Received December 17, 2019, accepted January 13, 2020, date of publication January 21, 2020, date of current

version January 30, 2020.

Digital Object Identifier 10.1109/ACCESS.2020.2968465

Delay-Optimized V2V-Based Computation

Offloading in Urban Vehicular Edge
Computing and Networks
CHEN CHEN 1 , (Senior Member, IEEE), LANLAN CHEN1 , LEI LIU 1 , (Student Member, IEEE),
SHUNFAN HE 2 , XIAOMING YUAN 3 , DAPENG LAN 4 , AND ZHUANG CHEN 5
1 StateKey Laboratory of Integrated Service Networks, Xidian University, Xi’an 710071, China
2 College of Computer Science, South-Central University for Nationalities, Wuhan 430073, China
3 Qinhuangdao Branch Campus, Northeastern University, Qinhuangdao 066004, China
4 Department of Informatics, University of Oslo, 0316 Oslo, Norway
5 School of Computer and Information Security, Guilin University of Electronic Technology, Guilin 541004, China

Corresponding author: Lei Liu ([email protected])

This work was supported in part by the National Key Research and Development Program of China under Grant 2018YFE0126000, in part
by the National Natural Science Foundation of China under Grant 61571338, Grant U1636209, Grant 61672131, and Grant 61901367,
in part by the Key Research and Development Plan of Shaanxi Province under Grant 2017ZDCXL-GY-05-01, Grant 2019ZDLGY13-04,
and Grant 2019ZDLGY13-07, in part by the Xi’an Key Laboratory of Mobile Edge Computing and Security under Grant
201805052-ZD3CG36, in part by the Innovation Fund of Xidian University under Grant 5001-20109195456, in part by the National Nature
Science Foundation of China under Grant 6190021180, in part by the Natural Science Foundation of China under Grant 61901367, and in
part by the Natural Science Foundation of China under Grant 61901099.

ABSTRACT The Internet of Vehicles (IoV) is an emerging paradigm, driven by recent advancements in
vehicular communications and networking. Meanwhile, the capability and intelligence of vehicles are being
rapidly enhanced, and this will have the potential of supporting a plethora of new exciting applications, which
will integrate fully autonomous vehicles, the Internet of Things (IoT), and the environment. In view of the
delay-sensitive property of these promising applications, as well as the high expense by using infrastructures
and roadside units (RSU), the task offloading among vehicles has gained enormous popularity considering
its free-of-charge and timely response. In this paper, by utilizing the gathering period of vehicles in urban
environment due to stopped by traffic lights or Area of Interest (AOI), a task offloading scheme merely
relying on vehicle-to-vehicle (V2V) communication is proposed by fully exploring the idle resources
of gathered vehicles for task execution. Through formulating the task execution as a Min-Max problem
among one task and several cooperative vehicles, the task executing time is optimized with the Max-Min
Fairness scheme, which is further solved by the Particle Swarm Optimization (PSO) Algorithm. Extensive
simulation demonstrate that our model could well meet the delay requirement of delay-sensitive application
by cooperative computing among vehicles.

INDEX TERMS Internet of vehicle, computing offloading, low latency, task allocation.

I. INTRODUCTION for achieving the coordinated development of human, vehi-

With the rapid development of social economy and the pro- cle, and environment, which can alleviate traffic congestion,
cess of industrialization, the number of vehicles has been enhance transportation efficiency [6]–[8].
growing rapidly in cities [1]–[5]. However, owing to limited However, there are still some technical problems to be
capacity, roads become crowded, which leads to a series of solved in the popularization and development of IoV [9],
issues, e.g., traffic accident, environmental pollution. Face [10], such as intelligent information processing [11], [12],
these relentless challenges, Internet of Vehicle (IoV) emerges that is, filtering and processing data in IoV to fetch out useful
to fully use the information and communication technologies information [6], [13], [14]. For the challenge of intelligent
information processing, it can be mainly summarized into
The associate editor coordinating the review of this manuscript and three points. First, the IoV is engaged to provide timely
approving it for publication was Lu Liu . services relevant to driving safety and traffic efficiency, which

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see http://creativecommons.org/licenses/by/4.0/
VOLUME 8, 2020 18863
 C. Chen et al.: Delay-Optimized V2V-Based Computation Offloading in Urban Vehicular Edge Computing and Networks

