10 1109@icc 2019 8761403

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 7

UAV-Aided Low Latency Mobile Edge Computing with

mmWave Backhaul
Ye Yu∗† , Xiangyuan Bu∗ , Kai Yang∗ , Hongyuan Yang‡ , and Zhu Han†
∗ School of Information and Electronics, Beijing Institute of Technology, Beijing 100081, China.
† Department of Electrical and Computer Engineering, University of Houston, Houston, TX 77004, USA.
† Department of Computer Science and Engineering, Kyung Hee University, Seoul, South Korea, 446-701.
‡ College of Instrumentation and Electrical Engineering, Jilin University, Changchun, China.

Abstract—Recently, unmanned aerial vehicle (UAV) has been In [2], the authors consider the UAV as a relay and minimize
considered as a promising technique in mobile edge computing the outage probability of the relay network, which the joint
networks, and enhances the performances of ultra-reliable and trajectory design and power control are involved. In [4], the
low-latency services. In this paper, we propose a UAV-aided low
latency mobile edge computing network with millimeter wave authors study the UAV-enabled information collection system
(mmWave) backhaul. There are two types of communication with the tradeoff of the energy efficiency, including the circular
links in our network, the UAV link and the ad hoc link. We flight and straight flight design. In [5], the authors give a link
jointly consider the network resource allocation and the UAV capacity analysis between the UAVs with 3-D trajectories. The
trajectory design in our proposed problem, which is a non-convex principles for U2U channels and UAV-to-ground link are given
mixed integer nonlinear programming (MINLP). For solving the
proposed problem, we give a novel algorithm framework. We and verified by the simulation results.
adopt generalized Benders decomposition (GBD) as the outer Mobile edge computing (MEC) is another promising tech-
loop algorithm to separate the integer variables and continuous nique in 5G. MEC can achieve lower latency and better
variables. In the inner loop, the continuous primal problem mobility by bringing the computing resources closer to the end
is solved by the joint alternating direction method of multi- users. MEC can significantly reduce the latency comparing to
pliers (ADMM), Dinkelbach algorithm and successive convex
approximation (SCA) algorithm. The simulation results show that traditional cloud computing. In 5G applications, many latency-
our proposed system architecture and algorithm framework can critical tasks have the milliseconds requirements to finish the
achieve low latency for the time-sensitive task in mobile edge task, such as virtual reality (VR), facial recognition and many
computing. other industrial IoT applications. Considering the potential
Index Terms—UAV; Mobile edge computing; mmWave; Gen- congestions and loss in the backhaul to the cloud, MEC shows
eralized Benders decomposition.
a tremendous advantage by processing beside the users. In [6],
the authors give a comprehensive survey of the research and
I. I NTRODUCTION development about this new technology. In [7], the authors
Contemporarily, unmanned aerial vehicle (UAV) becomes survey the influences and applications in the area of 3C
a compelling technique in 5G, with networking flexibility (Computing, Caching, and Communications), and the future
and economic saving. The emerging UAV communication challenges are presented as well. In [8], the authors jointly
networks have strikingly different architectures with many consider the computation offloading and resource allocation
other wireless networks. Although the benefits which UAV in the MEC network, and the authors transform the problem
brings are enormous, there are still many challenges for the into the convex problem and solve it distributely in an efficient
researchers. Firstly, in the UAV network, the network topology way. In [9], the authors propose an online learning algorithm
is changing time-to-time, which means the communication to solve the offloading and autoscaling in the energy harvesting
channel is highly dynamic. Despite the advantages of the enabled MEC network, and the convergence and complexity
line of sight channel between the UAV and the terrestrial are analyzed and proved.
equipments, the applications are harder to implement in such a Although the UAV communication and MEC networks have
network even comparing to the high-speed vehicular network been studied over years, there is little work for studying the
[1]. Furthermore, the capacity volume shortage for the UAV mixed integer programming problem in the UAV-aided MEC
is an obvious demerit for many heavy tasks for the 5G network. The main contributions of this paper are stated as
applications. In 5G, the ultra reliable and low latency are follows.
essential features for many time-sensitive services. Although • We propose a UAV-aided low latency MEC network with
UAV upgrades frequently, the high capacity UAV is still mmWave backhaul. We jointly consider the resource al-
expensive and even cannot apply to some scenarios [2]. In location problem and UAV trajectory design to minimize
addition, the energy management for the UAV is widely the total delay in the network. Both UAV link and ad hoc
studied, due to the limited battery and load of the UAV. In link are involved in reducing the latency.
[3], the authors designed the UAV trajectory to minimize the • We propose a novel algorithm framework for solving our
completion time of the file multicasting process while with proposed mixed integer nonlinear programming (MINLP)
the high possibility to recover the file at the ground terminals. problem. The generalized Benders decomposition per-

