1

SURVEY PAPER
Energy Consumption for Edge Computing in Internet of Things

Sawsan M. Mahmoud, Tawfeeq E. Abdoulabbas, Hanan A. Al-Jubouri†, Non membership type

ABSTRUCT Edge Computing has been recognized as a helpful battery life. In this paper, we will present a study on MEC
technology to improve computing performance and reduce power focusing on energy consumption, energy management, and
consumption in networks generally and in Internet of Things networks
as well. With the growth of using the Internet of Things in different areas
resource allocation. Figure 1 shows the organization of the
of applications, it is necessary to find the most suitable and optimal way study.
for using the Internet of Things devices within the required application.
Most Internet of Things devices depend on batteries or energy harvesters.
This led the hardware engineers to investigate technologies and methods
that provide low power and to find the required solutions for intelligent
energy management. This work aims to help the designers of edge
computing networks in developing applications of the Internet of Things
that use energy efficiently. In this article, recent methods of the energy
consumption management will be inspected that can be implemented in
edge computing based on the applications for Internet of Things devices.
Also, the article highlights some factors that affect the design of such
systems and can lead to higher power consumption. Furthermore, the
article will explore some mechanisms that can decrease the energy
consumption of the running tasks and may maximize the demand of IoT
devices. A recommendation is to put the energy management and
efficiency into account in the working area of cloud and edge data
centers.
Keywords: Energy consumption, Mobile Edge computing, Mobile
Cloud Computing, Energy Management, Resources Allocation.

1. Introduction
Fig.1 Organization of the study.
The main purpose of Edge Computing (EC) is to make
services of the Internet close to the users by putting the The attention of this survey is to try identifying research
infrastructures of IT near the network edge to behave like potential direction studies and exploit the proposed
tiny data centers [1]. EC infrastructure mostly integrated methodologies used, with addressing the challenges. In
with Internet of Things (IoT) to implement tasks. In IoT addition, finding the suitable solutions for each of these
architecture, most devices have limited power (battery directions as will be noted later. The fundamental
powered). All applications that may run on the EC will contributions of this paper are:
execute processes locally before connecting to the cloud. • Presenting a comprehensive study on several power
Thus, network problems like security and privacy issues management techniques suited to be implemented with
will be reduced. A client ought to transfer (upload) the data different IoT applications and the right choice in
by means of a channel that is shared by all the transferring selecting a power management technique.
clients. Interim, the nature of the channel is different for • Listing and highlighting some reasons that cause high
every client, which will limit the client's ability to power consumption
minimize the use of energy [1].
• Presenting some approaches and solutions that may
Mobile Cloud Computing (MCC) has presented for
reduce the consumption energy demanded for IoT.
applications that can perform their computation on the
The edge server is installed in between the mobile client
servers of the cloud. MCC is a centralized model that is
and the server near mobile proximity. The edge systems
used for mobile applications. With Mobile Edge
contain inadequate number of resources compared with the
Computing (MEC), the mobile devices suffer from some
data centers in the cloud. This will put users in a
constraints such as storage, computation power, and
competition in order to get their resources within their
proximity [2] [3] [4]. Most research on MEC has
concentrated on finding strategies for offloading in MEC
so that the energy consumption, average load and total cost
are minimized. There are some challenges and limitations
† The author is with the Computer Engineering Department,
College of Engineering, Mustansiriyah University such as providing limited power, limited storage, memory
constrain, and high energy consumption. In this study, we

Copyright © 20XX The Institute of Electronics, Information and Communication Engineers

will present comprehensive strategies on the energy Sensor Networks) and wearable devices that have become
consumption management. key components of numerous IoT frameworks. Similar to
The processing of computation task in MEC and IoT devices, the main problem with these wearables
transmitting application over IoT need to consume devices is limited resources and battery-operated
considerable energy, so the problem of reducing the energy dependence. A smart object consists of three units (power-
consumption must be studied for the two sides. consuming units) namely: MCU (Micro-Controller Unit),
sensors unit, and transceivers unit. Data transmission with
2. Energy Consumption IoT needs more energy than in the other units. The energy
consumption growth in IoT is presented in the work [11].