task from the task vehicle to service vehicles [14]. What

is more, in IoV, rapid vehicle movement results in frequent
network topology changes and shorter communication link
lifetime. So in the process of offloading, we need to consider
the node movement problem to prevent communication link
interruption [31].
Combine the above two considerations, in this paper,
we propose a cooperative task scheduling scheme in which
one task vehicle and multiple service vehicles jointly execute
the vehicle-carried task [14], [32]. The task vehicle com-
prehensively considers the computing power and maximum
service time of each service vehicle to decide what percent-
age of the task should be assigned to each service vehicle,
FIGURE 1. The scene of offloading computing. to minimize the task execution time. Here, the maximum
service time of each service vehicle depends not only on its
requires strong capability of computing and intelligence on residence time at the gathering point, but also on the relative
the edge [15]–[17]. Second, the strict latency requirement movement of the task vehicle and the service vehicle. Because
of applications often makes the V2V communication a pre- we must ensure that the task vehicle and each service vehicle
requisite, where a round-trip communication from infrastruc- are always within the communication range of each other. The
ture or RSU may fail the delay-sensitive services. Third, key contribution of this paper include:
in view of the high expense by using infrastructure or RSU • We propose the concept of "resource pool" formed by
to process and store the collected information, the V2V service vehicles that are slowly moving together. Based
communication is free-of-charge thus attracting more atten- on these idle resources that can be scheduled in resource
tions. Considering above concerns, the information process- pool, we propose a cooperative task scheduling scheme
ing cooperated by surrounding vehicles in a V2V pattern, to minimize the task execution time.
is very promising [2], [18], [19]. • We formulate the task execution time optimization
In addition, in urban environment, vehicles usually gather model as Min-Max problem under the permissible
together for various reasons, such as waiting for traffic lights, latency constraint. In addition, we also incorporate node
passing through toll stations, or attracted by an AOI. For mobility into problem formulation.
example, as shown in Fig.1, these service vehicles that are • For the Min-Max problem, we modify Max-Min Fair-
slowly moving together due to traffic lights will form a ness Algorithm and Particle Swarm Optimization (PSO)
resource pool for a short period of time. These vehicles may Algorithm respectively according to whether to call all
have rich and idle resource for task execution, i.e., infor- service vehicles, to find the best task allocation scheme
mation processing in IoV [20]. Therefore, we can certainly to minimize the task execution time. Simulation results
offload the task to these service vehicles, and make the task demonstrate the effectiveness of proposed schemes.
vehicle and several surrounding vehicles cooperatively com- The rest of this paper is organized as follows. The
pleting the vehicle-carried task to reduce the task execution system model which includes the mobility model, commu-
time [21]–[23]. nication model, computation model and problem formula-
There has been considerable amount of work focusing tion is presented in Section II. The problem solution is
on computation offloading [24]–[26]. The authors consid- given in Section III. We evaluate the performance of pro-
ered/analyzed the problem of binary offloading decision, posed schemes and provide illustrative results in Section IV.
where each user independently chooses whether to execute We conclude this paper in Section V. To make the readers eas-
the task locally or to offload the whole task to the edge ily follow, the notations used in our paper are listed in Table1.
servers, to minimize the task execution time. Studies, such
as Luo et al. [26] used dynamic programming and improved II. SYSTEM MODEL
greedy algorithm to assign different tasks to multiple edge We mainly focus on V2V-based task offloading model in
servers, to minimize the latency. Liu et al. [27] formulated order to minimize the needed time for task processing by
the multiple vehicles computation offloading problem as a leveraging all available idle resources in vehicular networks
multi-user computation offloading game problem, prove the as shown in Fig.2. The considered scenario includes one
existence of Nash equilibrium (NE) of the game and propose a task vehicle which has one task to execute and N service
distributed computation offloading algorithm to compute the vehicles which have idle computation resources, where N
equilibrium. Compared to binary offloading, it is more rea- vehicles are within the communication range of task vehi-
sonable to offload partially to service vehicles, since wireless cle. Specially, let N = {Vidle1 , Vidle2 , Vidle3 · · · VidleN }
channel resource is limited and local computing power should be the set of N service vehicles. We adopt three parameters
not be ignored [28]–[30]. Motivated by such considerations, {c, data, t max } to indicate the generated task of task vehicle.
in this paper, we dynamically offload parts of the computation Here, c (in CPU cycles per bit) is the computation resource

18864 VOLUME 8, 2020

C. Chen et al.: Delay-Optimized V2V-Based Computation Offloading in Urban Vehicular Edge Computing and Networks

TABLE 1. List of important notations. (IDM_IM) mobility model to accurately capture the realistic
mobility characteristics of moving vehicles [33], [34].
As for the behavior of a single car in straight road, it can
be described as follows. Assume the length of vehicle α is
lα , the position of vehicle α at time t is xα (t), the velocity of
vehicle α at time t is vα (t), the position of the leading vehicle
α −1 at time t is xα−1 (t), the velocity of vehicle α −1 at time t
is vα−1 (t). Note that the velocity here represents a vector. If it
is less than 0, the vehicle is moving toward the left; otherwise,
the vehicle is driving toward the right. The acceleration in the
IDM model can obtained as follows:
dvα (t) vα (t) s∗ (vα (t), 1vα (t)) 2
= aα [1 − ( ∗ )δ − ( ) ] (1)
dt vα sα (t)
where aα is maximum acceleration, δ is acceleration expo-
nent, the gap sα (t) := xα (t) − xα−1 (t) − lα , the velocity dif-
ference 1vα (t) := vα (t) − vα−1 (t), v∗α is desired velocity for
vanishing interactions(sα (t) → ∞). The desired minimum
gap s∗ (vα (t), 1vα (t)) can be obtained as follows:
s
00 vα (t)
s (vα (t), 1vα (t)) = sα + sα
∗ 0
+ vα (t)T
v∗α
vα (t)1vα (t)
+ √ (2)
2 aα bα
where s0α and s00α are jam distances, T is safe time headway
and bα is desired deceleration [35].
According to [35], we usually set s00α = 0, so we can get
vα (t)1vα (t)
s∗ (vα (t), 1vα (t)) ≈ s0α + vα (t)T + √ (3)
2 aα bα
According to [36], we can integrate Eq.(1) with a simple
Euler scheme using a coarse time discretization of 1t = 0.4s,
and get velocity and displacement calculation formula:
dvα (t)
vα (t + 1t) = vα (t) + 1t ·
dt
1 dvα (t)
required for processing the task, data (in bits) specifies the xα (t + 1t) = xα (t) + 1t · vα (t) + ( )(1t)2 . (4)
2 dt
task size, and t max represents the completion deadline for task Here, smaller value of 1t yields nearly indistinguishable
execution. Each service vehicle uploads its own information results [36]. By partitioning the interval [t, t + tα ] based on
to its nearby RSU, including vehicle ID, position, velocity step 1t, we can get the displacement 1Xα (tα ) as follows:
and available computation resource. The task vehicle can get
the information about these service vehicles from this RSU. 1Xα (tα )
Generally, the task vehicle can process its task locally or = xα (t + tα ) − xα (t)
by computation offloading. With the aim to reduce the task
= [xα (t + 1t) − xα (t)] + [xα (t + 21t) − xα (t + 1t)]
execution time, the task vehicle can offload part of its task to
surrounding vehicles which possess idle resources. Then, two + · · · + [xα (t +btα /1tc1t) − xα (t +(btα /1tc − 1)1t)]
problems arise for the task vehicle: (1) how to select service +[xα (t + tα ) − xα (t + [tα /1t]1t)] (5)
vehicles for computation offloading, and (2) how to offload
the computation task to these service vehicles. B. COMMUNICATION MODEL
When the task vehicle has a task to calculate, it needs to
A. MOBILITY MODEL assign the task to available service vehicles for task execu-
In our model, we assume that identical vehicles are driving tion time reduction. Receiving the assigned task by this task
on a two-way straight road, regardless of vehicle turning. vehicle, service vehicles are responsible for implementing
Vehicles can move toward the oppositive directions. We adapt the task. To this end, we adopt the WBSS (WAVE-based
Intelligent Driver Model with Intersection Management basic service set) communication mode after analyzing the

VOLUME 8, 2020 18865

 C. Chen et al.: Delay-Optimized V2V-Based Computation Offloading in Urban Vehicular Edge Computing and Networks

FIGURE 2. The illustration of system model, where b1 , b2 , b3, b4 indicates the offloaded task proportion.