978-1-5386-8088-9/19/$31.00 ©2019 IEEE


8$9
and the large task size, the mobile users cannot execute the
tasks on their own devices. Each user can choose to execute
the task by offloading to the edge cloud through the UAV or
through the ad hoc link. The binary decision variable is α.
:L)L If αn = 1, the nth user choose to transmit the task data to
PP:DYHEDFNKDXO the UAV and the UAV will send it back to the edge cloud for
processing. After the task is executed at the edge cloud, the
edge cloud will send the result data back to the UAV, and then
the UAV will transmit the data to the nth user. If αn = 0, the
8$9XVHU nth user will choose to offload the task by the ad hoc link
back to the edge node without the help from the UAV. After
%DVH6WDWLRQ
the processing by the edge node, the result will be sent back
$GKRFXVHU to the corresponding user by the original route. We assume
that the strategy of executing the task is certain, and the task
is not dividable for all the users.
For the path planning, we set up a three-dimensional Carte-
Fig. 1: System Model
sian coordinate to show the trajectory of the UAV. The location
at time t is formulated as q(t) = (x(t), y(t), z(t)). We assume
that the UAV keeps flying at a fixed altitude z(t) = H.
forms as the outer loop algorithm to separate the integer The total flying time for the UAV is T seconds, which is
variables and continuous variables. For the continuous divided into M frame with time length δ. Furthermore, the
primal problem in the inner loop, we propose a joint initial location and the landing location is predetermined by
alternating direction method of multipliers (ADMM), q(1) = (x(1), y(1), H) and q(M ) = (x(M ), y(M ), H). We
Dinkelbach algorithm and successive convex approxima- assume that the UAV has a fixed location during one frame.
tion (SCA) algorithm to deal with the nonlinear fractional The duration of one frame should be chosen small enough
programming problem. comparing to the total flying time T . Thus,
• We validate the proposed algorithm by simulation results.  the location for
the UAV at the mth frame is expressed as xU m
AV
, ym
U AV
,H .
The convergence performance for our algorithm and the Furthermore, the location of the nth userand the edge node
system properties are shown with analysis. 
is predetermined as (xuser , y user , 0) and xedge , y edge , 0 . In
The rest of the paper is organized as follows. In Section addition, the velocity of the UAV in the frame m is given as
II, we introduce our UAV-aided MEC network and formulate
q(m + 1) − q(m)
the delay minimization problem as an MINLP problem. In vm = . (1)
Section III, we propose our algorithm to solve the proposed δ
problem. In Section IV, we give the simulation results. At last, Therefore, the speed constraint for the UAV is
we conclude the paper in Section V.    
xU U AV 2 + y U AV − y U AV 2 ≤ δv max , ∀m,
m+1 − xm
AV
m+1 m
II. S YSTEM M ODEL AND P ROBLEM F ORMULATION (2)
In this section, we present our system model and problem
where v max is the maximum speed of the UAV.
formulation, including the analysis and transformation of the
The channels between the UAV and the users and the
original problem.
channels between the UAV and edge node are LoS (Line of
A. System Model Sight)-dominated. Thus, the path loss model is modeled by
In this paper, we consider the UAV-aided MEC network as the distance. The channel model between the nth user and the
shown in Fig. 1. N mobile users (MU) are distributed in the UAV at the mth frame is given by
cell. Different from the UAV in [10], we assume that the UAV 2U h0
n,m =
hU 2 2
cannot process the data. This is due to the limited load of the (xU AV− xuser )
+ (ym U AV − y user ) + H 2
m n n
UAV cannot afford too many delay sensitive tasks. Thus, the h0
only function for the UAV is collecting the data and sending = , (3)
||qm − qnuser ||2 + H 2
it back to the edge node through the mmWave backhaul. We
assume that the uplink and downlink channels between users where h0 is the channel power gain at the reference distance
and UAV are allocated the same B ct bandwidth by frequency of 1 m. Similarly, the channel model from the UAV to the
division duplex (FDD), and the uplink and downlink channels edge node is given by
between UAV and edge node are allocated B bk by FDD as 2E h0
well. hU
m = 2 2
(xU AV − xedge ) U AV − y edge ) + H 2
+ (ym
For the nth user, the task is modeled as {ωn , Ln }, where ωn m
h0
is the CPU cycles needed to compute the task and Ln is the == edge 2
. (4)
data length of the task. Due to the limited computing capacity ||qm − qn || + H2
In addition, we assume all the uplink and downlink channels
are reciprocal due to LoS.
For the ad hoc link, each user can choose a certain route min Dtotal (9)
to the edge node. The reason why is that in our model the α,L,q
 U AV AV 2

