5610 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 64, NO.

12, DECEMBER 2015

Toward Vehicle-Assisted Cloud

Computing for Smartphones
Hongli Zhang, Qiang Zhang, and Xiaojiang Du

Abstract—Mobile cloud computing is an emerging technology the application performance of mobile devices. As the comput-
for facilitating complex application execution on smartphones. ing capability of smartphones increases, mobile cloud comput-
Cloud services are utilized not only to speed up the running of ing can organize and make use of computation resources on
mobile applications but to save energy for smartphones as well.
In this paper, we propose to combine the vehicular cloud with surrounding smartphones to form the ad hoc virtual cloud [2].
the infrastructure-based cloud to expand the current available In addition to the ad hoc virtual cloud, cloud service providers
resources for task requests from smartphones. In our proposed also include a central cloud and a cloudlet that are referred to as
architecture, the vehicular cloud acts as a cloud service provider the infrastructure-based cloud. Cloudlets [3] are deployed near
for smartphones. Moreover, we propose a flexible offloading strat- Wi-Fi access points (APs) and cellular base stations to provide
egy (FOS) to carry out task migration. The vehicular cloud is
able to discover and utilize the underutilized resources in vehicles cloud services efficiently and decrease the network delay be-
to accomplish application offloading for smartphones. The FOS tween mobile users and the central cloud. Vehicles as mobile
estimates the efficiency of various cloud service providers based on devices can be also served by mobile cloud computing [4], [5].
current resource conditions and then selects the suitable cloud ser- For vehicles, cloud service providers include the central cloud,
vice provider to perform the requested task. Experimental results the roadside cloud, and the vehicular cloud [6]. Vehicular cloud
show that the proposed approach can improve the performance of
mobile applications on smartphones in terms of task response time is defined as [7] a group of largely autonomous vehicles whose
and energy consumption. corporate computing, sensing, communication, and physical
resources can be coordinated and dynamically allocated to au-
Index Terms—Cloudlet, mobile cloud computing, task migra-
tion, vehicular cloud. thorized users. Vehicles traveling on the road may meet various
events such as traffic congestion and traffic accidents. In such
I. I NTRODUCTION situations, the vehicular cloud has the potential to cooperate
with various authorities to solve problems that otherwise cannot

S MARTPHONES have become very popular because they

are portable and can run many kinds of applications. How-
ever, the portability of smartphones also limits their size and
be solved efficiently [8]. Vehicular cloud will play an important
role in implementing autonomous traffic, vehicle control, and
perception systems [9].
weight; hence, some resources on smartphones are limited, for To get good application performance of smartphones, appli-
example, computation resources and storage resources. Com- cation offloading is employed to speed up application execution
puting power and memory capacity of smartphones are gradually and save energy for a smartphone [10]. Offloading is a kind of
growing, but they still cannot meet the computation resource mechanism utilized to alleviate resource constraints of mobile
requirements of some mobile applications that are usually devices by migrating part or all of the tasks corresponding
computationally intensive. Many complex applications have to an application to resource-rich surrogates [11]. In mobile
poor performance when they are performed by smartphones, cloud computing, application offloading mainly means that
for instance, image processing, gaming, and so on. Due to the smartphones utilize cloud services to execute some tasks to
rich resources in the cloud platform, cloud computing is utilized implement good application performance. Two offloading tech-
to facilitate efficient application execution on smartphones. niques have been presented, including system-level offloading
Mobile cloud computing [1] is an emerging technology that and method-level offloading. For example, Clone Cloud [12]
integrates cloud computing and mobile computing to enhance and Cloudlet [13] are categorized as system-level offloading.
In system-level offloading, a cloned image of a smartphone
Manuscript received April 13, 2015; revised July 9, 2015; accepted is created by virtual machine technology and maintained in
September 2, 2015. Date of publication September 18, 2015; date of current
version December 14, 2015. This work was supported in part by the National a cloud platform. MAUI [14] is categorized as method-level
Basic Research Program of China (973 Program) under Grant 2011CB302605, offloading. In method-level offloading, program partitioning
by the National Natural Science Foundation of China under Grant 61202457, by and fine-grained code migration are implemented.
the National Science Foundation under Grant CNS-1022552 and Grant CNS-
1065444, and by the U.S. Army Research Office under Grant WF911NF-14-1- Some application offloading policies have been proposed
0518. The review of this paper was coordinated by the Guest Editors. to decide on whether the tasks should be migrated from a
H. Zhang and Q. Zhang are with the School of Computer Science and smartphone to a cloud node or be locally executed. However,
Technology, Harbin Institute of Technology, Harbin 150001, China (e-mail:
zhanghongli@hit.edu.cn; zhangqiang@pact518.hit.edu.cn). most research only focuses on the problem of whether the task
X. Du is with the Department of Computer and Information Sciences, Temple should be offloaded to a dedicated cloud node such as the
University, Philadelphia, PA 19122 USA (e-mail: dxj@ieee.org). central cloud or the cloudlet. When the dedicated cloud node
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. cannot satisfy the offloading requirements, the cloud services
Digital Object Identifier 10.1109/TVT.2015.2480004 become useless for smartphones. Therefore, it is necessary to
0018-9545 © 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
ZHANG et al.: TOWARD VEHICLE-ASSISTED CLOUD COMPUTING FOR SMARTPHONES 5611

