Constraint 4.4. Control theory-based algorithms
Satisfaction Problem (CSP) with multiple source of constraints such
as SLA, QoS, QoE, the heterogeneity of devices, the particularities of Originally, Control Theory was designed to regulate the behavior of
VMs and the dynamicity of the task generation process. CSP [265] dynamic systems and keep the system output(s) following the desired

16
F. Saeik et al. Computer Networks 195 (2021) 108177

control signal, also called a reference. Control theory relies on feedback the resource availability. In this case, an online task offloading algo-
mechanisms and is widely applied to computing systems. Because of rithm, which leverages Lyapunov optimization methods and utilizes the
its nature to rely on feedback, which measures the difference between current system information, can be used to predict the user’s resource
the actual output and the desired reference, control theory has been availability [121]. The benefits are two-fold. First, network operators
applied in the field of Edge Computing for implementing mechanisms with global network information are trusted to make the comprehen-
of efficient decision-making control [267–269], network design se- sive offloading decision for all the users. Second, the capabilities of
lection [270,271], time-critical systems [75,76,272–275], admission mobile devices are constantly improving and the multiplexing advan-
control [276], network management [277], switching Edge [278] and tage (due to the flexibility of resource availability between devices) can
network switching [279], among others. A summary of the control be exploited to enable the execution of collaborative tasks for a wide
theory approaches, along with the objective of each study, the algo- range of services.
rithm developed, type of offloading and the consideration of mobility,
is listed in Table 4. Moreover, Fig. 5 illustrates the distribution of the 4.4.4. Rest of the control approaches
key components of the existing control theory-based approaches. This paragraph includes the control-based studies that cannot be
classified in the previous categories. Dlamini et al. [277] developed an
4.4.1. Optimal control online algorithm for Edge network management based on predictive
The control theory foundations, specifically those based on linear control. This control mechanism aims at optimizing a two-objective
optimal control theory (i.e., Linear Quadratic Regulator (LQR) [281]), cost that includes energy consumption and QoS satisfaction. Sonmez
consider the design of a selection strategy [270] in a heterogeneous et al. [158] proposed a fuzzy workload orchestrator for the Edge
wireless network, with the objective to maximize the network resource Computing environment. Here, a set of fuzzy rules assigns the offloaded
utilization, while meeting the constraints of the supported services. requests to a computational unit in a hierarchical Edge Computing
Linear controllers are designed to meet the system’s constraints and architecture. Spatharakis et al. [275] proposed a switching offloading
QoS metrics [276]. The high efficiency of the LQR controllers guar- mechanism for robotic applications (i.e., path planning and localiza-
antees the control performance of the system. Skarin et al. [280] tion) within an Edge Computing setup. The offloading decision is
consider the LQR for MIMO linear systems. Control theory can be based on the uncertainty of the mobile robot’s position, the resource
deemed of utmost importance in UAVs-related task offloading where availability at the Edge servers and the complexity of the path planning.
LQR-based controllers are used in order to achieve robust adaptive In addition, control theory can be applied to address the task
attitude control [282]. allocation problem [268] with a novel integrated Top-Down Bottom-
Up (TDBU) approach. In particular, the top-down module (i) observes
4.4.2. State feedback control the bottom-up task preference decisions of the Edge devices and de-
Another advantage of the control-theoretic approach in task of-
cides the optimal task offloading strategy to ensure the overall system
floading is that it provides a methodology for the modeling, analysis
performance; (ii) leverages top-down incentive schemes to implicitly
and evaluation of the system. Avgeris et al. [276] proposed a control
guide the Edge devices to pick the tasks that are most likely to finish in
theory approach to study the adaptive resource allocation problem
time. Similarly, Wu et al. [269], proposed codeSpec, a decision making
for task offloading. The proposed two-level resource allocation and
control theory approach (on the Edge device) that periodically renews
admission management system for an Edge application cluster, gives
its offloading decisions, at code level, with nearby IoT devices, in real-
mobile users an alternative option for performing their tasks. The
time. CodeSpec, shifts the destined devices from inter-domain servers
proposed controllers allow mobile users to offload application-specific
to IoT devices nearby and only offloads binary code of user-specified re-
tasks within the coverage area. However, it should be noted that
gions across different instruction set architectures (e.g., x86_64, ARMv7
mobility of users within the proximity of the cluster is not taken
and IA-32), using control theory.
into consideration in this work. Kalatzis et al. [274] modeled the
performance of IoT-based applications with a switched system and
5. Open challenges and future directions
computed various equilibrium points that correspond to different op-
erating conditions. Based on these points, a simple scaling mechanism
was built to satisfy the varying workload demand. Extending [274], The concept of task offloading has evolved from the simple idea
SMOKE [100] is a scalable resource allocation mechanism for UAV- of migrating the computation intensive tasks of end-user devices at a
based forest fire detection. UAVs are able to offload images to Edge remote location. The concept of Edge Computing and the arrival of new
servers for further processing. In the case of wildfire, the workload of applications, enabled by recent trends in wireless communications such
the UAVs in the field increases significantly and the dynamic resource as the IoT and 5G, have introduced tremendous innovation opportuni-
allocation is essential to achieve the desired QoS. A group of linear ties for task offloading. However, new technical and business challenges
systems is used to model the container-based image processing services arise. This section discusses future research directions and open issues
and a state-feedback controller is designed to scale each container’s in the context of task offloading.
computing resources.
5.1. Heterogeneous networks
4.4.3. Lyapunov Optimization approaches
Lyapunov optimization algorithms provide a unique property of A Heterogeneous Network (HetNet) consists of a macro cell layout
finding the sufficient conditions for stability in dynamical systems. with some possible Low Power Nodes (LPNs) placed throughout their
Due to the stability theory of dynamical systems, Lyapunov-based coverage zones. Task offloading in HetNet is suited for three cases,
optimization algorithms can be used in order to study the task offload- as shown in Fig. 6: (i) Single-cell scenario, (ii) Contiguous cluster-cell
ing problem. In particular, for minimizing the energy consumption of scenario and (iii) Non-contiguous cluster-cell scenario.
mobile devices, there is a number of dynamic variables that need to be In the context of single-cell scenarios, new offloading decision vari-
fine-tuned and converged to an optimal value in order to minimize the ables can be the interference and less congested cells selection. In
total energy consumption. Thus, Lyapunov optimization can be used to the context of clustered-cells, the densest cell expands over several
find the necessary stability in the CPU-cycle frequency of the device, neighbor cells. The devices in the edge of the cell can extend the com-
transmission power, spectrum utilization and latency [139,170,178, munication capacity (as well as other types of network management,
191], while satisfying the necessary task execution constraints [115]. e.g., energy or mobility) through nearby devices, to neighboring cells.
Another variable that dynamically fluctuates during task offloading is In this case, the goal is to analyze the capability of a resource allocation

17
F. Saeik et al. Computer Networks 195 (2021) 108177

Table 4
Taxonomy of control theory-based task offloading algorithms.
Reference Optimization objective Algorithms Mobility Granularity
Delay Energy Bandwidth/ Load Deployment Optimal State feedback Lyapunov Rest control Static Mobile Partial Full Edge–
spectrum balancing cost control control approaches Cloud
[94] ✓ ✓ ✓ ✓ ✓ ✓
[100] ✓ ✓ ✓ ✓
[115] ✓ ✓ ✓ ✓ ✓
[121] ✓ ✓ ✓ ✓ ✓ ✓
[139] ✓ ✓ ✓ ✓ ✓
[158] ✓ ✓ ✓ ✓ ✓ ✓ ✓
[170] ✓ ✓ ✓ ✓ ✓
[178] ✓ ✓ ✓ ✓ ✓ ✓
[191] ✓ ✓ ✓ ✓ ✓ ✓
[268] ✓ ✓ ✓ ✓ ✓
[269] ✓ ✓ ✓ ✓ ✓ ✓ ✓
[274] ✓ ✓ ✓ ✓
[275] ✓ ✓ ✓ ✓
[276] ✓ ✓ ✓ ✓ ✓ ✓
[277] ✓ ✓ ✓ ✓
[280] ✓ ✓ ✓ ✓

Fig. 6. Overview of general heterogeneous networks scenarios.

technique to shift to a remote Edge location, involving inter-cell com- scenario, access networks are usually centralized, with all the traffic
munication and management aspects. A third scenario to be addressed going through a central node (i.e., a BS). Furthermore, by offloading
concerns the communication of devices (small-cell environments) that the tasks to a distributed Edge infrastructure, a separate backhaul
are not adjacent, i.e., scenarios where the clustered cells may or may connection is required which can increase the installation and energy
not be neighboring cells. This scenario assists in understanding how a costs for the mobile operators. Finding the right resources to offload the
resource allocation technique operates in terms of scalability, as well tasks to, in such distributed scenarios, is a critical objective, especially
as in terms of supporting challenges such as long delays, or network for heterogeneous networks. On top of that, the dynamic behavior
partitions (as the small-cells are not in contiguous clusters). of the user adds another level of complexity when the appropriate
Heterogeneous Dense Networks (HDHNs) consisting of clustered Edge site needs to be selected in order to achieve the task offload-
cells, require algorithms that are able to extend capacity over a distance ing objectives, such as low latency and resource optimization, while
of several cells towards the crowded cells. For the optimization of maintaining at the same time the user association information with
tasks with several small and distributed dense servers (either Edge or the adjacent BS. Developing a real-time task offloading algorithm that
Cloud), however, the algorithms should draw capacity from cells within considers the user interactions and a distributed Edge infrastructure, in
a short distance from the dense ones, such that only a few cells that are order to improve the service delivery in dynamic scenarios, is one of
located close to the dense cells are affected. Initial studies in this area the greatest currently open challenges. Moreover, in the heterogeneous
deal with simple scenarios of one end user and one Edge server in a networks, efficient real time allocation schemes that learn based on
single cell, or few end devices, one or more Edge servers and a central historic performance and adapt online to application’s statistics, are still
Cloud, while the results show the feasibility of task offloading with the in their infancy.
combination of Edge and Cloud communications [41,200]. However,
if multiple end-user devices reuse the spectrum to connect to multiple 5.3. Mobility-induced network dynamics
Edge and Cloud servers, imposing as a constraint to not degrade the
QoE and the service continuity, the effect of signaling overhead, smooth In many situations, the dynamic movement behavior of the users
handover and dynamic resource allocation on the offloading, becomes becomes the deciding factor on whether to offload the task or not.
more significant. Such effects have not been thoroughly explored in the Even though few existing researches aim to take mobility into ac-
aforementioned studies. count, the particular case is still considered an open challenge. For
example, developing algorithms by learning the user’s behavior and
5.2. Real-time distributed resource allocation network dynamics in parallel, in order to reduce communication and
computational costs, are of utmost importance for new and emerging
The optimization procedures during task offloading are primordial applications. Beyond 2020, there will be a growing demand for high
in order to handle crucial operations such as intelligent resource al- user mobility applications such as drone-based applications, high speed
location and service continuity, by making independent and rational trains, moving hotspots and 3D connectivity. Current solutions, would
strategic decisions and smartly adapting to the environment. In this be difficult to be applied in such extremes scenarios, not only in terms