mobile users are fixed at certain locations. Moreover, they s.t. xm+1 − xUm
use orthogonal channels to offload tasks which will not cause  U AV 
U AV 2 2
congestions. Thus, we do not consider the routing problem + ym+1 − ym ≤ (δv max ) , ∀m, (10)
in this paper. For the nth user, the transmission delay for the 
N

task is denoted as Dnadhoc . The routing problem will be further αn Ln,m ≤ C U AV , ∀m, (11)
studied in our future work. n=1
M

B. Problem Formulation αn Ln,m = Ln , ∀n, (12)


m=1
We aim to minimize the execution time of all the tasks in
q(1) = (x(1), y(1), H) , q(M ) = (x(M ), y(M ), H) ,
our mobile edge network. In general, the data size of the tasks
(13)
results are much smaller than the input data, and the downlink
rate is much higher than the uplink [11]. Furthermore, the 0 ≤ xU
m
AV
≤ r, 0 ≤ ym
U AV
≤ r, ∀m, (14)
downlink is similar to the uplink which can be added to Ln,m ≥ 0, ∀n, ∀m, (15)
our problem easily. For simplicity, we neglect the downlink αn ∈ {0, 1}, ∀n, (16)
procedure in this paper. For the nth UAV link user, the
transmission delay includes two parts. The first part is the
where pmax is the maximum transmit power for the nth user.
transmission delay between the nth user and UAV, which n
Constraint (10) indicates that the flying speed of the UAV
denotes as DnU 2U . The second part is the transmission delay
cannot exceed the maximum speed. Constraint (11) shows
through the mmwave backhaul, which denotes as DnU 2E . We
the total offloading task data size at the mth frame cannot
denote that the nth user transmit Ln,m bits of the task during
exceed the capacity of the UAV. Constraint (12) enforces the
the mth frame. Thus, the two kinds of transmission delay is
completion of all the tasks within M frames. Constraint (13)
given as
and (14) give the initial point, destination, and the flying

M
L range of the UAV. Constraint (15) indicates the data length is
DnU 2U =  n,m
pn hU 2U
, (5) nonnegative. Constraint (16) shows that the decision variable
m=1 B ct log2 1 + N0 B n,m
ct αn is a binary variable in our problem.
and

M
Ln,m III. P ROPOSED A LGORITHM
DnU 2E =  . (6)
pU AV hU 2E
m=1 B bk log2 1 + m
N0 B bk
In this section, we propose our algorithm to (9), which is
pn is the transmit power for the nth user. pU AV is the transmit an MINLP problem and will be handled iteratively in our
power for the UAV. We assume that both the users and UAV algorithm framework.
always transmit data by the maximum power to minimize the
delay. For the nth ad hoc link user, the delay is predefined as
Dnadhoc . For either UAV user or ad hoc user, the computing A. Generalized Benders Decomposition
duration at the edge node at the mth frame is given as
ωn The Generalized Benders Decommposition (GBD) can de-
Dnc = , (7) compose the problem into a primal problem and master
φn f e
problem and solve them iteratively. Thus, we deem the GBD
where f e is the total computing resource (CPU cycles per algorithm as the outer loop algorithm. We describe the proce-
second) at the edge node. φn is the proportion of computing dure of GBD in the following.
resource allocated to the nth task. 1) Initialization: First of all, we set the loop index l to 1.
Consequently, the total delay is given as The integer variable α is initialized to a feasible point. The
 maximum number of the iterations Lmax is set. The upper
N
 