IEEE 802.11p/WAVE (Wireless Access for Vehicular Envi- C. COMPUTATION MODEL

ronment) standard [37]. The task vehicle broadcasts a WBSS Before introducing the computation model, some assump-
announcement message, and the WAVE Service Advertise- tions are given:
ment(WSA) on the control channel (CCH), which contain • Assumption 1: the computation task can be divided into
all the information identifying the WAVE application and two or more parts.
the network parameters are necessary to join a WBSS [38], • Assumption 2: there are two options for task calcu-
[39]. Especially, WAVE application corresponding to the task lation. All of the computation task can be calculated
needs to be processed and these network parameters include locally, or part offloaded to multiple service vehicles for
the WBSS ID and the service channel (SCH) which this processing.
WBSS will use. By monitoring the CCH, service vehicles
are capable of knowing about the initiated WBSS by the 1) THE TASK EXECUTION TIME OF LOCAL COMPUTING
task vehicle and then join this WBSS by simply switching
For local task computing, we define f0 as the CPU computing
to the specified SCH. Switching between CCH and SCH
capability of the task vehicle. Then, the local execution time
adopts immediate channel access [37], which allows imme-
needed for implementing the task is expressed as follows:
diate communications access on a designated channel for an
extended period without consideration for time slot bound- tlocal = c/f0 (8)
aries. We use tCCH to indicate the time spent on the CCH for
the task vehicle and service vehicles. tSCH is the transmission
2) THE TASK EXECUTION TIME OF OFFLOADING
delay taken on the SCH for transmitting data, which can be
COMPUTING
written as
Denote bi ∈ [0, 1) be the task proportion which service
data vehicle i is responsible for processing. In other words, the task
tSCH = (6)
B log(1 + P·G
) offloading strategy is presented by b = {b1 , b2 , b3 · · · bN }.
σ2
Define fi as the CPU computing capability of service vehicle
where B is the bandwidth of the SCH that the WBSS initiated i. Then, the time needed for service vehicle i to execute its
by the task vehicle will use. P is the transmission power of the assigned task workload is obtained as follow:
task vehicle. G is the channel gain between the task vehicle
bi · c
and service vehicles. σ 2 is the noise power. tic = (9)
For simplicity, the time for the task vehicle and service fi
vehicles to access the CCH and SCH via the carrier sense mul- Based on the communication model described above,
tiple access protocol with collision avoidance (CSMA/CA) the data is broadcasted by the task vehicle. So, the time
mechanism is negligible, and the time spent on channel needed for the service vehicle to receives the data from the
switching, such as guard interval, also is negligible. So the task vehicle can be obtained as follows:
total time for transmitting data can be obtained as follows:
titr = t tr + δ (10)
t tr = tCCH + tSCH (7)
where δ is the propagation delay.
Because the calculation result after task processing is quite Thus, the total time for service vehicle i to process the
small [20], the link transmission delay of service vehicles assigned workload by the task vehicle is given as:
feeding back the calculation result to the task vehicle is
neglected. ti = tic + titr (11)

18866 VOLUME 8, 2020

C. Chen et al.: Delay-Optimized V2V-Based Computation Offloading in Urban Vehicular Edge Computing and Networks

On the other hand, the needed time for task vehicle to

process the rest task workload is written as:
N
X
t0 = c(1 − bi )/f0 (12)
i=1
According to Eq.(11) and Eq.(12), the total task execution
time of task vehicle is calculated as:
tedge = max(ti ), i = 0, 1, 2 · · · N (13)

D. PROBLEM FORMULATION
With the aim to minimize the task execution time, the follow-
FIGURE 3. The sketch of Constraint C3.
ing optimization problem is formulated:
minimize T = s · tlocal + (1 − s) · tedge
N For the optimal problem in Eq.(16), we adapt Max-Min
Fairness Algorithm to solve it. According to Max-Min Fair-
X
subject to C1 : bi < 1(0 ≤ bi < 1)
i=1
ness Algorithm, by reasonably adjusting the size of bi , we can
C2 : ti ≤ tres_i (i = 1, 2 · · · N ) guarantee each vehicle (including N service vehicles and the
task vehicle) has the same service time, i.e.,
C3 : |1Xi | ≤ 2 |x0 (t) − xi (t)| (i = 1, 2 · · · N )
(14) t0 = t1 = t2 = · · · = tN (17)

herein, s = 1 if the total task is implemented locally, s = According Eq.(17), we have:

0, otherwise. tres_i indicates the residence time of service c c
bequal_1 · + t1tr = bequal_2 · + t2tr
vehicle i in assembly point. x0 (t) and xi (t) represent the f1 f2
X-dimensional coordinate of task vehicle and service vehicle c
= · · · = bequal_N · + tNtr (18)
i at time t, respectively. 1Xi indicates the relative displace- fN
ment between task vehicle and service vehicle i after ti . Based on Eq.(18), we can get:
Constraint C1 corresponds to Assumption 2. Constraint
C2 guarantees that each service vehicle can finish the task fi t tr − titr
bequal_i = bequal_1 + 1 fi (i = 1, 2, · · · N ) (19)
computing before leaving the assembly point. Constraint f1 c
C3 ensures that the link between the task vehicle and its ser- Under the condition that t0 = t1 , the following equation
vice vehicle keeps available during task offloading. Specially, holds:
Constraint C3 is explained as follows: Due to the mobility, N
the value of 1Xi is changing with time as shown in Fig.3. c X c
(1 − bequal_i ) = bequal_1 · + t1tr (20)
Based on Eq.(5), we can get f0 f1
i=1
1Xi = 1Xi (ti ) − 1X0 (ti ) where bequal_i is the task segmentation ratio of each service
= [xi (t + ti ) − xi (t)] − [x0 (t + ti ) − x0 (t)]. (15) vehicle under the condition of average distribution.
By combining Eq.(19) and Eq.(20), we have:
In order to guarantee the success of task offloading, service
N
vehicle i should not move out of the communicate range of fi tr
c/f0 − t1tr − − titr )
P
task vehicle with dis as the radius. It is known that a straight f0 (t1
i=1
line is the shortest distance between two points. Thus, accord- bequal_1 = (21)
N
c c P fi
ing to the properties of isosceles triangle, the maximum value f1 + f0 · f1
of 1X is equal to 2 |x0 (t) − xi (t)|. i=1
As a result, the service time of each service vehicle is
III. PROBLEM SOLVING expressed as:
In this section, we present an algorithm to solve the problem c
in Eq.(14). t0 = t1 = t2 = · · · = tN = Tequal = bequal_1 + t1tr (22)
f1
However, due to the mobility, the link between the
A. THE OPTIMAL SOLUTION FOR SPECIAL CASE
task vehicle and its service vehicle may be not available
In this section, a special case of computation offloading
for such a long time. For Constraint C3, according to
is considered where all service vehicles participate in task
Eq.(4) and Eq.(5), we can see that the displacement and
processing.
time are quadratic inequality relations, which are solvable.
When s = 1, T = tlocal = c/f0 . If s = 0, then
We assume that the solution is [0, tconstraint_c3 ]. Then, com-
T = tedge = min{max(ti )} i = 0, 1, 2 · · · N (16) bining Constraint C2, we can get 0 < ti ≤ tmax _i

