SPECIAL SECTION ON GREEN COMMUNICATIONS AND NETWORKING FOR 5G WIRELESS

Received May 31, 2016, accepted June 16, 2016, date of publication August 26, 2016, date of current version October 6, 2016.
Digital Object Identifier 10.1109/ACCESS.2016.2597169

Energy-Efficient Offloading for Mobile Edge

Computing in 5G Heterogeneous Networks
KE ZHANG1 , YUMING MAO1 , SUPENG LENG1 , (Member, IEEE),
QUANXIN ZHAO1 , (Student Member, IEEE), LONGJIANG LI1 , (Member, IEEE),
XIN PENG2 , LI PAN2 , SABITA MAHARJAN3 , (Member, IEEE),
AND YAN ZHANG3 , (Senior Member, IEEE)
1 School of Communication and Information Engineering, University of Electronic Science and Technology of China, Chengdu 611731, China
2 College of Information and Communication Engineering, Hunan Institute of Science and Technology, Yueyang 414006, China
3 Simula Research Laboratory, Oslo 1364, Norway
Corresponding author: Y. Zhang ([email protected])
This work was supported in part by the National Natural Science Foundation of China under Grant 61273235,
and in part by the Research Council of Norway under Project 240079/F20.

ABSTRACT Mobile edge computing (MEC) is a promising paradigm to provide cloud-computing

capabilities in close proximity to mobile devices in fifth-generation (5G) networks. In this paper, we study
energy-efficient computation offloading (EECO) mechanisms for MEC in 5G heterogeneous networks.
We formulate an optimization problem to minimize the energy consumption of the offloading system, where
the energy cost of both task computing and file transmission are taken into consideration. Incorporating
the multi-access characteristics of the 5G heterogeneous network, we then design an EECO scheme, which
jointly optimizes offloading and radio resource allocation to obtain the minimal energy consumption under
the latency constraints. Numerical results demonstrate energy efficiency improvement of our proposed
EECO scheme.

INDEX TERMS Energy-efficiency, offloading, mobile edge computing, 5G.

I. INTRODUCTION devices to the cloud servers may cause delay fluctuation and
As smart mobile devices have seen advanced technology invoke extra transmission energy cost [7]. Thus, the compu-
and design, they facilitate us with a pervasive and powerful tation offloading efficiency can severely degrade.
platform to realize many novel mobile applications [1], [2]. Mobile Edge Computing (MEC) is envisioned as a promis-
Mobile applications, such as the interactive gaming, vir- ing approach to improve the offloading efficiency. In the
tual reality and natural language processing, typically MEC framework, cloud computing capabilities are provided
require intensive computation and result in high energy within the radio access network in close proximity to these
consumption [3]–[5]. However smart mobile devices have mobile devices [8]. In other words, with the aid of MEC,
limited computation capabilities and battery power. This con- mobile devices are enabled to offload their tasks to the
flict between the resource hungry applications and the limited MEC servers on the edge of the network, rather than utilizing
capability of the smart mobile devices brings in unprece- the servers in the core network. This MEC paradigm can
dented challenges to implement the novel mobile applications provide low latency, high bandwidth and computing agility
in an energy efficient manner. in the computation offloading process.
A new architecture and technology known as Mobile Cloud With the ever-growing energy consumption for infor-
Computing (MCC) has the potential to address the aforemen- mation and communication technology, the communication
tioned challenges. By migrating computational tasks from devices and infrastructure play an important role in global
the mobile devices to the infrastructure-based cloud servers, greenhouse gas emissions [9]. Therefore, the development of
MCC can improve the performance of mobile applications green 5G networks has become an important topic for the
and reduce the energy consumption of mobile devices [6]. design and implementation of future wireless communica-
However, the infrastructure-based cloud servers are always tions [10]. As MEC is a key component of 5G networks, the
located centrally in the core network and far away from energy efficiency has become a mainstream concern for the
the mobile devices. The long transmission from the mobile design of the MEC mechanism.

2169-3536 2016 IEEE. Translations and content mining are permitted for academic research only.
5896 Personal use is also permitted, but republication/redistribution requires IEEE permission. VOLUME 4, 2016
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
K. Zhang et al.: Energy-Efficient Offloading for Mobile Edge Computing in 5G Heterogeneous Networks

In this paper, we focus on the design of an energy-efficient MEC computation offloading schemes are described in
computation offloading mechanism for MEC in 5G hetero- Section V. Performance evaluation is presented in Section VI.
geneous networks. With the MEC computation offloading, Finally, we conclude our work in Section VII.
the energy consumption for accomplishing the computation
tasks includes two parts. The first part is the energy spent on II. RELATED WORK
transmitting the computation files to the MEC servers. Due The development of cloud computing and virtualization tech-
to the variable wireless channel states and the different sizes niques provides an efficient way to decouple the application
of the computation files, the energy consumption for trans- environment from the underlying hardware resources, and
mission may vary among the mobile devices. Furthermore, in thus greatly improves the utilization of available computing
the case where the mobile devices share the radio resources resources [14]–[18]. MCC, which has evolved from cloud
with each other, they may cause severe interference to each computing, is designed to address the computation require-
others [11]–[13]. The interference will decrease the transmis- ments of new smart mobile phone based applications [19].
sion rates for the files, and hence reduce the energy efficiency In recent years, several studies have addressed the mobile
of the MEC offloading. As a result, an efficient transmission computation offloading in the MCC scenario. In [20], the
control scheme is needed in the MEC offloading. The second authors formulated the computation offloading decision of
part is the energy spent on the computing. This amount mobile users as a decentralized game, and proposed a
mainly depends on the computation capabilities of the mobile game theoretic approach to achieve the efficient computation
devices and the MEC servers. offloading. Considering the local loads of mobile users and
Each mobile device can decide whether to offload its task the availabilities of cloudlets, the authors in [21] proposed a
to the MEC servers for remote computing or to accomplish Markov decision process based dynamic offloading scheme
the task locally on its own device. This decision is made for mobile users in an intermittently connected cloudlet sys-
by comparing the energy costs. However, due to the lim- tem. In [22], the authors studied the multi-user computation
ited communication resources, the interactivity between these partitioning problem in a large scale mobile cloud applica-
mobile devices in the transmission process may affect the tion scenario, and designed an offline heuristic algorithm to
transmission energy cost of each device. This effect makes minimize the average completion time for all users. In order
the MEC offloading couple with the wireless resource allo- to form an elastic mobile computing grid, the authors in [23]
cation. Furthermore, considering different QoS constraints proposed a resource provisioning framework for organizing
required for the computation tasks and the variable com- the heterogeneous devices in the vicinity. In [24], the authors
putation capabilities of these devices, archiving an energy- investigated the impacts of the geographical distribution of
efficient offloading by coordinating wireless transmission cloud resources on the cloud-based mobile augmentation
and task implementation among the mobile devices and the performance.
MEC servers is a challenging task. The cloud servers of a MCC are located in the core net-
In this paper, we design an energy-efficient MEC offload- work, which leads to high energy consumption by the mobile
ing mechanism for mobile devices in 5G heterogeneous devices for computation file transmission. Furthermore, the
networks. This mechanism minimizes the system energy con- latency caused by the transmission through wide area net-
sumption and concurrently ensures the latency constraints of works may seriously hamper the interactivity of the real-time
the computation tasks. The main contributions of this paper mobile applications. MEC is widely considered as a promis-
are as follows: ing technique to tackle these challenges. In MEC, services
• We present a multi-device computation offloading are hosted on the devices directly attached to radio access
framework for mobile edge cloud computing in network [25]. The proximity of the MEC servers results in the
5G heterogeneous networks. access to cloud functionalities with low transmission energy
• To cope with the multi-access characteristics of 5G het- and latency.
erogeneous networks, we formulate an energy-efficient There are a few studies on efficient computation offloading
optimization problem that minimizes the system energy mechanism of MEC. For instance, in [26], the authors inves-
consumption while satisfying the latency constraints. tigated the tradeoff between offloading computation tasks
• In order to overcome the complexity of solving the to infrastructure clouds and retaining them in mobile edge
optimization problem, we design a three-stage energy- clouds. In [27], the authors studied the dynamic service
efficient computation offloading scheme. In this scheme, migration problem in mobile edge clouds, and proposed a
through type classification and priority assignment for Markov decision process based sequential offloading deci-
the mobile devices, the optimization problem can be sion framework. The authors in [28] proposed a low complex-
solved in polynomial complexity. ity small cell cluster formation and load balancing scheme
The rest of the paper is organized as follows. In for the edge cloud in dense deployment 5G network. The
Section II, we review related work. In Section III, we present authors in [14] formulated the task offloading problem as
the framework of multi-device MEC offloading in a 5G a joint optimization of the radio resources together with
heterogeneous network. The energy-efficient optimization the computational resources, and proposed an iterative algo-
problem is formulated in Section IV. The energy-efficient rithm to solve the problem. In [29], the authors presented an