integrate various usable resources to enhance the scalability time is decided by the data size and the bandwidth. The work
of cloud services for smartphones. Once various computation argues that when the data size is large, offloading may not
resources are integrated and utilized, there will be multiple save energy. However, the problem of how to process the data-
cloud nodes to provide cloud services. When one of the cloud intensive task was not discussed. Moreover, the performance
nodes cannot satisfy the offloading requirements, other cloud metric of task response time was not analyzed.
nodes may be available for cloud services. Thus, the cloud ser- An optimization framework for energy-optimal application
vices are still useful for smartphones. In addition, there may execution [16] is proposed to optimally execute mobile ap-
be a situation where several cloud nodes satisfy the offloading plications in either the mobile device or the cloud clone. The
requirements. Then, how to select an optimal cloud node to corresponding model contains the mobile phone and a central
execute the application needs to be studied. cloud. The proposed method can calculate the minimum com-
In this paper, we propose to combine the vehicular cloud putation energy consumption and the minimum transmission
with the fixed central cloud and cloudlet to expand the current energy consumption with a time deadline on the mobile phone.
available resources for task requests from smartphones. In the The proposed framework decides on where to execute the
proposed cloud computing architecture, vehicular cloud near application according to the energy consumption conditions of
the cloudlet acts as a cloud service provider for smartphones. local execution and cloud clone. However, when the task is
When infrastructure-based cloud does not satisfy the offloading data intensive, the data transmission time may exceed the time
requirements, the proposed flexible offloading strategy (FOS) deadline due to wide area network (WAN) delay. The work did
discovers and utilizes the underutilized resources in vehicles to not consider this situation.
accomplish mobile application offloading. If both the central Cloudlet is designed to obtain the resource benefits of cloud
cloud and cloudlet satisfy the offloading requirements, the computing without incurring WAN delays and jitter corre-
FOS selects the optimal cloud node (central cloud or cloudlet) sponding to the central cloud [13]. A cloudlet can be viewed as
according to the energy consumption of the smartphone and the a cluster of multicore computers with Internet connectivity and
task response time. a high-bandwidth wireless local area network (WLAN), and
The rest of this paper is organized as follows. Section II the corresponding communication range is one hop. The com-
introduces the related work. Section III presents the system putation and storage resources in the cloudlet can be utilized
model and states the problem. Section IV describes the pro- to perform application offloading by nearby mobile computers.
posed flexible application offloading strategy for smartphones. However, if no cloudlet is available nearby, a mobile user cannot
Section V presents the numerical performance evaluation, and migrate the data-intensive task efficiently. The performance
Section VI concludes this paper. metric of energy consumption of smartphones was not analyzed.
Moreover, when many users have requests, the resources in
a cloudlet may be insufficient. This situation was not discussed.
II. R ELATED W ORK
A unified elastic computing platform [2] is proposed to sup-
Mobile application offloading and the corresponding offload- port application offloading for mobile devices and reduce en-
ing strategy have been investigated by researchers. Currently, ergy consumption of smartphones based on [16]. The proposed
the central cloud and cloudlet are cloud service providers most model consists of an infrastructure-based cloud and an ad hoc vir-
commonly used by smartphone users. Researchers have mainly tual cloud formed by a cluster of smartphones. The researchers
focussed on the analysis of energy consumption and data trans- presented an offloading policy to decide on where each task
mission time when a smartphone migrates a computationally of the application should be executed (i.e., on the standalone
intensive task to the cloud side. Our work considers the tasks smartphone, in the ad hoc virtual cloud, or in the infrastructure-
corresponding to different cases of data communication and based cloud). The offloading performance of the ad hoc virtual
computation workload. Furthermore, we analyze the offloading cloud was not analyzed and discussed. There is no comparison
performance corresponding to the central cloud and cloudlet between the infrastructure-based cloud and the ad hoc virtual
based on the related work. Finally, we select the suitable cloud in terms of energy consumption of smartphones and task
cloud node to perform the requested task using our proposed response time. The offloading policy only implemented the selec-
offloading strategy. Hence, we first introduce the related work tion between the infrastructure-based cloud and local execution.
on the analysis of offloading performance and the correspond- Most of the current work considered a single cloud service
ing offloading strategy. Then, we give the comparison between provider for smartphones. It is very possible that the single
our work and current methods. cloud service provider becomes useless because its resource
An analysis [15] on the energy saving of cloud computing for conditions are unsuitable for the requested task. For example,
mobile users is proposed. The researchers point out that cloud the central cloud is unsuitable for a data-intensive task, consid-
computing can potentially save energy for mobile users. How- ering the low data transmission rate of a WAN. In particular,
ever, not all applications are energy efficient when migrated to if the data-intensive task belongs to a real-time application,
the cloud. The corresponding model contains the mobile device the transmission delay of a WAN is possibly unacceptable
and a central cloud. The work focuses on the energy analysis for users. Although the cloudlet makes cloud services better
for mobile application offloading. When a smartphone uses for the data-intensive task, it also has limitations. Since the
application offloading, its energy consumption is mostly caused computation and storage resources in the cloudlet are limited,
by wireless transmission. Hence, the time to transmit data is a when a large number of users have requests, it is possible
key factor for energy consumption. Moreover, the transmission that the cloudlet does not have sufficient resources to provide
5612 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 64, NO. 12, DECEMBER 2015