VOLUME 8, 2020 18867

 C. Chen et al.: Delay-Optimized V2V-Based Computation Offloading in Urban Vehicular Edge Computing and Networks

(tmax _i := min{tres_i , tconstraint_c3 }). By ascending sort, we should offload the task to service vehicles based on the
tmax _1 ≤ tmax _2 ≤ tmax _3 ≤ · · · ≤ tmax _N . When compared final bi .
with Tequal , tmax _l < Tequal ≤ tmax _r l, r ∈ {1, 2, · · · , N }
and l < r. So, the task workloads assigned to Vidle1 , Vidle2 , B. THE FEASIBLE SOLUTION FOR GENERAL CASE
· · · , Vidlel are too big and need to be further adjusted. In this section, a more general case is considered where we
First, it is required to calculate all extra task offload- do not know whether each service vehicle will participate in
ing ratios. Take Vidle1 as an example. The maximum task task processing. In order to solve the problem in Eq.(16), 2N
offloading ratio that Vidle1 can undertake is bmax _1 , which program is needed, which brings difficulty to the analysis and
is obtained as follows: optimization of the optimal problem. With the aim to deal
c with this issue, we employ the PSO Algorithm to solve the
tmax _1 = bmax _1 · + t1tr
f1 problem in Eq.(16).
tmax _1 − t1tr According to the PSO Algorithm, define the position of
⇒ bmax _1 = f1 (23) the jth particle in the N-dimensional space at step k as Bkj =
c
(bkj1 , bkj2 , · · · , bkjN )T , j = 1, 2, · · · , l, the velocity as Vjk =
As a result, the extra task offloading ratio of Vidle1 is calcu-
lated as: (vkj1 , vkj2 , · · · , vkjN )T , j = 1, 2, · · · , l. The best position that jth
particle has experienced is pbestjk = (pkj1 , pkj2 , · · · , pkjN )T , j =
bextra_1 = bequal_1 − bmax _1 (24) 1, 2, · · · , l. The best position that the particle group has expe-
rienced is gbest k = (gk1 , gk2 , · · · , gkN )T . The fitness function
Correspondingly, all extra task offloading ratios can be given
is shown in Eq.(27):
by:
l
X Tjk = tedge (Bkj ) (j = 1, 2, · · · , l) (27)
bextra = bextra_i (25)
What is more, when updating the best position of each parti-
i=1
cle, we not only take into account the fitness function, but also
Then, the extra task offloading ratio bextra will be dis- take into account the number of service vehicles involved in
tributed to Vidlel+1 , Vidlel+2 , Vidlel+3 , · · · , VidleN and the the calculation. That is, when the fitness function is reduced,
task vehicle. Similar to the above uniform distribution, the number of service vehicles participating in the calculation
the following equation holds: must also be reduced.
fi The jth particle updates its velocity and position according
bequal2_i = bequal2_l+1 to the following formula:
fl+1
tr
tequal2_l+1 − titr Vjk+1 = wVjk + c1 r1 (pbestjk − Bkj ) + c2 r2 (gbest k − Bkj )
+ fi (i = l + 2, · · · N )
c Bk+1 = Bkj + Vjk+1 (28)
c tr j
bequal2_l+1 · + tl+1
fl+1 where w is inertia weight, c1 and c2 are learning weight,
N the elements of vector r1 and vector r2 are assigned values
c X
= (bextra − bequal2_i ) from a uniform random distribution between −1 and 1 which
f0
i=l+1 are independent with each other.
N However, we need to be aware that in the process of initial-
tr − fi tr
− titr )
P
bextra · c/f0 − tl+1 f0 (tl+1 ization and iteration according to Eq.(28), the position Bkj of
i=l+1
bequal2_l+1 = each particle must meet Constraint C1-C3. So, we make the
N
c c P fi
fl+1 + f0 · fl+1 corresponding adjustment to the PSO Algorithm. When we
i=l+1 find that the current particle’s location information does not
(26) meet the constraints, we will adjust according to the following
where bequal2_i is the task workload assigned to Vidlei , i = l+ steps:
1, l +2, · · · , N from bextra . The task offloading ratio assigned Step1. For Constraint C1, we count the sum of the
to Vidli is: bi = bequal_i + bequal2_i ,i = l + 1, l + 2, · · · N . N-dimensional positions of the current particles. If the sum
Similarly, we should check whether Vidli can undertake such is greater than 1, for the part beyond 1, the position of each
big task offloading ratio according to its maximum service dimension is reduced according to its proportion. Taking the
time. If no, we need to follow the above method to assign extra 1-dimensional position of jth particle bk+1 j1 as an example
segmentation ratio. Repeat the assignment scheme again and during the k-th iteration based on the above adjustment rules,
again until all service vehicles are not overloaded. After that, we can get the following equation:
We can get the final bi , i = 1, 2, 3, · · · N , and T = tedge = bkj1 XN
max(ti )i = 0, 1, 2 · · · N . bk+1 k
j1 = bj1 − ( bkji − 1) (29)
N
Finally, we compare the value of tlocal and tedge . If tlocal ≤ P
bkji i=1
tedge , it is better to adapt local computing (s = 1 ), otherwise, i=1

18868 VOLUME 8, 2020

C. Chen et al.: Delay-Optimized V2V-Based Computation Offloading in Urban Vehicular Edge Computing and Networks

Step2. For Constraint C2 and Constraint C3, based on

the result of the Step1 adjustment, we check each dimen-
sion position of the current particle in turn. If its value is
greater than its maximum value according Eq.(23), directly
let its value be equal the maximum value. The PSO Algo-
rithm implementation steps are shown in Algorithm 1.