VOLUME 4, 2016 5897

 K. Zhang et al.: Energy-Efficient Offloading for Mobile Edge Computing in 5G Heterogeneous Networks

analytical model for MEC cloud, and studied the performance

of MEC with the presence of user mobility. The authors
in [30] made a case study of the formation of a realtime con-
text aware ad-hoc collaboration system through combining
5G and MEC technologies. In [25], the authors studied the
virtual network embedding problems, and proposed network
virtualization in the context of MEC networks.
Due to the sharp increase of the energy consumption
and carbon emission of communication systems, the energy-
efficient performance metric has been a critical goal in the
design of the offloading mechanism in cloud-based net-
works [31]. Few studies have addressed the energy-efficient
computation offloading problem. For example, the authors
in [32] studied the energy-efficient offloading policies for
transcoding tasks in a mobile cloud system, and proposed
an online offloading algorithm with the objective to mini- FIGURE 1. The mobile edge computing offloading in a 5G heterogeneous
mize the energy consumption while achieving low latency. network.
To reduce energy cost, the authors in [33] presented a frame-
work of the joint optimization of the radio and computa-
tion resource for the energy-limited mobile terminals in a For each mobile device i (i ∈ N ), the task Ti can be
femto cell network. Based on the estimated execution time either executed locally on itself or on the MEC server via
and energy consumption in the computation offloading, the computation offloading. In our model, each task is atomic
authors in [34] proposed an offloading framework, which and can not be further divided. Let fiL and δiL denote the
reduces the energy consumption and shortens the response local computing ability and the energy consumption for one
time. In MEC, as the computation tasks can be offloaded to CPU cycle of device i, respectively. Thus, we can get the time
the cloud servers located close to the current position of the duration of the local execution of task Ti as
mobile devices, the energy consumption for data transmission
is mainly spent in the radio access network. To improve the tiL = ci /fiL . (1)
energy efficiency in MEC, the authors in [3] studied the multi-
The energy consumption of this local execution can be
user computation offloading problem in a multi-channel wire-
calculated as
less network, and designed a distributed offloading algorithm
through a game theoretic approach. Nevertheless, few work eLi = ci δiL . (2)
has taken into account the heterogeneous radio access net-
work in the MEC, and incorporated the multi-transmission In the 5G heterogeneous network, there is a Macro Base
mode selection into the energy-efficient offloading schemes. Station (MBS) equipped with an MEC server. The MEC
Different from these studies, in this paper, we concentrate server has the ability to run multiple computation tasks
on the computation offloading for MEC in 5G heterogeneous simultaneously. Besides the MBS, there is a Small Base Sta-
networks and propose the optimal offloading schemes to tion (SBS), whose service area is overlaid by that of the MBS.
improve the energy efficiency of the cloud computing system To reuse spectrum efficiently, both the MBS and the SBS
while guaranteeing the delay constraints of the computation operate in the same frequency band. The spectrum is divided
tasks. into K channels, which are denoted as K = {1, 2, ..., K }.
The bandwidth of each channel is identical, which is denoted
III. SYSTEM MODEL as W . In this paper, we focus on a multi-user OFDMA system
Fig. 1 shows the mobile devices offloading their computation in 5G networks, where each channel in the system is orthog-
tasks to the MEC server through a 5G heterogeneous network. onal to the others.
In the system, we consider a set of mobile devices, which is Between the SBS and the MBS, there is a backhaul. This
denoted as N = {1, 2, ..., N }. Each device has a computation backhaul relays the transmission from the SBS to the MBS.
task to be completed within a certain delay constraint. The We consider this backhaul is shared with other communica-
tasks include interactive gaming, natural language process- tion infrastructures. Thus, we ignore the power consumption
ing, image location and etc [35]. Each computation task can of this backhaul. The transmission bandwidth of the backhaul
be described in three terms as Ti = {di , ci , timax }, i ∈ N . For is limited. The transmission delay of the backhaul is propor-
task Ti , di is the size of the input data for the computation, tional to the length of the data with the scaling factor ϕ.
which may include program codes, input files etc. ci denotes When a mobile device chooses computing its task by the
the computing ability required for accomplishing this task, MEC server, the input data can be transmitted to the MEC
which is quantized by the number of CPU cycles. timax is the server through the MBS or the SBS. In the case that mobile
maximum latency required by the computation task. device i accesses the MBS on channel k, the obtained uplink