18
F. Saeik et al. Computer Networks 195 (2021) 108177

of accuracy but also in terms of minimum performance requirements are certain obstacles in achieving seamless connectivity and uninter-
(e.g over 500 km/h high speed mobility, high throughput and ultra low rupted access to an Edge server while moving. For example, network
latency). bandwidth and data exchange rates may vary or connection might be
lost. Thus, task offloading should be enhanced with fault tolerance
5.4. Node, resource and application heterogeneity techniques to guarantee the successful transmission and execution of
the task, as well as minimize application response time and energy
Another critical challenge is dealing the heterogeneity of the avail- consumption in end-user devices.
able infrastructure in terms of hardware and available resources. Both
mobile and Edge devices are characterized by a great heterogeneity in 5.8. Control-related challenges
terms of hardware, software and resource capabilities specifications.
Furthermore, the existence of a large range of applications with differ- Although control theory is widely applied in Cloud elasticity and
ent performance requirements, that are readily available at the same resource allocation problems [283], there are still open challenges on
end device, can affect or limit the efficiency of the task offloading task offloading and Edge Computing that can be addressed by control
solution used. All these factors are key components during task offload-
techniques. Apart from respecting the QoS requirements, control the-
ing. Thus, maintaining adequate service delivery and service continuity,
ory is able to guarantee important system properties, e.g., stability,
while addressing the task offloading in such a heterogeneous scenario,
positive invariant sets and ultimate boundedness [284], against the
as a research topic, is still in its infancy.
inherent system’s uncertainties and external disturbances. Intermittent
connectivity, the innate management features of the virtualization
5.5. Moving edge resources closer to the end devices technologies and the limited resources at the edge of the network,
lead to a highly volatile dynamic environment that necessitates ad-
There exist several situations in which task offloading is indispens- vanced modeling and control methodologies. Regarding the modeling
able but where the end devices are not able to directly offload data of the offloading-based applications, control theory provides many
to an Edge server (e.g., that are not in range or do not have enough
modeling alternatives involving switching systems [145,284], Linear
energy to reach it). In such cases, it may be useful to send off mobile
Parameter Varying (LPV) systems [285,286] and Fuzzy Takagi–Sugeno
Edge resources close to these end devices and adapt to their mobility
systems [287,288], that allow the natural incorporation of uncertainties
and needs. However, obviously, mobile Edge resources might not be
and disturbances in the performance model.
as powerful as stationary ones and might also be limited in terms of
the services they can offer and their autonomy. Thus, the challenges
imposed to task offloading here will be (i) how to trigger the sending 5.9. Controller design for cyber–physical systems (CPS)
of mobile resources, (ii) what kind of tasks to offload and (iii) from
what end devices. Another active and very practical challenge in the context of IoT
and mobile-enabled computing, is the controller design for cyber–
5.6. Security and privacy physical systems (CPS), found e.g., in manufacturing, transportation
and collaborative robotics. In the context of dynamic networks and
Naturally, task offloading involves a huge amount of data out- remote computing, a joint co-design decision making strategy for task
sourced to third party Edge infrastructures, thus security and privacy offloading, resource allocation and controller design for CPS, is more
concepts are of paramount importance. These concepts can be ad- appropriate when compared to separate layers of controllers for the
dressed from different angles, i.e., (i) end user device, (ii) Edge data infrastructure resources and the system to be controlled. This new
center and (iii) the actual data transmission over the network. Lately, a generation of controllers will be made possible by merging two sets
great increase in the variety of sophisticated attacks on end user devices of models, namely (a) the performance model described above and
has been observed, which constitutes the main target for attackers. (b) the process model (having, for example, variables related to po-
Regarding the Edge infrastructures, threats are mainly focused on the sition, orientation and velocity of mobile devices, lighting conditions,
data transmission between the different nodes of the network. Proposed room temperature and mode of operation of sensors). The co-designed
solutions include various steganography and homomorphic encryption controllers should address many of the non-idealities of the dynamic
techniques, as well as hardware-based secure execution. However, networks found during task offloading, where resources must be used
when used individually, most of these solutions have limitations in their parsimoniously, in balance with the constraints and the overall objec-
applications; e.g., encryption keys may be too large hence dramatically tive [289]. It is worth mentioning that in control engineering, shared
increasing the amount of transmitted and stored data, while computa- and imperfect communication networks between the controller and the
tion on encrypted data is still in early research stages. Undoubtedly, sensor/actuator have been studied extensively, generating the branch
Edge-related security and privacy threats are advancing in a quick of Networked Control Systems (NCS) [290]. Several developed methods
manner, making it challenging to adapt to and deflect. Centralized address time delays and packet dropouts of NCS, utilizing perturba-
monolithic security systems need to evolve as well into agile distributed tion theory, Lyapunov stability theory and hybrid systems analysis,
solutions that combine more than one techniques, to fit better to the including probabilistic methods involving Markov chains and stochas-
Edge Computing paradigm. Hence, task offloading solutions should tic automata [291–293]. A breakthrough will be the emergence of
be enhanced by taking into consideration security and confidentiality event-triggered and self-triggered control, that allows asynchronous
constraints. sampling, thus reducing the network traffic, while at the same time
providing guaranteed trade-offs of the degradation of the closed-loop
5.7. Fault tolerance system performance [294]. Consequently, in the context of task offload-
ing, a timely challenge is to provide co-design control formulations
Apart from security and privacy, an important factor contributing that develop dynamic task offloading as well as control design algo-
to building trust towards task offloading at the Edge is fault tolerance. rithms for CPS, taking simultaneously into account the schedulability,
As thoroughly discussed in the previous sections, mobility support is available and requested network resources, Edge resources and energy
one of the most important requirements during task offloading and consumption. It is anticipated that such control algorithms will improve
this is because autonomy of communication and freedom of movement performance, utilization of the underlying infrastructure as well as
are crucial criteria when it comes to users’ satisfaction. Still, there resilience and robustness of the systems to be controlled.

19
F. Saeik et al. Computer Networks 195 (2021) 108177

6. Conclusion References

In this paper, we have presented a detailed and comprehensive [1] A. Gupta, R.K. Jha, A survey of 5G network: Architecture and emerging
study of the task offloading problem and we have extensively ana- technologies, IEEE Access 3 (2015) 1206–1232.
[2] M. Series, IMT Vision–framework and overall objectives of the future
lyzed the main components and different computing paradigms of an development of IMT for 2020 and beyond, Recomm. ITU (2015) 2083–0.
end-to-end communication path. This path starts at the end device [3] M. Altamimi, A. Abdrabou, K. Naik, A. Nayak, Energy cost models of smart-
(i.e., mobile/IoT devices) and leverages the benefits of the added phones for task offloading to the cloud, IEEE Trans. Emerg. Top. Comput. 3
computational resources at the edge of the network, before ending (3) (2015) 384–398.
up to the Cloud. Nonetheless, offloading parts of an application from [4] X. Ma, Y. Zhao, L. Zhang, H. Wang, L. Peng, When mobile terminals meet the
cloud: computation offloading as the bridge, IEEE Netw. 27 (5) (2013) 28–33.
an end device to a remote Edge or Cloud site, arises a number of [5] W. Yu, F. Liang, X. He, W.G. Hatcher, C. Lu, J. Lin, X. Yang, A survey on the
challenges mostly related to the dynamic network behavior and the edge computing for the Internet of Things, IEEE Access 6 (2017) 6900–6919.
resource allocation problem to be solved. Between these challenges, the [6] P. Mach, Z. Becvar, Mobile edge computing: A survey on architecture and
type of end devices in terms of mobility is expected to be one of the computation offloading, IEEE Commun. Surv. Tutor. 19 (3) (2017) 1628–1656.
main characteristics of task offloading that can dictate the scheduling [7] C. Jiang, X. Cheng, H. Gao, X. Zhou, J. Wan, Toward computation offloading
in edge computing: A survey, IEEE Access 7 (2019) 131543–131558.
and allocation of the offloaded tasks. Prominent solutions addressing [8] Y. Mao, C. You, J. Zhang, K. Huang, K.B. Letaief, A survey on mobile edge
these challenges have given the necessary momentum to task offloading computing: The communication perspective, IEEE Commun. Surv. Tutor. 19
in order to be considered as a viable solution not only to existing, but (4) (2017) 2322–2358.
to emerging and upcoming applications such as immersive applications, [9] C.-H. Hong, B. Varghese, Resource management in fog/edge computing: a
survey on architectures, infrastructure, and algorithms, ACM Comput. Surv. 52
autonomous vehicles and robotics.
(5) (2019) 1–37.
The benefits of task offloading are numerous, allowing to increase [10] S. Wang, X. Zhang, Y. Zhang, L. Wang, J. Yang, W. Wang, A survey on mobile
the QoS and QoE of the application, while extending the battery edge networks: Convergence of computing, caching and communications, IEEE
lifetime of the end devices. To achieve that, specific emphasis is given Access 5 (2017) 6757–6779.
on the optimization models used along with the different objectives, [11] T.K. Rodrigues, K. Suto, H. Nishiyama, J. Liu, N. Kato, Machine learning meets
computation and communication control in evolving edge and cloud: Challenges
in order to partition and allocate the tasks of end devices into an end-
and future perspective, IEEE Commun. Surv. Tutor. (2019).
to-end communication path that creates a user to Cloud continuum. [12] J. Wang, J. Pan, F. Esposito, P. Calyam, Z. Yang, P. Mohapatra, Edge cloud
Since this problem contains various dynamic parameters in terms of offloading algorithms: Issues, methods, and perspectives, ACM Comput. Surv.
network and user behavior, appropriate AI and ML techniques to 52 (1) (2019) 1–23.
anticipate traffic demands have also been presented. Control theory [13] Q. Pham, F. Fang, V.N. Ha, M.J. Piran, M. Le, L.B. Le, W. Hwang, Z.
Ding, A survey of multi-access edge computing in 5G and beyond: Funda-
techniques have also been investigated as an alternative way to address
mentals, technology integration, and state-of-the-art, IEEE Access 8 (2020)
the uncertainties of this dynamic problem, with the aim to ensure the 116974–117017.
necessary stability of the task offloading systems. Our survey concluded [14] A. Yousefpour, C. Fung, T. Nguyen, K. Kadiyala, F. Jalali, A. Niakanlahiji, J.
with some interesting open challenges that will shape and transform the Kong, J.P. Jue, All one needs to know about fog computing and related edge
task offloading problem for future network scenarios and applications. computing paradigms: A complete survey, J. Syst. Archit. 98 (2019) 289–330.
[15] L. Lin, X. Liao, H. Jin, P. Li, Computation offloading toward edge computing,
These future directions emphasize on control co-design, dynamic and
Proc. IEEE 107 (8) (2019) 1584–1607.
real-time allocation in a heterogeneous Edge environment, and secure [16] X. Shan, H. Zhi, P. Li, Z. Han, A survey on computation offloading for mobile
and highly mobile network platforms. edge computing information, in: 2018 IEEE 4th International Conference on
Big Data Security on Cloud (BigDataSecurity), IEEE International Conference
CRediT authorship contribution statement on High Performance and Smart Computing, (HPSC) and IEEE International
Conference on Intelligent Data and Security (IDS), IEEE, 2018, pp. 248–251.
[17] Y. Ai, M. Peng, K. Zhang, Edge computing technologies for Internet of Things:
Firdose Saeik: Conceptualization, Investigation, Writing - original a primer, Digit. Commun. Netw. 4 (2) (2018) 77–86.
draft, Writing - review & editing. Marios Avgeris: Conceptualization, [18] W.Z. Khan, E. Ahmed, S. Hakak, I. Yaqoob, A. Ahmed, Edge computing: A
Investigation, Writing - original draft, Writing - review & editing. Dim- survey, Future Gener. Comput. Syst. 97 (2019) 219–235.
itrios Spatharakis: Conceptualization, Investigation, Writing - origi- [19] L. Mendiboure, M.-A. Chalouf, F. Krief, Edge computing based applications in
vehicular environments: Comparative study and main issues, J. Comput. Sci.
nal draft, Writing - review & editing. Nina Santi: Conceptualization,
Tech. 34 (4) (2019) 869–886.
Investigation, Writing - original draft, Writing - review & editing. [20] P. Mell, T. Grance, et al., The NIST Definition of Cloud Computing, Computer
Dimitrios Dechouniotis: Conceptualization, Investigation, Writing - Security Division, National Information Technology Laboratory, 2011.
original draft, Writing - review & editing. John Violos: Conceptualiza- [21] L. Mei, W.K. Chan, T. Tse, A tale of clouds: Paradigm comparisons and some
tion, Investigation, Writing - original draft, Writing - review & editing. thoughts on research issues, in: 2008 IEEE Asia-Pacific Services Computing
Conference, IEEE, 2008, pp. 464–469.
Aris Leivadeas: Conceptualization, Investigation, Writing - original
[22] N. Fernando, S.W. Loke, W. Rahayu, Dynamic mobile cloud computing: Ad hoc
draft, Writing - review & editing. Nikolaos Athanasopoulos: Con- and opportunistic job sharing, in: 2011 Fourth IEEE International Conference
ceptualization, Investigation, Writing - original draft, Writing - review on Utility and Cloud Computing, IEEE, 2011, pp. 281–286.
& editing. Nathalie Mitton: Conceptualization, Investigation, Writing [23] G. Huerta-Canepa, D. Lee, A virtual cloud computing provider for mobile
- original draft, Writing - review & editing. Symeon Papavassiliou: devices, in: Proceedings of the 1st ACM Workshop on Mobile Cloud Computing
& Services: Social Networks and beyond, ACM, 2010, p. 6.
Conceptualization, Investigation, Writing - original draft, Writing - [24] A.U.R. Khan, M. Othman, S.A. Madani, S.U. Khan, A survey of mobile cloud
review & editing. computing application models, IEEE Commun. Surv. Tutor. 16 (1) (2014)
393–413.
Declaration of competing interest [25] N.I.M. Enzai, M. Tang, A taxonomy of computation offloading in mobile
cloud computing, in: 2014 2nd IEEE International Conference on Mobile Cloud
Computing, Services, and Engineering, 2014, pp. 19–28.
The authors declare that they have no known competing finan- [26] M. Satyanarayanan, V. Bahl, R. Caceres, N. Davies, The case for vm-based
cial interests or personal relationships that could have appeared to cloudlets in mobile computing, IEEE Pervas. Comput. (2009).
influence the work reported in this paper. [27] F. Bonomi, R. Milito, J. Zhu, S. Addepalli, Fog computing and its role in the
Internet of Things, in; Proceedings of the First Edition of the MCC Workshop
on Mobile Cloud Computing, 2012, pp. 13–16.
Acknowledgment
[28] S. Antonio, Cisco delivers vision of fog computing to accelerate value from
billions of connected devices, Cisco (2014).
This work was supported in part by the CHIST-ERA-2018-DRUID- [29] O. Consortium, et al., Openfog reference architecture for fog computing, Archit.
NET project, France. Work. Group (2017).