Algorithm 1 PSO Algorithm to Solve the Problem in Equa-

tion(16) FIGURE 4. The scene of offloading computing.
Input:
The parameter of the task vehicle, f0 , X0 (t);
The parameter of all service vehicles, fi , Xi (t); TABLE 2. Summary of the simulation parameters.

The parameter of the task,{c, data, t max };

Output:
Get bi (gbest);
1: Set k = 1, iterative number K, error criterion δ,
tedge (gbest 0 ) = ∞;
2: According to Constraint C1-C3, randomly determine the
initial position and velocity of l particles;
3: Check and adjust the initial position of l particles;
4: while k ≤ K ||tedge (gbest k ) − tedge (gbest k−1 ) ≥ δ do
5: for j = 1, 2, · · · , l do
6: calculate fitness function Tjk = tedge (Bkj ), update
pbestjk ;
7: end for
8: Update gbest k
9: k =k +1
10: for j = 1, 2, · · · , l do First, we compare the performance of three schemes over
11: Update position and velocity of l particles according different number of service vehicles. The task execution
Eq.(28); time is greatly reduced in Fig.5(a) by the Max-Min Fairness
12: end for Algorithm and the PSO Algorithm, respectively. Because we
13: end while unload the part of the task to the service vehicles, and use the
14: return gbest K ; idle computing resources of the service vehicle to help calcu-
late the task. From Fig.5(a), we can see that with the increase
of the number of service vehicles, the task execution time
Through PSO Algorithm, we can get the value of bi and in the latter two schemes is slowly decreasing. Because the
tedge = max(ti ). Similar to the processing method in the more service vehicles, the more idle computing resources that
Special Case, we compare tlocal and tedge , and then decide can be scheduled for the task vehicle. The task vehicle can
whether to offload the task to service vehicles. unload a larger proportion of the task to the service vehicle.
Not only that, from Fig.5(a), at the beginning, by the Max-
IV. PERFORMANCE EVALUATION Min Fairness Algorithm and the PSO Algorithm, the task
In this section, we carry out simulations to evaluate the execution time is reduced to the same extent. But when
performance of the proposed scheme. In order to simulate the number of service vehicles reaches 7, the performance
realistic traffic, we use the TIGER map provided by The using the PSO Algorithm starts to be slightly lower than that
United States Census Bureau, and then adopt the IDM_IM using the Max-Min Fairness Algorithm. However, it can be
model provided by VanetMobiSim [34] to generate vehicle seen from Fig.5(b) that the PSO algorithm does not call all
track files. Then, we analyze the track files to obtain the service vehicles at the same time. The cost of calling an
moving parameters of vehicles in the resource pool, such as service vehicle is not considered in this paper. If the service
tmax _i , and use ns3 to obtain related parameters of the IEEE vehicle is called for a fee, the PSO Algorithm will greatly
802.11p/WAVE standard, such as reliable communication reduce the cost of the unloading calculation. On the other
distance dis. Based on the parameters above, we compare the hand, this paper does not consider the scenario of multiple
performance of three computing schemes: Local computing tasks. If there are multiple tasks in the future that need to
scheme, Offloading scheme with Max-Min Fairness Algo- be unloaded, we can also get better performance by using
rithm and Offloading scheme with PSO Algorithm. The used PSO Algorithm. Because each task only needs to be unloaded
main simulation parameters are listed in Table 2. to a part of service vehicles. The remaining service vehicles

VOLUME 8, 2020 18869

 C. Chen et al.: Delay-Optimized V2V-Based Computation Offloading in Urban Vehicular Edge Computing and Networks

FIGURE 5. Effect of three schemes. (a)The task execution time over different number of service vehicles; (b)The
number of service vehicles involved in the calculation over different number of service vehicles. The default
setting is c = 20 × 109 , data = 750KB, t max = 2.5, l = 150, K = 200. The value of fi is set according to Table 2.
The number marked on the curve indicates the number of service vehicles.

FIGURE 6. Effect of three schemes. (a)The task execution time over different size of the task; (b)The number of
service vehicles involved in the calculation over different size of the task. The default setting is N = 30,
data = 750KB, t max = 2.5, l = 150, K = 200. The value of fi is set according to Table 2.

can be scheduled for other tasks. Compared with the Max- always scheduled using the Max-Min Fairness Algorithm.
Min Fairness Algorithm, the PSO Algorithm does not call all With the PSO Algorithm, the number of service vehicles
service vehicles at the cost of not minimizing task execution scheduled fluctuates with the increasing the task size. But
time. Considering the performance of the task execution time the fluctuation range is small. And in the worst case, 18 ser-
and the number of service vehicles involved in the calculation, vice vehicles are called. In the best case, less than half
we can see that the PSO Algorithm is slightly better than the of the service vehicles are called. All the results show the
Max-Min Fairness Algorithm. third scheme using the PSO Algorithm has the best robust
Second, we compare the robustness of three schemes over performance.
different size of the task. In order to better compare the Next, we analyze the impact of relative movement between
performance of the three schemes, the task size we set is vehicles on the performance of these three schemes. Accord-
relatively large at the beginning. When the task size grows ing to the IDM_IM mobility model we use, the traffic light
linearly, the task execution time locally also increases linearly time, the initial position and initial velocity of the vehicles
in Fig.6(a). The reason is that only the computing resources will affect the relative displacement 1Xi between the task
of the task vehicle can be used in this case. The robust vehicle and each idle vehicle. Based on Constraint 3, the com-
performance of this scheme is the worst. For remaining two munication time tconstraint_c3 between them will be affected,
schemes, although the task execution time increases with further affects the value of tmax _i . In the simulation process,
the increase of the task size, the growth rate is much lower we get different tconstraint_c3 by setting different movement
than the task execution time locally. In addition, the per- parameters in the VanetMobiSim software, and then get dif-
formance using the PSO Algorithm is also still slightly ferent tmax _i . But we still set the abscissa of Fig.7 directly
lower than using the Max-Min Fairness Algorithm. Simi- to different tmax _i instead of different movement parameters.
larly, from Fig.6(b), we also find that all service vehicles are From Fig.7(a), we can see that with the increase of tmax _i ,