smartphone on a bus can directly connect to the bus through

Wi-Fi technology. Thus, in VACC, smartphones can access
three types of cloud nodes, including the central cloud, the
cloudlet, and the vehicular cloud.
Cloud nodes in VACC are heterogeneous; hence, their cor-
responding cloud services are different on resource conditions,
including computing power, memory capacity, network band-
width, network connectivity, and so on. Generally, a central
cloud has rich computation and storage resources, but it has
a low data transmission rate because it transfers data through a
WAN. Cloudlet has limited computation and storage resources,
but it has a high data transmission rate since it uses a WLAN
Fig. 1. VACC architecture. or a MAN for data transmission. Meanwhile, a cloudlet as a
public infrastructure has a high resource utilization rate due to
cloud services. Moreover, the one-hop communication range its convenience and a large number of service requests. Each
of the cloudlet is also limited. To efficiently offload the tasks node in a vehicular cloud has limited computation and storage
corresponding to different cases of data communication and resources, but the total computation and storage resources of a
computation workload to a suitable cloud service provider, our vehicular cloud are rich. In addition, the resources in a vehicular
work analyzes the offloading performance of various cloud cloud are dispersed, and the resource utilization rate is low.
nodes in terms of task response time and energy consumption. However, the vehicular cloud has a high data transmission rate
To relieve the limitation of the cloudlet, our work establishes a for smartphones because it employs WLAN or MAN commu-
communication model to expand the communication range of nications. According to the resource features of various cloud
the cloudlet and employs the resources in the vehicular cloud to nodes, we give the assumptions as follows. The computation
finish mobile application offloading for smartphones. Currently, and storage resources in the central cloud are unlimited, even
most research studies aim to perform complex applications on with heavy load of offloading requests from many users. The
vehicles and improve the intelligent transport system using the computation and storage resources in the cloudlet are limited.
vehicular cloud [4]–[8]. However, for an application on a vehi- When many users have requests, the resources in the cloudlet
cle, the performance of the application executed by the vehic- may be insufficient.
ular cloud was not analyzed. Our work analyzes the offloading VACC adopts the system-level offloading technique (i.e.,
performance of the vehicular cloud for smartphones. Based on virtual machine) for the central cloud and the cloudlet. We
the mobility and resource conditions of worker nodes in the ve- consider that, normally, the assigned virtual machine for a
hicular cloud, we design an algorithm to select a reliable worker smartphone cannot change its total amount of computation and
node for the requested task. When a smartphone user performs storage resources when it is working. For the vehicular cloud,
task migration, our work considers both the characteristics of VACC adopts the method-level offloading technique consider-
the requested task and the resource conditions of various cloud ing moderate computation and storage resources in vehicles.
nodes to select the suitable cloud node. Compared with the cur- For a mobile application, it may contain one or more tasks.
rent work, our work enhances the scalability and feasibility of This paper only discusses the case in which an application only
cloud services for smartphones in the aspect of task migration. contains one task, and the corresponding task cannot be divided
into several parts. Considering that only a task needs to be
migrated for a requested application, this paper only selects a
III. S YSTEM M ODEL AND P ROBLEM S TATEMENT
worker node (i.e., a vehicle) in the vehicular cloud to perform
Fig. 1 demonstrates the proposed vehicle-assisted cloud com- the requested task for smartphones. We will discuss the case
puting (VACC) architecture. VACC contains a central cloud, a in which an application contains multiple tasks in our future
cloudlet, and a vehicular cloud. Vehicular cloud is the combina- work, and these tasks will be divided and allocated to several
tion of cloud computing and vehicular networks. The original worker nodes in the vehicular cloud to implement distributed
aim of vehicular cloud is to exploit the underutilized resources and parallel computation.
in vehicles to provide better support for complex applica- Based on the given model and assumptions, we focus on
tions and services in vehicles and transport systems. VACC the following problems: which cloud node the requested task
transforms the conventional vehicular cloud model and aims should be migrated to and how to find a reliable worker node
to exploit the underutilized resources in vehicles to increase to execute the task when the vehicular cloud has been selected
the available resources of cloud services for smartphones. In as the cloud service provider. To solve the first problem, the
VACC smartphones, the vehicular cloud and the cloudlet are offloading performance of the central cloud and the cloudlet
in a metropolitan area network (MAN). The vehicular cloud is should be analyzed. If both the central cloud and the cloudlet
connected to a cloudlet through a WLAN created by a Wi-Fi satisfy the offloading requirements, the problem becomes the
AP. Smartphones can connect cloudlets through cellular third- selection between the central cloud and the cloudlet. If both
generation (3G)/fourth-generation (4G) networks or Wi-Fi APs; the central cloud and the cloudlet cannot satisfy the offloading
hence, smartphones can connect to the vehicular cloud by the requirements, the vehicular cloud should be selected as the
cloudlet. Apart from the aforementioned connection method, a cloud service provider. Compared with the first problem, the
ZHANG et al.: TOWARD VEHICLE-ASSISTED CLOUD COMPUTING FOR SMARTPHONES 5613