Different techniques were applied with energy Also, a report in [12] demonstrates the raising of the
consumption reduction topic, like Dynamic Voltage- energy consumption in the IoT and its services. The
Frequency Scaling (DVFS) or Dynamic Power authors of the report attempted to estimate the annual
Management (DPM) in [5]. With respect to DPM, the energy consumption for off-track powered IoT equipment
components which are currently idle are turned off, so the and the related gateways within the years (2015-2025) for
energy consumption can be reduced. Then the resources IoT applications such as power meters and smart light. In
will enter the hibernate mode. It only operates when the cloud-based mobile, when the resources are increased, the
idle time becomes rather long. On the other hand, DVFS response time becomes longer because the number of hubs
has been shown as a very promising model and it has the is increased. As a result, the execution time is increased
capabilities in saving energy [6]. DVFS worked under the and the mobile device battery level is deteriorated [13].
fact that energy saving in circuits like CMOS which is For this aspect one of the evaluation measurement factors
dependent on frequency (f) and voltage (V2). During task is Energy Efficiency (EE). In the next sections, the EE
execution, switching between the voltages/ frequencies can management, and resource allocation in MEC will be
lead to high reduction in power consumption and better explained.
performance is guaranteed at the same time [7].
Performance can also be enhanced when the numbers of 3. Energy Management
processors that are utilized by benefit from the idle
processor time (between the running tasks) in parallel so Energy management can be considered as a group of
that the utilized resources are increased. By using this procedures (standards) that control (switching) multiple
method, users may face little delay in the execution to sources of energy to reduce the energy consumption in the
obtain minimum energy consumption of the system. network, to ensure no deficient in energy at network node,
Therefore, combining DVFS approach and reducing the and no perpetually operational in the network. So that the
activity of the inefficient processors is a good method for application demands will be covered with the available
reducing the energy wasting and an efficient performance energy source, and the attained results is to improve
would be attained [7]. efficiency so that utilizing the optimum needed energy
An enhanced energy-efficient scheduling algorithm was sources with the required operation which reduce the
proposed in [8] to reduce power consumption of tasks. It is overall system cost. Heavy computation tasks and limited
primarily utilized to limit the recurrence non serious tasks energy in IoT applications are problems. Hence, different
in systems with heterogeneous distributed computing for approaches and schemes were proposed. For example,
parallel applications. In that work, the energy consumption taxonomy of energy management for devices like WSNs
can be decreased by minimizing the number of unaffected was introduced in a survey [14]. Also, there are
processors. Also, the algorithm proposed in [9] called investigations for energy consumption supplying schemes
DVFS empowered productive energy to make task process based on either batteries or with harvesting energy, with
work efficiently to attain the energy saving for parallel considering that the most disconnection energy
applications. With tracking the necessary workflow, only transference happened with the wireless a sensor nodes
effective tasks are chosen for processing in order for the which lead the necessary backup batteries.