Dtotal = αn DnU 2U + DnU 2E + Dnc bound U B(1) and lower bound LB(1) are set to positive
n=1 infinity and negative infinity, respectively.
N
  2) Primal Prolem: In the primal problem, we fix the integer
+ (1 − αn ) Dnadhoc + Dnc . (8) variable α to the initialized point. For each iteration, the
n=1
integer variable αl is fixed to the solution generated by the
In this paper, we aim to minimize the total task delay of the master problem from the l − 1 iteration. Thus, the primal
network. Thus, the minimization problem is formulated as problem becomes a continuous problem as
primal problem is infeasible. The corresponding Lagrangian
equations in the lth iteration are given as

N
 
min αn(l) DnU 2U + DnU 2E + Dnc (17) 
N
 
L,q
n=1 L(λl , αl ) = inf αn(l) DnU 2U + DnU 2E + Dnc
L,q
s.t. (10), n=1


N 
M
 AV 
U AV 2
αn(l) Ln,m ≤ C U AV , ∀m, (18) + λl1 (m)( xU
m+1 − xm
m=1
n=1
 U AV 
U AV 2 2
M + ym+1 − ym − (δv max ) )
αn(l) Ln,m = Ln , ∀n, (19) 
M 
N
m=1 + λl2 (m)( αn(l) Ln,m − C U AV )
(13), (14)(15). m=1 n=1
N 
M
The continuous primal problem is a non-convex problem, we + λl3 (n)( αn(l) Ln,m − Ln )
will solve the primal problem in the next subsection. n=1 m=1
3) Feasibility Check: If the primal problem is infeasible due M
to the integer solution from the master problem, the feasibility + λl4 (m)(xU
m
AV
− r)
problem needs to be solved which is given as m=1
M
+ λl5 (m)(ym
U AV
− r) (32)

M 
N
m=1
1 2 4
min (zm + zm + zm ) + zn3 (20)
L,q and
m=1 n=1
 U AV 
AV 2
 U AV 
U AV 2
s.t. xm+1 − xU + ym+1 − ym 
M
 AV 
m U AV 2
max 2 1
L̄(λ̄ , ᾱ ) = inf
l l
λ̄l1 (m)( xU
m+1 − xm
− (δv ) ≤ zm , ∀m, (21) L,q
m=1
 2

N
+ ym+1
U AV
− ym
U AV 2 1
− (δv max ) − zm )
αn(l) Ln,m −C U AV
≤ 2
zm , ∀m, (22)
n=1

M 
N
+ λ̄l2 (m)( 2
αn(l) Ln,m − C U AV − zm )
M
αn(l) Ln,m − Ln ≤ zn3 , ∀n, (23) m=1 n=1

m=1
N 
M

4 + λ̄l3 (n)( αn(l) Ln,m − Ln − zn3 )


xU
m
AV
− r ≤ zm , ∀m, (24) n=1 m=1
5
ym − r ≤ zm ,
U AV
∀m, (25) M
4
1
zm 2
, zm , zn3 , zm
4
, zm5
≥ 0, ∀m, ∀n, (26) + λ̄l3 (m)(xU
m
AV
− r − zm )
m=1
where 1
zm , 2
zm , zn3 ,
and 4
zm 5
zm
are auxiliary variables to show M
5
the violence of the constraints. By solving this feasibility prob- + λ̄l4 (m)(ym
U AV
− r − zm ), (33)
m=1
lem, the optimal dual variables and optimal primal solution
will be used in the master problem. which L(λl , αl ) is generated from the primal problem and
4) Master Problem: In the master problem, we fix the con- L̄(λ̄l , ᾱl ) is given from the feasibility problem. (30) and (31)
tinuous variables to the optimal point in the primal problem. compose the hyperplane space to cut the region of the master
Thus, the master problem is given as problem. In each iteration, new cuts will added to the former
hyperplane set to approximate and acquire the optimal point of
min η (27) the integer solution. The master problem is an MILP problem
α,η
with just one auxiliary continuous variable η. This smaller