second problem is more complex. To solve the second problem, TABLE I

N OTATION D EFINITION IN FOS
several factors should be considered. Due to the mobility of
worker nodes in the vehicular cloud, the connection between
a smartphone user and the vehicular cloud is not stable. If the
duration of the connection is too short, the data transmission
of the requested task possibly cannot be finished. Hence, we
should select the worker node with longer connected time.
However, the worker node with longer connected time may
not have sufficient resources. If the cloudlet uses unicast to
discover each worker node and collect the corresponding status
information, the total delay of multiple-worker-node discovery
is larger. Therefore, to find a qualified node, multiple worker
nodes need to be discovered by the cloudlet using multicast.
The time to discover multiple worker nodes impacts the task
response time for smartphone users. If we discover and collect
all nodes in the vehicular cloud, the probability of finding
a qualified node will be higher. However, the corresponding
task response time may not satisfy the offloading requirement.
Hence, we try to find a reliable worker node within a certain
amount of worker nodes in the vehicular cloud.

IV. F LEXIBLE O FFLOADING S TRATEGY

FOR S MARTPHONES

In VACC, the vehicular cloud plays an assistant role in cloud a bus, and the smartphone can directly connect to the vehicle
computing for smartphones. The vehicular cloud is exploited through Wi-Fi technology. In the first situation, the network be-
to relieve the limitation of the infrastructure-based cloud, in- tween the smartphone and the vehicular cloud is intermittently
cluding the WAN delay of the central cloud and the limited connected; hence, VACC needs to select a reliable worker node
available resources of the cloudlet. Hence, the offload scenar- (i.e., the lifetime of the connection between the smartphone
ios in VACC can be divided into two categories. In the first and the requested vehicle is long enough to receive the task) to
category, either the central cloud or the cloudlet can satisfy the implement task migration. Meanwhile, a reliable worker node
offloading requirements, and the requested task is migrated to also needs to satisfy the offloading requirements. In the second
one of them based on the offloading performance. In the second situation, the network connection between the smartphone and
category, both the central cloud and the cloudlet cannot satisfy the vehicle is relatively stable. If the vehicle satisfies both the
the offloading requirements, and the task is migrated to the offloading requirements and the connection requirement, VACC
vehicular cloud. For the central cloud, the FOS calculates the will employ it to execute the task.
energy consumption values of a smartphone and task response In the following sections, we analyze the offloading per-
time. For the cloudlet, the FOS judges whether the current avail- formance of the central cloud and the cloudlet. Moreover, we
able resources in the cloudlet satisfy the resource requirement design an algorithm to select the reliable worker node in the ve-
of the requested task. If yes, the FOS continues to calculate the hicular cloud. Table I shows the notation definition in the FOS.
energy consumption values of a smartphone and task response
time corresponding to the cloudlet. Then, the FOS determines A. Offloading Performance Analysis of the Central Cloud and
whether the central cloud and the cloudlet satisfy the require- the Cloudlet
ments of task response time and energy consumption. If both
The offloading performance of the central cloud and the
the central cloud and the cloudlet satisfy the offloading require-
cloudlet can be calculated as follows. The energy consumption
ments, the FOS will select the optimal cloud node (central cloud
by local execution is
or cloudlet) to execute the task. When only the central cloud
or the cloudlet satisfies the offloading requirements, the task Cu
Eu = Puc × . (1)
will be executed by the corresponding cloud node. If both the Su
central cloud and the cloudlet cannot satisfy the offloading The energy consumption of the smartphone corresponding to
requirements, the FOS will discover and utilize the vehicular the central cloud is
cloud to accomplish application offloading. Du Dr
The vehicular cloud in VACC contains two situations. A
c
Ecc = Ptx × c + Prx c
× c . (2)
Btx Brx
request from a smartphone should be processed in the corre-
sponding situation. In the first situation, the user is not on a The energy consumption of the smartphone corresponding to
bus, but rather, he/she may be walking on a street. His/her the cloudlet is
smartphone needs to connect to the cloudlet to discover and Du Dr
c
Ecl = Ptx × l + Prxc
× l . (3)
utilize a vehicular cloud. In the second situation, the user is on Btx Brx
5614 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 64, NO. 12, DECEMBER 2015