5898 VOLUME 4, 2016

K. Zhang et al.: Energy-Efficient Offloading for Mobile Edge Computing in 5G Heterogeneous Networks

data transmission rate can be shown as given as

XK
M pM M
i gi riM = M
ai,2,k ri,k
ri,k = W log2 (1 + S + σ2
), (3) k=1
Ii,k K
X pM M
i gi
= ai,2,k W log2 (1 + ). (6)
where pM
i is the power of mobile device i transmitting data
N
al,3,k pSl gM + σ2
P
k=1
to MBS in a unit channel. The transmission power can be l
l=1,l6=i
determined by the MBS through some power control mech-
anisms [36]. gM The total energy consumption in this case can be shown as
i is the channel gain between mobile user i
i = bi pi di /ri + ci δ ,
eM
S denotes the interference at the MBS on M M M R
and the MBS. Ii,k (7)
channel k, which is caused by the other devices’ uplink trans-
where δ R is the energy cost of the MEC server for implement-
mission to the SBS on the same channel. σ 2 is the background
ing a unit CPU cycle. As the MEC server always has higher
noise power.
computation energy efficiency than the mobile devices, we
Similarly, we can give the uplink transmission rate of
consider that δ R < δ L . bM i is the number of the chan-
device i accessing to the SBS on the channel k as
nels utilized by device i for transmitting data to the MBS,
PK
S pSi gSi and bMi = k=1 ai,2,k .
ri,k = W log2 (1 + ). (4)
M
Ii,k + σ2 Similarly, the time cost for the case where device i chooses
offloading the task through the SBS can be given by
tiS = di /riS + di ϕ + ci /f0R , (8)
IV. PROBLEM FORMULATION
In this paper, we focus on the energy efficiency of the com- where ϕ is a coefficient representing backhaul transmission
putation offloading network, and aim to minimize the sys- time delay for a unit data. The uplink transmitting rate riS
tem energy consumption under the latency constraints of the in (8) is shown as
computation tasks. The energy consumption includes both the XK
computational energy and the communication energy. riS = S
ai,3,k ri,k
k=1
Due to the variable computing and communication abilities K
X pSi gSi
of these mobile devices, for each device, the energy consump- = ai,3,k W log2 (1 + ). (9)
N
tion of computing the task locally may be higher or lower than
l gl + σ
k=1
P
al,2,k pM S 2
when the task is executed on the MEC server. Thus, in order l=1,l6=i
to reduce the energy cost, each mobile device should decide
The total energy cost for offloading the task through the
whether to offload the task in an energy-efficient manner.
SBS is
Considering device i is able to offload Ti in two ways,
namely through the MBS and the SBS, we denote ai,j,k as the eSi = bSi pSi di /riS + ci δ R , (10)
task offloading decisions of the device i, where ai,j,k = {0, 1}.
where bSi is number of the channels occupied by device i, and
ai,j,k = 1 means mobile device i chooses mode j to accom-
bSi = K
P
plish the task Ti , and the computation data is transmitted k=1 ai,3,k .
To minimize the total energy consumption of the system,
through channel k. ai,j,k = 0, otherwise. Here, i ∈ N , and
the optimization problem is mathematically modeled as
k ∈ K. Let j = {1, 2, 3} denote the chosen modes. They
are computing locally, transmitting through the MBS and N K
X di X
transmitting through the SBS, respectively. It should be noted min (ai,1,1 eLi + si,2 (pM
i ai,2,k + ci δ R )
{ai,j,k }
i=1
riM k=1
that as there is no channel in the local computing mode, the
item k is meaningless when j = 1. Thus we take ai,1,1 = 1 as K
di X
the indicator that device i selects local computation. + si,3 (pSi ai,3,k + ci δ R ))
riS k=1
For the computation task offloading to the MEC server,
some extra energy and time cost are incurred by the wireless s.t. C1 : ai,1,1 · tiL ≤ timax , i ∈ N
uplink transmission. In the case that device i offloads its task di
C2 : riM ≥ max , i∈N
through the MBS to the MEC server, the total time duration ti − ci /f0R
can be calculated as di
C3 : riS ≥ max , i∈N
tiM = di /riM + ci /f0R , (5) ti − ci /f0R − di ϕ
K
X K
X
where f0R is the computing ability of the MEC server. To C4 : ai,1,k · ai,2,k = 0, i∈N
concentrate our studies on the effects of the 5G heterogeneous k=1 k=1
network on the computation offloading, we consider f0R is a XK XK
constant for each offloading task. riM is the total uplink rate C5 : ai,2,k · ai,3,k = 0, i∈N
of the data transmitting from device i to the MBS. riM can be k=1 k=1

VOLUME 4, 2016 5899

 K. Zhang et al.: Energy-Efficient Offloading for Mobile Edge Computing in 5G Heterogeneous Networks

K K
X X problem (12) can be taken as a special maximum cardinality
C6 : ai,1,k · ai,3,k = 0, i∈N
bin packing problem, and is proved NP-hard [3], [37].
k=1 k=1
To obtain the sub-optimal solution of (12), we propose
N
X a scheme named Energy-Efficient Computation Offload-
C7 : ai,j,k = 1, i ∈ N , j ∈ {2, 3}
ing (EECO) to decide the tasks’ implementation modes and
i=1
N X M
allocate radio resources to the offloading devices. In order to
X solve (12) efficiently, the EECO scheme is divided into three
C8 : ai,j,k ≤M , i ∈ N , j ∈ {2, 3}
stages, which are stated as follows.
i=1 k=1
• Stage 1: Mobile device classification. The mobile
C9 : ai,j,k = {0, 1}, i ∈ N , j ∈ {1, 2, 3}, k ∈ K.
devices are classified into three types according to their
(11) time and energy cost features of the task computing
process.
In (11), si,j = 1( K k=1 ai,j,k > 0), where j = {2, 3}.
P
• Stage 2: Priority determination. We derive the priorities
Here, 1(x) is an indicator function which equals 1 if x is of the devices, which choose offloading their tasks to the
true and 0 otherwise. Constraints C1∼C3 ensure the latency MEC server. The priorities are used for radio resource
requirements of the three possible implementation ways allocation, and determined by the wireless communica-
of task i, respectively. Constraints C4∼C6 state that only tion states and the task requirements.
one implementation way can be selected for each device i. • Stage 3: Radio resource allocation. In this stage, the
According to constraint C7, one channel can only be allocated channels of MBS and SBS are allocated to the mobile
to at most one device. Constraint C8 indicates that the total devices according to the priorities determined in stage 2.
channels occupied by the mobile devices are less than the In the above stages, stage 2 and 3 operate iteratively, until
maximum number of channels possessed by the MBS or a convergence criterion is satisfied. In the following of this
the SBS. section, we present the EECO scheme in detail.
Lemma 1: The optimization problem (11) can be equally
transformed to the problem shown as follows. A. MOBILE DEVICE CLASSIFICATION
Based on the tasks’ latency constraints and the comparison
N K K of the energy costs between different task implementation
X 0 di X S di
X
min (ai,1,1 eLi + pM
i ai,2,k + pi S ai,3,k ) modes, we classify the mobile devices into three types.
{ai,j,k } riM k=1 ri k=1
i=1 The first type of devices is defined as the devices, which
s.t. C1 ∼ C9, (12) should compute their tasks on the MEC server. We denote the
0 set of the devices of this type as GR . For a device with limited
where eLi = eLi − ci δ R . computation resource, which cannot satisfy the latency con-
Proof: ThePobjective function of (11) can be rewrit- straint of the task, the device needs to choose offloading the
ten as min{ai,j,k } N L0
i=1 (ai,1,1 + si,2 + si,3 )ci δ + (ai,1,1 ei +
R
task to the MEC server. Thus, we can get that if tiL > timax ,
d i PK S d i PK a device i belongs to GR , i ∈ N .
si,2 pM
i rM k=1 ai,2,k + si,3 pi S k=1 i,3,k ). According to
ri
i The second type is defined as the devices should compute
the definition of si,j and constraint C7 in (11), we can get the task on their local equipments. We denote the device set of
ai,1,1 + si,2 + si,3 = 1. Thus, (ai,1,1 + si,2 + si,3 )ci δ R this type as GL . The condition used to determine the devices
can be omitted in the objective function, as it is a constant. belonging to this type is given as follows.
Furthermore, according to constraints C4 ∼ C6 in (11), we
1: If tiL ≤ timax and eLi < min{pM n0 ,
 M
Theorem i
can see that at a given time, for each device i, one and only
pSi nS0 } + ci δ R , device i belongs to GL (i ∈ N ), where
 
