Received December 24, 2017, accepted February 6, 2018, date of current version March 16, 2018.

Digital Object Identifier 10.1109/ACCESS.2018.2808411

An Offloading Strategy in Mobile Cloud

Computing Considering Energy and
Delay Constraints
VENUS HAGHIGHI AND NAGHMEH S. MOAYEDIAN , (Member, IEEE)
Department of Electrical and Computer Engineering, Isfahan University of Technology, Isfahan 8415683111, Iran
Corresponding author: Naghmeh S. Moayedian (moayedian@cc.iut.ac.ir)

ABSTRACT With the expansion of wireless networks throughout the world and the high growth rate of
the use of sophisticated programs in mobile devices, users’ expectations for the services provided by these
devices have increased. Mobile devices have some limitations, such as battery life time and processing
power for delivering all types of services to users. In recent years, mobile cloud computing, which is a
phenomenal branch of cloud computing, has achieved considerable evolution in the computing community.
By considering the advantage of offloading to the cloud, the limitations of mobile devices can be overcome
to a great extent. A mobile device can be converted to a powerful device by applying cloud resources.
The outstanding challenges in offloading are finding an optimum solution for the offloading problem to
overcome these limitations. In this paper, offloading is modeled via a mathematical graph where both Wi-Fi
and 3G links are topics of concern. Finding the best solution for offloading is equivalent to finding the
constrained shortest path in this graph. By considering the K-LARAC and M-LARAC heuristic algorithms,
a new heuristic algorithm is introduced to find the optimized path that can assess energy and delay, at a
minimum, financial cost. This path is an appropriate solution for the offloading problem. The obtained results
indicate that the designed algorithm can find an arbitrary approximation solution for the offloading problem
with low complexity in comparison to existing algorithms.

INDEX TERMS Call graph, Delay, Energy consumption, K-shortest path, LARAC, Mobile cloud
computing, Offloading.

I. INTRODUCTION
In today’s world, mobile devices have become very pop-
ular for network users because mobility is an important
criterion for purchasing computing equipment/components.
Computing-intensive applications have significantly expanded
to a degree that the growth rate of complex programs has
become faster than the growth rate of computing resources
of mobile devices. Despite the mobility advantages among
network users, mobile devices have some limitations. Conse-
quently, enhancing some characteristics of mobile devices to
satisfy users’ expectations is not possible. The most impor-
tant limitations of mobile devices are the CPU processing
power and battery lifetime. Accordingly, an approach that can
overcome these limitations while maintaining the features of
mobile devices is required. Mobile cloud computing provides
a platform to achieve these objectives.
Mobile cloud computing has recently become a new model
in distributed computing [1], [2]. By development of mobile
computing, cloud services and integrated access to typical
networks (Wi-Fi and 3G), the computing industry has sig-
nificantly changed. Therefore, many researchers believe that
mobile cloud computing is a great evolution in this con-
text. In general, resources such as networks, servers, storage,
applications, and services are provided as computing facili-
ties that can be leased and released by users via a network
in a cloud computing environment. The definition of cloud
computing according to the National Institute of Standards
and Technology (U.S.NIST) is
‘Cloud computing is a model for enabling convenient,
on-demand network access to a shared pool of configurable
computing resources that can be rapidly provisioned and
released with minimal management effort or service provider
interaction’ [3].
Mobile cloud computing is an extension of cloud com-
puting that provides cloud services via mobile devices and
wireless networks’ interfaces. Offloading to the cloud is one
of the best advantages of mobile cloud computing. Offloading
is defined as transferring the computing intensive operations