The task response time corresponding to the central cloud is duration of the connection will be greater than the estimated
value; hence, this situation does not destroy the normal task
Du Cu Dr
Tcc = c + + c . (4) migration. When a vehicle accelerates, the actual duration of the
Btx Scc Brx connection is close to the estimated value. If the task is executed
The task response time corresponding to the cloudlet is by a worker node in the vehicular cloud, the corresponding
energy consumption of the smartphone is
Du Cu Dr Du Dr
Tcl = + + l . (5) × v + Prx × B v .
c c
l
Btx Scl Brx Evc = Ptx (7)
Btx rx
If the central cloud can satisfy the conditions of Ecc < Eu Equation (7) denotes the energy consumption when the smart-
and Tcc ≤ T at the same time, it is considered to be available to phone uses 3G/4G networks to connect to the vehicular cloud.
perform application offloading. If the smartphone uses the Wi-Fi AP to connect to the nearby
If the current available resources in the cloudlet are sufficient vehicular cloud, the corresponding energy consumption is
for the requested task and the cloudlet can satisfy the conditions
Du Dr
of Ecl < Eu and Tcl ≤ T at the same time, it is considered to w
Evc = Ptx × v
w
+ Prx × v . (8)
be available to perform application offloading. Btx Brx
When there are two available cloud nodes (i.e., central cloud The task response time corresponding to vehicle i can be
and cloudlet), the FOS needs to select the optimal cloud node. If calculated as follows:
Ecc ≤ Ecl , then the central cloud is selected to execute the re-
Du Cu Dr
v + S i + B v + m × Tv
quested task. Otherwise, the optimal cloud node is the cloudlet. i
Tvc = (9)
When there is only one available cloud node (i.e., central cloud Btx vc rx
or cloudlet), the corresponding cloud node is selected. where Tv denotes the time to establish the connection between
When both the central cloud and the cloudlet are unavailable a vehicle and the cloudlet.
for the requested task, the FOS will discover and utilize the Based on the offloading requirements and the estimated
vehicular cloud to execute the task. duration of the connection between the vehicular cloud and the
smartphone, we design Algorithm 1 to select a reliable worker
B. Worker Node Selection in Vehicular Cloud node, and the selected worker node will execute the task. In the
first situation of the vehicular cloud, if the time to transfer the
In the first situation of the vehicular cloud, first, a smartphone input data to a worker node is less than the estimated duration
sends a request to the cloudlet and then utilizes the AP con- of the connection, we consider that the worker node satisfies
nected with the cloudlet to discover the vehicular cloud. The the connection requirement. After the task has been finished,
request from a smartphone contains the energy consumption if the worker node is out of the communication range of the AP,
Eu , the required response time T , the required memory size the result can be transferred to the cloudlet through multihop
M , and the set {Du , Cu , Dr }. The total number of nodes in the vehicular networks. Then, the cloudlet can return the result to
vehicular cloud is n. The cloudlet randomly selects m vehicles the smartphone through 3G/4G networks or Wi-Fi technology.
in the communication range of APs as candidate worker nodes. The values of energy consumption and task response time in
Then, some status information of m vehicles is transferred to Algorithm 1 can be calculated by (7)–(9). For brevity, the
the cloudlet. The status information of vehicle i consists of the equations do not appear in Algorithm 1.
current speed Vi , the direction θi , the available memory Mvi , the
i
CPU clock frequency Svc , and the coordinate (Xi , Yi ). After Algorithm 1
the status information of a vehicle is received by the cloudlet,
the corresponding connection is successfully established. INPUT: W : {w1 , w2 , . . . , wm }, S : {t1 , t2 , . . . , tm }, T , Eu ,
Next, the cloudlet estimates the duration of the connection Du , Cu , Dr , Svc : {Svc1
, Svc2 m
, . . . , Svc }, Mv :
between the smartphone and the vehicular cloud. Because the {Mv , Mv , . . . , Mv }
1 2 m

network connection between the smartphone and the cloudlet 1: for (i = 1; i <= m − 1; i + +)
is stable, the previously estimated duration is equal to the du- 2: Ts = ti ;
ration of the connection between the cloudlet and the candidate 3: k = 0;
worker node in the vehicular cloud. For vehicle i, the estimated 4: for (j = 1; j <= m − i; j + +)
duration can be obtained by the following equation: 5: if (Ts < tj+1 )
 6: Ts = tj+1 ;
2
R = (Xi + ti × Vix − Xs )2 + (Yi + ti × Viy − Ys ) . (6) 7: k = j + 1;
8: endif
In (6), Vix = (Vi +Vm ) sin θi /2, and Viy = (Vi +Vm ) cos θi /2. 9: endfor
Vm represents the permitted maximum speed on the road. 10: if (Ts > ti )
(Xs , Ys ) denotes the coordinate of the AP. R denotes the 11: switch (ti , tk );
communication range of the AP. We use the average value of the 12: switch (wi , wk );
i k
current speed and maximum speed to calculate time ti . When 13: switch (Svc , Svc );
a vehicle decelerates or keeps the current speed, the actual 14: switch (Mv , Mkv );
i
ZHANG et al.: TOWARD VEHICLE-ASSISTED CLOUD COMPUTING FOR SMARTPHONES 5615