energy to be reduced. This algorithm meets the service Many schemes on the basis of energy consumption are
level agreement which is based on its performance. developed for efficient energy management. Examples of
Usually, the runtime of a mobile device depends on its such schemes are: duty cycling approach, data loading and
battery limits, the quantity of tasks that are accomplished transfer approaches. In order to reach to an efficient energy
per cycle, and the size of the traded data. Also, when utilization in duty cycling approach, nodes are changed
transferring the data to the data centers, which consumes a between active and inactive (sleep) modes [14] . Also,
lot of energy. In MEC, the power load is allocated between there are low power scheme techniques such as deep sleep
the nodes of the system; the requested power that is state, where the MCU is inter with this state when it needs
transferred can be reduced by compressing the data. In to perform a task. An inner clock may be used to trigger
general, with the MEC paradigm, at first, only part of the the wake-up signal (to be look like a deep sleep of a
computation can be transferred towards the data collection known length of time) or – an external interrupt (wake-up
sources, and then these data are sent to the server attached event). In addition, there are other approaches such as
at data centers [10]. clock gating which pause the clock for the unneeded
WSN (Wireless Sensor Networks), WBSN (Wireless Body subsystems the current running task. When disabling the
IEICE TRANS. ELECTRON., VOL.XX-X, NO.X XXXX XXXX
3

clock, no transition happens and subsequently the dynamic offloading via different wireless channels such as WiFi,
power is in a relation with the internal activities of Bluetooth, or Direct WiFi. A Multi-criteria Offloading
combinational circuit will be decreased that is because no Decision method is used by this platform that depends on
need to deliver energy for these subsystems [14]. Recently, the following parameters (energy consumption, execution
these computation and energy issues have been solved by time reduction, resource availability, network conditions,
using MEC and wireless power transfer approaches [15]. and user preferences). The energy is conserved by pre-
Power gating is also an approach used to reduce power figuring accessible resources, using Cost Model Processing
consumption by disconnecting the supply voltage from the and Estimation Social Relations. Precog. (which is
circuit, therefore, in the current CMOS circuit technology; expanding cache) are used in [20] to reduce the response
there is almost no leakage and no dynamic power latency of the amount of data needed with pattern
consumption [10]. A Body Basis Multiplexer BBMUX recognition packages that is offload the related subscribed
denominated implements a dynamic power and is used to computing (edge / cloud) so any pattern distinction will
allow choosing the back basis for each cluster as different be inspected and perfected previously. Tasks offloading is
workloads need different computational power. In order to one of the ways of energy consumption reduction for some
switch from ordinary mode to super mode, it is possible to applications or services. Also the energy consumption is
allow switching from one reverse bias value to another reduced by offloading (more favorable way) over wireless
which increases the speed handling capability. The Power communication even when data, and computational
Management Unit (PMU) manages these transitions and is demand are increased [21].
transparent to the end user. The main sources of power
consumption are: sensor scanning, reading/writing from/to 4.2 Resource Management and Cooperated devices
memory, sending/receiving information.
Energy management can be combined with resource
4. Energy Management Methods sharing and with device cooperation for the purpose of
saving energy by tasks segregation (or offloading) at edge
The energy management platform is used by EC for networks. A practicable sharing can be used with remote
different applications and field areas. Mainly the platform resources even when they are busy or not available.
is preferred to characterize the following [16]: Already edge tasks require significant energy i.e. there are
• Open source programming. constraints violating demand. For example, in [22] PCloud
• scalability and adaptability can use resources from nearby devices so it can reduce
• Help users to implement the functions of energy both task execution time and energy consumption. It is like
management with the suitable monitoring service that providing a special own cloud servicers to edge-nodes
can decrease the cost and involved time. when the cloud resources cannot be reached or accessed.
The following are the major approaches that are The process has been performed in a privacy-preserving
implemented to achieve high EE in (edge/ fog) computing manner. PCloud idea is also deduced offloading mobile
systems:- applications as proposed in [23].
A memory replication architecture (REPLISON) is
4.1 Computation Offloading proposed in [24] to achieve computation offloading at the
network edge where IoT devices are increased.
The offload computation is utilized to make little demand REPLIISON improves the performance of memory
of energy with limited energy sources whether it can be replication using an optimized protocol named Long Term
implemented on any IoT device, in the cloud, or edge Evolution (LTE). In this protocol, Device-To-Device
servers. For example, in [17], Drop Computing is used to (D2D) connection has been implemented to support
access resources that are demanded with executing the massive simultaneous access of IoT devices (memory
proximity-based network. Most contents are close to the revoking will be swapped with them). The energy
user, and even with the gathered user requests, the consumption is increased with the increase of active
computation stay achieved efficiently. Also, the content devices (compared with a separated device replicate
offloaded from the cloud to be at the buffer attached to the transmission), which is the main problem of memory
edge of the cloud or any IoT equipment and transfer to replication architecture. It has been shown that when
serve others i.e. it still available when demanded. Besides, reducing the replicate transferring size, the REPLISON has
the applications are divided into jobs, where it can be better energy consumed per device than the conventional
processed either locally at the network edge or postponed LTE. With further evaluation results.
