IEEE INTERNET OF THINGS JOURNAL, VOL. 7, NO.

4, APRIL 2020 3147

Joint Task Offloading and Resource Allocation in

UAV-Enabled Mobile Edge Computing
Zhe Yu, Student Member, IEEE, Yanmin Gong , Member, IEEE, Shimin Gong , Member, IEEE,
and Yuanxiong Guo , Senior Member, IEEE

Abstract—Mobile edge computing (MEC) is an emerging of data are collected from these devices at an exponential
technology to support resource-intensive yet delay-sensitive appli- rate [1] and need to be transported from place to place for
cations using small cloud-computing platforms deployed at the intelligent decision making, which has generated tremendous
mobile network edges. However, the existing MEC techniques
are not applicable to the situation where the number of mobile burden on our wireless communication infrastructure with the
users increases explosively or the network facilities are spar- limited radio spectrum. It is estimated that 25 billion Internet
ely distributed. In view of this insufficiency, unmanned aerial of Things (IoT) devices will be in use by 2025 [2], and such
vehicles (UAVs) have been employed to improve the connectiv- multitudes of wireless connected devices are enabling many
ity of ground Internet of Things (IoT) devices due to their high compelling new applications, such as real-time video analyt-
altitude. This article proposes an innovative UAV-enabled MEC
system involving the interactions among IoT devices, UAV, and ics [3], [4], augmented/virtual reality [5], and smart cities [6],
edge clouds (ECs). The system deploys and operates a UAV prop- which are computation intensive and delay sensitive and rely
erly to facilitate the MEC service provisioning to a set of IoT on our ability to quickly process the data and extract use-
devices in regions where the existing ECs cannot be accessible ful information, precluding the traditional cloud-based data
to IoT devices due to terrestrial signal blockage or shadowing. processing paradigm [7].
The UAV and ECs in the system collaboratively provide MEC
services to the IoT devices. For optimal service provisioning in Mobile edge computing (MEC), in which computing and
this system, we formulate an optimization problem aiming at min- storage resources are placed at the mobile network edges
imizing the weighted sum of the service delay of all IoT devices (e.g., cellular base stations or WiFi access points) [8]–[10],
and UAV energy consumption by jointly optimizing UAV posi- has emerged as a prospective solution to resolve the network
tion, communication and computing resource allocation, and task latency issue by pushing the frontier of data and services away
splitting decisions. However, the resulting optimization problem is
highly nonconvex and thus, difficult to solve optimally. To tackle from centralized cloud to the edge of the network, thereby
this problem, we develop an efficient algorithm based on the enabling data analytics and functional operation in the prox-
successive convex approximation to obtain suboptimal solutions. imity to the data sources. By moving resources to the network
Numerical experiments demonstrate that our proposed collabo- edge, close to where the data are being generated and acted
rative UAV-EC offloading scheme largely outperforms baseline upon, MEC can bring many benefits to users, such as lower
schemes that solely rely on UAV or ECs for MEC in IoT.
service latency, reduced network congestion, and better ser-
Index Terms—Mobile edge computing (MEC), resource man- vice quality. Meanwhile, resource management becomes a key
agement, successive convex approximation, unmanned aerial problem in MEC due to the much limited resources compared
vehicles (UAVs).
to remote clouds and the tight coupling of communication
and computing. There has been substantial research on MEC
I. I NTRODUCTION resource management with the goal of optimizing system
latency [11]–[14], energy consumption [15]–[17], and overall
S THE number of wireless connected devices contin-
A ues to grow vastly and rapidly, an enormous amounts
cost of system latency and/or energy consumption [18]–[21].
However, all of these studies assume wired or dedicated wire-
Manuscript received September 11, 2019; revised December 18, 2019; less connections with sufficient bandwidth among distributed
accepted December 28, 2019. Date of publication January 10, 2020; date edge resources deployed in a fixed fashion. Particularly, the
of current version April 14, 2020. The work of Yanmin Gong was supported existing MEC techniques are not applicable to the situation
in part by the U.S. National Science Foundation under Grant CNS-1850523.
(Corresponding author: Yuanxiong Guo.) where the number of mobile users (MUs) increases explo-
Zhe Yu is with the School of Electrical and Computer Engineering, sively or the network facilities are sparely distributed [22].
Oklahoma State University, Stillwater, OK 74078 USA (e-mail: In view of this insufficiency, wireless networks enabled by
[email protected]).
Yanmin Gong is with the Department of Electrical and Computer unmanned aerial vehicles (UAVs) have recently been proposed
Engineering, University of Texas at San Antonio, San Antonio, TX 78249 as a promising solution to improve the connectivity of ground
USA (e-mail: [email protected]). IoT devices.
Shimin Gong is with the School of Intelligent Systems Engineering,
Sun Yat-sen University, Guangzhou 510275, China (e-mail: UAVs, especially, low-cost quadcopters, are undergoing an
[email protected]). explosive growth and a major regulation relaxation nowadays
Yuanxiong Guo is with the Department of Information Systems and Cyber and have been widely used in civilian domains, such as traffic
Security, University of Texas at San Antonio, San Antonio, TX 78249 USA
(e-mail: [email protected]). monitoring [23], public safety [24], search and rescue [25], and
Digital Object Identifier 10.1109/JIOT.2020.2965898 reconnaissance over disaster rescue and recovery [26]. UAVs
2327-4662 
c 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: National Chung Cheng University. Downloaded on September 11,2024 at 06:22:36 UTC from IEEE Xplore. Restrictions apply.
3148 IEEE INTERNET OF THINGS JOURNAL, VOL. 7, NO. 4, APRIL 2020