20
F. Saeik et al. Computer Networks 195 (2021) 108177

[30] R. Cisco, M.Y. Upc, M. Nemirovsky, Fog computing, in: Proc. Cloud Assist. [56] U. Drolia, R. Martins, J. Tan, A. Chheda, M. Sanghavi, R. Gandhi, P.
Serveys Eur. Conf. Bled, 2012, pp. 1–15. Narasimhan, The case for mobile edge-clouds, in: 2013 IEEE 10th International
[31] F. Bonomi, R. Milito, P. Natarajan, J. Zhu, Fog computing: A platform for Conference on Ubiquitous Intelligence and Computing and 2013 IEEE 10th
Internet of Things and analytics, in: Big Data and Internet of Things: A Roadmap International Conference on Autonomic and Trusted Computing, IEEE, 2013,
for Smart Environments, Springer, 2014, pp. 169–186. pp. 209–215.
[32] M. Chiang, T. Zhang, Fog and IoT: An overview of research opportunities, IEEE [57] J. Dolezal, Z. Becvar, T. Zeman, Performance evaluation of computation
Internet Things J. 3 (6) (2016) 854–864. offloading from mobile device to the edge of mobile network, in: 2016 IEEE
Conference on Standards for Communications and Networking, CSCN, IEEE,
[33] X. Hou, Y. Li, M. Chen, D. Wu, D. Jin, S. Chen, Vehicular fog computing: A
2016, pp. 1–7.
viewpoint of vehicles as the infrastructures, IEEE Trans. Veh. Technol. 65 (6)
(2016) 3860–3873. [58] I.-S. Comşa, G.-M. Muntean, R. Trestian, An innovative machine-learning-based
scheduling solution for improving live UHD video streaming quality in highly
[34] S.A. Soleymani, A.H. Abdullah, M. Zareei, M.H. Anisi, C. Vargas-Rosales, M.K.
dynamic network environments, IEEE Trans. Broadcast. (2020).
Khan, S. Goudarzi, A secure trust model based on fuzzy logic in vehicular ad
[59] S. Mangiante, G. Klas, A. Navon, Z. GuanHua, J. Ran, M.D. Silva, VR is on
hoc networks with fog computing, IEEE Access 5 (2017) 15619–15629.
the edge: How to deliver 360 videos in mobile networks, in: Proceedings of
[35] M.S. Elbamby, M. Bennis, W. Saad, Proactive edge computing in latency-
the Workshop on Virtual Reality and Augmented Reality Network, 2017, pp.
constrained fog networks, in: 2017 European Conference on Networks and
30–35.
Communications, EuCNC, IEEE, 2017, pp. 1–6.
[60] J. Chakareski, VR/AR immersive communication: Caching, edge computing, and
[36] D. You, T.V. Doan, R. Torre, M. Mehrabi, A. Kropp, V. Nguyen, H. Salah, G.T. transmission trade-offs, in: Proceedings of the Workshop on Virtual Reality and
Nguyen, F.H. Fitzek, Fog computing as an enabler for immersive media: Service Augmented Reality Network, 2017, pp. 36–41.
scenarios and research opportunities, IEEE Access 7 (2019) 65797–65810. [61] Q. Liu, S. Huang, J. Opadere, T. Han, An edge network orchestrator for mobile
[37] M. Iorga, L. Feldman, R. Barton, M. Martin, N. Goren, C. Mahmoudi, The Nist augmented reality, in: IEEE INFOCOM 2018-IEEE Conference on Computer
Definition of Fog Computing, Tech. Rep., National Institute of Standards and Communications, IEEE, 2018, pp. 756–764.
Technology, 2017. [62] E. Fraedrich, R. Cyganski, I. Wolf, B. Lenz, User Perspectives on Au-
[38] Y.C. Hu, M. Patel, D. Sabella, N. Sprecher, V. Young, Mobile edge computing—A tonomous Driving: A Use-Case-Driven Study in Germany, Geographisches
key technology towards 5G, ETSI White Paper 11 (11) (2015) 1–16. Institut, Humboldt-Universität zu Berlin, 2016.
[39] M. ETSI, Mobile edge computing (mec); framework and reference architecture, [63] U. Puetzschler, LTE and Car2x: Connected cars on the way to 5G, in: Mobile
in: ETSI, DGS MEC, vol. 3, 2016. Broadband SIG, vol. 6, 2016.
[40] D. Sabella, A. Vaillant, P. Kuure, U. Rauschenbach, F. Giust, Mobile-edge [64] Y. Sahni, J. Cao, S. Zhang, L. Yang, Edge mesh: A new paradigm to
computing architecture: The role of MEC in the Internet of Things, IEEE enable distributed intelligence in Internet of Things, IEEE Access 5 (2017)
Consumer Electron. Mag. 5 (4) (2016) 84–91. 16441–16458.
[41] K. Zhang, Y. Mao, S. Leng, Q. Zhao, L. Li, X. Peng, L. Pan, S. Mahar- [65] X. Wang, Z. Ning, L. Wang, Offloading in internet of vehicles: A fog-enabled
jan, Y. Zhang, Energy-efficient offloading for mobile edge computing in 5G real-time traffic management system, IEEE Trans. Ind. Inf. 14 (10) (2018)
heterogeneous networks, IEEE Access 4 (2016) 5896–5907. 4568–4578.
[66] G. Hu, W.P. Tay, Y. Wen, Cloud robotics: architecture, challenges and
[42] I. Ketykó, L. Kecskés, C. Nemes, L. Farkas, Multi-user computation offloading
applications, IEEE Netw. 26 (3) (2012) 21–28.
as multiple knapsack problem for 5G mobile edge computing, in: 2016
[67] B. Kehoe, S. Patil, P. Abbeel, K. Goldberg, A survey of research on cloud robotics
European Conference on Networks and Communications, EuCNC, IEEE, 2016,
and automation, IEEE Trans. Autom. Sci. Eng. 12 (2) (2015) 398–409.
pp. 225–229.
[68] S. Jordan, T. Haidegger, L. Kovács, I. Felde, I. Rudas, The rising prospects
[43] K. Zhang, Y. Mao, S. Leng, A. Vinel, Y. Zhang, Delay constrained offloading
of cloud robotic applications, in: 2013 IEEE 9th International Conference on
for mobile edge computing in cloud-enabled vehicular networks, in: 2016 8th
Computational Cybernetics, ICCC, IEEE, 2013, pp. 327–332.
International Workshop on Resilient Networks Design and Modeling, RNDM,
[69] D. Song, A.K. Tanwani, K. Goldberg, B. Siciliano, Networked-, Cloud-and
IEEE, 2016, pp. 288–294.
Fog-Robotics, Springer, 2019.
[44] T. Taleb, S. Dutta, A. Ksentini, M. Iqbal, H. Flinck, Mobile edge computing [70] A.K. Tanwani, R. Anand, J.E. Gonzalez, K. Goldberg, RILaaS: Robot Inference
potential in making cities smarter, IEEE Commun. Mag. 55 (3) (2017). and Learning as a Service, IEEE Robotics Autom. Lett. (2020).
[45] T.X. Tran, M.-P. Hosseini, D. Pompili, Mobile edge computing: Recent efforts [71] K. Bekris, R. Shome, A. Krontiris, A. Dobson, Cloud automation: Precomputing
and five key research directions, IEEE COMSOC MMTC Commun.-Frontiers roadmaps for flexible manipulation, IEEE Robot. Autom. Mag. 22 (2) (2015)
(2017). 41–50.
[46] T. Taleb, K. Samdanis, B. Mada, H. Flinck, S. Dutta, D. Sabella, On multi-access [72] J. Glover, D. Rus, N. Roy, Probabilistic models of object geometry for
edge computing: A survey of the emerging 5G network edge cloud architecture grasp planning, in: Proceedings of Robotics: Science and Systems IV, Zurich,
and orchestration, IEEE Commun. Surv. Tutor. 19 (3) (2017) 1657–1681. Switzerland, 2008, pp. 278–285.
[47] S. Kekki, W. Featherstone, Y. Fang, P. Kuure, A. Li, A. Ranjan, D. Purkayastha, [73] L. Riazuelo, J. Civera, J.M. Montiel, C2tam: A cloud framework for cooperative
F. Jiangping, D. Frydman, G. Verin, et al., MEC In 5G networks, ETSI White tracking and mapping, Robot. Auton. Syst. 62 (4) (2014) 401–413.
Paper 28 (2018) 1–28. [74] L. Turnbull, B. Samanta, Cloud robotics: Formation control of a multi robot sys-
[48] V. Bahl, Emergence of micro datacenter (cloudlets/edges) for mobile computing, tem utilizing cloud infrastructure, in: 2013 Proceedings of IEEE Southeastcon,
in: Microsoft Devices & Networking Summit 2015, 2015. IEEE, 2013, pp. 1–4.
[49] European Telecommunication Standards Institute (ETSI), Multi-access Edge [75] G. Mohanarajah, V. Usenko, M. Singh, R. D’Andrea, M. Waibel, Cloud-based
Computing (MEC), https://www.etsi.org/technologies/multi-access-edge- collaborative 3D mapping in real-time with low-cost robots, IEEE Trans. Autom.
computing. Sci. Eng. 12 (2) (2015) 423–431.
[76] N. Tian, A.K. Tawani, K. Goldberg, S. Sojoudi, Motion segmentation and
[50] E. Miluzzo, T. Wang, A.T. Campbell, Eyephone: activating mobile phones
synthesis for latency mitigation in a cloud robotic tele-operation system, in:
with your eyes, in: Proceedings of the Second ACM SIGCOMM Workshop on
International Symposium on Robotics Research, 2019.
Networking, Systems, and Applications on Mobile Handhelds, ACM, 2010, pp.
[77] S.L. Bowman, N. Atanasov, K. Daniilidis, G.J. Pappas, Probabilistic data asso-
15–20.
ciation for semantic slam, in: 2017 IEEE International Conference on Robotics
[51] R. Kemp, N. Palmer, T. Kielmann, H. Bal, Cuckoo: a computation offloading
and Automation, ICRA, IEEE, 2017, pp. 1722–1729.
framework for smartphones, in: International Conference on Mobile Computing,
[78] L. Riazuelo, M. Tenorth, D. Marco, M. Salas, L. Mosenlechner, L. Kunze,
Applications, and Services, Springer, 2010, pp. 59–79.
M. Beetz, J. Tardos, L. Montano, J. Montiel, Roboearth web-enabled and
[52] R. Kemp, N. Palmer, T. Kielmann, H. Bal, Opportunistic communication for knowledge-based active perception, in: IROS Workshop on AI-Based Robotics,
multiplayer mobile gaming: Lessons learned from photoshoot, in: Proceedings 2013.
of the Second International Workshop on Mobile Opportunistic Networking, [79] A.K. Tanwani, N. Mor, J. Kubiatowicz, J.E. Gonzalez, K. Goldberg, A fog
ACM, 2010, pp. 182–184. robotics approach to deep robot learning: Application to object recognition and
[53] S. Bohez, J. De Turck, T. Verbelen, P. Simoens, B. Dhoedt, Mobile, collaborative grasp planning in surface decluttering, in: 2019 International Conference on
augmented reality using cloudlets, in: 2013 International Conference on MOBILe Robotics and Automation, ICRA, IEEE, 2019, pp. 4559–4566.
Wireless MiddleWARE, Operating Systems, and Applications, IEEE, 2013, pp. [80] A.W. Services, AWS robomaker [online], 2021, https://aws.amazon.com/
45–54. robomaker/. (Accessed 20 January 2021).
[54] T. Verbelen, P. Simoens, F. De Turck, B. Dhoedt, Leveraging cloudlets for [81] Google, Cloud robotics core [online], 2021, https://googlecloudrobotics.github.
immersive collaborative applications, IEEE Pervas. Comput. 12 (4) (2013) io/core/, (Accessed 20 January 2021).
30–38. [82] Rapyuta robotics [online], 2021, https://www.rapyuta-robotics.com. (Accessed
[55] J.-M. Chung, Y.-S. Park, J.-H. Park, H. Cho, Adaptive cloud offloading of aug- 20 January 2021).
mented reality applications on smart devices for minimum energy consumption, [83] J. Dilley, B. Maggs, J. Parikh, H. Prokop, R. Sitaraman, B. Weihl, Globally
Ksii Trans. Internet Inf. Syst. 9 (8) (2015). distributed content delivery, IEEE Internet Comput. 6 (5) (2002) 50–58.