one of these three items of the objective function is a positive
value, and the other two are zeros. Then, we can remove si,j di
nM
0 = , (13)
(j = {2, 3}) from the function.  (timax − ci /f0R )wlog2 (1 + pM
i gi /σ )
M 2

and
V. ENERGY-EFFICIENT COMPUTATION
di
OFFLOADING SCHEMES nS0 = , (14)
In this section, we focus on solving the optimization prob- (timax − ci /f0R − di ϕ)wlog2 (1 + pSi gSi /σ 2 )
lem (12). We design energy-efficient MEC cloud offloading where d·e is the ceil function.
and radio resource allocation schemes in the 5G heteroge- Proof: If device i chooses offloading its task to the
neous networks. MEC server through the MBS, the highest transmission rate
i gi /σ ).
per channel it can obtain is ri,max M = wlog2 (1 + pM M 2
In the computation offloading process, the mobile devices
max
To satisfy the latency constraint ti , the least number
choose their tasks’ implementation modes through the binary
strategies {ai,j,k }. The decisions of the {ai,j,k } not only of channels needed by device i is nM 0 = di /((timax −
depend on the tasks’ delay constraints, but also on the ci /f0 )wlog2 (1 + pi gi /σ )). Thus, we can get the least
R M M 2

transmission interference between the mobile devices and energy for device
 M  i to offload task Ti to the MEC server via the
the limited resources of the radio access networks. Thus, the MBS is pM i n0 +c i δ R . Similarly, we can get the least energy

5900 VOLUME 4, 2016

K. Zhang et al.: Energy-Efficient Offloading for Mobile Edge Computing in 5G Heterogeneous Networks

Algorithm 1 The Algorithm for Classifying the Mobile which device i can gain Signal to Interference plus Noise
Devices Ratios (SINRs) above a given threshold θi .
Initialization: In Definition 1, the threshold θi can be calculated as
Mobile device set: N = {1, 2, · · · , N };
Transmission power to MBS: {pM i }, i ∈ N ;
θi = 2r̄i W − 1, (15)
Transmission power to SBS: {pSi }, i ∈ N ;
where r̄i is device i’s additional required data rate to transmit
Computation tasks: Ti = {di , ci , timax }, i ∈ N ;
the task file within delay constraint, based on the obtained
Categorized device sets: GL = GO = GR = ∅.
channels in the last stage iteration. We can get r̄i =
1: For Each device i ∈ N Do
di /timax − ri,a . Here, ri,a is the total transmission rate on the
2: Calculate tiL and eLi according to (1) and (2), respec-
channels which have been already allocated to device i.
tively;
Let S denote the wireless communication state of the cur-
3: if (tiL > timax ) then
rent priorities determination stage. As in the EECO scheme,
4: i ⇒ GR ;
the stage 2 and 3 operate iteratively, S is updated in each
5: else if (tiL ≤ timax ) then
iteration. Given a state S, the number of qualified channels
if (eLi < min{pM n0 , pi n0 } + ci δ R ) then
 M S  S
6: i for device i accessing the MBS can be expressed as
7: i ⇒ GL ; X
8: end if hM
i =
M
1(γi,k ≥ θi ), (16)
9: else 0
k∈KM
10: i ⇒ GO ;
where KM 0 is the set of the available channels of MBS. γ M is
11: end if i,k
12: End For the SINR of device i transmitting in channel k of the MBS,
Output: GL , GR and GO . which can be shown as γi,k M = pM gM /(I S + σ 2 ).
i i i,k
Similarly, we can get the number of qualified channels for
device i accessing the SBS as
X
consumed hSi = S
1(γi,k ≥ θi ), (17)
  to offload task Ti to the MEC server via the SBS
as pSi nS0 + ci δ R . Thus, we get the conclusion that given the k∈KS0
condition tiL ≤ timax , if the energy spent on local computing
where KS0 is the set of the unoccupied channels of SBS,
is less than the minimal value of the energy consumed for
and γi,k
S = pS gS /(I M + σ 2 ).
i i i,k
offloading, device i chooses to complete its task locally. 
Besides the two types mentioned before, the third type of Definition 2: The priority of device i in the process of the
the mobile devices is denoted as GO . The devices belong to GO radio resource allocation is defined as
can either decide to implement their tasks locally or to offload timax hi Ē
the tasks to the MEC server. The decisions mainly depend on pi = α1 ( ) + α2 (
) + α3 ( L ), (18)
T̄max H̄ ei − ci δ R
the wireless communication states.
The complete device classification process is illustrated in where T̄max =
P max , H̄ = P
i∈GO ti i∈GO hi , and Ē =
Algorithm 1. L − c δ R ). α (j = {1, 2, 3}) is a coefficient, 0 ≤
P
i∈GO i(e i j
αj ≤ 1 and 3j=1 αj = 1. The device with the less pi value
P
B. MOBILE DEVICE PRIORITY DETERMINATION has the higher priority.
Considering the limited capacity of the radio resources and The priority definition jointly takes the delay constraint,
the transmission interference between the mobile devices, we radio resources and the offloading energy gain into consider-
set different priorities for the devices in the radio resource ation. In (18), the first item indicates the effect of the delay
allocation process. constraint on the priority. The device with more critical delay
For the devices belonging to GR , due to their inadequate constraint should have the higher priority. The second item
computation capabilities, the computation tasks need to be in (18) states the effect of the radio resource availability on the
offloaded to the MEC server. The radio resource allocation of priority. The device that has less qualified channels should be
GR should have the highest priority. However, for the devices allocated radio resources preferentially. Otherwise, the device
belonging to GO , their tasks can be either offloaded to the may fail to transmit the task file to the MEC server within
MEC server or the tasks can be executed locally. Thus, to the delay constraint, due to the insufficient radio resources.
reduce the energy consumption of the offloading system and The third item in (18) means that the device with higher
to utilize the radio resource more efficiently, the devices in computing energy difference between the local computing
GO should be assigned different priorities. Before introducing and MEC server computing should have the higher priority.
the priority determination algorithm, we first present two key It is worth noting that as the devices of GO can either choose
definitions as follows. the MBS or the SBS to transmit the task file, the hi in (18)
should be either hM S
Definition 1: The qualified channels for device i is i or hi , when device i has decided to access
the unoccupied channels of the MBS or the SBS, from the MBS or the SBS, respectively. However, before device i

VOLUME 4, 2016 5901

 K. Zhang et al.: Energy-Efficient Offloading for Mobile Edge Computing in 5G Heterogeneous Networks

Algorithm 2 The Algorithm for the Device Priority Algorithm 3 The Algorithm of Resource Allocation
Determination Initialization:
Initialization: Categorized device set: GO ;
Categorized device set: GO ; Computation tasks: Ti = {di , ci , timax }, i ∈ GO ;
Computation tasks: Ti = {di , ci , timax }, i ∈ GO ; Allocated channels for devices: {Ai }, i ∈ GO ;
Allocated channels for devices: Ai , i ∈ GO ; MBS channel set: KM = {1, 2, ..., K };
MBS channel set: KM = {1, 2, ..., K }; SBS channel set: KS = {1, 2, ..., K };
SBS channel set: KS = {1, 2, ..., K }; MBS’s unoccupied channel set: KM 0 ;
Wireless communication state: S; SBS’s unoccupied channel set: KS0 ;
Priority set: P = ∅. Wireless communication state: S.
1: For Each device i ∈ GO Do 1: Set the temporary set GO 0 =G ;
O
2: if Ai = ∅ then 2: while GO 0 6 = ∅ Do
Calculate hM S
3: i and hi according to (16) and (17), 3: Select the device i, where i = arg min{pi }, i ∈ GO ;
respectively; i
4: if Ai = ∅ then
i , hi };
4: Set hi = max{hM S