18870 VOLUME 8, 2020

C. Chen et al.: Delay-Optimized V2V-Based Computation Offloading in Urban Vehicular Edge Computing and Networks

FIGURE 7. Effect of three schemes. (a)The task execution time over different service time of service vehicles;
(b)The number of service vehicles involved in the calculation over different service time of service vehicles. The
default setting is c = 10 × 109 , data = 750KB, t max = 2.5, l = 150, K = 200. The value of fi is set according to
Table 2.

FIGURE 8. Effect of the PSO Algorithm. (a)The task execution time over different number of iterations; (b)The
number of service vehicles involved in the calculation over different number of iterations. The default setting is
N = 30, data = 750KB, t max = 2.5, l = 150. The value of fi is set according to Table 2. The number marked on the
curve indicates the coordinates.

the task execution time in the scheme using the Max-Min V. CONCLUSION
Fairness Algorithm and the PSO Algorithm keeps decreasing In this paper, to reduce the expensive charge from cellular
until it is stable. Because the value of tmax _i is greater, the task or infrastructure communication, and guarantee the desired
vehicle could allocate a larger proportion of the task to the latency for delay-sensitive applications in IoV, a task offload-
service vehicles for reducing the task execution time. From ing scheme merely relying on V2V communication is pro-
Fig.7(b), the number of service vehicles called by the PSO posed. By formulating the one task and multiple cooperative
Algorithm fluctuates around 18, and nearly half of the service offloading pattern into a Min-Max problem, the total task
vehicles are not scheduled. execution time can be optimized with affordable commu-
Finally, we analyze the performance of the scheme using nication and computing overhead. Numerical results have
the PSO Algorithm. In Fig.8(a), when the number of iter- demonstrated that our scheme could outperform some state-
ations reaches 91, the task execution time is reduced to of-the-art strategies in terms of task execration time and con-
a minimum value, which is much smaller than the task sumed resources. In addition, the introduced PSO algorithm
execution time locally. This shows that using the PSO also contributes a lot to the convergence speed and accuracy
Algorithm does not require iterating many times to achieve of our envisioned problem. Our future work will investigate
optimal performance. The task vehicle does not need to the possibility to enroll multiple tasks into the offloading
spend too much time cost to schedule service vehicles with procedure using V2V communication, and analyze the impact
the PSO Algorithm. Through Fig.8(b), we can see that as of high mobility of vehicles on the offloading performance.
the number of iterations increases, the PSO Algorithm opti-
mizes the task execution time while the number of ser- REFERENCES
[1] H. Cao, S. Wu, G. S. Aujla, Q. Wang, L. Yang, and H. Zhu,
vice vehicles scheduled is also optimal. This shows the ‘‘Dynamic embedding and quality of service driven adjustment for
PSO Algorithm can obtain good performance in the iterative cloud networks,’’ IEEE Trans. Ind. Informat., to be published, doi:
process. 10.1109/tii.2019.2936074.