not only provide extended coverage over wide geographical and computation energy consumption, by jointly optimizing
areas but also possess unique characteristics like fast deploy- the task splitting decisions, UAV placement, communication
ment, easy programmability, and high scalability. Various bandwidth allocation, and computation resource allocation at
payloads, such as IoT sensors (including cameras), miniatur- the UAV and ECs.
ized base stations, and embedded computing modules can be The above-formulated optimization problem is challeng-
mounted on UAVs to enable different sensing, communica- ing to solve due to the nonconvex objective function and
tion, and computing tasks [27], [28]. In particular, reliable and constraints. To tackle that challenge, we implement an effi-
cost-effective wireless communication solutions for multitudes cient algorithm by means of successive convex approximation
of real-world scenarios can be offered by UAVs if properly [37], [38]. The basic idea of the proposed algorithm is to com-
deployed and operated [28]. UAVs can act as wireless relays or pute a suboptimal solution of the original nonconvex problem
aerial base stations for improving connectivity and extending by solving a sequence of convex subproblems where the non-
coverage of ground wireless devices since the high altitude of convex objective function and constraints are replaced by
UAV enables wireless devices to effectively establish line-of- suitable convex approximants. We first convert the nonconvex
sight (LoS) communication links thus mitigating the potential objective function and constraints into suitable convex approx-
signal blockage and shadowing. imants by introducing the initial feasible solutions, while the
However, most prior works in the area of the UAV-enabled local first-order behavior of the original nonconvex problem is
wireless networks (e.g., [29]–[32]) ignore the computing preserved. Then, we iteratively compute the local optimum of
capability provided by UAVs and mainly focus on their the resulting convex problem by updating the initial feasible
communication aspect, and only a very few recent stud- solutions until a stationary solution of the original nonconvex
ies [22], [33]–[36] start to consider computing with UAVs’ problem is found. The convergence of the proposed algorithm
onboard resources. Hu et al. [22], Jeong et al. [33], and is guaranteed if the step-size rule and termination criterion are
Zhou et al. [34] only considered communication and compu- properly chosen.
tation interactions between two types of entities where ground The main contributions of this article are summarized as
MUs offload the tasks to UAV for computation. Asheralieva follows.
and Niyato [35] proposed a game-theoretic and reinforce- 1) We propose a novel UAV-enabled MEC system where a
ment learning approach in investigating the cooperation among UAV is deployed to facilitate the provisioning of MEC
UAVs and ground base stations. Hu et al. [36] studied a new services to IoT devices that cannot directly access ECs
UAV-enabled MEC system with interactions among a UAV, a on the ground due to terrestrial signal blockage and
set of ground user equipments, and an access point. To the best shadowing.
of our knowledge, UAV-enabled MEC systems involving MUs, 2) Considering the stringent quality-of-service requirement
UAVs, and edge clouds (ECs) with UAV-EC collaboration have of MEC services and the limited battery size of UAV,
not been studied. we formulate the joint IoT task offloading and UAV
In this article, we envision an innovative UAV-enabled MEC placement under the proposed system as an optimization
system where IoT devices offload computing tasks to ECs out- problem with the goal of minimizing the service delay
side their communication range with the assistance of UAV, of IoT devices and maximizing the energy efficiency of
which are endowed with computing capability, to take the ben- UAV.
efits of collaboration among UAV and ECs. Specifically, we 3) Given the nonconvexity of the formulated optimization
consider the regions where the terrestrial wireless communi- problem, we reformulate it into tractable one using
cation between IoT devices and ground cellular base stations successive convex approximation, and then develop an
or WiFi access points cannot be effectively established due to efficient algorithm to find the suboptimal approximate
signal blockage and shadowing. Therefore, a UAV is deployed solutions to the problem.
and operated to facilitate MEC service provisioning to a set 4) We conduct extensive simulations to evaluate the
of stationary IoT devices in such regions. The IoT devices performance of our proposed collaborative UAV-EC
perform some sensing tasks and need to process the gener- scheme. Numerical experiments demonstrate that our
ated data quickly. We assume that the sensing data analysis proposed collaborative UAV-EC offloading scheme
is not performed locally due to limited onboard communi- largely outperforms baseline schemes that solely rely on
cation, computing, and storage (CCS) resources but we seek UAV or ECs for MEC in IoT.
to utilize those from the UAV and existing ground ECs. The The remainder of this article is organized as follows.
UAV, equipped with miniaturized base stations and embedded Related work is reviewed in Section II. In Section III,
computing modules, is placed properly to collect the gen- we describe the system model and then formulate the
erated sensing data from IoT devices and then, can further optimal IoT task offloading processes as a nonconvex
forward the computation tasks to more resourceful ground optimization problem. In Section IV, we reformulate the origi-
ECs nearby. We formulate the IoT task offloading process as nal problem as an approximated convex optimization problem
a nonconvex optimization problem aiming at minimizing the and then solve it by means of successive convex approxima-
weighted sum of the service delay of all IoT devices con- tion. The simulation results based on real-world traces are
sisting of task offloading delay and computation delay, and presented in Section V. Finally, the conclusion is given in
UAV energy consumption consisting of transmission energy Section VI.