2169-3536
2018 IEEE. Translations and content mining are permitted for academic research only.
VOLUME 6, 2018 Personal use is also permitted, but republication/redistribution requires IEEE permission. 11849
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 V. Haghighi, N. S. Moayedian: Offloading Strategy in Mobile Cloud Computing Considering Energy and Delay Constraints
from a mobile device to the cloud and returning the results
to the mobile device. Offloading to the cloud can satisfy the
objective of mobile device advantages while removing limita-
tions of resource shortages. Due to the advantages of offload-
ing, the authors of this paper are motivated to present this
study. The following offloading advantages are presented:
a. Increases the computing capability of mobile devices,
where any complex program that needs a significant
amount of computational resources can be run in the
cloud and obtain results. A shared pool of resources that
supports any complex program with minimal manage-
ment effort is available; therefore, a powerful compu-
tation device that supports a basic mobile device.
b. Improves mobile device energy consumption [4].
c. Reduces delay due to a processing unit.
d. Provides integrated access at any time and place to the
cloud resources.
Note that the application of cloud services and offloading
operations is not always efficient. Low bandwidth, consumed
energy and delay caused by data transmission are some draw-
backs of offloading, which motivates the researchers to find
solutions to overcome these drawbacks.
The focus of this paper is the two most important mobile
device limitations of energy consumption and delay with
respect to the least financial cost. The financial cost parameter
in the use of cloud services and data transmission via Wi-Fi
or 3G networks are important challenges for network clients.
An effective offloading solution with the least financial cost
to satisfy the energy and delay constraints for a mobile device
in a simultaneous manner is assessed in this paper.
The remainder of this paper is organized as follows: The
literature review is run in section 2; an offloading architec-
ture and the descriptive profilers are presented in section 3;
offloading operations as a directed graph that computes
energy, delay and financial cost parameters are modeled in
section 4; the offloading problem is defined and formu-
lated and a new K-M-LARAC is obtained and proposed
in section 5; the online offloading algorithm is introduced
in section 6; the numerical results to assess this proposed
offloading architecture is presented in section 7; and the
conclusions and suggestions are presented in section 8.
II. RELATED STUDIES
Many attempts are made to achieve the offloading solution.
Most of the proposed solutions are based on one of the three
following methods: the first approach relies on the program-
mer who should determine how to partition a program and
how to adapt the program partitioning to change the network
conditions for offloading. This approach is fine-grained and
significantly improves energy consumption and the perfor-
mance of the mobile device [5]–[8].
The second approach comprises all or part of an individual
program that is offloaded to the cloud in an automated man-
ner. In this approach, programs do not need to be modified
for offloading operation, which reduces the burden on the pro-
grammer. This approach is coarse-grained [9], [10]. The third
11850
approach proposes an architecture that includes the benefits
of both the first approach and the second approach, in which
several cloud-assist mobile platforms are proposed. An archi-
tecture for offloading solutions based on the .NET framework
MAUI is proposed by [11]. MAUI supports fine-grained
program offloading to maximize energy-saving with minimal
burden on the programmer. The MAUI solution is derived via
an optimization engine. The optimization problem is solved
using an integer linear programming (ILP) solver. Another
proposed project is the CloneCloud [12], [13], which is a
system that transforms parts of a program to the cloud in an
automated manner. An offline static analysis is performed for
different conditions and dynamic profiling for program parti-
tioning. The optimization problem is solved by the standard
integer linear programming solver. CloneCloud constructs
the pre-computed partitioning database of an applicable pro-
gram to determine which part of the program can be offloaded
to the cloud. Thinkair [14] is a framework that seeks to
overcome the disadvantages of the MAUI and CloneCloud
methods. An integer linear programming solver has expo-
nential complexity, which increases when the size of the
inputs increases. Previously proposed methods have focused
on presenting an architecture for implementing an offloading
mechanism in the cloud. Some studies have focused on find-
ing an optimum offloading mechanism as follows:
In [15] and [16], the energy consumption of mobile devices
is optimized, where two optimization problems for making
offloading decisions are solved: the solution to the first prob-
lem is to reconfigure the CPU clock frequency to locally
execute a program, and the solution to the second problem is
to schedule the task of offloading to the cloud side. Offloading
decisions are made by comparing the energy consumption for
local task execution and the energy required for offloading
a task. A sequence of tasks in linear topology to minimize
the energy consumption on a mobile device is considered
in [17] and [18] while meeting a time deadline. The opti-
mization problem is solved by applying the Lagrangian
Relaxation-based Aggregate Cost (LARAC) [19] algorithm
in an approximation. Sometimes the approximate solution
is not optimal. The feasibility of both mobile computation
offloading and mobile data backups by considering band-
width and energy consumption of offloading are explored
in [20]. The energy consumption of a mobile device is
minimized while meeting the given program deadline based
on Lyapunov optimization [21]. Zhang and Wen assess the
energy-efficient collaborative task execution to minimize the
energy consumption of mobile devices, for which a mobile
application is considered as a general topology to model
the offloading decision. The optimization problem is solved
by the LARAC algorithm. In [23], a distributed offload-
ing mechanism that minimizes the energy consumption of
a mobile device is presented. In [24], an energy-efficient
dynamic offloading and resource scheduling policy is formu-
lated to reduce energy consumption subject to a completion
deadline. The energy- efficiency cost minimization problem
is solved while meeting the task-dependency requirements
VOLUME 6, 2018
V. Haghighi, N. S. Moayedian: Offloading Strategy in Mobile Cloud Computing Considering Energy and Delay Constraints
FIGURE 1. Offloading architecture.
and the completion deadline constraint. To solve the opti-
mization problem, they propose a distributed algorithm that
consists of three sub-algorithms for computation offloading
selection, clock frequency control and transmission power
allocation. Balancing the tradeoff betweenenergy efficiency
and responsiveness of mobile applications by developing
application-aware wireless transmission scheduling algo-
rithms is proposed in [25]. Xia et al. [26] optimized the exe-
cution delay by setting a threshold. If the predicted execution
delay is greater than this threshold, then the application will
be locally executed; otherwise, it will be offloaded to the
cloud. Shah-Mansouri et al. [27] designed a task scheduler
and determined the optimal pricing strategy for a cloud ser-
vice provider. They formulated a utility maximization prob-
lem, in which the energy, delay and price of cloud services
are concerns. This problem is solved by convexification and
primal-dual methods. In [28], the offloading problem is for-
mulated as a joint optimization problem, in which both the
energy consumption of a mobile device and the execution
delay in a mobile application are minimized. A heuristic
algorithm for the general case of the problem is proposed
and solved. Geng et al. [29] applied the Dijkstra algorithm
to find an energy efficient offloading solution for a cellular
network, where no mobile device limitation is considered for
the offloading problem.
In this paper, first, an offloading architecture with a focus
on extracting a program partitioning mechanism is proposed.
Second, an offloading problem is modeled as a directed graph
with both Wi-Fi and 3G networks’ links. Energy consump-
tion and delay caused by processing, which are the two
most important limitations of mobile devices, are consid-
ered. Third, an offloading problem is formulated to find the
constraints shortest path to minimize the financial cost as a
subject. Last, the optimal approximation solution is extracted
via this newly proposed algorithm (K-M-LARAC), which is
subject to current network conditions.
III. OFFLOADING ARCHITECTURE
The different parts of the offloading architecture are shown
in figure 1 and described as follows. Different parts of the
offloading architecture exist in both device and cloud sides;
because the offloading strategy can change at runtime due to
VOLUME 6, 2018
different networks’ conditions. So we have to obtain valid
data from profilers in both sides, in order to achieve new
offloading strategy for the remaining tasks.
A. DEVICE PROFILER
Some information about the mobile device, such as the CPU
clock frequency and CPU power consumption in active and
IDLE modes, are labeled fdevice , pactive IDLE
device , and pdevice . The cloud
CPU clock frequency is labeled fcloud .
B. PROGRAM PROFILER
A program profiler provides information about a specific pro-
gram, including the CPU cycles that are needed to complete
a certain method in the program. A program can be shown
as a directed call graph, where nodes represent methods and
edges represent the relations among methods and the way
they invoke each other. A call graph shows the program
execution trend. The methods that are capable of offloading
are designated by a programmer.
The call graph, number of CPU cycles and amount of input
and output data for each method execution are captured by the
program profiler. The input and output data for each method