VOLUME 8, 2020 18871

 C. Chen et al.: Delay-Optimized V2V-Based Computation Offloading in Urban Vehicular Edge Computing and Networks

[2] S. Mao, S. Leng, J. Hu, and K. Yang, ‘‘Energy-efficient transmission [25] J. Du, L. Zhao, X. Chu, F. R. Yu, J. Feng, and C.-L. I, ‘‘Enabling low-
schemes for cooperative wireless powered cellular networks,’’ IEEE Trans. latency applications in LTE—A based mixed fog/cloud computing sys-
Green Commun. Netw., vol. 3, no. 2, pp. 494–504, Jun. 2019. tems,’’ IEEE Trans. Veh. Technol., vol. 68, no. 2, pp. 1757–1771, Feb. 2019.
[3] H. Cao, Y. Zhu, G. Zheng, and L. Yang, ‘‘A novel optimal mapping algo- [26] J. Luo, X. Deng, H. Zhang, and H. Qi, ‘‘Ultra-low latency service provision
rithm with less computational complexity for virtual network embedding,’’ in edge computing,’’ in Proc. IEEE Int. Conf. Commun. (ICC), May 2018,
IEEE Trans. Netw. Service Manag., vol. 15, no. 1, pp. 356–371, Mar. 2018. pp. 1–6.
[4] S. He, W. Tian, J. Zhang, K. Li, M. Zhang, and R. Zhu, ‘‘A high [27] Y. Liu, S. Wang, J. Huang, and F. Yang, ‘‘A computation offloading
efficient approach for power disturbance waveform compression in the algorithm based on game theory for vehicular edge networks,’’ in Proc.
view of heisenberg uncertainty,’’ IEEE Trans. Ind. Inf., vol. 15, no. 5, IEEE Int. Conf. Commun. (ICC), May 2018, pp. 1–6.
pp. 2580–2591, May 2019. [28] X. Hou, Y. Li, M. Chen, D. Wu, D. Jin, and S. Chen, ‘‘Vehicular fog
[5] X. Yuan, C. Li, Q. Ye, K. Zhang, N. Cheng, N. Zhang, and X. Shen, computing: A viewpoint of vehicles as the infrastructures,’’ IEEE Trans.
‘‘Performance analysis of IEEE 802.15.6-based coexisting mobile WBANs Veh. Technol., vol. 65, no. 6, pp. 3860–3873, Jun. 2016.
with prioritized traffic and dynamic interference,’’ IEEE Trans. Wireless [29] S. Li, Q. Ni, Y. Sun, G. Min, and S. Al-Rubaye, ‘‘Energy-efficient resource
Commun., vol. 17, no. 8, pp. 5637–5652, Aug. 2018. allocation for industrial cyber-physical IoT systems in 5G era,’’ IEEE
[6] Y. Dai, D. Xu, S. Maharjan, G. Qiao, and Y. Zhang, ‘‘Artificial intelligence Trans. Ind. Informat., vol. 14, no. 6, pp. 2618–2628, Jun. 2018.
empowered edge computing and caching for Internet of vehicles,’’ IEEE [30] J. Zhou, Y. Sun, Q. Cao, S. Li, H. Xu, and W. Shi, ‘‘QoS-based robust
Wireless Commun., vol. 26, no. 3, pp. 12–18, Jun. 2019. power optimization for SWIPT NOMA system with statistical CSI,’’ IEEE
[7] M. M. K. Tareq, O. Semiari, M. A. Salehi, and W. Saad, ‘‘Ultra reliable, low Trans. Green Commun. Netw., vol. 3, no. 3, pp. 765–773, Sep. 2019,
latency vehicle-to-infrastructure wireless communications with edge com- doi: 10.1109/tgcn.2019.2914736.
puting,’’ in Proc. IEEE Global Commun. Conf. (GLOBECOM), Dec. 2018, [31] L. Liu, C. Chen, Q. Pei, S. Maharjan, and Y. Zhang, ‘‘Vehicular edge
pp. 1–7. computing and networking: A survey,’’ 2019, arXiv:1908.06849. [Online].
[8] F. Yang, J. Li, T. Lei, and S. Wang, ‘‘Architecture and key technologies Available: https://arxiv.org/abs/1908.06849
for Internet of vehicles: A survey,’’ J. Commun. Inf. Netw., vol. 2, no. 2, [32] K. Zhang, Y. Mao, S. Leng, Y. He, and Y. Zhang, ‘‘Mobile-edge computing
pp. 1–17, Jun. 2017. for vehicular networks: A promising network paradigm with predictive off-
[9] C. Chen, Y. Jin, Q. Pei, and N. Zhang, ‘‘A connectivity-aware intersection- loading,’’ IEEE Veh. Technol. Mag., vol. 12, no. 2, pp. 36–44, Jun. 2017.
based routing in VANETs,’’ EURASIP J. Wireless Commun. Netw., [33] N. Ababneh and H. Labiod, ‘‘A performance analysis of VANETs rout-
vol. 2014, no. 1, p. 42, 2014. ing protocols using different mobility models,’’ in Proc. IEEE Int. Conf.
[10] C. Chen, Q. Pei, and X. Li, ‘‘A GTS allocation scheme to improve multiple- Wireless Commun., Netw. Inf. Secur. (WCNIS), Jun. 2010, pp. 498–502.
access performance in vehicular sensor networks,’’ IEEE Trans. Veh. Tech- [34] J. Härri, F. Filali, C. Bonnet, and M. Fiore, ‘‘VanetMobiSim: Generating
nol., vol. 65, no. 3, pp. 1549–1563, Mar. 2016. realistic mobility patterns for VANETs,’’ in Proc. 3rd Int. Workshop Veh.
[11] J. Guan, R. Lai, and A. Xiong, ‘‘Wavelet deep neural network for stripe Ad Hoc Netw., 2006, pp. 96–97.
noise removal,’’ IEEE Access, vol. 7, pp. 44544–44554, 2019. [35] M. Treiber, A. Hennecke, and D. Helbing, ‘‘Congested traffic states in
[12] Y. Yang, D. Li, and Z. Duan, ‘‘Chinese vehicle license plate recognition empirical observations and microscopic simulations,’’ Phys. Rev. E, Stat.
using kernel-based extreme learning machine with deep convolutional Phys. Plasmas Fluids Relat. Interdiscip. Top., vol. 62, no. 2, p. 1805, 2000.
features,’’ IET Intell. Transp. Syst., vol. 12, no. 3, pp. 213–219, Apr. 2018. [36] M. Treiber and D. Helbing, ‘‘Explanation of observed features of self-
[13] Y. Dai, D. Xu, S. Maharjan, Z. Chen, Q. He, and Y. Zhang, ‘‘Blockchain organization in traffic flow,’’ 1999, arXiv:cond-mat/9901239. [Online].
and deep reinforcement learning empowered intelligent 5G beyond,’’ IEEE Available: https://arxiv.org/abs/cond-mat/9901239
Netw., vol. 33, no. 3, pp. 10–17, May 2019. [37] IEEE Standard for Wireless Access in Vehicular Environments (WAVE)—
[14] Y. Dai, D. Xu, S. Maharjan, and Y. Zhang, ‘‘Joint load balancing and Multi-Channel Operation, IEEE Standard 1609.4-2016, 2016, pp. 4–1609.
offloading in vehicular edge computing and networks,’’ IEEE Internet [38] M. Amadeo, C. Campolo, and A. Molinaro, ‘‘Enhancing IEEE 802.11
Things J., vol. 6, no. 3, pp. 4377–4387, Jun. 2019. p/WAVE to provide infotainment applications in VANETs,’’ Ad Hoc Netw.,
[15] S. Xiao-Hong, ‘‘Key technology and its application of IoV,’’ Commun. J., vol. 10, no. 2, pp. 253–269, 2012.
vol. 4, pp. 47–50, 2013. [39] C. Campolo and A. Molinaro, ‘‘Data rate selection in WBSS-based
[16] X. Xu, Y. Xue, L. Qi, Y. Yuan, X. Zhang, T. Umer, and S. Wan, ‘‘An edge IEEE 802.11 p/WAVE vehicular ad hoc networks,’’ in Proc. 7th Int.
computing-enabled computation offloading method with privacy preser- Symp. Commun. Syst., Netw. Digit. Signal Process. (CSNDSP), 2010,
vation for Internet of connected vehicles,’’ Future Gener. Comput. Syst., pp. 412–416.
vol. 96, pp. 89–100, Jul. 2019.
[17] Z. Duan, Y. Yang, K. Zhang, Y. Ni, and S. Bajgain, ‘‘Improved deep hybrid
networks for urban traffic flow prediction using trajectory data,’’ IEEE
Access, vol. 6, pp. 31820–31827, 2018.
[18] S. Mao, S. Leng, J. Hu, and K. Yang, ‘‘Power minimization resource
allocation for underlay MISO-NOMA SWIPT systems,’’ IEEE Access,
vol. 7, pp. 17247–17255, 2019.
[19] P. Guo, W. Hou, L. Guo, W. Sun, C. Liu, H. Bao, L. H. K. Duong, and
W. Liu, ‘‘Fault-tolerant routing mechanism in 3D optical network-on-chip
based on node reuse,’’ IEEE Trans. Parallel Distrib. Syst., vol. 31, no. 3, CHEN CHEN (Senior Member, IEEE) received
pp. 547–564, Mar. 2020, doi: 10.1109/tpds.2019.2939240. the B.Eng., M.Sc., and Ph.D. degrees in telecom-
[20] C. Li, S. Wang, X. Huang, X. Li, R. Yu, and F. Zhao, ‘‘Parked vehicu- munication from Xidian University, Xi’an, China,
lar computing for energy-efficient Internet of vehicles: A contract theo- in 2000, 2006, and 2008, respectively. He was a
retic approach,’’ IEEE Internet Things J., vol. 6, no. 4, pp. 6079–6088, Visiting Professor with the EECS in the Univer-
Aug. 2019, doi: 10.1109/jiot.2018.2869892. sity of Tennessee and the CS in the University of
[21] S. Mao, S. Leng, K. Yang, Q. Zhao, and M. Liu, ‘‘Energy efficiency and California. He is currently an Associate Profes-
delay tradeoff in multi-user wireless powered mobile-edge computing sys- sor with the Department of Telecommunication,
tems,’’ in Proc. IEEE Global Commun. Conf. (GLOBECOM), Dec. 2017,
Xidian University. He is also the Director of the
pp. 1–6.
Xi’an Key Laboratory of Mobile Edge Computing
[22] J. Guo, B. Song, F. R. Yu, Y. Chi, and C. Yuen, ‘‘Fast video frame
correlation analysis for vehicular networks by using CVS-CNN,’’ IEEE
and Security, and the Director of the Intelligent Transportation Research
Trans. Veh. Technol., vol. 68, no. 7, pp. 6286–6292, Jul. 2019. Laboratory, Xidian University. He has authored/coauthored three books, over
[23] J. Guo, B. Song, Y. He, F. R. Yu, and M. Sookhak, ‘‘A survey on 90 scientific articles in international journals and conference proceedings.
compressed sensing in vehicular infotainment systems,’’ IEEE Commun. He has contributed to the development of five copyrighted software systems
Surveys Tuts., vol. 19, no. 4, pp. 2662–2680, 4th Quart., 2017. and invented 90 patents. He is also a member of ACM, Chinese Institute
[24] Y. Wang, Y. Zhang, M. Sheng, and K. Guo, ‘‘On the interaction of video of Electronics and a Senior Member of China Computer Federation (CCF).
caching and retrieving in multi-server mobile-edge computing systems,’’ He also serves as a General Chair, a PC Chair, and a Workshop Chair or a
IEEE Wireless Commun. Lett., vol. 8, no. 5, pp. 1444–1447, Oct. 2019. TPC Member of a number of conferences.