Authorized licensed use limited to: National Chung Cheng University. Downloaded on September 11,2024 at 06:22:36 UTC from IEEE Xplore. Restrictions apply.
YU et al.: JOINT TASK OFFLOADING AND RESOURCE ALLOCATION IN UAV-ENABLED MEC 3149

II. R ELATED W ORK B. UAV-Enabled MEC Networks

In this section, we review the prior works most rel- Extensive research efforts have been made from the
evant to our article from two aspects: 1) resource academia to employ UAVs as different kinds of wireless
management in MEC and 2) UAV-enabled MEC communication platforms [39]. For instance, UAVs equipped
networks. with base stations can be flexibly deployed at specific areas
to provide reliable uplink and downlink communication for
ground users. They can also serve as the mobile relaying nodes
to connect two or more distant users [29], [30]. Moreover,
A. Resource Management in MEC
UAVs can assist with information dissemination or data col-
There is a rich literature on resource management in MEC lection by flying over the specific areas [31], [32]. However,
that aims at optimizing system latency [11]–[14], energy con- prior works in the area of the UAV-enabled wireless networks
sumption [15]–[17], and overall cost of system latency and/or ignore the computing capability provided by UAVs and mainly
energy consumption [18]–[21]. The tradeoff problem is stud- focus on their communication aspect, and only a very few
ied in [11] for computing networks with fog node cooperation recent studies [22], [33]–[36] start to consider computing with
aiming at minimizing the response time of fog nodes under UAVs’ onboard resources. Hu et al. [22] investigated joint
a given power efficiency constraint. Xu et al. [12] stud- offloading and trajectory design for a MEC system where
ied the joint service caching and task offloading problem in a UAV endowed with computing capability is deployed to
the dense network aiming at minimizing computation latency serve the task offloading of MUs, aiming at minimizing the
while keeping the total computation energy consumption low. sum of the maximum delay among all the users in each
Chen and Hao [13] investigated the MEC task offloading time slot. Jeong et al. [33] studied the joint optimization of
problem in the software-defined ultradense network aiming path planning and bit allocation for an MEC system where
at minimizing the total task duration under energy budget a UAV-mounted cloudlet is deployed to provide offloading
constraints. Ren et al. [14] investigated a joint communica- opportunities to MUs, aiming at minimizing the mobile energy
tion and computation resource allocation problem under the consumption while satisfying the quality-of-service require-
collaboration of cloud and edge computing for minimizing ments of offloaded applications. Zhou et al. [34] formulated
the system delay of all mobile devices. Sardellitti et al. [15] the computation rate maximization problem under both par-
formulated the multicell MEC task offloading problem as a tial and binary task offloading schemes in a UAV-enabled
joint optimization of radio and computation resources aiming MEC wireless-powered system where the UAV can simulta-
at minimizing the overall users’ energy consumption, while neously transmit energy and perform computation. However,
meeting latency constraints. Zhang et al. [16] proposed an these works only consider communication and computation
energy-efficient offloading scheme for MEC in 5G heteroge- interactions between two types of entities where ground MUs
neous networks by formulating the optimization problem with offload the tasks to UAV for computation. Besides, Asheralieva
the objective of minimizing the total system energy consump- and Niyato [35] presented a game-theoretic and reinforce-
tion. You et al. [17] studied the resource allocation problem ment learning framework to study the computation offloading
for a multiuser MEC offloading system based on TDMA problem in UAV-enabled MEC networks with multiple service
and OFDMA with the objective to minimize the weighted providers where UAV-based privately owned base stations are
sum of mobile energy consumption. Chen et al. [18] for- interacting with terrestrial privately owned and operator con-
mulated a multiuser computation offloading game to study trolled base stations. Hu et al. [36] considered a UAV-aided
the energy-delay tradeoff problem in a mobile-edge cloud MEC system where the cellular-connected UAV is served as
computing architecture. Chen et al. [19], [20] jointly opti- a mobile computing server as well as a relay to help the user
mized the offloading decisions of all users and computing equipments complete their computing tasks or further offload
access point and resource allocation aiming at minimiz- their tasks to the AP for computing.
ing the overall energy cost and the maximum delay among To the best of our knowledge, UAV-enabled MEC systems
all users. Zhang et al. [21] proposed a distributed joint involving MUs, UAVs, and ECs have not been studied.
computation offloading and resource allocation optimization Different from [35] which focuses on the user’s perspective,
scheme in heterogeneous networks with MEC to minimize we optimize the UAV energy-efficiency and IoT task service
the overhead of local energy consumption and execution time latency from the system operator’s perspective with UAV-EC
cost. collaboration. Different from [36] which focuses on a single
However, all of these studies assume wired or dedi- EC, we consider the scenario where multiple ECs and the UAV
cated wireless connections with sufficient bandwidth among collaboratively provide MEC services to the IoT devices.
distributed edge resources deployed in a fixed fashion.
Particularly, the existing MEC techniques are not applicable to
the situation where the number of MUs increases explosively
or the network facilities are sparely distributed [22]. In view III. S YSTEM M ODEL AND P ROBLEM F ORMULATION
of the above limitations, we propose to deploy and operate In this section, we first introduce the system model for
a UAV to assist the IoT task offloading processes in a MEC the UAV-enabled MEC system. After that, we formulate an
system where ECs cannot be accessible to IoT devices due to optimization problem to model the optimal UAV-enabled IoT
terrestrial signal blockage and shadowing. task offloading process.