5: else if Ai ∈ KM then 5: Under the given state S, choose the channels with
the highest SINR from KM 0 and K0 . The selected
6: Get hi = hMi according to (16);
S
7: else if Ai ∈ KS then channels are denoted as k0 and k0S , respectively.;
M

8: Get hi = hSi according to (17); 6: Compute the transmission rate obtained from k0M
9: end if and k0S . The rates are denoted as ri,0 M and r S ,
i,0
10: Calculate pi according to (18); respectively.;
i /ri,0 ≥ pi /ri,0 && ei < ei then
7: if pM M S S M L
11: End For
k0M ⇒ Ai and KM 0 = K0 \k M ;
Output: The priority set for the devices P = {pi }, i ∈ GO . 8: M 0
9: else if pi /ri,0 < pSi /ri,0
M M S && eS < eL then
i i
10: k0S ⇒ Ai and KS0 = KS0 \k0S ;
11: end if
has decided accessing a base station, there is no restriction for 12: else if Ai ∈ KM then
its access selection. Thus, it chooses the base station with the 13: Choose the channel k with the highest SNIR in KM 0 ;

i , hi }.
more qualified channels, i.e., hi = max{hM S M
14: Compute the energy costs ei with the channels Ai
We present this priority determination process and eM
0
i with the channels {Ai ∪ k};
in Algorithm 2. 0
15: if eM
i < eMi < ei then
L

16: k ⇒ Ai and KM 0 = K0 \k;

M
C. RADIO RESOURCE ALLOCATION 17: end if
In the third stage, the channels of both the MBS and the SBS 18: else if Ai ∈ KS then
are allocated to the devices based on the determined priorities. 19: Choose the channel k with the highest SNIR in KS0 ;
To ensure the fairness between these devices, each device can 20: Compute the energy costs eSi with the channels Ai
0
get at most one channel in a stage iteration. and eSi with the channels {Ai ∪ k};
In the allocation process, if a device has not decided to 21:
0
if ei < eSi < eLi then
S
access any base station, the device can make the decision 22: k ⇒ Ai and KS0 = KS0 \k;
based on the comparison of the energy savings between 23: end if
offloading the task through the MBS and the SBS. 24: end if
For the device that has decided access the MBS, it can 25: GO0 = GO0 \i;
only choose a channel from the ones belonging to the MBS. 26: end while
Given device i accesses the MBS, as the transmitting power Output: The updated wireless communication state S.
pMi is identical for each channel of the MBS, device i should
choose the channel that has the highest SINR. The reason is
that higher SINR leads to shorter transmission time and less
transmission energy consumption. The new selected channel For the devices that have chosen to access the SBS, the
improves the total transmission rate of device i with the channel allocation is in a similar way. We show the complete
increase of the transmission power on this channel. However, radio resource allocation process in Algorithm 3.
the improving transmission rate reduces the transmission
time, which may decrease the total transmission energy. Thus, D. MAIN PROCESS OF EECO
we should compare the energy cost between the case with The detail of the proposed EECO scheme is illustrated in
the new selected channel and the case without it. If the new Algorithm 4. At first, we classify the mobile devices into
selected channel brings higher energy cost, it should not be three types. For the devices belonging to GR , as they cannot
allocated to device i. accomplish the tasks on their own devices under the latency

5902 VOLUME 4, 2016

K. Zhang et al.: Energy-Efficient Offloading for Mobile Edge Computing in 5G Heterogeneous Networks

Algorithm 4 The Main Process of the Energy-Efficient Com- channel resources. The iteration process is terminated when
puting Offloading Scheme there is no device in GO or no available radio resource.
Initialization: Theorem 2: The computation of EECO scheme has a poly-
Mobile device set: N = {1, 2, · · · , N }; nomial complexity.
Computation tasks: Ti = {di , ci , timax }, i ∈ N ; Proof: Firstly, the mobile device classification in Line 1
Allocated channels for devices: {Ai } = ∅, i ∈ N ; of Algorithm 4 has N iterations by categorizing each device in
MBS channel set: KM = {1, 2, · · · , K }; mobile device set N . Secondly, the priority determination in
SBS channel set: KS = {1, 2, · · · , K }; Line 5 has |GO | iterations for assigning priorities for devices
MBS’s unoccupied channel set: KM 0 =K ;
M in GO . After that, radio resource allocation is applied in Line 6
SBS’s unoccupied channel set: KS0 = KS ; with |GO | iterations in total. The number of iterations of the
Wireless communication state: S; while-loop is limited by the process of removing devices in
final final 0 and K0 . The analysis
Final determined device sets: GL = GR0 = ∅. GO and the eliminating channels in KM S
1: Clarify the mobile devices into sets GL , GR and GO of the iterative process is given as follows. On one hand, a
according to Algorithm 1; channel is allocated to a device and removed from KM 0 or K0
S
2: Allocate radio resources to the device set GR , and offload in Line 8, 10, 16 and 22 of Algorithm 3, taking at most k + K
their tasks to the MEC server; iterations of the while-loop in Algorithm 4 in terms of remov-
3: Update KM 0 and K0 ;
S ing channels. On the other hand, each device is removed in
4: while GO 6 = ∅ || { KM 0 6 = ∅ && 0 6 = ∅} Do
KS Algorithm 4 Line 11 and 13, giving |GO | iterations of the
5: Get priority set P according to Algorithm 2; while-loop in terms of deleting devices. To sum, the iteration
6: Based on priority set P, allocate radio resource to the of the while-loop is (|GO | + |GO | + |GO |)max(|GO |, K + K ).
device set GO according to Algorithm 3, and get the Then, the computational complexity of EECO can be given
allocated channels A0i for device i, i ∈ GO ;
7: For Each device i ∈ GO Do O(N + (|GO | + |GO | + |GO |)max(|GO |, K + K ))
8: if A0i 6 = Ai && tiM ≤ timax (tiS ≤ timax ) then = O(max(|GO |2 + N , |GO |K + N )). (19)
9: Ai = A0i ;
10: else if A0i == Ai && tiM ≤ timax (tiS ≤ timax ) then 
final
11: GO = GO \i and i ⇒ GR0 ;
12: else if A0i == Ai && tiM > timax (tiS > timax ) then VI. NUMERICAL RESULTS
final In this section, we evaluate the performance of our pro-
13: GO = GO \i, i ⇒ GL , and put the channels in
posed EECO scheme through the extensive simulations.
Ai to the unoccupied channel set;
We consider there are an MBS and an SBS that cover a
14: end if
0 and K0 ; 1000m×1000m area. Each base station has K = 50 channels
15: Update the channel sets KM S and the channels belonging to one base station are orthogonal.
16: End For
There is an MEC server located in the MBS, whose com-
17: end while
final putation capability is 4GHz/sec. The energy consumption
18: For device set GL , implement their computing tasks
for the MEC is given as δ R =1W/GHz [38]. The backhaul
locally;
final time delay coefficient ϕ is set as 0.0001 sec/KB. N = 50
19: For device set GR0 , offload their computing tasks to the
mobile devices are randomly scattered over the area. The
MEC server.
number of CPU cycles required by the computation tasks
of the devices are randomly distributed between 0.1 and
1 GHz. The corresponding computation file size is randomly
constraint, we allocate the channels to the devices with the distributed between 300 and 800KB. The mobile devices’
highest priority. Then, we update the channel allocation state, latency requirements are randomly distributed between
and the scheme goes into the iterative process. In each iter- 0.5 and 1 second.
ation, the devices of GO are assigned priorities according to Fig. 2 indicates the energy consumption performance of the
Algorithm 2 and allocated channels according to Algorithm 3. system with three different task implementation schemes in
At the end of each iteration, each device of GO is checked in terms of the number of the mobile devices. The energy con-
terms of its channel resource assignment. In every iteration of sumption includes both the energy cost spent on computation
EECO, each device chooses the channel that brings the high- and file transmission. In this figure, the energy consumption
est reduction in the energy consumption. In a given iteration, of all three schemes increase as the number of devices grows.
if a device cannot get a channel that decreases its energy cost, The energy consumption of the scheme without offloading
it is less likely that the device will obtain a suitable channel is higher than that of the other two schemes in all cases,
in the next iteration. Thus, we remove the devices from GO especially with large number of devices. It is worth noting
whose channel assignments are identical in two consecutive that some tasks’ delay constraints may not be satisfied when
iterations. The removed devices are classified into GL or GR all the tasks are implemented locally on the devices. Here
according to the offloading time delay based on their assigned we use the performance of this scheme as a benchmark to