N
s.t. αn L(l)
n,m ≤ C
U AV
, ∀m, (28) scale MILP problem can be solved with many solvers easily
n=1 [12].
M 5) Bounds Update: Then, we need to update the upper
αn L(l)
n,m = Ln , ∀n, (29) bound and lower bound to check the convergence of the
m=1 algorithm. We define the stopping criterion as the difference
η ≥ L(λk , αk ), k ∈ Ω, (30) between the upper bound and lower bound under a predefined
0 ≥ L̄(λ̄ , ᾱ ),
i i
i ∈ Φ, (31) threshold. In the lth iteration, we update the upper bound as
U B(l) = min(U B(l − 1), f (l)), (34)
where (30) is the optimality cut and (31) is the feasibility cut.
Ω is the set of all the iteration indices for the primal problem where f (l) is the optimal objective function after solving the
is feasible. Φ is the set of all the iteration indices for the primal problem or the feasibility problem. We define the lower
Algorithm 1 Generalized Benders Decomposition The corresponding augmented Lagrangian function with the
1: Initialize: loop index l, maximum number of the iterations equality constraints is given as
Lmax , U B(1) = ∞, LB(1) = −∞, threshold  
N  (l)
M
αn Ln,m
2: while U B(l) − LB(l) ≥ ε or l < Lmax do L=  
pn hU 2U 
n,m (q )
n=1 m=1 B log2 1 +
3: Primal Problem / Feasibility Problem ct
N0 B ct
4: if (17) is feasible then

N 
M
α L
(l)
5: Solve the primal problem in (17) and obtain the +  n n,m 
optimal solution {Ll , q l } and corresponding pU AV hU 2E (q)
n=1 m=1 B log2 1 +
6: bk m
bk
N0 B
7: Lagrangian multiplier λl
8: else if (17) is infeasible then 
N 
M 
M
 
+ λL
n,m (Ln,m − Ln,m ) + λqm (qm − qm )
9: Solve the feasibility problem in (20) and obtain the
n=1 m=1 m=1
10: optimal solution and the corresponding Lagrangian
ρ N M
ρ  
M
11: multiplier λ̄l + ||Ln,m − Ln,m ||2 + ||q − qm ||2 .
12: end if 2 n=1 m=1 2 m=1 m
13: Master Problem (42)
14: Add new cut (optimality cut or feasibility cut) to
Thus, the primal problem can be decoupled as
15: the master problem in (27) and solve it

N  (l)
16: Bound Update M
αn Ln,m
17: Update the upper bound and lower bound as (34) min  
 
L ,q pn hU 2U 
n,m (q )
n=1 m=1 B log2 1 +
ct
ct
18: and (35), respectively N0 B
19: l =l+1 
N 
M 
M
 
20: end while + λL
n,m Ln,m + λqm qm
n=1 m=1 m=1

ρ N M
ρ  
M
+ ||Ln,m − L̃n,m ||2 + ||q − q̃m ||2
bound as 2 n=1 m=1 2 m=1 m
(43)
LB(l) = max(LB(l − 1), η). (35)  2
 
s.t. m+1 − xm
xU AV U AV
The upper bound and lower bound update in each iteration  2
until the algorithm converge. U AV  U AV  2
+ ym+1 − ym ≤ (δv max ) , ∀m, (44)
After checking the bounds, the iteration will go to solve
the primal problem again. The whole procedure of the GBD Ln,m ≥ 0, ∀n ∀m, (45)
algorithm is described in Algorithm 1. 
N
αn(l) Ln,m ≤ C U AV , ∀m, (46)
B. Joint ADMM, Dinkelbach Algorithm, and SCA Algorithm n=1
M
Although we decouple the integer variables and continuous αn(l) Ln,m = Ln , ∀n, (47)
variables by the GBD, the proposed primal problem is still a m=1
nonconvex programming problem. Thus, we propose an inner and
algorithm to tackle the problem.