15: endif TABLE II

S YSTEM PARAMETERS
16: if (M < Mvi &&Evc < Eu &&Tvc i
< T &&(Du /Btx v
) < ti)
17: node wi is released as the reliable worker node;
18: return;
19: endif
20: if (i = m − 1)
21: if (M < Mvm && < Eu &&Tvc m
≤ T &&(Du /Btx v
) < tm )
22: node wm is selected as the reliable worker node;
23: return;
24: endif
25: endif
26: endfor

The set W denotes the identities corresponding to m can-

didate worker nodes. The set S denotes the estimated dura-
tions of the connections corresponding to candidate worker
nodes. Algorithm 1 sorts candidate worker nodes for (m − 1)
rounds, and the worker node with the longest connection du-
ration is found after the first round. Then, Algorithm 1 judges
whether the current worker node satisfies the offloading require-
ments and the connection requirement. If it does, the current
worker node is selected as the reliable worker node; otherwise,
Algorithm 1 continues to sort and judge whether the next node
that the there is no qualified worker node in randomly selected
satisfies the required conditions. If none of the m candidate
m vehicles. Then, the algorithm can run again to find a reliable
worker nodes satisfy the required conditions, then the user can
worker node. However, the time to discover m vehicles at the
send a request to the cloudlet for vehicular cloud services again.
first time will increase the task response time. Hence, if the al-
When the cloudlet receives the same request for the second time,
gorithm has run several times and cannot find a reliable worker
it will randomly select m vehicles, except the selected vehicles
node, it will become useless due to the accumulated response
at the first time. Then, Algorithm 1 will be performed again.
time although a qualified worker node appears next time.
In the second situation of the vehicular cloud (i.e., a user is on
a bus), at first, the smartphone sends a request to the current bus
x, and the request message is the same as that in the first situ- V. P ERFORMANCE E VALUATION
ation. Then, the FOS judges whether the current bus x satisfies A. Experimental Setting
the offloading requirements. Since the smartphone is directly
connected to the current bus x through Wi-Fi technology, the Our experiments are the theoretical performance evaluation
corresponding energy consumption of the smartphone is based on our numerical calculations. To implement better the-
oretical performance evaluation, we need to select the suitable
Du Dr parameters that conform to the practical situations. Generally,
w
Evc = Ptx × v
w
+ Prx × v + Ev . (10)
Btx Brx if a user wants to use the central cloud, first, he/she needs to
buy some resources in the central cloud according to his/her
The task response time corresponding to the current bus x is
demands. Therefore, if the user usually has medium or low
x Du Cu Dr demands, the corresponding resources (e.g., computing power
Tvc = v + S x + B v + Tv . (11) and memory capability) may be moderate or even limited such
Btx vc rx
as a common personal computer. In this situation, the user’s
In (10), Ev denotes the energy consumption of the smart- smartphone acts as a thin client, and the central cloud acts
phone corresponding to the connection established between the as a personal computer controlled by the smartphone. If the
smartphone and bus x. In addition, the user needs to offer the user often has high demands, then the corresponding resources
estimated duration tx . If bus x satisfies both the offloading re- should be more powerful. Of course, buying powerful resources
x
quirements and the condition of Tvc < tx , the task will be exe- requires bigger expenses. Hence, in our experiments, we not
cuted by bus x. Otherwise, the smartphone will send a request to only select the parameters corresponding to high performance
the cloudlet for the vehicular cloud services in the first situation. but also use the parameters of medium and low performance for
The time complexity of Algorithm 1 is O(m2 ) smaller than the central cloud. For the networking parameters, we consider
O(n2 ). The main functionality of Algorithm 1 is to select a that the WAN has a low data transmission rate and that the link
worker node that satisfies the offloading requirements and the between the smartphone and the base station has a high data
connection requirement for the requested task. Algorithm 1 is transmission rate by 3G communication technology.
able to find a reliable worker node to execute the requested We set the same CPU clock frequency and available memory
task if there is one or more qualified worker nodes in randomly space for each candidate worker node in the vehicular cloud,
selected m vehicles, but it also has limitations. It is possible and we assume that there exists one or more available worker
5616 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 64, NO. 12, DECEMBER 2015

Fig. 2. Task response time evaluation for case 1. (a) Central cloud is powerful. (b) Cloudlet is powerful.