“offloaded” to be closer to their demand data or their Techniques that attempt to minimize latency and at the
resources devices. In this way the performance will be same time maintain constant energy consumption were
improved, that because the reliability of the system will applied [25] [26. For EE, researchers make an effort to get
increase, energy consumption will decrease, as well as the little consumption of power in the networks and especially
cached context of devices and networks resources when on network nodes (hop). For example, in [25], a physical
captured by sensors [18]. interference model with a data aggregation algorithm has
A platform named mCloud is used in [19] that consider been proposed for a multi-hop WSN. In that work, the
 IEICE TRANS. ELECTRON., VOL.XX-X, NO.X XXXX XXXX
4

algorithm reduces latency while keeping constant energy support the AI task operation, the authors, suggested
consumption. So, the network utility maximization (NUM) providing four layers computing systems in an intelligent
can be optimized with dynamic routing rechargeable industrial environment. Then, an algorithm with two
sensor networks RSNs, it has been joined between the flow phases has designed to schedule edge layer resources
rate (which is depend on the link capacity and sampling taking the latency restriction into account. Using these
rate) and the constraints of the energy consumption rate strategies, it has been demonstrated an excellent real-time,
should be neither too large such that the source depletes good performance of computing services and energy
the battery and stops working (aggressive case), nor too consumption in these environments with EC. Also, a VM
small such that the battery level reaches maximum and (Virtual Machine) placement/scheduling can improve the
misses recharging opportunities (conservative case). EE if an appropriate scheduling is used to provide the
Another work in [26] proposes a distributed algorithm that needed power resource freely with the system
optimizes the (NUM) in dynamic-routing RSNs and infrastructure. Finally, to improve the energy consumption,
considers the flow and energy as constraints. The flow an application (in terms of VMs) as multi-components with
constraint:- help of graph theory [16] was introduced.
Neurosurgeon [29] presents a technique that make
(Incoming flow rate + sampling rate) ≤ (Outgoing flow scheduling to divide the computation of Deep Neural
rate) ……(1) Network (DNN) between attached devices (like mobile)
and data centers in the granular layer of the neural
(The flow rate) ≤ (the link capacity) ……. (2) network. With performing impactful processing of
resources in the cloud edge; which reduce the energy
The energy consumption rate must not be high that cause consumption, high traffic throughput and low computing
the source to depress the battery and shut down (aggressive latency. Latency is definitely not an independent metric in
case), and also not too low and lose battery charge EC; rather, it is related to consumption of energy, as more
(conservative case). The proposed algorithm can optimize effective network nodes that in turn, could serve other
(energy management, rate control and routing) and can users and interaction client traffic quicker. [16]. In [30]
reach the globally optimal solution. hierarchical structure is depended as well as primitive
decision model to solve computation offloading problem.
5. Task scheduling This is developed as an advanced decision model. The
decision model can operate on the intrinsic hierarchy, and
In current data centers, there is an enormous growth in an online scheduling algorithm has been done for the
energy consumption, and the price of consumption has incoming requests. In traditional solutions, it is assumed
been very high. It becomes a big problem for the data that any requests can be expected by the edge servers.
center supercomputers and public cloud systems. For Unfortunately, in reality, congestion happens unexpectedly.
example, the authors in [27] show that Data Center (DC) Second types of solutions try to find suitable methods to
may need about 100,000,000 kWh in a year when it has manage and schedule all requests. The scheduling process
50,000 computing nodes. This equalizes consumption turns out to be a separated job, and some active modelling
around one hander thousand residential yearly. This becomes expected. To avoid congestion, the solution was
explosion has led to a focus on developing models such as to take into account the case of the related-hope server, so,
green computing and finding a way to schedule tasks to the initial decision is taken by the sender [30].
minimize the consumption of energy.