N 
M
α L
(l)
We introduce new auxiliary variables as min  n n,m 
L,q pU AV hU 2E (q)
n=1 m=1 B
bk log
2 1+
m
Ln,m = Ln,m ,
bk
∀n, m, (36) N0 B

q  = q, ∀m. (37) 
N 
M 
M
− λL
n,m Ln,m − λqm qm
n=1 m=1 m=1
Therefore, the primal problem can be reformulate as
⎛ ⎞ ρ N M
ρ  
M

N M  + ||L̃n,m − Ln,m ||2 + ||q̃ − qm ||2


L 2 n=1 m=1 2 m=1 m
αn(l) ⎝ ⎠
n,m
min  (38)
L,q pn hU 2U
n,m (q )

(48)
n=1 m=1 B ct log
2 1 + N0 B ct
⎛ ⎞ s.t. (10), (13), (14), (15), (18), (19),
N M
L
αn(l) ⎝  ⎠ (39)
n,m
+  where L̃n,m , L̃n,m q̃m and q̃m

are the fixed value generated
pU AV hU 2E
n,m (q)
n=1 m=1 B bk log
2 1 + N0 B bk from the last variables updating. After solving above problems
s.t. (10), (13), (14), (15), (18), (19), sequentially, the dual variables update as
Ln,m = Ln,m , ∀n, m, (40) λL 
n,m (j + 1) = λn,m (j) + ρ(Ln,m − Ln,m )
L
(49)
q  = q, ∀m. (41) 
λpm (j + 1) = λpm (j) + ρ(qm − qm ) (50)
where j is the loop index for the ADMM. Algorithm 2 Joint ADMM, Dinkelbach Algorithm and SCA
Both problems (43) and (48) has the fractional forms. We Algorithm
need to transform these problems into processable forms. 1: Initialization
According to [13], our problems have the form of sum- 2: while The ADMM stopping criterion is not satisfied do
of-ratios problem. Therefore, we can further transform the 3: Update the primal and dual variables sequentially by
problems as solving (43), (48), (49) and (50)
4: 1. Solve (43):

N 
M
5: while The Dikelbach algorithm stopping criterion is
min
 
αn(l) Ln,m not satisfied do
L ,q
n=1 m=1
6: 1). Solve (51):
N  M
pn h U 2U 
n,m (q ) while The SCA algorithm stopping criterion is not
 7:
− θn,m B ct log2 1+
n=1 m=1
N0 B ct satisfied do
8: .
1 Solve the convex approximated problem
N  M 
M
  according to (53)
+ λL
n,m Ln,m + λqm qm 9:
n=1 m=1 m=1 10: .
2 Update the local points
N  end while
ρ  
M M 11:
ρ
+ ||Ln,m − L̃n,m ||2 + ||q − q̃m ||2 12: 2). Update the auxiliary variable
2 n=1 m=1 2 m=1 m  α(l) L
13: θn,m =  n n,mU 2U   , ∀n, ∀m
pn h (q )
(51) B ct log2 1+ Nn,m B ct
0
s.t. (44), (45), (46), (47), 14: end while
15: 2. Solve (48):
and 16: while The Dikelbach algorithm stopping criterion is
not satisfied do