21
F. Saeik et al. Computer Networks 195 (2021) 108177

[84] C. Papagianni, A. Leivadeas, S. Papavassiliou, A cloud-oriented content delivery [111] F. Liu, Z. Huang, L. Wang, Energy-efficient collaborative task computation
network paradigm: Modeling and assessment, IEEE Trans. Dependable Secure offloading in cloud-assisted edge computing for IoT sensors, Sensors 19 (5)
Comput. 10 (5) (2013) 287–300. (2019) 1105.
[85] K. Bilal, A. Erbad, M. Hefeeda, Crowdsourced multi-view live video streaming [112] M. Chen, B. Liang, M. Dong, Multi-user multi-task offloading and resource
using cloud computing, IEEE Access 5 (2017) 12635–12647. allocation in mobile cloud systems, IEEE Trans. Wireless Commun. 17 (10)
[86] J.O. Fajardo, I. Taboada, F. Liberal, Improving content delivery efficiency (2018) 6790–6805.
through multi-layer mobile edge adaptation, IEEE Netw. 29 (6) (2015) 40–46. [113] Y. Cao, T. Jiang, C. Wang, Optimal radio resource allocation for mobile task
[87] T.X. Tran, P. Pandey, A. Hajisami, D. Pompili, Collaborative multi-bitrate video offloading in cellular networks, IEEE Netw. 28 (5) (2014) 68–73.
caching and processing in mobile-edge computing networks, in: 2017 13th [114] Y. Cao, C. Long, T. Jiang, S. Mao, Share communication and computation
Annual Conference on Wireless on-Demand Network Systems and Services, resources on mobile devices: a social awareness perspective, IEEE Wirel.
WONS, IEEE, 2017, pp. 165–172. Commun. 23 (4) (2016) 52–59.
[88] J. Ren, G. Yu, Y. Cai, Y. He, Latency optimization for resource allocation [115] D. Huang, P. Wang, D. Niyato, A dynamic offloading algorithm for mobile
in mobile-edge computation offloading, IEEE Trans. Wireless Commun. 17 (8) computing, IEEE Trans. Wireless Commun. 11 (6) (2012) 1991–1995.
(2018) 5506–5519. [116] S. Sardellitti, S. Barbarossa, G. Scutari, Distributed mobile cloud comput-
[89] Y.-H. Kao, B. Krishnamachari, M.-R. Ra, F. Bai, Hermes: Latency optimal ing: Joint optimization of radio and computational resources, in: 2014 IEEE
task assignment for resource-constrained mobile computing, IEEE Trans. Mob. Globecom Workshops, GC Wkshps, IEEE, 2014, pp. 1505–1510.
Comput. 16 (11) (2017) 3056–3069. [117] O. Munoz, A. Pascual-Iserte, J. Vidal, Optimization of radio and computational
[90] S. Yang, Y. He, X. Zheng, Fovr: Attention-based VR streaming through resources for energy efficiency in latency-constrained application offloading,
bandwidth-limited wireless networks, in: 2019 16th Annual IEEE International IEEE Trans. Veh. Technol. 64 (10) (2014) 4738–4755.
Conference on Sensing, Communication, and Networking, SECON, IEEE, 2019, [118] J. Plachy, Z. Becvar, E.C. Strinati, Dynamic resource allocation exploiting
pp. 1–9. mobility prediction in mobile edge computing, in: 2016 IEEE 27th Annual Inter-
[91] J. Gubbi, R. Buyya, S. Marusic, M. Palaniswami, Internet of Things (IoT): A national Symposium on Personal, Indoor, and Mobile Radio Communications,
vision, architectural elements, and future directions, Future Gener. Comput. PIMRC, IEEE, 2016, pp. 1–6.
Syst. 29 (7) (2013) 1645–1660. [119] T. Truong-Huu, C.-K. Tham, D. Niyato, To offload or to wait: An opportunistic
[92] I. Farris, L. Militano, M. Nitti, L. Atzori, A. Iera, Federated edge-assisted mobile offloading algorithm for parallel tasks in a mobile cloud, in: 2014 IEEE 6th
clouds for service provisioning in heterogeneous IoT environments, in: 2015 International Conference on Cloud Computing Technology and Science, IEEE,
IEEE 2nd World Forum on Internet of Things, WF-IoT, IEEE, 2015, pp. 591–596. 2014, pp. 182–189.
[93] C. Zhu, X. Li, L. Song, L. Xiang, Development of a theoretically based thermal [120] M. Chen, Y. Hao, Y. Li, C.-F. Lai, D. Wu, On the computation offloading at
model for lithium ion battery pack, J. Power Sources 223 (2013) 155–164. ad hoc cloudlet: architecture and service modes, IEEE Commun. Mag. 53 (6)
[94] Y. Nan, W. Li, W. Bao, F.C. Delicato, P.F. Pires, Y. Dou, A.Y. Zomaya, Adaptive (2015) 18–24.
energy-aware computation offloading for cloud of things systems, IEEE Access [121] L. Pu, X. Chen, J. Xu, X. Fu, D2D fogging: An energy-efficient and incentive-
5 (2017) 23947–23957. aware task offloading framework via network-assisted D2D collaboration, IEEE
[95] I. Farris, L. Militano, M. Nitti, L. Atzori, A. Iera, Mifaas: A mobile-IoT-federation- J. Sel. Areas Commun. 34 (12) (2016) 3887–3901.
as-a-service model for dynamic cooperation of IoT cloud providers, Future [122] S. Zhou, Y. Sun, Z. Jiang, Z. Niu, Exploiting moving intelligence: Delay-
Gener. Comput. Syst. 70 (2017) 126–137. optimized computation offloading in vehicular fog networks, IEEE Commun.
[96] S.K. Sharma, X. Wang, Live data analytics with collaborative edge and cloud Mag. 57 (5) (2019) 49–55.
processing in wireless IoT networks, IEEE Access 5 (2017) 4621–4635. [123] W. Li, X. You, Y. Jiang, J. Yang, L. Hu, Opportunistic computing offloading in
[97] A. Yousefpour, G. Ishigaki, R. Gour, J.P. Jue, On reducing IoT service delay edge clouds, J. Parallel Distrib. Comput. 123 (2019) 69–76.
via fog offloading, IEEE Internet Things J. 5 (2) (2018) 998–1010. [124] Y. Sun, S. Zhou, J. Xu, EMM: Energy-aware mobility management for mobile
[98] K. Kim, C.S. Hong, Optimal task-UAV-edge matching for computation offloading edge computing in ultra dense networks, IEEE J. Sel. Areas Commun. 35 (11)
in uav assisted mobile edge computing, in: 2019 20th Asia-Pacific Network (2017) 2637–2646.
Operations and Management Symposium, APNOMS, IEEE, 2019, pp. 1–4. [125] Z. Wang, Z. Zhao, G. Min, X. Huang, Q. Ni, R. Wang, User mobility aware task
[99] S. Chen, Q. Wang, J. Chen, T. Wu, An intelligent task offloading algorithm assignment for mobile edge computing, Future Gener. Comput. Syst. 85 (2018)
(iTOA) for UAV network, in: 2019 IEEE Globecom Workshops, GC Wkshps, 1–8.
IEEE, 2019, pp. 1–6. [126] G. Ahani, D. Yuan, BS-assisted task offloading for D2D networks with pres-
[100] M. Avgeris, D. Spatharakis, D. Dechouniotis, N. Kalatzis, I. Roussaki, S. ence of user mobility, in: 2019 IEEE 89th Vehicular Technology Conference
Papavassiliou, Where there is fire there is smoke: a scalable edge computing (VTC2019-Spring), IEEE, 2019, pp. 1–5.
framework for early fire detection, Sensors 19 (3) (2019) 639. [127] Z. Ning, P. Dong, X. Wang, J.J. Rodrigues, F. Xia, Deep reinforcement learning
[101] J. Wang, M. Conrad Meyer, Y. Wu, Y. Wang, Maximum data-resolution for vehicular edge computing: An intelligent offloading system, ACM Trans.
efficiency for fog-computing supported spatial big data processing in disaster Intell. Syst. Technol. (TIST) 10 (6) (2019) 1–24.
scenarios, IEEE Trans. Parallel Distrib. Syst. 30 (2019) 1826–1842. [128] T.M.T. Do, O. Dousse, M. Miettinen, D. Gatica-Perez, A probabilistic kernel
[102] J. Zhou, J. Fan, J. wang, J. Zhu, Task offloading for social sensing applications method for human mobility prediction with smartphones, Pervasive Mob.
in mobile edge computing, in: 2019 Seventh International Conference on Comput. 20 (2015) 13–28.
Advanced Cloud and Big Data, CBD, IEEE, 2019, pp. 333–338. [129] B. Li, H. Zhang, H. Lu, User mobility prediction based on Lagrange’s inter-
[103] G. Ananthanarayanan, P. Bahl, P. Bodík, K. Chintalapudi, M. Philipose, L. polation in ultra-dense networks, in: 2016 IEEE 27th Annual International
Ravindranath, S. Sinha, Real-time video analytics: The killer app for edge Symposium on Personal, Indoor, and Mobile Radio Communications, PIMRC,
computing, Computer 50 (10) (2017) 58–67. IEEE, 2016, pp. 1–6.
[104] J. Wang, J. Pan, F. Esposito, P. Calyam, Z. Yang, P. Mohapatra, Edge cloud [130] H.A. Alameddine, S. Sharafeddine, S. Sebbah, S. Ayoubi, C. Assi, Dynamic task
offloading algorithms: Issues, methods, and perspectives, ACM Comput. Surv. offloading and scheduling for low-latency IoT services in multi-access edge
52 (1) (2019) 2:1–2:23, http://doi.acm.org/10.1145/3284387. computing, IEEE J. Sel. Areas Commun. 37 (3) (2019) 668–682.
[105] T.X. Tran, A. Hajisami, P. Pandey, D. Pompili, Collaborative mobile edge [131] S. Wu, W. Xia, W. Cui, Q. Chao, Z. Lan, F. Yan, L. Shen, An efficient offload-
computing in 5G networks: New paradigms, scenarios, and challenges, IEEE ing algorithm based on support vector machine for mobile edge computing
Commun. Mag. 55 (4) (2017) 54–61. in vehicular networks, in: 2018 10th International Conference on Wireless
[106] J. Violos, S. Pelekis, A. Berdelis, S. Tsanakas, K. Tserpes, T. Varvarigou, Communications and Signal Processing, WCSP, IEEE, 2018, pp. 1–6.
Predicting visitor distribution for large events in smart cities, in: 2019 IEEE [132] M. Maanoja, M. Weckström, Location services, Google Patents, US Patent 7,
International Conference on Big Data and Smart Computing, BigComp, 2019, 069, 023, 2006.
pp. 1–8 (ISSN: 2375-9356). [133] S. Deng, L. Huang, D. Hu, J.L. Zhao, Z. Wu, Mobility-enabled service selection
[107] T.X. Tran, D. Pompili, Joint task offloading and resource allocation for multi- for composite services, IEEE Trans. Serv. Comput. 9 (3) (2016) 394–407.
server mobile-edge computing networks, IEEE Trans. Veh. Technol. 68 (1) [134] S. Deng, L. Huang, J. Taheri, J. Yin, M. Zhou, A.Y. Zomaya, Mobility-aware
(2018) 856–868. service composition in mobile communities, IEEE Trans. Syst. Man Cybern.:
[108] J. Violos, E. Psomakelis, K. Tserpes, F. Aisopos, T. Varvarigou, Leveraging Syst. 47 (3) (2017) 555–568.
user mobility and mobile app services behavior for optimal edge resource [135] S. Gambs, M.-O. Killijian, M.N. del Prado Cortez, Next place prediction using
utilization, in: Proceedings of the International Conference on Omni-Layer mobility Markov chains, in: Proceedings of the First Workshop on Measurement,
Intelligent Systems, COINS ’19, Association for Computing Machinery, Crete, Privacy, and Mobility, 2012, pp. 1–6.
Greece, 2019, pp. 7–12. [136] K. Lee, I. Shin, User mobility model based computation offloading decision for
[109] L. Yang, H. Zhang, M. Li, J. Guo, H. Ji, Mobile edge computing empowered mobile cloud, J. Comput. Sci. Eng. 9 (3) (2015) 155–162.
energy efficient task offloading in 5G, IEEE Trans. Veh. Technol. 67 (7) (2018) [137] X. Sun, N. Ansari, Adaptive avatar handoff in the cloudlet network, IEEE Trans.
6398–6409. Cloud Comput. 7 (3) (2019) 664–676.
[110] Z. Xu, X. Liu, G. Jiang, B. Tang, A time-efficient data offloading method with [138] M. Chen, Y. Hao, M. Qiu, J. Song, D. Wu, I. Humar, Mobility-aware caching
privacy preservation for intelligent sensors in edge computing, EURASIP J. and computation offloading in 5G ultra-dense cellular networks, Sensors 16 (7)
Wireless Commun. Networking 2019 (1) (2019) 1–12. (2016) 974.