5.2 Efficient Energy Task Scheduling Algorithm
5. 1 Resource Task scheduling
In the cloud environment, efficient task scheduling tries to
The actual CPU utilizing for the required servers is not get a minimal completion time and also should increase the
high if it was compared with energy consumption for the utilization of system resources and moreover minimize the
exact task needed a management is necessary for that consumption of energy [27]. Software program named
purposes. In [27], it has been shown that over six month’s (directed cyclic graph scheduling) is used for solving most
period, the average CPU utilization was between 10-50 of the static scheduling problems as the work in [31]. In
percent for more than 5,000 servers of their maximum this case, the application tasks have represented by nodes,
utilization. Research also showed that for most data and the inter-task data dependencies have represented as
centers, the ratio of energy used for data center and IT load edges. ECR (Energy Consumption Ratio) is the main
consumption, named Power Usage Effectiveness (PUE) indication of the calculation, and it denotes the total energy
are inadequate. For parallel applications, there is a criterion consumed by the processor task execution on the critical
to assess mechanisms, which is focused on reducing the path. The execution measurements and performance are
schedule length rather than encouraging the use of green determined with four factors: ECR, available used
computing system [27]. resource, average running time, and the rate of saving
In [28], scheduling method is done by applying resource energy. Also, an Edge Node Resource Management
management with intelligent treatments designed to fulfill (ENORM) framework is proposed and developed in [32].
the real-time industrial system supported by EC. To The authors introduced another design and deployed a
IEICE TRANS. ELECTRON., VOL.XX-X, NO.X XXXX XXXX
5

system to connect the data center in the cloud and edge [37].
nodes. For edge nodes, ENORM provides offloaded • Idle Power Consumption: Sometimes the IoT service is
workloads. Low upward and dynamic auto-expansion not used for some time i.e. it is inactive, the server can
instruments have incorporated to make resource managing be used for other jobs in the data centres of the cloud.
(adding or eliminating) by the ENORM, more effective on This means the server is being idle and the server’s
edge nodes. For asynchronous MEC offloading systems, a power consumption can be reduced. If servers in the
resource management mechanism is proposed to save the Edge or Fog data centres are idle for a long time, the IoT
energy in [33]. The general arrival data series is analyzed service will not be energy efficient compared to that in
to determine the suited data partitioning way and time- the cloud.
sharing approach. Then, using the block coordinate descent • Application Type: There are some parameters that affect
method, the total energy consumption of the mobile is the energy consumption in some applications as stated in
reduced. [38]. These are: how many downloads and uploads are
In [34], a full-duplex virtual framework for small mobile made and how much data is preloaded. In the cloud,
networks based on heterogeneous service caching in EC there are some IoT applications that are considered as
was developed. The authors developed a typical scheme energy efficient utilization compared with that in the
for resource allocation where full duplex communication Edge / Fog where huge amount of updates and power
and caching method are considered. The scheme also takes are needed.
into consideration power control, user correlation, and • Network Management and Infrastructure type:
caching, offloading computation besides resource Configuring and managing EC infrastructures to provide
allocation. In [35] a multi-user MEC system has been services to IoT devices is not an easy task. For example,
studied based on two schemes: TDMA -OFDMA (Time in [39] a model has proposed a scenario called optimal
Division Multiple Access) - (Orthogonal Frequency placement scenario where VM placement is allowed at
Division Multiple Access). In that work, some cases were relays. It was shown that the total IoT devices’ power
considered limited or unlimited CC. For TDMA scheme consumption can be considerably reduced when the
MEC offloading systems, the authors proposed a resource distribution of mini clouds is optimal.
allocation approach depending on redefining the priority of
the offloading function and updating the policy of the 7. Conclusions and Recommendations
previous threshold for limited CC. After that, based on the
approximated offloading priority, an algorithm is proposed This paper has presented a comprehensive literature review
for sub-optimal resource allocation (more details are of research on different energy management techniques.