N 
M
min αn(l) Ln,m 17: 1). Solve (52):
L,q 18: while The SCA algorithm stopping criterion is not
n=1 m=1
M  2E
 satisfied do
pU AV hU
m (q) .
− θm B bk
log2 1+ 19: 1 Solve the convex approximated problem
m=1
N0 B bk 20: according to (54)
N M 
M 21: .
2 Update the local points
− λL
n,m Ln,m − λqm qm 22: end while
n=1 m=1 m=1 23: 2). Update the auxiliary variable
α(l) L
ρN M
ρ  
M
24: θm =  n n,m  , ∀m
+ ||L̃n,m − Ln,m ||2 + ||q̃ − qm ||2 B bk log2 1+
pU AV hU 2E (q)
m
2 n=1 m=1 2 m=1 m N0 B bk
25: end while
(52) 26: 3. Update the dual variables (49) and (50)
s.t. (10), (13), (14), (15), (18), (19), 27: end while


where θn,m and θm are auxiliary variables. Although the ob-
jective function is a parametric polynomial function instead of and
fraction, it is still not a convex function. The composite func-
tion with log functions is the reason why it loses the convexity. F 2 (q) =
However, wecan note that these log functions are convex γ2 log2 e(||qm − qnedge ||2 − ||qm
i
− qnedge ||2 )
2 
with
 2  −
  i − q edge ||2 )(H 2 + ||q i − q edge ||2 + γ )
respect to xUm
AV
− xuser
n + ym U AV
− ynuser and (H 2 + ||qm n m n 2
 2  U AV 2  γ2
+ log2 (1 + 2 ), (54)
xUm
AV
− xedge + ym − y edge , respectively. The H + ||qm i − q user ||2
n
first order Taylor expansion will give the tight lower bound U AV

of the convex function. Thus, we adopt successive convex where γ1 = Np0nBh0ct and γ2 = pN0 B bk
h0
. Therefore, the problems
approximation (SCA) to approximate the non-convex function. become convex problems, which can be solved efficiently by
The corresponding log functions are approximated with the CVX.
given local point q i and q i in the ith iteration of SCA as IV. S IMULATION R ESULTS
In this section, we evaluate the performance of our pro-
F 1 (q  ) =

posed algorithm. We assume there are 10 users randomly and
γ1 log2 e(||qm − qnuser ||2 − ||qm i
− qnuser ||2 ) uniformly distributed in a square with the 200 m side length.

(H 2 + ||qm i − q user ||2 )(H 2 + ||q i − q user ||2 + γ )
n m n 1 The bandwidth for users B ct is allocated as 20 MHz, and the
γ1 bandwidth for the mmWave backhaul B bk is 1 GHz. The The
+ log2 (1 + 2 i − q user ||2
), (53)
H + ||qm n maximum speed of the UAV is set to 40 m/s. The flying height
200 V. C ONCLUSIONS
Optimized UAV trajectory
180 UAV users
Ad hoc users
In this paper, we propose a UAV-Aided low latency mobile
160 edge computing network with mmWave backhaul. The users
140 have two communication links to compute the tasks, which are
120 the UAV link and the ad hoc link. In order to minimize the
total latency, we jointly consider the resource allocation and
y(m)

100

80
trajectory design in our problem. Then, we propose a novel
60
algorithm framework to solve the proposed MINLP problem.
40
The outer loop algorithm is the generalized Benders decom-
position, and the inner loop algorithm is the joint ADMM,
20
Dinkelbach algorithm, and SCA algorithm. The simulation
0
0 20 40 60 80 100
x(m)
120 140 160 180 200 results show the good performances of our proposed network
and algorithm.
Fig. 2: The optimized UAV trajectory and the locations of
ACKNOWLEDGEMENT
users
This work is partially supported by National Natu-
ral Science Foundation of China under Grant 61771054,
20 61801025, National Key R&D Program of China (Grant No.
Total delay for ad hoc links
2018YFC0603204), and US MURI AFOSR MURI 18RT0073,
Total delay for UAV links NSF CNS-1717454, CNS-1731424, CNS-1702850, CNS-
15
1646607.
R EFERENCES
Delay(s)