VOLUME 4, 2016 5903

 K. Zhang et al.: Energy-Efficient Offloading for Mobile Edge Computing in 5G Heterogeneous Networks

FIGURE 2. The energy consumption of the offloading system with

different schemes. FIGURE 3. The number of the device sets GL , GO and GR with different
backhaul time delay coefficient.

measure the energy cost reduction gained by the other two

schemes.
Compared to the scheme without offloading, Fig. 2 shows
that the average rate of the decrease in energy consumption
in our proposed EECO scheme is 18%. From this figure, we
can see that the energy consumption difference between the
scheme without offloading and the schemes with offloading
doesn’t continue to increase with the growth of the number
of the mobile devices. This is because that the energy con-
sumption for file transmission may vary with the increase
of the device number, and there is no linear relationship
between them. Furthermore, we can see from the figure that
the energy consumption of EECO and that of the offloading
scheme without device priorities are almost the same when
FIGURE 4. The energy consumption of the offloading system with
the number of device is below 30. However, the difference different backhaul time delay coefficient.
becomes visible as the number increases. The reason is that,
when there are few devices in the network, the radio resource
is sufficient for the offloading devices. With the increase
in the number of device, the channel contention occurs.
The offloading scheme without device priorities allocates the
channels to the offloading devices randomly. On the contrary,
in EECO scheme, the channels are preferentially allocated to
the devices, which may result in higher energy cost reduction.
Fig. 3 shows the number of the mobile devices of the three
types with different time delay coefficient ϕ. In Fig. 3, when
ϕ ≤ 0.0003 sec/KB, there is no change in the number of
devices. This is because that the low transmission delay of
the backhaul may not cause the total time cost of the task
accomplishment breaks the delay constraints of the tasks.
With the increase of ϕ, higher transmission delay is posed on
FIGURE 5. The number of the offloading devices with different MEC
the backhaul. Thus, the offloading devices that can not use the server CPU capabilities.
SBS for transmitting files under the delay constraint choose to
transmit through the MBS. However, due to the radio resource
limitation of the MBS, the number of devices offloading the of the system increases with higher ϕ. This can be explained
tasks via MBS reaches the maximum value when ϕ ≥ 0.0007. as follows. First, when ϕ increases, the devices located
The left devices that can neither offload tasks via MBS nor via near the SBS have to transmit through the MBS due to the
SBS should implement their tasks locally. time delay constraint. For these devices, transmitting via the
Fig. 4 indicates the energy consumption of the system with MBS causes higher transmission energy cost compared to
different time delay coefficient ϕ. The energy consumption transmitting through the SBS. Second, the number of devices