execution are the program states that the program needs to
continue remote execution. The program states, for example,
contain the CPU registered data, CPU counter and stack data.
In this paper, the number of CPU cycles and the amount of
input data and amount of output data for the execution of
the ith method of the program are denoted by Ci , Ii , and Oi ,
respectively.
C. NETWORK PROFILER
Mobile devices can use different technologies to access dif-
ferent network infrastructures. Two extensively employed
network infrastructures are 3G and Wi-Fi. Use of a 3G net-
work can be advantageous due to its extensive coverage in the
absence of Wi-Fi. However, note that 3G network provides
less bandwidth than the time consumed by a Wi-Fi network.
In addition, the interfaces connected to these two networks
significantly differ in energy consumption and charging. The
energy consumption of these two networks’ interfaces are
assessed as follows:
11851
 V. Haghighi, N. S. Moayedian: Offloading Strategy in Mobile Cloud Computing Considering Energy and Delay Constraints
• POWER MODEL FOR 3G NETWORK
The energy consumption in a 3G cellular network is related
to two main factors: the energy consumption for data trans-
mission, which is fully dependent on the amount of trans-
mitted data, the power level of transmission and the Radio
Resource Control (RRC) protocol. RRC is in charge of con-
trolling channel allocation and estimating radio frequency
(RF) power consumption based on a timer. The state machine
for the RRC protocol in a 3G cellular network is shown
in figure 2. When no data is transmitted, the network interface
state is changed to IDLE. When some data are ready to be
transmitted, the network interface state is changed to either
a Dedicated Channel (DCH) or a Forward Access Chan-
nel (FACH). In the DCH mode, a dedicated channel with
high bandwidth and minimum delay is allocated for data
transmission. This mode has the highest power consumption,
whereas a channel in the FACH mode is shared with other
devices. When the network traffic is minimal, FACH is used
to transmit data. The power consumption in the FACH mode
is half of the power consumption DCH mode. The power
consumption in the IDLE mode is approximately one percent
of the power consumption in the DCH mode [30], [31].
FIGURE 2. State machine of RRC protocol in cellular network.
The state transition in the network interface is controlled by
a timer. When the power state is in the DCH mode and trans-
mission has ended, the state remains in the DCH mode for
a period of time set by the timer. After this time expires, the
state is changed to the FACH mode. The time that the network
interface waits for the change after completing transmission is
named Tail time. Note that no data is transmitted at Tail time
but a reasonable amount of energy is consumed due to high
power consumption in the DCH mode. Regardless of this fact,
use of this timer has two benefits: it reduces the delay of the
state transition from the IDLE mode to the DCH mode, and it
decreases the overhead of the signaling operation in channel
allocation when switching between two power states [31].
Almost 60% of the total energy consumption in a 3G network
is related to the Tail time [30].
In this paper, the 3G transmission, reception and IDLE
power are labeled p3G 3G 3G
t , pr , and pIDLE , respectively. The Tail
time is referred to as ttail and its related power is referred to as
ptail . The 3G network interface bit rate is referred to as R3G .
11852
• POWER MODEL FOR Wi-Fi NETWORK
The consumed energy in Wi-Fi networks depends on the
operational states and the amount of transmitted data. A Wi-Fi
interface has three operational states by default: transmission,
reception and IDLE. In the ‘‘802.11 PSM’’ standard, another
state named Sleep exists [32]. When the interface is in the
Sleep state, it wakes up based on a predefined period to check
the input traffic. If no input traffic exists, it reverts to the Sleep
state. Energy consumption in the Sleep state is noticeably
lower than the energy consumption in other states. The Sleep
state considerably decreases the efficiency since the traffic
that arrives in this state must be buffered (or discarded if the
buffer is full) until the time expires and the interface exits
the Sleep mode to manipulate incoming data. This mecha-
nism causes additional delay for reduction of which a new
version of the mentioned standard is introduced as ‘‘adaptive
802.11 PSM’’. This commercial version works as follows:
after data transmission, the network interface remains in the
IDLE state for a period of time named ‘‘PSM timeout’’. PSM
timeout can vary in different devices. In the PSM timeout,
if no traffic arrives for transmission, the state of the network
interface is the Sleep state [32]. The state machine for the
Wi-Fi network interface is schemed in figure 3.
FIGURE 3. State machine for Wi-Fi interface transmission.
The Wi-Fi transmission, reception, IDLE and Sleep power
wifi wifi wifi wifi
are referred to as pt , pr , pIDLE , and psleep , respectively.
The ‘‘PSM timeout’’ and Wi-Fi transmission bit rate are
referred to as ttimeout and Rwifi , respectively.
D. FINANCIAL COST PROFILER
The financial cost includes the money to be paid for using
cloud services, Wi-Fi and 3G networks. The financial cost is
considered to be one of the most significant parameters for
network clients. The cost of transmitting one bit in Wi-Fi
and 3G networks is labeled α and β, respectively, in this
paper. The cost of a cloud service is calculated based on the
consumed CPU cycles; this cost is denoted γ . In the proposed
architecture, α, β, and γ are obtained by the cost profiler.
E. COST GRAPH CONSTRUCTION AND OFFLOADING
STRATEGY EXTRACTION
The offloading operation is modeled in a cost graph. The
collected data are the inputs required to construct the cost
graph. The optimum solution for the offloading is calculated
VOLUME 6, 2018
V. Haghighi, N. S. Moayedian: Offloading Strategy in Mobile Cloud Computing Considering Energy and Delay Constraints
FIGURE 4. Sample call graph and the cost graph corresponding to the sample
call graph.
after the cost graph is constructed. Extracting the offloading
strategy from this cost graph is the most important part of
this architecture and the focus of this paper. This solution
indicated which part of the selected program is executed in
the cloud and which part is executed in the mobile device;
it is referred to as the offloading strategy.
Consider a computing-intensive program which is going
to be run. This program has three steps. The first step is
collecting some information from user. In the second step,
computing-intensive function is run. The third step displays
results on the monitor. The mobile device has low battery
and the specific program must be run in the tight deadline.
Wireless network transmission rate is good for offloading.
By applying the offloading architecture which is shown
in figure 1 the specific program is divided into two parts
which include locally and remotely executions. The extract
offloading strategy is as follows: The first and third steps
are run locally in the mobile device and the second step
is offloaded to the cloud. This offloading solution satisfies
energy and deadline limitations with the least financial cost.
IV. COST GRAPH CONSTRUCTION
The offloading operation is modeled in a graph. This cost
graph consists of all possible solutions for the offloading
operation. The cost graph is constructed based on the program
call graph. This graph is constructed with a set of paths,
each of which declare a certain offloading strategy. Two
execution states are assumed for each method in the call
graph, as shown in figure 4. The first state represents the local
execution of the method in a mobile device, and the second
state represents the remote execution of the method on the
cloud side. In other words, states 1, 2, . . . , n − 1, n represent
local execution and states 1∗ , 2∗ , . . . , n − 1∗ , n∗ represent
execution in the cloud. Some methods cannot be executed in
the cloud since they need at least one resource of the mobile
device. In this case, only one state exists for the method in
the cost graph, which is related to its local execution. These
VOLUME 6, 2018
methods are designated by the programmer. Subsequently,
these states in the cost graph are related by some links. Each
link in the cost graph represents three parameters: financial
cost m, energy e and delay d. These parameters are related
to the required data transfer and execution of the next state
in the cost graph. These parameters are computed in the
computation phase subsection. In some cases, more than one
link exists between two states due to the existence of both 3G
and Wi-Fi networks. This type of link is named a hyperlink
and is represented by the dashed lines in the cost graph.
In general there are two types of offloading modes. The
first mode is on-the-spot. An on-the-spot offloading mode
is applied in this paper because current smartphones use the
on-the-spot mode by default. Using on-the-spot offloading,
the data transmission related to the offloading process can
switch to the cloud using a Wi-Fi network in an automated
sense when it is available. When users exit the Wi-Fi coverage
area, the remainder of the data is transmitted via cellular
networks. The other offloading mode is the Delayed mode.
Using Delayed offloading, data transmission must be finished
by a deadline. In this case, when Wi-Fi networks are regained
by the deadline, data transmission is resumed with Wi-Fi until
all data are transmitted. If the data transfer is not completed
by the assigned deadline, it will be completed via the cellular
network [33].
A. COMPUTATION PHASE
Three parameters of energy, delay and financial cost are
obtained for each link in the cost graph. A cost vector
cost (l) = (m (l) , e (l) , d(l)) for each link l, which includes
these parameters, is defined. Assume that method i calls
method j, which is shown by l(i, j). The three parameters
of link l are computed for transferring data and executing
method j. In general, four cases exists for the execution of the
two methods i and j based on whether they run in the cloud
or a mobile device:
11853
 V. Haghighi, N. S. Moayedian: Offloading Strategy in Mobile Cloud Computing Considering Energy and Delay Constraints