nodes. In the theoretical performance evaluation, we do not

consider the mobility of vehicles when the reliable worker
node performs the requested task, and we mainly test the
computation and communication performance of the vehicular
cloud services.
We refer to the system parameters in [17] and [18] to eval-
uate the theoretical performance. Table II shows the specific
parameters in our experiments. Ml denotes the current available
memory space of the cloudlet. We test four cases for various
requested tasks with different workload and resource conditions
of cloud nodes. In each case, different computing capabilities
of the central cloud are evaluated. Meanwhile, we test the
scenario where the resources in the cloudlet are insufficient for
the requested task. For the vehicular cloud, we mainly test the
performance using 3G networks.
Fig. 3. Energy consumption evaluation for case 1.
Case 1: Cu = (200, 400, 600, 800, 1000) M cycles. Du = (10,
20, 30, 40, 50) KB. Dr = (4, 8, 12, 16, 20) KB.
of response time. The cloudlet has a high data transmission
T = (0.4, 0.8, 0.12, 0.16, 0.20) s.
rate; hence, the transmission delay is shorter, and the energy
Case 2: Cu = (100, 200, 300, 400, 500) M cycles. Du = (15,
consumption of the smartphone is less than that of the central
30, 45, 60, 75) KB. Dr = (5, 10, 15, 20, 25) KB.
cloud. Finally, the FOS selects the cloudlet to execute the task.
T = (0.2, 0.4, 0.6, 0.8, 1.0) s.
In case 2, the computing workload is lighter than in case 1.
Case 3: Cu = (100, 200, 300, 400, 500) M cycles. Du = (100,
The available memory space of the cloudlet is not sufficient. As
200, 300, 400, 500) KB. Dr = (40, 80, 120, 160,
shown in Fig. 4, the vehicular cloud has a shorter response time
200) KB. T = (1.0, 2.0, 3.0, 4.0, 5.0) s.
than the required time. When the CPU of the central cloud is
Case 4: Cu = (200, 400, 600, 800, 1000) M cycles. Du = (200,
more powerful, the central cloud has a shorter response time,
400, 600, 800, 1000) KB. Dr = (40, 80, 120, 160,
but it is still longer than the required time. Both the central
200) KB. T = (2.0, 4.0, 6.0, 8.0, 10.0) s.
cloud and the vehicular cloud have less energy consumption
than the local execution. However, the vehicular cloud has a
high data transmission rate; hence, the transmission delay is
B. Experimental Results
shorter, and the energy consumption is less than that of the
The theoretical performance of the FOS corresponding to central cloud. Finally, the FOS selects the vehicular cloud to
four cases is shown in Figs. 2–7. In case 1, the workload of execute the task.
data transmission is relatively light. As shown in Figs. 2 and 3, In case 3, the workload of data transmission is heavier than in
both the central cloud and the cloudlet have a shorter response case 1, which has a heavier computing workload. As shown in
time than the required time. When the CPU of the central Figs. 5 and 6, the cloudlet has a shorter response time than the
cloud is more powerful than that of the cloudlet, the central required time. The central cloud cannot satisfy the offloading
cloud has a shorter response time. However, when the cloudlet’s requirements in terms of response time and energy consump-
CPU is more powerful, the cloudlet is more efficient in terms tion. Finally, the FOS selects the cloudlet to execute the task.
ZHANG et al.: TOWARD VEHICLE-ASSISTED CLOUD COMPUTING FOR SMARTPHONES 5617

Fig. 4. Performance evaluation for case 2. (a) Task response time. (b) Energy consumption of the smartphone.

Fig. 5. Task response time evaluation for case 3. (a) Central cloud is powerful. (b) Cloudlet is powerful.

the low data transmission rate. Meanwhile, the central cloud

cannot satisfy the offloading requirements in terms of energy
consumption. The vehicular cloud has less energy consumption
than the local execution. Finally, the FOS selects the vehicular
cloud to execute the task.

VI. C ONCLUSION
In this paper, we have proposed combining the vehicular
cloud with the fixed central cloud and the cloudlet to expand the
current available resources for task requests from smartphones.
In the proposed VACC, the vehicular cloud acts as a cloud
service provider for smartphones. To gain the energy saving of
smartphones and accomplish tasks in the required time, an FOS
is proposed. If both the central cloud and the cloudlet do not
Fig. 6. Energy consumption evaluation for case 3. satisfy the offloading requirements, then the FOS utilizes the
underutilized resources in vehicles to accomplish application
In case 4, the workload of data transmission is heavier than offloading. If both the central cloud and the cloudlet satisfy
in case 3. The available memory space of the cloudlet is not the offloading requirements, the FOS selects the optimal cloud
sufficient. As shown in Fig. 7, the vehicular cloud has a shorter node (central cloud or cloudlet) to execute the task. Numerical
response time than the required time, but the response time experimental results show that the FOS can employ the vehicu-
of the central cloud is longer than the required time due to lar cloud to accomplish application offloading for smartphones
5618 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 64, NO. 12, DECEMBER 2015

Fig. 7. Performance evaluation for case 4. (a) Task response time. (b) Energy consumption of the smartphone.