10
[1] L. Gupta, R. Jain, and G. Vaszkun, “Survey of important issues in
uav communication networks,” IEEE Communications Surveys Tutorials,
vol. 18, no. 2, pp. 1123–1152, Nov. 2016.
5 [2] V. Roberge, M. Tarbouchi, and G. Labonte, “Comparison of parallel
genetic algorithm and particle swarm optimization for real-time uav path
planning,” IEEE Transactions on Industrial Informatics, vol. 9, no. 1,
0 pp. 132–141, Feb. 2013.
10 20 30 40 50 [3] Y. Zeng, X. Xu, and R. Zhang, “Trajectory design for completion
Task data length(Mbits) time minimization in uav-enabled multicasting,” IEEE Transactions on
Fig. 3: The total delay of UAV links and ad hoc links Wireless Communications, vol. 17, no. 4, pp. 2233–2246, Apr. 2018.
[4] D. Yang, Q. Wu, Y. Zeng, and R. Zhang, “Energy tradeoff in ground-
to-uav communication via trajectory design,” IEEE Transactions on
Vehicular Technology, vol. 67, no. 7, pp. 6721–6726, Jul. 2018.
[5] X. Yuan, Z. Feng, W. Xu, W. Ni, J. A. Zhang, Z. Wei, and R. P.
of the UAV is set to 100 m. The total flying time for UAV Liu, “Capacity analysis of uav communications: Cases of random
is 20 seconds, and the frame number M is assumed as 30. trajectories,” IEEE Transactions on Vehicular Technology, vol. 67, no. 8,
pp. 7564–7576, Aug. 2018.
The power density of the noise is set as N0 = −174 dbm/Hz. [6] N. Abbas, Y. Zhang, A. Taherkordi, and T. Skeie, “Mobile edge
Other parameters are randomly generalized based on the users computing: A survey,” IEEE Internet of Things Journal, vol. 5, no. 1,
location. All the convex problems are solved by MATLAB pp. 450–465, Feb. 2018.
[7] S. Wang, X. Zhang, Y. Zhang, L. Wang, J. Yang, and W. Wang, “A
CVX with MOSEK solver. survey on mobile edge networks: Convergence of computing, caching
In Fig. 2, the optimized UAV trajectory and the locations and communications,” IEEE Access, vol. 5, pp. 6757–6779, Mar. 2017.
of users are shown. We set the edge node location as (200, 0). [8] C. Wang, C. Liang, F. R. Yu, Q. Chen, and L. Tang, “Computation
offloading and resource allocation in wireless cellular networks with
The initial point and destination for UAV is the same at (0,0). mobile edge computing,” IEEE Transactions on Wireless Communica-
First, the UAV fly with high speed, which is due to there is tions, vol. 16, no. 8, pp. 4924–4938, Aug. 2017.
no users around the initial point. Then, the UAV tends to stay [9] J. Xu, L. Chen, and S. Ren, “Online learning for offloading and autoscal-
ing in energy harvesting mobile edge computing,” IEEE Transactions on
longer with the UAV users in the center of the area. The users Cognitive Communications and Networking, vol. 3, no. 3, pp. 361–373,
closed to the edge node are optimized as the ad hoc users, Sep. 2017.
which are more efficient to transmit through the ad hoc links. [10] S. Jeong, O. Simeone, and J. Kang, “Mobile edge computing via a uav-
mounted cloudlet: Optimization of bit allocation and path planning,”
When fly back to the destination, the UAV increases its speed IEEE Transactions on Vehicular Technology, vol. 67, no. 3, pp. 2049–
again. In general, the UAV avoids the unnecessary area where 2063, Mar. 2018.
close to the ad hoc users. [11] M. Chen and Y. Hao, “Task offloading for mobile edge computing in
software defined ultra-dense network,” IEEE Journal on Selected Areas
In Fig. 3, the total delay of the network is shown, which in Communications, vol. 36, no. 3, pp. 587–597, Mar. 2018.
includes both the delay of UAV links and ad hoc links. When [12] D. W. K. Ng, Y. Wu, and R. Schober, “Power efficient resource allocation
for full-duplex radio distributed antenna networks,” IEEE Transactions
the task data length increases, all kinds of delays increase. on Wireless Communications, vol. 15, no. 4, pp. 2896–2911, Apr. 2016.
However, the UAV links delay increases much more slower, [13] K. Shen and W. Yu, “Fractional programming for communication
which gains the benefits from the trajectory optimization. For systemspart i: Power control and beamforming,” IEEE Transactions on
Signal Processing, vol. 66, no. 10, pp. 2616–2630, May 2018.
the heavy task, the delay for ad hoc will dominate.

You might also like