Case 1 (Both Methods i and j Are Executed in a Mobile device .

e(CloudExe)j = d(CloudExe)j × pidle (12)
Device): Here, e (l) is the energy consumption by a mobile
device for executing method j in the mobile device; it is Unlike case 1 and case 2, the hyperlink is a permanent com-
labeled e (DeviceExe)j . Parameter d(l) is the delay caused ponent of the cost graph, which enables the use of Wi-Fi or a
by the execution of method j in the mobile device; it is 3G network for transmission.
labeled d(DeviceExe)j . Parameter m (l) is the financial cost Wi-Fi: If the input data for running method j in the cloud
that corresponds to the execution of method j in the mobile are transmitted by a Wi-Fi network, then the following calcu-
device; it is labeled m(DeviceExe)j . Consequently, lations are performed and
wifi
m (l) = m(DeviceExe)j = 0, (1) m (SendData)j = Ij × α, (13)
wifi wifi −1
d (l) = d(DeviceExe)j = Cj × fdevice
−1
, (2) d(SendData)j = Ij × (R ) , (14)

device , (3)
wifi wifi wifi
e (l) = e(DeviceExe)j = d(DeviceExe)j × pactive e (SendData)j = d (SendData)j × pt
wifi
ttimeout × pidle ,

are yielded. 

d (CloudExe)j > ttimeout

Case 2 (Both Methods i and j Are Executed in the Cloud): 
Here, e (l) is the energy consumption by the mobile device + wifi (15)

 d (CloudExe)j × pidle ,
during the period of time that method j is executed on the 
d (CloudExe)j ≤ ttimeout

cloud side, is labeled e (CloudExe)j . Parameter d(l) is the
delay caused by the execution of method j on the cloud side; it are yielded.
is labeled d(CloudExe)j . The parameter m (l) is the financial The energy consumption of the Wi-Fi network interface in
cost that corresponds to the execution of method j on the cloud the Sleep mode is not considered.
side; it is labeled m(CloudExe)j . Assume that a mobile device 3G: If the input data for running method j on the cloud
is in the IDLE mode when method j is executed on the cloud side are transmitted by the 3G network, then the following
side. Consequently, equations are made and
m (l) = m(CloudExe)j = Cj × γ , (4) j = Ij × β,
m (SendData)3G (16)
d (l) = d(CloudExe)j = −1
Cj × fcloud , (5) d(SendData)3G = Ij × (R ) 3G −1
, (17)
j
e (l) = e(CloudExe)j = device ,
d(CloudExe)j × pidle (6) e (SendData)3G = d (SendData)3G 3G
j j × pt
ttail × ptail ,

are yielded. 

Case 3 (Method i is Executed in a Mobile Device and 
 d(CloudExe)j > ttail
Method j is Executed in the Cloud): In this case, the energy + (18)
d(CloudExe) j ) × ptail ,
consumption e (l) consists of two parts: The consumed




energy by a mobile device for sending the required input d(CloudExe)j ≤ ttail
data for the execution of method j in the cloud is labeled are yielded.
e (SendData)j , and the mobile device energy consumption The energy consumption of the 3G network interface in the
during the time of executing method j on the cloud side is IDLE mode is not considered.
labeled e (CloudExe)j . The total delay d (l) is the sum of the Case 4 (Method i is Executed on the Cloud Side and
delay caused by sending the input data for method j to the Method j is Executed in a Mobile Device): Here, e (l) consists
cloud, which is labeled d(SendData)j and the delay caused of the energy consumption by a mobile device for receiving
by the execution of method j in the cloud, which is labeled the output data of method i that are required for executing
d (CloudExe)j . The total financial cost m (l) is the sum of the method j in the mobile device; it is labeled e (ReceiveData)i ,
financial cost for sending the input data of method j to the and the consumed energy by the mobile device for executing
cloud, which is labeled m (SendData)j , and the financial cost method j in the mobile device is labeled e (DeviceExe)j . The
that corresponds to the execution method j in the cloud, which total delay d (l) is the sum of the delay caused by receiving the
is labeled m (CloudExe)j . The cost vector cost(l) for link l is output data of method i in the mobile device, which is labeled
expressed as: d(ReceiveData)i , and the delay caused by the execution of
method j in the device, which is labeled by d (DeviceExe)j .
m (l) = m (CloudExe)j + m (SendData)j , (7)
The financial cost m (l) is the sum of the financial cost caused
d (l) = d (CloudExe)j + d(SendData)j , (8) by receiving the output data of method i in the mobile device,
e (l) = e (CloudExe)j + e (SendData)j , (9) which is labeled m (ReceiveData)i , and the financial cost that
corresponds to the execution of method j in the mobile device,
where which is labeled m (DeviceExe)j . The cost vector cost(l) for
m(CloudExe)j = Cj × γ , (10) the link l is expressed as follows:
d(CloudExe)j = −1
Cj × fcloud , (11) m (l) = m (ReceiveData)i + m (DeviceExe)j , (19)

11854 VOLUME 6, 2018

V. Haghighi, N. S. Moayedian: Offloading Strategy in Mobile Cloud Computing Considering Energy and Delay Constraints

e (l) = e (DeviceExe)j + e (ReceiveData)i , (20) the method set V and edge set E described in section 4 and
d (l) = d (DeviceExe)j + d(ReceiveData)i , (21) n = |V | and m = |E| be the number of nodes and edges.
Assume that P(S, D) is the set of all paths from the source
where node S to the destination node D in the cost graph. Then,
a feasible path named p∗ with a minimum financial cost is
m (DeviceExe)j = 0, (22)
the optimal solution among all feasible paths. The offloading
d (DeviceExe)j = −1
Cj × fdevice , (23) operation can be formulated as the constrained shortest path
decice ,
d(DeviceExe)j × pactive
e(DeviceExe)j = (24) problem as
min m (p) ,
are yielded. pP(S,D)
Unlike case 1 and case 2, the hyperlink is a permanent subject to : e (p) ≤ B,
component of the cost graph; which enables the use of a Wi-Fi d (p) ≤ 1, (31)
or 3G network for transmission.
Wi-Fi: If the output data of method i for running method where m(p), e (p) and d (p) are the financial cost of path p,
j in the mobile device are received by a Wi-Fi network, then the energy consumed by the mobile device and the delay that
the following calculations are performed and corresponds to the path p in the cost graph, respectively.
wifi This optimization problem is NP-complete [34]; therefore,
m (ReceiveData)i = Oi × α, (25) an algorithm that ends at the theoretical optimum solution
wifi −1 in the polynomial time cannot be found. An approximate
d(ReceiveData)i = Oi × (Rwifi ) (26)
solution for problem (31) is achieved using the Lagrangian
wifi wifi
e (ReceiveData)i = d (ReceiveData)i × pwifir relaxation technique. Assume that L(λ1 , λ2 ) is the Lagrangian

wifi function defined by

 ttimeout × pidle ,
L (λ1 , λ2 ) = min {m (p)+λ1 e(p)+λ2 d (p)}−λ1 B−λ2 1,

d (DeviceExe)j > ttimeout



pP(S,D)
+ wifi


 d (DeviceExe)j × pidle , (32)
d (DeviceExe)j ≤ ttimeout

where λ1 and λ2 are the Lagrange multipliers. Due to the non-



(27) convex nature of the original problem (31), the Lagrangian

function L (λ1 , λ2 ) is a lower bound to problem (31) for any
are yielded. λ1 ≥ 0 and λ2 ≥ 0. The best lower bound of the problem (31)
The energy consumption of the Wi-Fi network interface in is
the Sleep mode is not considered.
3G: If the output data of method i for running method j in max L(λ1 , λ2 ). (33)
λ1 ,λ2 ≤0
the mobile device are received by the 3G network, then the
L(λ1 , λ2 ) is a piecewise linear function. The following func-
following calculations are made and
tion is defined as
i = Oi × β,
m(ReceiveData)3G (28)
Mλ1 ,λ2 (p) = m (p) + λ1 e (p) + λ2 d(p). (34)
3G −1
d(ReceiveData)3G
i = Oi × (R ) , (29)
Minimizing Mλ1 ,λ2 (p) is equivalent to finding the shortest
e (ReceiveData)3G
i = d (ReceiveData)3G i × pr
3G
path from the source node S to the destination node D for
ttail × ptail , a given λ = (λ1 , λ2 ) in the cost graph.