22
F. Saeik et al. Computer Networks 195 (2021) 108177

[139] J. Xu, L. Chen, P. Zhou, Joint service caching and task offloading for mobile [165] P. Di Lorenzo, S. Barbarossa, S. Sardellitti, Joint optimization of radio resources
edge computing in dense networks, in: IEEE INFOCOM 2018-IEEE Conference and code partitioning in mobile edge computing, 2013, arXiv preprint arXiv:
on Computer Communications, IEEE, 2018, pp. 207–215. 1307.3835.
[140] D. Wang, Z. Liu, X. Wang, Y. Lan, Mobility-aware task offloading and migration [166] W. Labidi, M. Sarkiss, M. Kamoun, Energy-optimal resource scheduling and
schemes in fog computing networks, IEEE Access 7 (2019) 43356–43368. computation offloading in small cell networks, in: 2015 22nd International
[141] Y. Shi, S. Chen, X. Xu, MAGA: A mobility-aware computation offloading Conference on Telecommunications, ICT, IEEE, 2015, pp. 313–318.
decision for distributed mobile cloud computing, IEEE Internet Things J. 5 (1) [167] M. Kamoun, W. Labidi, M. Sarkiss, Joint resource allocation and offloading
(2017) 164–174. strategies in cloud enabled cellular networks, in: 2015 IEEE International
[142] F. Yu, H. Chen, J. Xu, DMPO: Dynamic mobility-aware partial offloading in Conference on Communications, ICC, IEEE, 2015, pp. 5529–5534.
mobile edge computing, Future Gener. Comput. Syst. 89 (2018) 722–735. [168] S. Sardellitti, G. Scutari, S. Barbarossa, Joint optimization of radio and compu-
[143] V.A. Siris, D. Kalyvas, Enhancing mobile data offloading with mobility predic- tational resources for multicell mobile-edge computing, IEEE Trans. Signal Inf.
tion and prefetching, ACM SIGMOBILE Mob. Comput. Commun. Rev. 17 (1) Process. Netw. 1 (2) (2015) 89–103.
(2013) 22–29. [169] W. Labidi, M. Sarkiss, M. Kamoun, Joint multi-user resource scheduling and
[144] C. Yang, Y. Liu, X. Chen, W. Zhong, S. Xie, Efficient mobility-aware task computation offloading in small cell networks, in: 2015 IEEE 11th Inter-
offloading for vehicular edge computing networks, IEEE Access 7 (2019) national Conference on Wireless and Mobile Computing, Networking and
26652–26664. Communications, WiMob, IEEE, 2015, pp. 794–801.
[145] D. Spatharakis, I. Dimolitsas, D. Dechouniotis, G. Papathanail, I. Fotoglou,
[170] Y. Mao, J. Zhang, K.B. Letaief, Dynamic computation offloading for mobile-edge
P. Papadimitriou, S. Papavassiliou, A scalable edge computing architecture
computing with energy harvesting devices, IEEE J. Sel. Areas Commun. 34 (12)
enabling smart offloading for location based services, Pervasive Mob. Comput.
(2016) 3590–3605.
67 (2020) 101217.
[171] K. Singh, A. Awasthi, Quality, Reliability, Security and Robustness in
[146] G. Papathanail, I. Fotoglou, C. Demertzis, A. Pentelas, K. Sgouromitis, P.
Heterogeneous Networks, Springer, 2013.
Papadimitriou, D. Spatharakis, I. Dimolitsas, D. Dechouniotis, S. Papavassiliou,
[172] A. Kiani, N. Ansari, Optimal code partitioning over time and hierarchical
COSMOS: An orchestration framework for smart computation offloading in edge
cloudlets, IEEE Commun. Lett. 22 (1) (2017) 181–184.
clouds, in: NOMS 2020-2020 IEEE/IFIP Network Operations and Management
[173] Y. Mao, J. Zhang, K.B. Letaief, Joint task offloading scheduling and transmit
Symposium, IEEE, 2020, pp. 1–6.
power allocation for mobile-edge computing systems, in: 2017 IEEE Wireless
[147] M. Akter, F.T. Zohra, A.K. Das, Q-MAC: Qos and mobility aware optimal re-
Communications and Networking Conference, WCNC, IEEE, 2017, pp. 1–6.
source allocation for dynamic application offloading in mobile cloud computing,
in: 2017 International Conference on Electrical, Computer and Communication [174] J. Zhang, X. Hu, Z. Ning, E.C.-H. Ngai, L. Zhou, J. Wei, J. Cheng, B. Hu,
Engineering, ECCE, IEEE, 2017, pp. 803–808. Energy-latency tradeoff for energy-aware offloading in mobile edge computing
[148] C. Puliafito, E. Mingozzi, C. Vallati, F. Longo, G. Merlino, Companion fog networks, IEEE Internet Things J. 5 (4) (2017) 2633–2645.
computing: Supporting things mobility through container migration at the edge, [175] J. Guo, Z. Song, Y. Cui, Z. Liu, Y. Ji, Energy-efficient resource allocation for
in: 2018 IEEE International Conference on Smart Computing, SMARTCOMP, multi-user mobile edge computing, in: GLOBECOM 2017-2017 IEEE Global
2018, pp. 97–105. Communications Conference, IEEE, 2017, pp. 1–7.
[149] W. Junior, A. França, K. Dias, J.N. de Souza, Supporting mobility-aware [176] F. Samie, V. Tsoutsouras, L. Bauer, S. Xydis, D. Soudris, J. Henkel, Computation
computational offloading in mobile cloud environment, J. Netw. Comput. Appl. offloading and resource allocation for low-power IoT edge devices, in: 2016
94 (2017) 93–108. IEEE 3rd World Forum on Internet of Things, WF-IoT, IEEE, 2016, pp. 7–12.
[150] L. Yang, J. Cao, H. Cheng, Y. Ji, Multi-user computation partitioning for [177] A. Leivadeas, C. Papagianni, S. Papavassiliou, Going Green with the Networked
latency sensitive mobile cloud applications, IEEE Trans. Comput. 64 (8) (2014) Cloud: Methodologies and Assessment, John Wiley & Sons, Inc., 2015.
2253–2266. [178] S. Li, N. Zhang, S. Lin, L. Kong, A. Katangur, M.K. Khan, M. Ni, G. Zhu, Joint
[151] S. Bi, Y.J. Zhang, Computation rate maximization for wireless powered mobile- admission control and resource allocation in edge computing for Internet of
edge computing with binary computation offloading, IEEE Trans. Wireless Things, IEEE Netw. 32 (1) (2018) 72–79.
Commun. 17 (6) (2018) 4177–4190. [179] L. Tong, Y. Li, W. Gao, A hierarchical edge cloud architecture for mobile com-
[152] T. Zhao, S. Zhou, X. Guo, Z. Niu, Tasks scheduling and resource allocation in puting, in: IEEE INFOCOM 2016-the 35th Annual IEEE International Conference
heterogeneous cloud for delay-bounded mobile edge computing, in: 2017 IEEE on Computer Communications, IEEE, 2016, pp. 1–9.
International Conference on Communications, ICC, IEEE, 2017, pp. 1–7. [180] A. Leivadeas, G. Kesidis, M. Ibnkahla, I. Lambadaris, VNF placement
[153] Q. Liu, T. Han, N. Ansari, Joint radio and computation resource management optimization at the edge and cloud, Future Internet 11 (3) (2019) 69.
for low latency mobile edge computing, in: 2018 IEEE Global Communications [181] A. Leivadeas, G. Kesidis, M. Falkner, I. Lambadaris, A graph partitioning game
Conference, GLOBECOM, IEEE, 2018, pp. 1–7. theoretical approach for the VNF service chaining problem, IEEE Trans. Netw.
[154] L. Yang, J. Cao, Y. Yuan, T. Li, A. Han, A. Chan, A framework for partitioning Serv. Manag. 14 (4) (2017) 890–903.
and execution of data stream applications in mobile cloud computing, ACM [182] A. Leivadeas, M. Falkner, I. Lambadaris, M. Ibnkahla, G. Kesidis, Balancing
SIGMETRICS Perform. Eval. Rev. 40 (4) (2013) 23–32. delay and cost in virtual network function placement and chaining, in: 2018
[155] J. Liu, Y. Mao, J. Zhang, K.B. Letaief, Delay-optimal computation task schedul- IEEE International Conference on Network Softwarization and Workshops,
ing for mobile-edge computing systems, in: 2016 IEEE International Symposium NetSoft, IEEE, 2018, pp. 433–440.
on Information Theory, ISIT, IEEE, 2016, pp. 1451–1455. [183] I. Skarga-Bandurova, M. Derkach, I. Kotsiuba, The information service for
[156] M. Jia, W. Liang, Z. Xu, M. Huang, Y. Ma, Qos-aware cloudlet load balancing
delivering arrival public transport prediction, in: 2018 IEEE 4th International
in wireless metropolitan area networks, IEEE Trans. Cloud Comput. (2018).
Symposium on Wireless Systems Within the International Conferences on
[157] C. Wang, C. Liang, F.R. Yu, Q. Chen, L. Tang, Computation offloading and
Intelligent Data Acquisition and Advanced Computing Systems, IDAACS-SWS,
resource allocation in wireless cellular networks with mobile edge computing,
2018, pp. 191–195.
IEEE Trans. Wireless Commun. 16 (8) (2017) 4924–4938.
[184] A. Hameed, A. Leivadeas, IoT Traffic multi-classification using network and
[158] C. Sonmez, A. Ozgovde, C. Ersoy, Fuzzy workload orchestration for edge
statistical features in a smart environment, in: EEE International Workshop on
computing, IEEE Trans. Netw. Serv. Manag. (2019).
Computer Aided Modeling and Design of Communication Links and Networks,
[159] C.-F. Liu, M. Bennis, H.V. Poor, Latency and reliability-aware task offloading
CAMAD, IEEE, 2020, pp. 1–7.
and resource allocation for mobile edge computing, in: 2017 IEEE Globecom
[185] W. Sun, J. Liu, Y. Yue, AI-enhanced offloading in edge computing: when
Workshops, GC Wkshps, IEEE, 2017, pp. 1–7.
machine learning meets industrial IoT, IEEE Netw. 33 (5) (2019) 68–74.
[160] L. Wang, L. Jiao, J. Li, M. Mühlhäuser, Online resource allocation for
arbitrary user mobility in distributed edge clouds, in: 2017 IEEE 37th Interna- [186] A. Botchkarev, A new typology design of performance metrics to measure
tional Conference on Distributed Computing Systems, ICDCS, IEEE, 2017, pp. errors in machine learning regression algorithms, Interdisciplinary J. Inf. Knowl.
1281–1290. Manag. 14 (2019) 045–076.
[161] S. Cao, X. Tao, Y. Hou, Q. Cui, An energy-optimal offloading algorithm of [187] J.-O. Palacio-Niño, F. Berzal, Evaluation metrics for unsupervised learning
mobile computing based on hetnets, in: 2015 International Conference on algorithms, 2019, arXiv preprint arXiv:1905.05667.
Connected Vehicles and Expo, ICCVE, IEEE, 2015, pp. 254–258. [188] J. Lever, M. Krzywinski, N. Altman, Classification evaluation, Nature Methods
[162] Y. Zhao, S. Zhou, T. Zhao, Z. Niu, Energy-efficient task offloading for mul- 13 (8) (2016) 603–604, https://www.nature.com/articles/nmeth.3945.
tiuser mobile cloud computing, in: 2015 IEEE/CIC International Conference on [189] Y. Wang, M. Sheng, X. Wang, L. Wang, J. Li, Mobile-edge computing: Partial
Communications in China, ICCC, IEEE, 2015, pp. 1–5. computation offloading using dynamic voltage scaling, IEEE Trans. Commun.
[163] Y. Wang, M. Sheng, X. Wang, L. Wang, W. Han, Y. Zhang, Y. Shi, Energy- 64 (10) (2016) 4268–4282.
optimal partial computation offloading using dynamic voltage scaling, in: 2015 [190] M. Deng, H. Tian, B. Fan, Fine-granularity based application offloading policy
IEEE International Conference on Communication Workshop, ICCW, IEEE, 2015, in cloud-enhanced small cell networks, in: 2016 IEEE International Conference
pp. 2695–2700. on Communications Workshops, ICC, IEEE, 2016, pp. 638–643.
[164] C. You, K. Huang, Multiuser resource allocation for mobile-edge computation [191] Y. Mao, J. Zhang, S. Song, K.B. Letaief, Power-delay tradeoff in multi-
offloading, in: 2016 IEEE Global Communications Conference, GLOBECOM, user mobile-edge computing systems, in: 2016 IEEE Global Communications
IEEE, 2016, pp. 1–6. Conference, GLOBECOM, IEEE, 2016, pp. 1–6.