shown in [35]). For OFDMA scheme MEC offloading We can conclude that in each management technique there
systems, a resource allocation problem is done as an are treatments and evaluations on the modeling for better
integer optimization problem, and to optimize resource reliability not only for one system but also with other types
allocation priorities, TDMA method is used with unlimited of systems or other parts of systems. We also conclude that
CC capabilities. The resource allocation problem is also any technique that leads to reduction in power
addressed in [36] to allocate IoT needs in many wireless consumption will enhance the performance for example
MEC systems having only one access point used. In that minimizing the number of the inefficient processors is a
study, a dynamic game model called Stackelberg is good way in reducing the energy wasting and an efficient
proposed for the nodes in the IoT devices in order to find performance would be attained. Also compressing the data
an optimal resource allocation. The fundamental objective in MEC is a good choice to reduce the requested power
is to find the best value of cost function while transferring that can be transferred.
energy with managing device offloading in relation to the Power gating is also a convenient approach used to
access point. minimize power consumption by disconnecting the supply
voltage from the circuit, Therefore, in the current CMOS
6. Main Factors for Energy Consumption circuit technology, there is very little leakage and no
dynamic power consumption. [10].
It is concluded that there are important factors that have a Below are some recommended issues which can be drawn
significant impact on the energy consumption that should from previous works:
be taken into account while designing computer networks, • The energy consumption due to different part in the
some of them are listed below: networks like IoT, transmission media and energy
• Access Network: for energy consumption, it is important supply should be consider during system design .
to well consider the choice of the right technology to • Further investigation on offloading in EC to increase
access the network. The average of service bits in uplink energy saving is recommended while taking some
and downlink traffic, and how many users and services challenges into consideration like program partitions
and their data can be used to pass through the network and task scheduling, sophisticated resource
equipment in a specific time. It can be concluded that managements to attain minimal energy consumption in
IoT services can consume a small amount of energy the same time.
when the EE mechanism is based on a good choice of • We also advocate considering the possibility of using
access network and when sleep mode is implemented
 IEICE TRANS. ELECTRON., VOL.XX-X, NO.X XXXX XXXX
6

more than one energy source (or hybrid techniques) 781–786., 2012.
comprising the ability of providing the new energy
[10] M. Capra, R. Peloso, G. Masera, M. Ruo Roch, and M. Martina,
sources, especially for the small sensors or edge devices,
instead of using another supplying way to prevent "Edge computing: A survey on the hardware requirements in the
depending only on the signal source so the network internet of things world,," Futur. Internet, vol. 11, no. 4, p. 100,
lifetime will increase. 2019.
This review can be considered a valuable guideline to the [11] K.Mohammed S. Huq, Sh.Mumtaz, J. Bachmatiuk,. J.
future academic research community and industrial works.
Rodriguez, "Green HetNet CoMP: Energy efficiency analysis
Acknowledgments and optimization," IEEE Trans. Veh. Technol, vol. 64, no. 10, p.
pp. 4670–4683, 2014.
We would like to express gratitude toward Mustansiriyah [12] F. Jalali, S. Khodadustan, C. Gray, K. Hinton, and F. Suits,
University in Baghdad/Iraq www.uomustansiriyah.edu.iq
"Greening iot with fog: A survey," 2017.
for support of this work.
[13] Keqin Li, Fellow, "Computation offloading strategy
References optimization with multiple heterogeneous servers in mobile edge
computing," 2019.
[1] L. Lin, X. Liao, H. Jin, and P. Li, "Computation offloading
[14] J. A. Khan, H. K. Qureshi, and A. Iqbal,, "Energy management
toward edge computing," Proc. IEEE, vol. 107, no. 8, p. 1584–
in wireless sensor networks: A survey," Comput. Electr. Eng,
1607, 2019.
vol. 41, p. 159–176, 2015.