d (DeviceExe)j > ttail A new heuristic algorithm is proposed for solving prob-


+ (30) lem (33) based on the multidimensional Lagrangian relax-
d (DeviceExe) j × ptail ,
ation aggregate cost (M-LARAC) [35] and the Lagrangian


d (DeviceExe)j ≤ ttail

relaxation aggregate cost using the k-shortest paths
are yielded. (K-LARAC) [36] algorithms; both costs are extracted from
The energy consumption of the 3G network interface in the the LARAC [19] algorithm.
IDLE mode is not considered.
A. MULTIDIMENSIONAL LARAC ALGORITHM
V. PROBLEM FORMULATION USING K-SHORTEST PATH
The problem of finding the optimal solution for offloading A new K-M-LARAC algorithm is proposed for solving
operations is converted to find the shortest path in the pre- the optimization problem (33) based on two algorithms:
viously mentioned cost graph. This path must acquire the M-LARAC and K-LARAC. This algorithm maximizes the
minimum financial cost while maintaining the total energy Lagrangian function L(λ1 , λ2 ) in a heuristic manner. This
consumption and completion time below the given values. algorithm is run in several steps; at every step, the k short-
Assume that B is an energy consumption constraint and 1 is a est paths p1 , p2 , . . . , pk in the cost graph for a given λ =
completion time constraint. Let G(V , E) be a cost graph with (λ1 , λ2 ), which are denoted by Pλ , are found. A set of

VOLUME 6, 2018 11855

 V. Haghighi, N. S. Moayedian: Offloading Strategy in Mobile Cloud Computing Considering Energy and Delay Constraints

B. DESCRIPTION OF K-M-LARAC
The steps of the K-M-LARAC algorithm are described as
follows:
1. Whether a solution for the problem exists is verified.
By assuming λ = (λ1 → ∞, λ2 → ∞), the shortest
path for λ is obtained by the Dijkstra algorithm and
named pλ . If pλ does not satisfy B and constraints,
no feasible path for the problem exists.
2. Here, λ is set to λ = λm = (0, 0). Next, the shortest
FIGURE 5. Planes corresponded to three sample paths. path for λm is obtained via the Dijkstra algorithm,
where the metric function Mλ1 ,λ2 in (34) is applied
and named pm . This path is optimized for the financial
cost. If pm satisfies both the 1 and B constraints, then
pm is the optimum solution for this problem. Then,
the algorithm ends; otherwise, step three is run.
3. Here, k shortest paths are calculated for λm , where
set Pλm includes these k shortest paths. If Pλm is the
mixed − type, then
p∗λm = arg min m(p), (35)
FIGURE 6. Lagrangian function that corresponds to the paths from pPB,1
λm
figure 5.
which will become the optimum path and exit the algo-
rithm. Otherwise, if Pλm is not the mixed − type, then
pλm becomes
Pλ = {p1 , p2 , . . . , pk is found, where Mλ (p1 ) ≤ Mλ (p2 ) ≤
pλm = arg min m(p), (36)
. . . Mλ (pk ). Assume that PBλ is a subset of Pλ , including all pPλm
paths that only satisfy the energy constraint B, and P1 λ is a which will become the optimum path for the financial
subset of Pλ with all paths that only satisfy the delay con- cost. Therefore, it does not satisfy one or both of the 1
straint 1. Assume that PB,1 λ is a subset of Pλ , which involves and B constraints.
paths that suit both the 1 constraint and the B constraint. The 4. Here, the k shortest paths are calculated for λe =
shortest path for a given λ = (λ1 , λ2 ) is labeled pλ . (λ1 → ∞, λ2 = 0). The set Pλe includes these k
The set Pλ is classified into nine types according to 1 shortest paths. If Pλe is the mixed − type, then
and B requirements; five types are important. For the first
type, which is named pN −type , no path satisfies the 1 and p∗λe = arg min m(p), (37)
B constraints. For the second type, which is named pB−type , pPB,1
λe

all paths only suit the B constraint. For the third type, which which will become the optimum path and exit the algo-
is named p1−type , all paths satisfies the 1 constraint. All rithm. Otherwise, if Pλe is not the mixed − type, then
paths satisfy both 1 and B constraints for the fourth type Pλ , pλe becomes
which is named pB,1−type . In the last type, which is named
pmixed−type , some paths are feasible, whereas other paths are pλe = arg min m(p), (38)
pPλe
not feasible.
which will become the path with an energy constraint
The λ1 space and λ2 space can be separated based on Pλ
that has the minimum financial cost.
types. For small values of λ1 and λ2 , Pλ will be the pN −type .
5. Here, the k shortest paths are calculated for λd = (λ1 =
An increase in λ1 would push Pλ closer to pB−type , which indi-
0, λ2 → ∞). The set Pλd includes these k shortest
cates that the energy parameter is more dominant. An increase
paths. If Pλd is the mixed − type, then
in λ2 would push Pλ closer to p1−type , which indicates
that delay is the dominant parameter. With a simultaneous p∗λd = arg min m(p), (39)
increase in λ1 and λ2 values, Pλ tends to be pmixed−type , and pPB,1
λ d
Pλ is pB,1−type . When a Pλ set of type pmixed−type is achieved,
which will become the optimum path and exit the algo-
an optimized path is obtained. If appropriate values are cho-
rithm. Otherwise, if Pλd is not the mixed − type, then
sen for the energy and delay constraints, achieving a Pλ
pλd becomes
set of pmixed−type becomes possible. Three sample paths are
shown in figure 5; their corresponding Lagrangian function pλd = arg min m(p), (40)
pPλd
is shown in figure 6. Each path yields a plane that matches
F (λ1 , λ2 ) = m (p) + λ1 (e (p) − B) + λ2 (d (p) − 1) with λ1 which will become the path with a delay constraint that
and λ2 parameters. has the minimum financial cost.