5904 VOLUME 4, 2016

K. Zhang et al.: Energy-Efficient Offloading for Mobile Edge Computing in 5G Heterogeneous Networks

that implement tasks locally increases. Considering that the [8] M. Patel, B. Naughton, C. Chan, N. Sprecher, S. Abeta, and A. Neal,
computation energy efficiency of the MEC server is higher ‘‘Mobile-edge computing,’’ Mobile-Edge Comput.-Introductory, Tech.
Rep., Sep. 2014.
than that of the devices, some extra energy should be con- [9] K. Yang, Q. Yu, S. Leng, B. Fan, and F. Wu, ‘‘Data and energy integrated
sumed. communication networks for wireless big data,’’ IEEE Access, vol. 4,
Fig. 5 compares the impacts of the MEC server CPU capa- pp. 713–723, 2016.
[10] K. M. S. Huq, S. Mumtaz, J. Bachmatiuk, J. Rodriguez, X. Wang, and
bility on the number of the offloading devices with different R. L. Aguiar, ‘‘Green HetNet CoMP: Energy efficiency analysis and
MEC server energy cost. The number of devices that choose optimization,’’ IEEE Trans. Veh. Technol., vol. 64, no. 10, pp. 4670–4683,
to offload their tasks to the MEC server decreases with the Oct. 2015.
[11] P. Guo, J. Wang, X. H. Geng, C. S. Kim, and J.-U. Kim, ‘‘A variable
increase in the server computation energy cost. When energy threshold-value authentication architecture for wireless mesh networks,’’
cost reaches 6 W/GHz, since more computation energy is J. Internet Technol., vol. 15, no. 6, pp. 929–936, Nov. 2014.
consumed for the task implemented on the MEC server than [12] C. Shao, S. Leng, Y. Zhang, A. Vinel, and M. Jonsson, ‘‘Performance
analysis of connectivity probability and connectivity-aware MAC protocol
on the local devices, no device offloads its task to the MEC
design for platoon-based VANETs,’’ IEEE Trans. Veh. Technol., vol. 64,
server. As the CPU capability gets lower, the number of no. 12, pp. 5596–5609, Dec. 2015.
offloading devices reduces. It is worth noting that the offload- [13] Q. Wang, S. Leng, H. Fu, and Y. Zhang, ‘‘An IEEE 802.11p-based mul-
ing device numbers with CPU capabilities of 4 GHz/sec and tichannel MAC scheme with channel coordination for vehicular ad hoc
networks,’’ IEEE Trans. Intell. Transp. Syst., vol. 13, no. 2, pp. 449–458,
3 GHz/sec are the same in this figure. The reason is that both Jun. 2012.
these CPU capabilities can satisfy the delay constraints of the [14] S. Sardellitti, G. Scutari, and S. Barbarossa, ‘‘Joint optimization of radio
offloaded tasks. To this end, we can draw a conclusion that the and computational resources for multicell mobile-edge computing,’’ IEEE
Trans. Signal Inf. Process. Netw., vol. 1, no. 2, pp. 89–103, Jun. 2015.
energy cost of the MEC server directly affects the offloading [15] Z. Xia, X. Wang, X. Sun, and Q. Wang, ‘‘A secure and dynamic multi-
devices. However, the offloading choices of the devices can keyword ranked search scheme over encrypted cloud data,’’ IEEE Trans.
only be influenced when the MEC server CPU capability is Parallel Distrib. Syst., vol. 27, no. 2, pp. 340–352, Feb. 2016.
[16] B. Gu, V. S. Sheng, K. Y. Tay, W. Romano, and S. Li, ‘‘Incremental support
below a certain threshold. vector learning for ordinal regression,’’ IEEE Trans. Neural Netw. Learn.
Syst., vol. 26, no. 7, pp. 1403–1416, Jul. 2015.
VII. CONCLUSION [17] Y. Ren, J. Shen, J. Wang, J. Han, and S. Lee, ‘‘Mutual verifiable provable
data auditing in public cloud storage,’’ J. Internet Technol., vol. 16, no. 2,
In this paper, we investigated the MEC offloading mecha- pp. 317–323, Mar. 2015.
nisms in 5G heterogeneous networks. In order to improve [18] Z. Fu, X. Sun, Q. Liu, L. Zhou, and J. Shu, ‘‘Achieving efficient cloud
the energy efficiency of the offloading system, we formu- search services: Multi-keyword ranked search over encrypted cloud data
supporting parallel computing,’’ IEICE Trans. Commun., vol. E98-B, no. 1,
lated a problem to minimize the energy consumption of the pp. 190–200, Jan. 2015.
computation task implementation together with that of the [19] A. ur Rehman Khan, M. Othman, F. Xia, and A. N. Khan, ‘‘Context-aware
communication process. To solve the problem more effi- mobile cloud computing and its challenges,’’ IEEE Cloud Comput., vol. 2,
no. 3, pp. 42–49, May/Jun. 2015.
ciently, we proposed an EECO scheme, which jointly opti-
[20] X. Chen, ‘‘Decentralized computation offloading game for mobile
mizes the computation offloading decisions and the radio cloud computing,’’ IEEE Trans. Parallel Distrib. Syst., vol. 26, no. 4,
resource allocation strategies to minimize the system energy pp. 974–983, Apr. 2015.
cost under the delay constraints. In addition, we conducted a [21] Y. Zhang, D. Niyato, and P. Wang, ‘‘Offloading in mobile cloudlet systems
with intermittent connectivity,’’ IEEE Trans. Mobile Comput., vol. 14,
simulation study, which clearly displays the energy efficiency no. 12, pp. 2516–2529, Dec. 2015.
enhancement in our proposed EECO scheme. [22] L. Yang, J. Cao, H. Cheng, and Y. Ji, ‘‘Multi-user computation partitioning
for latency sensitive mobile cloud applications,’’ IEEE Trans. Comput.,
vol. 64, no. 8, pp. 2253–2266, Aug. 2015.
REFERENCES [23] H. Viswanathan, E. K. Lee, I. Rodero, and D. Pompili, ‘‘Uncertainty-aware
[1] Z. Fu, K. Ren, J. Shu, X. Sun, and F. Huang, ‘‘Enabling personal- autonomic resource provisioning for mobile cloud computing,’’ IEEE
ized search over encrypted outsourced data with efficiency improve- Trans. Parallel Distrib. Syst., vol. 26, no. 8, pp. 2363–2372, Aug. 2015.
ment,’’ IEEE Trans. Parallel Distrib. Syst., vol. 27, no. 9, pp. 2546–2559, [24] S. Abolfazli, Z. Sanaei, M. Alizadeh, A. Gani, and F. Xia, ‘‘An exper-
Sep. 2016. imental analysis on cloud-based mobile augmentation in mobile cloud
[2] B. Fan, S. Leng, and K. Yang, ‘‘A dynamic bandwidth allocation algorithm computing,’’ IEEE Trans. Consum. Electron., vol. 60, no. 1, pp. 146–154,
in mobile networks with big data of users and networks,’’ IEEE Netw., Feb. 2014.
vol. 30, no. 1, pp. 6–10, Jan./Feb. 2016. [25] M. T. Beck and M. Maier, ‘‘Mobile edge computing: Challenges for future
[3] X. Chen, L. Jiao, W. Li, and X. Fu, ‘‘Efficient multi-user computation virtual network embedding algorithms,’’ in Proc. Adv. Eng. Comput. Appl.
offloading for mobile-edge cloud computing,’’ IEEE/ACM Trans. Netw., Sci., Aug. 2014, pp. 65–70.
vol. PP, no. 99, Oct. 2015. [26] U. Drolia et al., ‘‘The case for mobile edge-clouds,’’ in Proc. IEEE 10th
[4] J. Shen, H. Tan, J. Wang, J. Wang, and S. Lee, ‘‘A novel rout- Int. Conf. Ubiquitous Intell. Comput., 10th Int. Conf. Auto. Trusted Com-
ing protocol providing good transmission reliability in underwater sen- put. (UIC/ATC), Dec. 2013, pp. 209–215.
sor networks,’’ J. Internet Technol., vol. 16, no. 1, pp. 171–178, [27] S. Wang, R. Urgaonkar, M. Zafer, T. He, K. Chan, and K. K. Leung,
Jan. 2015. ‘‘Dynamic service migration in mobile edge-clouds,’’ in Proc. IFIP Netw.
[5] Z. Pan, Y. Zhang, and S. Kwong, ‘‘Efficient motion and disparity esti- Conf., May 2015, pp. 1–9.
mation optimization for low complexity multiview video coding,’’ IEEE [28] J. Oueis, E. C. Strinati, and S. Barbarossa, ‘‘The fog balancing: Load
Trans. Broadcast., vol. 61, no. 2, pp. 166–176, Jun. 2015. distribution for small cell cloud computing,’’ in Proc. Veh. Technol. Conf.,
[6] E. Meskar, T. D. Todd, D. Zhao, and G. Karakostas, ‘‘Energy efficient May 2015, pp. 1–6.
offloading for competing users on a shared communication channel,’’ in [29] T. Taleb and A. Ksentini, ‘‘An analytical model for follow me cloud,’’ in
Proc. IEEE Int. Conf. Commun., Jun. 2015, pp. 3192–3197. Proc. IEEE Global Commun. Conf., Dec. 2013, pp. 1291–1296.
[7] B. Gu and V. S. Sheng, ‘‘A Robust regularization path algorithm for [30] S. Nunna et al., ‘‘Enabling real-time context-aware collaboration through
υ-support vector classification,’’ IEEE Trans. Neural Netw. Learn. Syst., 5G and mobile edge computing,’’ in Proc. 12th Int. Conf. Inf. Technol.-New
vol. PP, no. 99, Feb. 2016. Generat., Apr. 2015, pp. 601–605.