[2] S. Wang, Y. Zhao, L. Huang, J. Xu, and C.-H. Hsu, "QoS
[15] B. Liu, H. Xu, and X. Zhou,, "Resource allocation in wireless-
prediction for service recommendations in mobile edge
powered mobile edge computing systems for internet of things
computing," J. Parallel Distrib. Comput., vol. 127, p. 134–144,
applications," Electronics 2019, 8, 206;
2019.
doi:10.3390/electronics8020206
[3] H. Badri, D. Grosu, T. Bahreini, and K. Yang, "Poster: Risk-
www.mdpi.com/journal/electronics, vol. 8, no. 2, p. 206, 2019.
based Optimization of Resource Provisioning in Mobile Edge
[16] Y. Zhao, W. Wang, Y. Li, C. C. Meixner, M. Tornatore, and J.
Computing.," in 2018 IEEE/ACM Symposium on Edge
Zhang, "Edge computing and networking: A survey on
Computing (SEC), Seattle, WA, USA, 2018.
infrastructures and applications," IEEE Access,, vol. 7, p.
[4] Grosu, T. Bahreini and D., "Efficient placement of multi-
101213–101230, 2019.
component applications in edge computing systems,," in
[17] A. Rudenko, P. Reiher, G. J. Popek, and G. H. Kuenning,
Proceedings of the Second ACM/IEEE Symposium on Edge
"Saving portable computer battery power through remote
Computing, 2017.
process execution," ACM SIGMOBILE Mob. Comput.
[5] Braun T., Siegel, H., Beck, N., Boloni, L., Maheswaran, M.,
Commun. Rev., vol. 2, no. 1, p. 19–26, 1998.
Reuther, A., Robertson, J., et al, "A comparison of eleven static
[18] R.-I. Ciobanu, C. Negru, F. Pop, C. Dobre, C. X.
heuristics for mapping a class of independent tasks onto
Mavromoustakis, and G. Mastorakis, "Drop computing: Ad-hoc
heterogeneous distributed computing systems," J. Parallel
dynamic collaborative computing," Futur. Gener. Comput. Syst,
Distrib. Comput., vol. 61, no. 6, p. 810–837, 2001.
vol. 92, p. 889–899, 2019.
[6] D. Zhu, R. Melhem, and B. R. Childers, "Scheduling with
[19] Cuervo, E., "MAUI: making smartphones last longer with code
dynamic voltage/speed adjustment using slack reclamation in
offload," in Proceedings of the 8th international conference on
multiprocessor real-time systems," IEEE Trans. parallel Distrib.
Mobile systems, applications, and services, p. 49–62, 2010.
Syst., vol. 14, no. 7, p. 686–700, 2003.
[20] Barbera, M.V., Kosta, S., Mei, A., Stefa, J, "To offload or not to
[7] Y. Zhang, X. Hu, and D. Z. Chen, "Task scheduling and voltage
offload? the bandwidth and energy costs of mobile cloud
selection for energy minimization," 183–188, 2002.
computing," 2013.
[8] Z. Ning, J. Huang, X. Wang, J. J. P. C. Rodrigues, and L. Guo,
[21] B.-G. Chun, S. Ihm, P. Maniatis, M. Naik, and A. Patti,
"Mobile edge computing-enabled Internet of vehicles: Toward
"Clonecloud: elastic execution between mobile device and
energy-efficient scheduling," IEEE Netw, vol. 33, no. 5, p. 198–
cloud," 2011.
205, 2019.
[22] M. Jang, K. Schwan, K. Bhardwaj, A. Gavrilovska, and A.
[9] Q. Huang, S. Su, J. Li, P. Xu, K. Shuang, and X. Huang,
Avasthi, "Personal clouds: Sharing and integrating networked
"Enhanced energy-efficient scheduling for parallel applications
resources to enhance end user experiences," IEEE INFOCOM
in cloud," 2012 12th IEEE/ACM International Symposium on
2014-IEEE Conference on Computer Communications, p. 2220–
Cluster, Cloud and Grid Computing (ccgrid 2012), vol. 12, p.
IEICE TRANS. ELECTRON., VOL.XX-X, NO.X XXXX XXXX
7

2228, 2014. 10th International Conference on Ubiquitous Intelligence and

[24] S. Abdelwahab, B. Hamdaoui, M. Guizani, and T. Znati, [36] F. Fang, Y. Xu, Z. Ding, C. Shen, M. Peng, and G. K.