11856 VOLUME 6, 2018

V. Haghighi, N. S. Moayedian: Offloading Strategy in Mobile Cloud Computing Considering Energy and Delay Constraints

6. In this step, the three primary paths of pλm , pλe and

pλd obtained from previous steps are concerning. The
function L (λ1 , λ2 ) will be maximized for λ1 and λ2
based on (43) and (44). These values of λ1 and λ2
are new candidates in achieving the optimum path.
The k shortest paths for the new λ = (λ1 , λ2 ) are
calculated by the metric function Mλ1 ,λ2 in (34). The
set Pλ includes these k shortest paths. If Pλ is the
mixed − type, then p∗λ is the optimum path based on
the following equation and exits the algorithm:
p∗λ = arg min m(p), (41)
pPB,1
λ

Otherwise, if Pλ is not the mixed − type, pλ becomes:

pλ = arg min m(p). (42)
pPλ
7. Here, either pλm , pλe or pλd are replaced with pλ accord-
ing to normal vectors of the paths. Each of the primary
paths pλm , pλe or pλd , with a minimum angle between
its normal vector and its new pλ , will be replaced with
pλ . Steps 6 and 7 are repeated until the mixed − type
set is achieved.
The running time of K-M-LARAC algorithm depends on
the number of iterations and the computation during each
iteration. The LARAC algorithm terminates after O(mlog3 m),
which indicates that the running time of the algorithm
is O(m2 log4 m) [19]. The complexity of M-LARAC is FIGURE 7. K-M-LARAC algorithm.
O(c3 m2 log4 m) [35] and parameter c is the number of the con-
sidered metrics. Three metrics exist in the problem introduced
in (31): financial cost, energy consumption and delay. Each
iteration consists of construction a heap-ordered data struc-
ture, finding first k paths and sorting these paths. Therefore,
the worst time complexity of the K-M-LARAC algorithm is
O(c3 m2 log4 m. (m + nlogn + KlogK )) [36].

VI. EXTRACTING ONLINE OFFLOADING STRATEGY

The transmission rate for the Wi-Fi and 3G networks fluctu-
ates; their availability varies over time and sites. Therefore,
the information provided by the network profiler is not valid,
and the offloading solution can change over different condi-
tions. Extracting the offloading strategy from the cost graph
should become adapted to network changes.
The pseudo code for the K-M-LARAC algorithm is shown
in figure 7, and the online algorithm for extracting the
offloading solution from the cost graph is shown in figure 8.
The extracting online offloading strategy is as follows: first,
the provided information is obtained from profilers; second,
the cost graph G∗ is constructed according to the call graph
G; and third, the parameters of energy, delay and financial
cost are calculated for each link inG∗ , The offloading solu-
tion is achieved by the K-M-LARAC algorithm in the cost
graph G∗ . This process with the new information from the
network profiler is repeated in each stage of G∗ . If networks’
conditions encounter numerous changes, then a new path may
be obtained for the offloading operation, which is adapted to
the new networks’ conditions.
FIGURE 8. Extracting offloading strategy online.
VII. NUMERICAL RESULTS
The potency of the K-M-LARAC algorithm in finding an
optimal solution in an offloading operation is evaluated.
A typical smartphone with the hardware characteristics listed
in table 1 is considered [37], [38]. This information is related
to processor, Wi-Fi and 3G networks’ interfaces for mobile
device and cloud processor capabilities. Here, a cloud is
considered as a desktop computer with a high processing
capability.
Some parts of a face recognition application is applied to
evaluate the benefits obtained from the K-M-LARAC algo-
rithm [39]. The call graph that corresponds to the parts of the
face recognition application is shown in figure 9. Each circle
represents a method and the number of CPU cycles that are

VOLUME 6, 2018 11857

 V. Haghighi, N. S. Moayedian: Offloading Strategy in Mobile Cloud Computing Considering Energy and Delay Constraints

FIGURE 9. Face recognition application call graph.

TABLE 1. Hardware characteristics for the mobile device and cloud.
needed to execute this method, and each edge is specified by
the required input and output data for offloading the method.
Three strategies are considered for offloading to the cloud:
the application is executed in the mobile device, where no
offloading operation exists and the purposed application is
locally executed. The application is executed on the cloud
side, where the offloading strategy is the predefined cloud
execution and an appropriate offloading strategy is obtained
according to this proposed offloading architecture and the
K-M-LARAC algorithm at k = 10 and k = 20. These three
execution strategies are compared with the optimal solution
because the K-M-LARAC algorithm sometimes finds the
approximate solution. The optimum solution is obtained via
a brute force search.
These three strategies are applied in different conditions
for Wi-Fi and 3G networks. The quality of the networks is the
primary requirement for the offloading operation. The energy
and delay constraints are effective in extracting the offloading
solution. Therefore, these three strategies are applied at differ-
ent ranges of energy and delay constraints subject to various
networks’ conditions. The energy and delay constraints are
divided into three ranges: high range, medium range and low
range.
At any given time, various applications are being run
in a mobile device, the energy consumption of which is
unknown. Only one specified remote application (face recog-
nition application) is running in the mobile device with a
known time execution and energy consumption rate. There-
fore, the values of the energy and delay constraints are deter-
mined according to the local execution of the face recognition
application in the mobile device. The energy consumption,
delay and financial cost for different offloading strategies
subject to various ranges of energy and delay constraints are
shown in figure 10, where both the Wi-Fi and 3G rates are
appropriate. The larger is the number of limitations in the
energy and delay constraints, the closer is the K-M-LARAC
solution to the cloud execution strategy.
Finding the solution that is similar to the cloud execution
strategy indicates that a greater financial cost is necessary to
overcome the energy and delay constraints. When the energy
and delay limitations are reduced, the K-M-LARAC solution
tends to be the device execution strategy, which demon-
strates that the mobile device executes the application when
no battery and time limitations exist because device exe-
cution has no financial cost. Sometimes, the K-M-LARAC
algorithm finds the solution close to the optimum because
K-M-LARAC finds the lower bound for this problem and
achieves the best solution that is close to the lower bound.
If the K-M-LARAC approximate solution is the optimum,
the lower bound is close to the optimum solution. The Wi-Fi
network has a lower financial cost and a higher bit rate than
the financial cost and bit rate of the 3G network. When
the Wi-Fi and 3G rates are appropriate, the K-M-LARAC
algorithm uses the Wi-Fi interface for the offloading oper-
ation, as shown in figure 10. When the Wi-Fi rate is low,
the K-M-LARAC uses the 3G network for the offloading










m pλe d pλm − d pλd + m pλm d pλd − d pλe + m pλd d pλe − d pλm
λ1 :








 (43)
d pλe e pλm − e pλd + d pλm e pλd − e pλe + d pλd e pλe − e pλm










m pλe e pλm − e pλd + m pλm e pλd − e pλe + m pλd e pλe − e pλm
λ2 :








 (44)
e pλe d pλm − d pλd + e pλm d pλd − d pλe + e pλd d pλe − d pλm

11858 VOLUME 6, 2018

V. Haghighi, N. S. Moayedian: Offloading Strategy in Mobile Cloud Computing Considering Energy and Delay Constraints

FIGURE 10. Energy consumption, delay and financial cost for different offloading strategies for very good network conditions.
The horizontal lines show the limitations of the offloading problem.