18872 VOLUME 8, 2020

C. Chen et al.: Delay-Optimized V2V-Based Computation Offloading in Urban Vehicular Edge Computing and Networks

LANLAN CHEN received the B.Eng. degree in XIAOMING YUAN received the B.E. degree
communication engineering from the Heifei Uni- in electronics and information engineering from
veesity of Techology, Heifei, China, in 2013. She Henan Polytechnic University, China, in 2012, and
is currently pursuing the master’s degree in trans- the Ph.D. degree in communication and infor-
portation information engineering and control with mation system from Xidian University, China,
Xi’an University. Her research interests include in 2018. From 2016 to 2017, she was a Vis-
vehicular ad hoc networks, named data networks, iting Scholar with the Broadband Communica-
mobile edge computing, and the Internet of Thing. tions Research Group, Department of Electrical
and Computer Engineering, University of Water-
loo, Waterloo, ON, Canada. She is currently with
Northeastern University at Qinhuangdao. Her research interests include
cloud/edge computing, medium access control and performance analysis for
wireless body area networks, and the Internet of Things.

LEI LIU (Student Member, IEEE) received the DAPENG LAN received the B.Eng. degree in
B.Eng. degree in communication engineering microelectronics from Sun Yat-sen University,
from Zhengzhou University, Zhengzhou, China, Guangzhou, China, in 2014, the M.Sc. degree in
in 2010, and the M.Sc. degree in communica- ICT innovation from the KTH Royal Institute of
tion engineering from Xidian University, Xi’an, Technology, Stockholm, Sweden, in 2016, and
China, in 2013. He is currently pursuing the Ph.D. the M.Sc. degree in innovation in information
degree with Xidian University. He has published and communication technology from the Techni-
more than ten journals and conference publica- cal University of Berlin, Berlin, Germany, in 2017.
tions. His research interests include vehicular ad In 2017, he was with InnoEnergy, where he was
hoc networks, intelligent transportation, mobile involved in smart building energy management in
edge computing, and the Internet of Thing. Sweden. He was a thesis Student with the ABB Corporate Research Center,
Västerås, Sweden, in 2016. He is currently a Ph.D. Research Fellow with
the Department of Informatics, University of Oslo, Norway. His research
interests include fog computing, the Internet of Things, and distributed
systems.

SHUNFAN HE was born in Wuhan, Hubei, China, ZHUANG CHEN received the B.S. degree in Inter-
in 1984. He received the Ph.D. degree in power net of Things engineering from the Qingdao Uni-
electrical engineering from the Huazhong Univer- versity of Science and Technology, China, in 2017.
sity of Science and Technology, Wuhan, China, He is currently pursuing the M.S. degree with
in 2013. He is currently with the College of the School of Computer and Information Secu-
Computer Science, South-Central University for rity, Guilin University of Electronic Technology,
Nationalities, Wuhan, China. He has authored China. He is also a Visiting Student with the
more than ten articles in international journals Department of Informatics, University of Oslo.
in the areas of electromagnetic compatibility, big His current research interests include mobile edge
data processing, electronic transducers, and power computing, multimedia cache, wireless networks,
disturbance analysis. His current research interests include electromagnetic deep reinforcement learning, and blockchain.
compatibility, big data processing, and power disturbance analysis.

VOLUME 8, 2020 18873