"Replisom: Disciplined tiny memory replication for massive iot Karagiannidis, "Optimal task assignment and power allocation
devices in lte edge cloud," IEEE Internet Things, vol. 3, no. 3, p. for NOMA mobile-edge computing networks," in 2019 IEEE
327–338, 2015. Globecom Workshops (GC Wkshps), Waikoloa, HI, USA, 2019.

[25] H. Li, C. Wu, D. Yu, Q.-S. Hua, and F. C. M. Lau,, [37] C. Gray, R. Ayre, K. Hinton, and R. S. Tucker, "Power
"Aggregation latency-energy tradeoff in wireless sensor consumption of IoT access network technologies," in IEEE
networks with successive interference cancellation," IEEE International Conference on Communication Workshop
Trans. Parallel Distrib. Syst., vol. 24, no. 11, p. 2160–2170, (ICCW), 2015, pp. 2818–2823., 2015.
2012. [38] F. Jalali, A. Vishwanath, J. De Hoog, and F. Suits,
[26] R. Deng, H. Liang, J. Yong, B. Chai, and T. Yang, "Distributed "Interconnecting Fog computing and microgrids for greening
rate control, routing, and energy management in dynamic IoT," IEEE Innovative Smart Grid Technologies-Asia (ISGT-
rechargeable sensor networks," Peer-to-Peer Netw. Appl., vol. Asia),, p. 693–698, 2016.
10, no. 3, p. 425–439, 2017. [39] Al-Azez, Z. T., A. Q. Lawey, and J. M. El-Gorashi, T. E.,
[27] J. Song, T.-T. Li, Z.-X. Yan, J. Na, and Z.-L. Zhu, "Energy- Elmirghani, "Virtualization framework for energy efficient IoT
efficiency model and measuring approach for cloud computing," networks," in IEEE 4th International Conference on Cloud
Ruanjian Xuebao/Journal Softw, vol. 23, no. 2, p. 200–214, Networking (CloudNet), 2015, pp. 74–77., 2015.
2012.
[28] X. Li- J. Wan, H.-N. Dai, M. Imran, M. Xia, and A. Celesti, "A
hybrid computing solution and resource scheduling strategy for
edge computing in smart manufacturing," IEEE Trans. Ind.
Informatics, vol. 15, no. 7, p. 4225–4234, 2019.
[29] Y. Kang et al, "Neurosurgeon: Collaborative intelligence
between the cloud and mobile edge," ACM SIGARCH Comput.
Archit. News, vol. 45, no. 1, p. 615–629, 2017.
[30] X. Chen, L. Jiao, W. Li, and X. Fu, "Efficient multi-user
computation offloading for mobile-edge cloud computing,"
IEEE/ACM Trans. Netw, vol. 24, no. 5, p. 2795–2808, 2018.
[31] Hölzle, L. A. Barroso and U., "The case for energy-proportional
computing," Computer (Long. Beach. Calif), vol. 40, no. 12, p.
33–37, 2007.
[32] N. Wang, B. Varghese, M. Matthaiou, and D. S. Nikolopoulos,
"ENORM: A framework for edge node resource management,"
IEEE TRANSACTIONS ON SERVICES COMPUTING,DOI
10.1109/TSC.2017.2753775, 2017.
[33] J. Niu, W. Song, and M. Atiquzzaman, "Bandwidth-adaptive
partitioning for distributed execution optimization of mobile
applications," J. Netw. Comput. Appl., vol. 37, p. 334–347,
2014.
[34] Z. Tan, F. R. Yu, X. Li, H. Ji, and V. C. M. Leung,, "Virtual
Resource Allocation for Heterogeneous Services in Full Duplex-
Enabled SCNs With Mobile Edge Computing and Caching,"
IEEE Transactions on Vehicular Technology, doi:
10.1109/TVT.2017.27640, vol. 67, no. 2, p. 1794–1808, 2018.
[35] Drolia, et al U., "The case for mobile edge-clouds," in IEEE