FIGURE 11. Energy consumption, delay and financial cost for different offloading strategies for moderate network conditions.
The horizontal lines represent the limitations of offloading problem.

operation. The Wi-Fi and 3G rates are average and the
K-M-LARAC algorithm uses the Wi-Fi network for the
offloading operation, as shown in figure 11.
When the K-M-LARAC lower bound is far from the opti-
mum solution, the obtained approximate solution is far from
the optimum solution. An increase in the K parameter in the
K-M-LARAC algorithm will yield a better lower bound that is
close the optimum solution. The experimental results indicate
the impact of the K parameter in finding the optimum solu-
tion. The K-M-LARAC at K = 20 finds the better solution

VOLUME 6, 2018 11859

 V. Haghighi, N. S. Moayedian: Offloading Strategy in Mobile Cloud Computing Considering Energy and Delay Constraints
compared with the K-M-LARAC at K = 10. These results
indicate that this proposed K-M-LARAC algorithm finds the
approximate solution with the minimum financial cost, which
satisfies the energy and delay constraints. By setting the
K parameter for the specified program, the optimum solution
for the offloading problem subject to different networks’
conditions is achieved.
We consider the mobile device which has the average
battery and the face recognition application which should be
executed at a moderate deadline. Networks conditions are
good for offloading. To clarify the obtaining results in this
section, we show the offloading strategy which is achieved
by K-M-LARAC algorithm in the figure 12. Some of the
methods of face recognition application which are shown
in figure 9 are offloaded to the cloud side.
FIGURE 12. Offloading strategy when networks conditions are good and
there are moderate limitations for energy and delay parameters.
VIII. CONCLUSION
Smartphones have become a computational tool for executing
sophisticated programs. To assure the advantages of portabil-
ity and overcome the limitations of energy consumption and
the delay caused by processing, an efficient solution is sought
for offloading operations. In this paper, a new architecture is
proposed for offloading operations.
The most important part of this architecture is model-
ing the offloading operation as a cost graph by consid-
ering both Wi-Fi and 3G links and extracting the proper
offloading solution from this graph. The new heuristic algo-
rithm K-M-LARAC is proposed to extract the approximate
solution from the cost graph with polynomial complexity.
K-M-LARAC, in some conditions, achieves the optimum
solution, whereas an arbitrary approximate solution is found
for offloading operations in other conditions. An appropriate
K parameter set-up, according to a specific program, would
enable the optimum solution to be obtained.
In future studies, the K parameter would be set in
K-M-LARAC based on a specific program. The linear call
graph is considered for the program, whereas the call graph
can be extended to nonlinear methods in the program. Energy
can be conserved in the period of the Tail time by responding
to several incoming requests and offloading several methods
in a simultaneous manner.
11860
REFERENCES
[1] A. R. Khan, M. Othman, S. A. Madani, and S. U. Khan, ‘‘A survey of
mobile cloud computing application models,’’ IEEE Commun. Surveys
Tuts., vol. 16, no. 1, pp. 393–413, 1st Quart., 2014.
[2] M. Armbrust et al., ‘‘A view of cloud computing,’’ Commun. ACM, vol. 53,
no. 4, pp. 50–58, Apr. 2010.
[3] T. Dillon, C. Wu, and E. Chang, ‘‘Cloud computing: Issues and chal-
lenges,’’ in Proc. IEEE Int. Conf. Adv. Inf. Netw. Appl. (AINA), Apr. 2010,
pp. 27–33.
[4] K. Kumar and Y.-H. Lu, ‘‘Cloud computing for mobile users: Can
offloading computation save energy?’’ Computer, vol. 43, no. 4,
pp. 51–56, Apr. 2010.
[5] J. Flinn, D. Narayanan, and M. Satyanarayanan, ‘‘Self-tuned remote exe-
cution for pervasive computing,’’ in Proc. 8th Workshop Hot Topics Oper.
Syst. (HotOS), Elmau, Germany, May 2001, pp. 61–66.
[6] J. Flinn, S. Park, and M. Satyanarayanan, ‘‘Balancing performance, energy,
and quality in pervasive computing,’’ in Proc. 22nd Int. Conf. Distrib.
Comput. Syst. (ICDCS), Vienna, Austria, Jul. 2002, pp. 217–226.
[7] R. Balan, J. Flinn, M. Satyanarayanan, S. Sinnamohideen, and H.-I. Yang,
‘‘The case for cyber foraging,’’ in Proc. 10th ACM SIGOPS Eur. Workshop,
Saint-Emilion, France, Sep. 2002, pp. 87–92.
[8] R. K. Balan, M. Satyanarayanan, S. Y. Park, and T. Okoshi, ‘‘Tactics-
based remote execution for mobile computing,’’ in Proc. 3rd Int. Conf.
Mobile Syst., Appl., Services (MobiSys), San Francisco, CA, USA, 2003,
pp. 273–286.
[9] S. Osman, D. Subhraveti, G. Su, and J. Nieh, ‘‘The design and implemen-
tation of Zap: A system for migrating computing environments,’’ in Proc.
5th Symp. Oper. Syst. Design Implement. (OSDI), 2002, pp. 361–376.
[10] M. Satyanarayanan, P. Bahl, R. Caceres, and N. Davies, ‘‘The case for
VM-based cloudlets in mobile computing,’’ IEEE Pervasive Comput.,
vol. 8, no. 4, pp. 14–23, Oct./Dec. 2009.
[11] E. Cuervo et al., ‘‘MAUI: Making smartphones last longer with code
offload,’’ in Proc. 8th Int. Conf. Mobile Syst., Appl., Services (MobiSys),
Mar. 2010, pp. 49–62.
[12] B.-G. Chun and P. Maniatis, ‘‘Augmented smartphone applications
through clone cloud execution,’’ in Proc. 8th Workshop Hot Topics Oper.
Syst. (HotOS), Monte Verità, Switzerland, May 2009, p. 8.
[13] B.-G. Chun, S. Ihm, P. Maniatis, M. Naik, and A. Patti, ‘‘CloneCloud:
Elastic execution between mobile device and cloud,’’ in Proc. 6th Eur.Conf.
Comput. Syst. (EuroSys), 2011, pp. 301–304.
[14] S. Kosta, A. Aucinas, P. Hui, R. Mortier, and X. Zhang, ‘‘ThinkAir:
Dynamic resource allocation and parallel execution in the cloud for
mobile code offloading,’’ in Proc. IEEE IEEE Conf. Comput. Commun.,
Mar. 2012, pp. 945–953.
[15] Y. Wen, W. Zhang, and H. Luo, ‘‘Energy-optimal mobile application
execution: Taming resource-poor mobile devices with cloud clones,’’ in
Proc. Annu. IEEE INFOCOM: Mini-Conf., Mar. 2012, pp. 2716–2720.
[16] W. Zhang, Y. Wen, K. Guan, D. Kilper, H. Luo, and D. O. Wu, ‘‘Energy-
optimal mobile cloud computing under stochastic wireless channel,’’ IEEE
Trans. Wireless Commun., vol. 12, no. 9, pp. 4569–4581, Sep. 2013.
[17] W. Zhang, Y. Wen, and D. O. Wu, ‘‘Energy-efficient scheduling policy
for collaborative execution in mobile cloud computing,’’ in Proc. IEEE
INFOCOM, Apr. 2013, pp. 190–194.
[18] W. Zhang, Y. Wen, and D. O. Wu, ‘‘Collaborative task execution in
mobile cloud computing under a stochastic wireless channel,’’ IEEE Trans.
Wireless Commun., vol. 14, no. 1, pp. 81–93, Jan. 2015.
[19] A. Juttner, B. Szviatovski, I. Mecs, and Z. Rajko, ‘‘Lagrange relaxation
based method for the QoS routing problem,’’ in Proc. IEEE INFOCOM,
vol. 2. Apr. 2001, pp. 859–868.
[20] M. V. Barbera, S. Kosta, A. Mei, and J. Stefa, ‘‘To offload or not to offload?
The bandwidth and energy costs of mobile cloud computing,’’ in Proc.
IEEE INFOCOM, Apr. 2013, pp. 1285–1293.
[21] D. Huang, P. Wang, and D. Niyato, ‘‘A dynamic offloading algorithm
for mobile computing,’’ IEEE Trans. Wireless Commun., vol. 11, no. 6,
pp. 1991–1995, Jun. 2012.
[22] W. Zhang and Y. Wen, ‘‘Energy-efficient task execution for application
as a general topology in mobile cloud computing,’’ IEEE Trans. Cloud
Comput., to be published.
[23] H. Al-Shatri, S. Müller, and A. Klein, ‘‘Distributed algorithm for
energy efficient multi-hop computation offloading,’’ in Proc. IEEE ICC,
Kuala Lumpur, Malaysia, May 2016, pp. 1–6.
[24] S. Guo, B. Xiao, Y. Yang, and Y. Yang, ‘‘Energy-efficient dynamic offload-
ing and resource scheduling in mobile cloud computing,’’ in Proc. IEEE
INFOCOM, San Francisco, CA, USA, Apr. 2016, pp. 1–9.
VOLUME 6, 2018
V. Haghighi, N. S. Moayedian: Offloading Strategy in Mobile Cloud Computing Considering Energy and Delay Constraints
[25] L. Tong and W. Gao, ‘‘Application-aware traffic scheduling for workload
offloading in mobile clouds,’’ in Proc. IEEE INFOCOM, San Francisco,
CA, USA, Apr. 2016, pp. 1–9.
[26] F. Xia, F. Ding, J. Li, X. Kong, L. T. Yang, and J. Ma, ‘‘Phone2Cloud:
Exploiting computation offloading for energy saving on smartphones in
mobile cloud computing,’’ Mobile Cloud Comput. Inf. Syst. Frontiers,
vol. 16, no. 1, pp. 95–111, Mar. 2014.
[27] H. Shah-Mansouri, V. W. Wong, and R. Schober, ‘‘Joint optimal pricing
and task scheduling in mobile cloud computing systems,’’ IEEE Trans.
Wireless Commun., vol. 16, no. 8, pp. 5218–5232, Aug. 2017.
[28] X. Wang, J. Wang, X. Wang, and X. Chen, ‘‘Energy and delay tradeoff for
application offloading in mobile cloud computing,’’ IEEE Syst. J., vol. 11,
no. 2, pp. 858–867, Jun. 2017.
[29] Y. Geng, W. Hu, Y. Yang, W. Gao, and G. Cao, ‘‘Energy-efficient compu-
tation offloading in cellular networks,’’ in Proc. IEEE 23rd Int. Conf. Netw.
Protocols (ICNP), Nov. 2015, pp. 145–155.
[30] N. Balasubramanian, A. Balasubramanian, and A. Venkataramani,
‘‘Energy consumption in mobile phones: A measurement study and impli-
cations for network applications,’’ in Proc. 9th ACM SIGCOMM Conf.
Internet Meas., Nov. 2009, pp. 280–293.
[31] F. Qian, Z. Wang, A. Gerber, Z. M. Mao, S. Sen, and O. Spatscheck,
‘‘Characterizing radio resource allocation for 3G networks,’’ in Proc. 10th
ACM SIGCOMM Conf. Internet Meas. (IMC), Nov. 2010, pp. 137–150.
[32] Y. Xiao et al., ‘‘Modeling energy consumption of data transmission over
Wi-Fi,’’ IEEE Trans. Mobile Comput., vol. 13, no. 8, pp. 1760–1773,
Aug. 2014.
[33] K. Lee, J. Lee, Y. Yi, I. Rhee, and S. Chong, ‘‘Mobile data offloading: How
much can WiFi deliver?’’ in Proc. 6th Int. ACM Conf., Nov. 2010, p. 26.
[34] M. R. Garey and D. S. Johnson, Computers and Intractability: A Guide to
the Theory of NP-Completeness. New York, NY, USA: Freeman, 1979.
[35] K. Stachowiak and P. Zwierzykowski, ‘‘Lagrangian relaxation and linear
intersection based QoS routing algorithm,’’ Int. J. Electron. Telecommun.,
vol. 58, no. 4, pp. 307–314, Dec. 2012.
[36] Z. Jia and P. Varaiya, ‘‘Heuristic methods for delay constrained least cost
routing using κ-shortest-paths,’’ IEEE Trans. Autom. Control, vol. 51,
no. 4, pp. 707–712, Apr. 2006.
VOLUME 6, 2018
[37] A. Carroll and G. Heiser, ‘‘An analysis of power consumption in a smart-
phone,’’ in Proc. USENIX Annu. Tech. Conf., vol. 14. Jun. 2010, p. 21.
[38] L. Zhang et al., ‘‘Accurate online power estimation and automatic
battery behavior based power model generation for smartphones,’’
in Proc. IEEE/ACM/IFIP Int. Conf. Hardw./Softw. Codesign Syst.
Synth. (CODES+ISSS), Oct. 2010, pp. 105–114.
[39] Y. Zhang, H. Liu, L. Jiao, and X. Fu, ‘‘To offload or not to offload:
An efficient code partition algorithm for mobile cloud computing,’’
in Proc. IEEE 1st Int. Conf. Cloud Netw. (CLOUDNET), Nov. 2012,
pp. 80–86.
VENUS HAGHIGHI was born in Roudan, Iran,
in 1990. She received the B.S. degree in computer
engineering from the Shahid Bahonar University
of Kerman in 2012, and the M.S. degree from the
Isfahan University of Technology in 2016.
NAGHMEH S. MOAYEDIAN (S’06–M’11) was
born in Pasau, Germany, in 1977. She received
the B.S., M.S., and Ph.D. degrees in electrical
engineering from the Isfahan University of Tech-
nology, Isfahan, Iran, in 1999, 2002, and 2009,
respectively.
Since 2010, she has been with the ECE
Department, Isfahan University of Technology.
Her research interests include cloud networks,
nanonetworks, and wireless networks.
11861