VOLUME 4, 2016 5905

 K. Zhang et al.: Energy-Efficient Offloading for Mobile Edge Computing in 5G Heterogeneous Networks

[31] M. Peng, K. Zhang, J. Jiang, J. Wang, and W. Wang, ‘‘Energy- QUANXIN ZHAO (S’–) is currently pursuing the
efficient resource assignment and power allocation in heterogeneous cloud Ph.D. degree with the University of Electronic
radio access networks,’’ IEEE Trans. Veh. Technol., vol. 64, no. 11, Science and Technology of China. His research
pp. 5275–5287, Nov. 2015. interests include resource, spectrum, energy,
[32] W. Zhang, Y. Wen, and H.-H. Chen, ‘‘Toward transcoding as a service: and networking in broadband wireless access
Energy-efficient offloading policy for green mobile cloud,’’ IEEE Netw., networks, next generation mobile networks, and
vol. 28, no. 6, pp. 67–73, Nov./Dec. 2014. software defined wireless networks.
[33] O. Muñoz, A. Pascual-Iserte, and J. Vidal, ‘‘Optimization of radio and
computational resources for energy efficiency in latency-constrained
application offloading,’’ IEEE Trans. Veh. Technol., vol. 64, no. 10,
pp. 4738–4755, Oct. 2015.
[34] Y.-D. Lin, E. T.-H. Chu, Y.-C. Lai, and T.-J. Huang, ‘‘Time-and-energy-
aware computation offloading in handheld devices to coprocessors and
clouds,’’ IEEE Syst. J., vol. 9, no. 2, pp. 393–405, Jun. 2015.
[35] J. Li, X. Li, B. Yang, and X. Sun, ‘‘Segmentation-based image copy-move
forgery detection scheme,’’ IEEE Trans. Inf. Forensics Security, vol. 10,
no. 3, pp. 507–518, Mar. 2015.
[36] M. Xiao, N. B. Shroff, and E. K. P. Chong, ‘‘A utility-based power-control
scheme in wireless cellular systems,’’ IEEE/ACM Trans. Netw., vol. 11, LONGJIANG LI (M’12) received the B.S. degree
no. 2, pp. 210–221, Apr. 2003. in computer science and engineering from Xid-
[37] K.-H. Loh, B. Golden, and E. Wasil, ‘‘Solving the maximum cardinality ian University, Xi’an, in 2001, the M.S. degree
bin packing problem with a weight annealing-based algorithm,’’ in Oper- in computer science and engineering from the
ations Research and Cyber-Infrastructure, vol. 47. USA: Springer, 2009, Xi’an Telecommunication Institute, in 1998, and
pp. 147–164. the Ph.D. degree in computer science and technol-
[38] M. Wittmann, G. Hager, T. Zeiser, J. Treibig, and G. Wellein, ‘‘Chip- ogy from Shanghai Jiao Tong University, China,
level and multi-node analysis of energy-optimized lattice Boltzmann
in 2007. He is an Associate Professor with the
CFD simulations,’’ Concurrency Comput., Pract. Exper., vol. 28,
Department of Communication and Information
pp. 2295–2315, May 2016.
Engineering, University of Electronic Science and
Technology of China. His research interest is in the domain of commu-
nication and information engineering, especially with regard to mobile ad
KE ZHANG is currently pursuing the Ph.D. degree
hoc networks (including VANETs/WSNs), unmanned aerial vehicles, and
with the University of Electronic Science and
distributed virtual reality.
Technology of China. He is a Lecturer with the
University of Electronic Science and Technology
of China. He is currently a Visiting Researcher
with the Simula Research Laboratory, Fornebu,
Norway. His research interests include energy
management and communications in the smart
grid, design and optimization of next-generation
wireless networks, and Internet of Things.
XIN PENG received the bachelor’s degree in com-
munication engineering, the master’s degree in
YUMING MAO is a Professor with the Uni-
communication and information system, and the
versity of Electronic Science and Technology of
Ph.D. degree in computer science from Hunan
China. He is the Chairman of the Department
University, Changsha, China, in 2003, 2008,
of Network Engineering. His main research area
and 2011, respectively. He is an Associate Profes-
includes broadband communication network, net-
sor with the College of Information and Communi-
work organization and protocol analysis, TCP/IP
cation Engineering, Hunan Institute of Science and
technology, network management and protocol,
Technology. His research interests include wire-
routing protocol, and network engineering. He was
less sensor networks and cyber physical systems.
recipient of several awards, including the first
grade, second grade, and third grade awards of the
Ministry of Electronic Industry for science and technology progress, the
second grade National Award for science and technology progress.

SUPENG LENG (M’06) received the Ph.D. degree

from Nanyang Technological University (NTU),
Singapore. He is a Professor with the School LI PAN received the M.S. degree in computer
of Communication and Information Engineering, software and theory from Sun Yat-Sen University,
University of Electronic Science and Technol- Guangzhou, China, in 2004, and the Ph.D. degree
ogy of China, Chengdu, China. He has been a in computer applied technology from Tongji Uni-
Research Fellow with the Network Technology versity, Shanghai, China, in 2009. He is cur-
Research Center, NTU. He has published over rently an Associate Professor with the Department
100 research papers. His research interests include of Information and Communication Engineering,
resource, spectrum, energy, routing, and network- Hunan Institute of Science and Technology, China,
ing in broadband wireless access networks, vehicular networks, Internet of and also with the Key Laboratory of Optimiza-
Things, next-generation mobile networks, and smart grids. He serves as an tion and Control for Complex Systems, College of
Organizing Committee Chair and a Technical Program Committee Member Hunan Province, Yueyang, China. He has authored over 20 papers. His cur-
for many international conferences, and a Reviewer for over ten international rent research interests include Petri nets, formal method, and computational
research journals. intelligence.

5906 VOLUME 4, 2016

K. Zhang et al.: Energy-Efficient Offloading for Mobile Edge Computing in 5G Heterogeneous Networks

SABITA MAHARJAN (S’09–M’13) received the YAN ZHANG (M’05–SM’10) received the Ph.D.
M.Eng. degree in wireless communication from degree from the School of Electrical and Elec-
the Antenna and Propagation Laboratory, Tokyo tronics Engineering, Nanyang Technological Uni-
Institute of Technology, Tokyo, Japan, in 2008, and versity, Singapore. He is currently the Head of
the Ph.D. degree in network and distributed sys- the Department with the Department of Net-
tems from the University of Oslo, Oslo, Norway, works, Simula Research Laboratory, Fornebu,
and the Simula Research Laboratory, Fornebu, Norway, and an Adjunct Associate Professor
Norway, in 2013. She is currently a Post-Doctoral with the Department of Informatics, University of
Fellow with the Simula Research Laboratory. Her Oslo, Oslo, Norway. His current research interests
current research interests include wireless net- include wireless networks and reliable and secure
works, network optimization, security, game theory, smart grid communi- cyber physical systems (e.g., healthcare, transport, and smart grids). He is
cations, and cyber physical systems. an Associate Editor being on the Editorial Boards of a number of well-
established scientific international journals, e.g., Wiley Wireless Communi-
cations and Mobile Computing. He also serves as a Guest Editor of the
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, the IEEE Communications
Magazine, IEEE WIRELESS COMMUNICATIONS, and the IEEE TRANSACTIONS ON
DEPENDABLE AND SECURE COMPUTING. He serves as a Chair or a Technical
Program Committee Member of numerous international conferences. He has
received seven Best Paper Awards. He is a Senior Member of the IEEE
Communications and Vehicular Technology Societies, and a fellow of the
Institution of Engineering and Technology.

VOLUME 4, 2016 5907