when the infrastructure-based cloud does not satisfy the [17] D. Huang, P. Wang, and D. Niyato, “A dynamic offloading algorithm
offloading requirements. Therefore, energy savings can be for mobile computing,” IEEE Trans. Wireless Commun., vol. 11, no. 6,
pp. 1991–1995, Jun. 2012.
achieved smartphones. [18] W. Zhang, Y. Wen, J. Wu, and D. Wu, “Energy-efficient scheduling policy
for collaborative execution in mobile cloud computing,” in Proc. IEEE
Infocom, 2013, pp. 190–94.
R EFERENCES
[1] D. Huang, T. Xing, and H. Wu, “Mobile cloud computing service mod-
els: A user-centric approach,” IEEE Netw., vol. 27, no. 5, pp. 6–11, Hongli Zhang received the B.S. degree in computer
Sep./Oct. 2013. science from Sichuan University, Chengdu, China, in
[2] W. Zhang, Y. Wen, J. Wu, and H. Li, “Toward a unified elastic computing 1994 and the Ph.D. degree in computer science from
platform for smartphones with cloud support,” IEEE Netw., vol. 27, no. 5, Harbin Institute of Technology (HIT), Harbin, China,
pp. 34–40, Sep./Oct. 2013. in 1999.
[3] M. Felemban, S. Basalamah, and A. Ghafoor, “A distributed cloud ar- She is a Professor with the Department of Com-
chitecture for mobile multimedia services,” IEEE Netw., vol. 27, no. 5, puter Science and Technology, HIT, and the Vice
pp. 20–27, Sep./Oct. 2013. Director of the National Computer Information Con-
[4] R. Hussain, J. Son, H. Eun, S. Kim, and H. Oh, “Rethinking vehicular tent Security Key Laboratory. Her research interests
communications: Merging VANET with cloud computing,” in Proc. IEEE include network and information security, network
4th Int. Conf. Cloud Comput. Technol. Sci., 2012, pp. 606–609. measurement, and cloud computing.
[5] M. Gerla, “Vehicular cloud computing,” in Proc. 11th Annu. Med-Hoc-
Net, 2012, pp. 152–155.
[6] R. Yu, Y. Zhang, W. Xia, and K. Yang, “Toward cloud-based vehicu-
Qiang Zhang received the B.E. degree in informa-
lar networks with efficient resource management,” IEEE Netw., vol. 27,
tion security from Harbin Institute of Technology
no. 5, pp. 48–55, Sep./Oct. 2013.
(HIT), Harbin, China, in 2009 and the M.S. de-
[7] S. Olariu, M. Eltoweissy, and M. Younis, “Towards autonomous vehicular
gree from Harbin Engineering University in 2012.
clouds,” ICST Trans. Mobile Commun. Appl., vol. 11, no. 7–9, pp. 1–11,
He is currently working toward the Ph.D. degree
2011.
with the Department of Computer Science and Tech-
[8] S. Olariu, I. Khalil, and M. Abuelela, “Taking VANET to the clouds,” Int.
nology, HIT.
J. Pervasive Comput. Commun., vol. 7, no. 1, pp. 7–21, 2011.
His research interests include cloud computing
[9] M. Whaiduzzaman, M. Sookhak, A. Gani, R. Buyya, “A survey on vehicu-
and wireless networks.
lar cloud computing,” J. Netw. Comput. Appl., vol. 40, no. 1, pp. 325–344,
Apr. 2014.
[10] E. Tilevich and Y. Kwon, “Cloud-based execution to improve mobile appli-
cation energy efficiency,” Computer, vol. 47, no. 1, pp. 75–77, Jan. 2014.
[11] S. Ou, K. Yang, and A. Liotta, “An adaptive multi-constraint partition-
ing algorithm for offloading in pervasive systems,” in Proc. IEEE 4th Xiaojiang Du received the B.E. degree from
Annu. Int. Conf. Pervasive Comput. Commun., Mar. 2006, Pisa, Italy, Tsinghua University, Beijing, China, in 1996 and
pp. 116–125. the M.S. and Ph.D. degrees from the University of
[12] B. Chun, S. Ihm, P. Maniatis, M. Naik, and A. Patti, “Clonecloud: Elastic Maryland, College Park, MD, USA, in 2002 and
execution between mobile device and cloud,” in Proc. 6th Conf. Comput. 2003, respectively, all in electrical engineering.
Syst., New York, NY, USA, 2011, pp. 301–314. He is currently an Associate Professor with the
[13] M. Satyanarayanan, R. C. P. Bahl, and N. Davies, “The case for VM-based Department of Computer and Information Sciences,
cloudlets in mobile computing,” IEEE Pervasive Comput., vol. 8, no. 4, Temple University, Philadelphia, PA, USA. He has
pp. 14–23, Oct.–Dec. 2009. published over 160 journal and conference papers in
[14] E. Cuervo et al., “MAUI: Making smartphones last longer with code his areas of interest. His research interests include
offload,” in Proc. Int. Conf. Mobile Syst., Appl., Serv., 2010, pp. 49–62. security, systems, wireless networks, and computer
[15] K. Kumar and Y. Lu, “Cloud computing for mobile users: Can offloading networks.
computation save energy?” Computer, vol. 43, no. 4, pp. 51–56, Apr. 2010. Dr. Du has been awarded more than $4M in research grants from the
[16] Y. WenW. Zhang, K. Guan, D. Kilper, and H. Luo, “Energy-optimal exe- National Science Foundation, the Army Research Office, and the Air Force
cution policy for a cloud-assisted mobile application platform,” Nanyang Research Laboratory. He serves on the Editorial Board of four international
Technol. Univ., Singapore, Tech. Rep., Sep. 2011. journals.