23
F. Saeik et al. Computer Networks 195 (2021) 108177

[192] C. You, K. Huang, H. Chae, B.-H. Kim, Energy-efficient resource allocation [219] D. Rahbari, M. Nickray, Task offloading in mobile fog computing by
for mobile-edge computation offloading, IEEE Trans. Wireless Commun. 16 (3) classification and regression tree, Peer-to-Peer Netw. Appl. 13 (1) (2020)
(2016) 1397–1411. 104–122.
[193] M.-H. Chen, B. Liang, M. Dong, Joint offloading and resource allocation for [220] W. Junior, E. Oliveira, A. Santos, K. Dias, A context-sensitive offloading system
computation and communication in mobile cloud with computing access point, using machine-learning classification algorithms for mobile cloud environment,
in: IEEE INFOCOM 2017-IEEE Conference on Computer Communications, IEEE, Future Gener. Comput. Syst. 90 (2019) 503–520.
2017, pp. 1–9. [221] H. Bashir, S. Lee, K.H. Kim, Resource allocation through logistic regression
[194] M.-H. Chen, B. Liang, M. Dong, A semidefinite relaxation approach to mobile and multicriteria decision making method in IoT fog computing, Trans. Emerg.
cloud offloading with computing access point, in: 2015 IEEE 16th International Telecommun. Technol. n/a (n/a) (2019) e3824, _eprint: https://onlinelibrary.
Workshop on Signal Processing Advances in Wireless Communications, SPAWC, wiley.com/doi/pdf/10.1002/ett.3824, https://onlinelibrary.wiley.com/doi/abs/
IEEE, 2015, pp. 186–190. 10.1002/ett.3824.
[195] Y. Liu, F.R. Yu, X. Li, H. Ji, V.C. Leung, Hybrid computation offloading in fog [222] Y. Jararweh, M.B. Issa, M. Daraghmeh, M. Al-Ayyoub, M.A. Alsmirat, Energy
and cloud networks with non-orthogonal multiple access, in: IEEE INFOCOM efficient dynamic resource management in cloud computing based on logistic
2018-IEEE Conference on Computer Communications Workshops, INFOCOM regression model and median absolute deviation, Sustain. Comput.: Inform. Syst.
WKSHPS, IEEE, 2018, pp. 154–159. 19 (2018) 262–274.
[196] Q.-V. Pham, T. Leanh, N.H. Tran, B.J. Park, C.S. Hong, Decentralized computa- [223] J.C. Guevara, R.d.S. Torres, N.L.S. da Fonseca, On the classification of fog
tion offloading and resource allocation for mobile-edge computing: A matching computing applications: A machine learning perspective, J. Netw. Comput.
game approach, IEEE Access 6 (2018) 75868–75885. Appl. 159 (2020) 102596.
[197] Z. Ning, P. Dong, X. Kong, F. Xia, A cooperative partial computation offloading [224] L. Liu, D. Qi, N. Zhou, Y. Wu, A task scheduling algorithm based on classifi-
scheme for mobile edge computing enabled Internet of Things, IEEE Internet cation mining in fog computing environment, Wirel. Commun. Mob. Comput.
Things J. (2018). (ISSN: 1530-8669) 2018 (2018) e2102348, https://www.hindawi.com/journals/
[198] J. Du, L. Zhao, J. Feng, X. Chu, Computation offloading and resource allocation wcmc/2018/2102348/.
in mixed fog/cloud computing systems with min-max fairness guarantee, IEEE [225] K. Kim, J. Lynskey, S. Kang, C.S. Hong, Prediction based sub-task offloading
Trans. Commun. 66 (4) (2018) 1594–1608. in mobile edge computing, in: 2019 International Conference on Information
[199] S. Li, Y. Tao, X. Qin, L. Liu, Z. Zhang, P. Zhang, Energy-aware mobile edge Networking, ICOIN, 2019, pp. 448–452 (ISSN: 1976-7684).
computation offloading for IoT over heterogenous networks, IEEE Access 7 [226] F. Farahnakian, P. Liljeberg, J. Plosila, LiRCUP: Linear regression based cpu
(2019) 13092–13105. usage prediction algorithm for live migration of virtual machines in data
[200] H. Guo, J. Liu, J. Zhang, Computation offloading for multi-access mobile edge centers, in: 2013 39th Euromicro Conference on Software Engineering and
computing in ultra-dense networks, IEEE Commun. Mag. 56 (8) (2018) 14–19. Advanced Applications, 2013, pp. 357–364 (ISSN: 2376-9505).
[201] X. Chen, L. Jiao, W. Li, X. Fu, Efficient multi-user computation offloading [227] R.A.C. da Silva, N.L.S.d. Fonseca, Resource allocation mechanism for a fog-cloud
for mobile-edge cloud computing, IEEE/ACM Trans. Netw. 24 (5) (2015) infrastructure, in: 2018 IEEE International Conference on Communications, ICC,
2795–2808. 2018, pp. 1–6 (ISSN: 1938-1883).
[202] Y. Liu, C. Xu, Y. Zhan, Z. Liu, J. Guan, H. Zhang, Incentive mechanism for
[228] R. Hu, J. Jiang, G. Liu, L. Wang, CPU load prediction using support vector
computation offloading using edge computing: A Stackelberg game approach,
regression and Kalman Smoother for Cloud, in: 2013 IEEE 33rd International
Comput. Netw. 129 (2017) 399–409.
Conference on Distributed Computing Systems Workshops, 2013, pp. 88–92
[203] H. Guo, J. Liu, Collaborative computation offloading for multiaccess edge
(ISSN: 2332-5666).
computing over fiber–wireless networks, IEEE Trans. Veh. Technol. 67 (5)
[229] F. Farahnakian, T. Pahikkala, P. Liljeberg, J. Plosila, Energy aware consolidation
(2018) 4514–4526.
algorithm based on k-nearest neighbor regression for cloud data centers, in:
[204] M. Zeng, Y. Li, K. Zhang, M. Waqas, D. Jin, Incentive mechanism design
2013 IEEE/ACM 6th International Conference on Utility and Cloud Computing,
for computation offloading in heterogeneous fog computing: A contract-based
2013, pp. 256–259.
approach, in: 2018 IEEE International Conference on Communications, ICC,
[230] H. Cheng, W. Xia, F. Yan, L. Shen, Balanced clustering and joint re-
2018, pp. 1–6.
sources allocation in cooperative fog computing system, in: 2019 IEEE Global
[205] J. Du, E. Gelenbe, C. Jiang, H. Zhang, Y. Ren, Contract design for traffic
Communications Conference, GLOBECOM, 2019, pp. 1–6 (ISSN: 2576-6813).
offloading and resource allocation in heterogeneous ultra-dense networks, IEEE
[231] M. Bouet, V. Conan, Mobile edge computing resources optimization: A
J. Sel. Areas Commun. 35 (11) (2017) 2457–2467.
geo-clustering approach, IEEE Trans. Netw. Serv. Manag. 15 (2) (2018)
[206] Z. Hu, Z. Zheng, L. Song, T. Wang, X. Li, UAV offloading: Spectrum trading con-
787–796.
tract design for UAV-assisted cellular networks, IEEE Trans. Wireless Commun.
[232] Y. Li, N.T. Anh, A.S. Nooh, K. Ra, M. Jo, Dynamic mobile cloudlet clustering for
17 (9) (2018) 6093–6107.
fog computing, in; 2018 International Conference on Electronics, Information,
[207] Y. Li, J. Zhang, X. Gan, L. Fu, H. Yu, X. Wang, A contract-based incentive
and Communication, ICEIC, 2018, pp. 1–4.
mechanism for delayed traffic offloading in cellular networks, IEEE Trans.
[233] G. Li, Q. Lin, J. Wu, Y. Zhang, J. Yan, Dynamic computation offloading
Wireless Commun. 15 (8) (2016) 5314–5327.
based on graph partitioning in mobile edge computing, IEEE Access 7 (2019)
[208] M.G.C. Resende, P. Pardalos, Handbook of Optimization in Telecommunication,
185131–185139.
Springer, 2006.
[209] D. Monderer, L.S. Shapley, Potential games, Games Economic Behav. 14 (1) [234] L. Hu, Y. Tian, J. Yang, T. Taleb, L. Xiang, Y. Hao, Ready player one: UAV-
(1996) 124–143. clustering-based multi-task offloading for vehicular VR/AR gaming, IEEE Netw.
[210] Y. Zhang, M. Pan, L. Song, Z. Dawy, Z. Han, A survey of contract theory-based 33 (3) (2019) 42–48.
incentive mechanism design in wireless networks, IEEE Wirel. Commun. 24 (3) [235] X. Liu, J. Yu, J. Wang, Y. Gao, Resource allocation with edge computing in IoT
(2017) 80–85. networks via machine learning, IEEE Internet Things J. 7 (4) (2020) 3415–3426.
[211] M. Gendreau, J.Y. Potvin, Handbook of Metaheuristics, Int. Series in Operations [236] A.A.A. Gad-Elrab, A.Y. Noaman, Fuzzy clustering-based task allocation approach
Research & Management Science, 2010. using bipartite graph in cloud-fog environment, in: Proceedings of the 16th
[212] K.B. Letaief, W. Chen, Y. Shi, J. Zhang, Y.-J.A. Zhang, The roadmap to 6G: AI EAI International Conference on Mobile and Ubiquitous Systems: Computing,
empowered wireless networks, IEEE Commun. Mag. 57 (8) (2019) 84–90. Networking and Services, MobiQuitous ’19, Association for Computing Machin-
[213] F.A. Salaht, F. Desprez, A. Lebre, C. Prud’Homme, M. Abderrahim, Service ery, New York, NY, USA, 2019, pp. 454–463, https://doi.org/10.1145/3360774.
placement in fog computing using constraint programming, in: 2019 IEEE 3360833.
International Conference on Services Computing, SCC, IEEE, 2019, pp. 19–27. [237] I. Ullah, H.Y. Youn, Task classification and scheduling based on K-means clus-
[214] A.I. Orhean, F. Pop, I. Raicu, New scheduling approach using reinforcement tering for edge computing, Wirel. Pers. Commun. 113 (4) (2020) 2611–2624,
learning for heterogeneous distributed systems, J. Parallel Distrib. Comput. 117 https://doi.org/10.1007/s11277-020-07343-w.
(2018) 292–302. [238] A. Bozorgchenani, D. Tarchi, G.E. Corazza, An energy-aware offloading clus-
[215] Y.-R. Shiue, K.-C. Lee, C.-T. Su, Real-time scheduling for a smart factory using tering approach (eaoca) in fog computing, in: 2017 International Symposium
a reinforcement learning approach, Comput. Ind. Eng. 125 (2018) 604–614. on Wireless Communication Systems, ISWCS, 2017, pp. 390–395 (ISSN:
[216] S. Wang, T. Tuor, T. Salonidis, K.K. Leung, C. Makaya, T. He, K. Chan, Adaptive 2154-0225).
federated learning in resource constrained edge computing systems, IEEE J. Sel. [239] S. Yu, X. Wang, R. Langar, Computation offloading for mobile edge computing:
Areas Commun. 37 (6) (2019) 1205–1221. A deep learning approach, in: 2017 IEEE 28th Annual International Symposium
[217] L.C. Monteiro, F.C. Delicato, L. Pirmez, P.F. Pires, C. Miceli, Dpcas: Data on Personal, Indoor, and Mobile Radio Communications, PIMRC, 2017, pp. 1–6
prediction with cubic adaptive sampling for wireless sensor networks, in: (ISSN: 2166-9589).
International Conference on Green, Pervasive, and Cloud Computing, Springer, [240] D.S. Rani, M. Pounambal, Deep learning based dynamic task offloading in
2017, pp. 353–368. mobile cloudlet environments, Evol. Intell. (ISSN: 1864-5917) (2019) https:
[218] B. Yang, X. Cao, J. Bassey, X. Li, T. Kroecker, L. Qian, Computation offloading //doi.org/10.1007/s12065-019-00284-9.
in multi-access edge computing networks: A multi-task learning approach, in: [241] L. Huang, X. Feng, A. Feng, Y. Huang, L.P. Qian, Distributed deep learning-
ICC 2019 - 2019 IEEE International Conference on Communications, ICC, 2019, based offloading for mobile edge computing networks, Mob. Netw. Appl. (2018)
pp. 1–6 (ISSN: 1938-1883). https://doi.org/10.1007/s11036-018-1177-x.

24
F. Saeik et al. Computer Networks 195 (2021) 108177

[242] X. Wang, X. Wei, L. Wang, A deep learning based energy-efficient computational [265] C. Lecoutre, Constraint Networks: Targeting Simplicity for Techniques and
offloading method in internet of vehicles, China Commun. 16 (3) (2019) 81–91. Algorithms, John Wiley & Sons, 2013.
[243] Z. Ali, L. Jiao, T. Baker, G. Abbas, Z.H. Abbas, S. Khaf, A deep learning ap- [266] F. Rossi, P. Van Beek, T. Walsh, Handbook of Constraint Programming, Elsevier,
proach for energy efficient computational offloading in mobile edge computing, 2006.
IEEE Access 7 (2019) 149623–149633. [267] X. Yin, A. Jindal, V. Sekar, B. Sinopoli, A control-theoretic approach for
[244] X. Zhu, S. Chen, S. Chen, G. Yang, Energy and delay co-aware computation dynamic adaptive video streaming over HTTP, in: Proceedings of the 2015
offloading with deep learning in fog computing networks, in: 2019 IEEE 38th ACM Conference on Special Interest Group on Data Communication, 2015, pp.
International Performance Computing and Communications Conference, IPCCC, 325–338.
2019, pp. 1–6 (ISSN: 2374-9628). [268] D.Y. Zhang, D. Wang, An integrated top-down and bottom-up task allocation
[245] L. Huang, S. Bi, Y.J. Zhang, Deep reinforcement learning for online computation approach in social sensing based edge computing systems, in: IEEE INFOCOM
offloading in wireless powered mobile-edge computing networks, IEEE Trans. 2019-IEEE Conference on Computer Communications, IEEE, 2019, pp. 766–774.
Mob. Comput. (2019) 1. [269] C. Wu, Y. Zhang, Y. Deng, Toward fast and distributed computation migration
[246] R. Zhao, X. Wang, J. Xia, L. Fan, Deep reinforcement learning based mobile system for edge computing in IoT, IEEE Internet Things J. 6 (6) (2019)
edge computing for intelligent Internet of Things, Phys. Commun. (2020) 10041–10052.
101184. [270] Y. Li, Edge Computing-Based Access Network Selection for Heterogeneous
[247] K. Zhang, Y. Zhu, S. Leng, Y. He, S. Maharjan, Y. Zhang, Deep learning Wireless Networks (Ph.D. dissertation), Universite Rennes, 2017.
empowered task offloading for mobile edge computing in urban informatics, [271] A. Koike, N. Higo, Y. Sueda, Proxy-based network function to assist robotic
IEEE Internet Things J. 6 (5) (2019) 7635–7647. feedback control system, in: 2018 IEEE International Symposium on Local and
[248] H. Lu, C. Gu, F. Luo, W. Ding, X. Liu, Optimization of lightweight task Metropolitan Area Networks, LANMAN, IEEE, 2018, pp. 116–118.
offloading strategy for mobile edge computing based on deep reinforcement [272] S. Host, W. Tärneberg, P. Ödling, M. Kihl, M. Savi, M. Tornatore, Network
learning, Future Gener. Comput. Syst. 102 (2020) 847–861, http://www. requirements for latency-critical services in a full cloud deployment, in: 2016
sciencedirect.com/science/article/pii/S0167739X19308209. 24th International Conference on Software, Telecommunications and Computer
[249] J. Baek, G. Kaddoum, Heterogeneous task offloading and resource allocations Networks, SoftCOM, IEEE, 2016, pp. 1–5.
via deep recurrent reinforcement learning in partial observable multi-fog [273] M. Maheswaran, T. Yang, S. Memon, A fog computing framework for au-
networks, IEEE Internet Things J. (2020) 1. tonomous driving assist: architecture, experiments, and challenges, 2019, arXiv
[250] J. Liu, X. Wei, T. Wang, J. Wang, An ant colony optimization fuzzy clustering preprint arXiv:1907.09454.
task scheduling algorithm in mobile edge computing, in: J. Li, Z. Liu, H.
[274] N. Kalatzis, M. Avgeris, D. Dechouniotis, K. Papadakis-Vlachopapadopoulos,
Peng (Eds.), Security and Privacy in New Computing Environments, in: Lecture
I. Roussaki, S. Papavassiliou, Edge computing in IoT ecosystems for UAV-
Notes of the Institute for Computer Sciences, Social Informatics and Telecom-
enabled early fire detection, in: 2018 IEEE International Conference on Smart
munications Engineering, Springer International Publishing, Cham, 2019, pp.
Computing, SMARTCOMP, IEEE, 2018, pp. 106–114.
615–624.
[275] D. Spatharakis, M. Avgeris, N. Athanasopoulos, D. Dechouniotis, S. Papavassil-
[251] M.K. Hussein, M.H. Mousa, Efficient task offloading for IoT-based applica-
iou, A switching offloading mechanism for path planning and localization in
tions in fog computing using ant colony optimization, IEEE Access 8 (2020)
robotic applications, in: 2020 International Conferences on Internet of Things
37191–37201.
(IThings) and IEEE Green Computing and Communications (GreenCom) and
[252] T.G. Rodrigues, K. Suto, H. Nishiyama, N. Kato, A PSO model with VM
IEEE Cyber, Physical and Social Computing (CPSCom) and IEEE Smart Data
migration and transmission power control for low service delay in the
(SmartData) and IEEE Congress on Cybermatics (Cybermatics), IEEE, 2020, pp.
multiple cloudlets ECC scenario, in: 2017 IEEE International Conference on
77–84.
Communications, ICC, IEEE, 2017, pp. 1–6.
[276] M. Avgeris, D. Dechouniotis, N. Athanasopoulos, S. Papavassiliou, Adaptive
[253] L.N.T. Huynh, Q.-V. Pham, X.-Q. Pham, T.D.T. Nguyen, M.D. Hossain, E.-N.
resource allocation for computation offloading: A control-theoretic approach,
Huh, Efficient computation offloading in multi-tier multi-access edge computing
ACM Trans. Internet Technol. (TOIT) 19 (2) (2019) 1–20.
systems: A particle swarm optimization approach, Appl. Sci. 10 (1) (2020) 203,
[277] T. Dlamini, Á.F. Gambín, D. Munaretto, M. Rossi, Online resource management
https://www.mdpi.com/2076-3417/10/1/203.
in energy harvesting BS sites through prediction and soft-scaling of computing
[254] Y. Zhang, Y. Liu, J. Zhou, J. Sun, K. Li, Slow-movement particle swarm
resources, in: 2018 IEEE 29th Annual International Symposium on Personal,
optimization algorithms for scheduling security-critical tasks in resource-limited
Indoor and Mobile Radio Communications, PIMRC, IEEE, 2018, pp. 1820–1826.
mobile edge computing, Future Gener. Comput. Syst. 112 (2020) 148–161,
[278] L. Zhang, S. Wang, R.N. Chang, QCSS: a QoE-aware control plane for adaptive
http://www.sciencedirect.com/science/article/pii/S0167739X19333552.
streaming service over mobile edge computing infrastructures, in: 2018 IEEE
[255] D. Zhang, F. Haider, M. St-Hilaire, C. Makaya, Model and algorithms for the
International Conference on Web Services, ICWS, IEEE, 2018, pp. 139–146.
planning of fog computing networks, IEEE Internet Things J. 6 (2) (2019)
[279] F. Basic, A. Aral, I. Brandic, Fuzzy handoff control in edge offloading, in: 2019
3873–3884.
IEEE International Conference on Fog Computing, ICFC, IEEE, 2019, pp. 87–96.
[256] A.A. Al-habob, O.A. Dobre, A. Garcia Armada, Sequential task scheduling for
mobile edge computing using genetic algorithm, in: 2019 IEEE Globecom [280] P. Skarin, J. Eker, M. Kihl, K. Årzén, Cloud-assisted model predictive control,
Workshops, GC Wkshps, 2019, pp. 1–6. in: 2019 IEEE International Conference on Edge Computing, EDGE, 2019, pp.
[257] V. Yadav, B.V. Natesha, R.M.R. Guddeti, GA-PSO: Service allocation in fog 110–112.
computing environment using hybrid bio-inspired algorithm, in: TENCON 2019 [281] R.E. Kalman, et al., Contributions to the theory of optimal control, Bol. Soc.
- 2019 IEEE Region 10 Conference, TENCON, 2019, pp. 1280–1285 (ISSN: Mat. Mexicana 5 (2) (1960) 102–119.
2159-3450). [282] Z. Zhou, J. Yu, X. Dong, Q. Li, Z. Ren, Robust adaptive attitude control of the
[258] Z. Wang, Z. Zhao, G. Min, X. Huang, Q. Ni, R. Wang, User mobility aware task quad-rotor UAV based on the LQR and NESO technique, in: 2018 IEEE 14th
assignment for mobile edge computing, Future Gener. Comput. Syst. 85 (2018) International Conference on Control and Automation, ICCA, IEEE, 2018, pp.
1–8, http://www.sciencedirect.com/science/article/pii/S0167739X17318587. 745–750.
[259] M.B. Kamal, N. Javaid, S.A.A. Naqvi, H. Butt, T. Saif, M.D. Kamal, Heuristic [283] A. Ullah, J. Li, Y. Shen, A. Hussain, A control theoretical view of cloud
min-conflicts optimizing technique for load balancing on fog computing, in: elasticity: taxonomy, survey and challenges, Cluster Comput. 21 (4) (2018)
Advances in Intelligent Networking and Collaborative Systems, 2019, pp. 1735–1764.
207–219. [284] D. Dechouniotis, N. Leontiou, N. Athanasopoulos, G. Bitsoris, S. Denazis, ACRA:
[260] G. Amarasinghe, M.D. de Assun¸ c ao, A. Harwood, S. Karunasekera, A data A unified admission control and resource allocation framework for virtualized
stream processing optimisation framework for edge computing applications, in: environments, in: 2012 8th International Conference on Network and Service
2018 IEEE 21st International Symposium on Real-Time Distributed Computing, Management (Cnsm) and 2012 Workshop on Systems Virtualiztion Management
ISORC, 2018, pp. 91–98 (ISSN: 2375-5261). (Svm), IEEE, 2012, pp. 145–149.
[261] F. Ait Salaht, F. Desprez, A. Lebre, C. Prud’homme, M. Abderrahim, Service [285] P.S. Saikrishna, R. Pasumarthy, N.P. Bhatt, Identification and multivariable
placement in fog computing using constraint programming, in: 2019 IEEE gain-scheduling control for cloud computing systems, IEEE Trans. Control Syst.
International Conference on Services Computing, SCC, 2019, pp. 19–27 (ISSN: Technol. 25 (3) (2016) 792–807.
2474-2473). [286] D. Dechouniotis, N. Leontiou, N. Athanasopoulos, A. Christakidis, S. Denazis,
[262] S. Vakilinia, D. Qiu, M.M. Ali, Optimal multi-dimensional dynamic resource al- A control-theoretic approach towards joint admission control and resource
location in mobile cloud computing, EURASIP J. Wireless Commun. Networking allocation of cloud computing services, Int. J. Netw. Manag. 25 (3) (2015)
2014 (1) (2014) 201, https://doi.org/10.1186/1687-1499-2014-201. 159–180.
[263] M.C. Calzarossa, L. Massari, D. Tessera, Workload characterization: A survey [287] S. Rashidi, S. Sharifian, Cloudlet dynamic server selection policy for mobile
revisited, ACM Comput. Surv. 48 (3) (2016) 48:1–48:43, https://doi.org/10. task off-loading in mobile cloud computing using soft computing techniques, J.
1145/2856127. Supercomput. 73 (9) (2017) 3796–3820.
[264] I. Bouras, F. Aisopos, J. Violos, G. Kousiouris, A. Psychas, T. Varvarigou, G. [288] N. Leontiou, D. Dechouniotis, S. Denazis, S. Papavassiliou, A hierarchical
Xydas, D. Charilas, Y. Stavroulas, Mapping of quality of service requirements control framework of load balancing and resource allocation of cloud computing
to resource demands for iaas, in: CLOSER, 2019. services, Comput. Electr. Eng. 67 (2018) 235–251.

25
F. Saeik et al. Computer Networks 195 (2021) 108177

[289] D. Dechouniotis, N. Athanasopoulos, A. Leivadeas, N. Mitton, R. Jungers, S. John Violos is research associate in the Dept. of Soft-
Papavassiliou, Edge computing resource allocation for dynamic networks: The ware Engineering and Information Technology at ETS. His
DRUID-net vision and perspective, Sensors 20 (8) (2020) 2191. previous positions were research associate at National Tech-
[290] W. Zhang, M.S. Branicky, S.M. Phillips, Stability of networked control systems, nical University of Athens, sessional lecturer at Harokopio
IEEE Control Syst. Mag. 21 (1) (2001) 84–99. University of Athens and visiting lecturer at National and
[291] J.P. Hespanha, P. Naghshtabrizi, Y. Xu, A survey of recent results in networked Kapodistrian University of Athens. He was a member in
control systems, Proc. IEEE 95 (1) (2007) 138–162. the European Commission’s Digital Single Market working
[292] B. Sinopoli, L. Schenato, M. Franceschetti, K. Poolla, M.I. Jordan, S.S. Sastry, group on the code of conduct for switching and porting
Kalman filtering with intermittent observations, IEEE Trans. Autom. Control 49 data between cloud service providers. His research interests
(9) (2004) 1453–1464. include Deep Learning, Machine Learning, Cloud and Edge
computing.
[293] D. Simon, D. Robert, O. Sename, Robust control/scheduling co-design: applica-
tion to robot control, in: 11th IEEE Real Time and Embedded Technology and
Applications Symposium, IEEE, 2005, pp. 118–127.
[294] W. Heemels, K.H. Johansson, P. Tabuada, An introduction to event-triggered
and self-triggered control, in: 2012 IEEE 51st IEEE Conference on Decision and Aris Leivadeas is currently an Assistant Professor with the
Control, CDC, IEEE, 2012, pp. 3270–3285. Dept. of Software and Information Technology Engineering
at ETS. From 2015 to 2018 he was a postdoc in the Dept.
of SCE, at Carleton University. In parallel, Aris worked as
an intern at Ericsson and then at Cisco in Ottawa, Canada.
Firdose Saeik is currently pursuing a Ph.D. at ETS. Canada. He received his diploma in ECE from University of Patras in
He received the M.Sc. degree from VIT university, India, in 2008, the M.Sc. degree in Engineering from King’s College
2010. He has more than 9 years of work experience with London in 2009, and the Ph.D degree in ECE from NTUA in
multiple roles as System Engineer, Junior Researcher, and 2015. His research interests include Cloud Computing, IoT,
Research Engineer. His research interests are in the field and network optimization and management. He received
of Mobile Edge Computing, Quality of Experience, Next- the best paper award in ICPE’18 and the best presentation
generation mobile services and applications such as Virtual award in HPSR’20.
Reality (VR) and Augmented Reality (AR).

Nikolaos Athanasopoulos is a Lecturer at the School of

Electronics, Electrical Engineering and Computer Science
Marios Avgeris is currently a Ph.D student in the NET- at Queens University Belfast. He received a Diploma and
MODE Lab at the National Technical University of Athens a Ph.D. in Electrical and Computer Engineering from the
(NTUA). He received his Diploma in Electrical & Com- University of Patras, Greece and has held postdoctoral
puter Engineering (ECE) from NTUA, Greece, in 2016. researcher positions at TU/e and UCLouvain. He has been
His research interests are control theory, edge and cloud an IKY and a Marie Curie Fellow. His interests are in control
computing, IoT, semantic web technologies and network theory, focusing on hybrid systems and set-based methods
monitoring. with applications in edge/cloud computing and resource
allocation.

Nathalie Mitton received the M.Sc. and PhD. degrees in

Dimitrios Spatharakis is currently a Ph.D. student in the Computer Science from INSA Lyon in 2003 and 2006
NETMODE Lab at the National Technical University of respectively. She received her Habilitation à diriger des
Athens (NTUA). He received a Diploma in Electrical & recherches (HDR) in 2011 from Université Lille 1. She
Computer Engineering (ECE) from NTUA, Greece, in 2018. is currently an Inria full researcher since 2006 and from
His research interests focus on IoT, cyber–physical systems, 2012, she is the scientific head of the Inria FUN team
edge computing and cloud computing which is focused on small computing devices like elec-
tronic tags and sensor networks. Her research interests
focus on self-organization from PHY to routing for wireless
constrained networks. She has published her research in
more than 30 international revues and more than 100
international conferences. She is involved in the setup
of the FIT IoT LAB platform (http://fit-equipex.fr/, https:
Nina Santi is a Ph.D. student under the supervision of //www.iot-lab.info), the H2020 CyberSANE and VESSE-
Nathalie Mitton in the Inria FUN team. Their focus is on DIA projects and in several program and organization
small computing devices like electronic tags and sensor committees such as Infocom 2021&2020&2019, PerCom
networks. She has received the M.Sc. degrees in Computer 2020&2019, DCOSS 2021&2020&2019, Adhocnow (since
Science from University of Lille, France, in 2020. 2015), ICC (since 2015), Globecom (since 2017), VTC (since
2016), etc. She also supervises several Ph.D. students and
engineers.

Symeon Papavassiliou is currently a professor in the

School of Electrical and Computer Engineering at the Na-
tional Technical University of Athens (NTUA). From 1995
Dimitrios Dechouniotis is currently research associate with to 1999, he was a senior technical staff member at AT&T
NETMODE Lab of the National Technical University of Laboratories, New Jersey. In August 1999 he joined the ECE
Athens (NTUA). From 2007 to 2016, he was non-tenured Dept at the New Jersey Institute of Technology, USA, where
Lecturer at the EE Dept. of Technical Educational Institute he was an associate professor until 2004. He has an estab-
of Western Greece, Greece. He received his diploma in ECE lished record of publications in his field of expertise, with
from University of Patras in 2004, the M.Sc. degree in more than 350 technical journal and conference published
Control Systems and Robotics from NTUA in 2009, and the papers, while he has received several scientific awards and
Ph.D. degree in ECE from University of Patras in 2014. distinctions. His main research interests lie in the areas
His research interests lie in the area of cloud computing, of optimization and performance evaluation of mobile and
Internet of Things, mobile cloud computing and control distributed systems, wireless networks and complex systems.
theory.

26