Authorized licensed use limited to: National Chung Cheng University. Downloaded on September 11,2024 at 06:22:36 UTC from IEEE Xplore. Restrictions apply.
3150 IEEE INTERNET OF THINGS JOURNAL, VOL. 7, NO. 4, APRIL 2020

TABLE I
L IST OF N OTATIONS

Fig. 1. Illustration of an exemplary UAV-enabled MEC system with N MUs,

J ECs, and a UAV.

A. System Model
In this article, we consider the UAV-enabled MEC system
as depicted in Fig. 1, which consists of a set of ground MUs1
i ∈ N = {1, 2, . . . , N}, a UAV, and a set of ground ECs j ∈
J = {1, 2, . . . , J}. The UAV is deployed to facilitate the MEC
service provisioning for ground MUs who cannot establish
wireless communication with nearby cellular base stations or
WiFi access points due to signal blockage and shadowing. In
this scenario, ground-to-air (G2A) uplink communication is
from MUs to the UAV while air-to-ground (A2G) downlink
communication is from the UAV to ECs, which form a 3-D
wireless communication network. For ease of reference, we
list important notations in Table I.
We assume that the UAV is equipped with certain CCS
resources but subject to the size, weight, and power (SWAP)
limitations. The ECs are composed of ground MEC servers co-
located with cellular base stations or WiFi access points that
have more CCS resources compared to the UAV and MUs.
Each MU i has periodical computation-intensive tasks to per-
form, which are modeled as a triplet Wi = Li , Ci , λi , where
Li (in bits) denotes the input data size for processing the task,
Ci (in CPU cycles/bit) denotes the number of CPU cycles
required to process 1-bit of task data, and λi (in unit of #task
per second) denotes the arrival rate of tasks.
In this article, we use the 3-D Cartesian coordinate system
to represent the locations of MUs, UAV, and ECs. The posi-
tion of the UAV is denoted by QUAV = (xUAV , yUAV , H),
where the height H is assumed to be fixed while the hor-
izontal coordinates xUAV and yUAV affect the channel gain
during data communication processes and need to be opti-
mized in our problem. Besides, we assume the positions of
MU i and EC j are fixed in our model, which are denoted as gain from MU i to the UAV can be described by the free-space
QMU
i = (xiMU , yMU
i , 0) and Qj = (xj , yj , 0), respectively.
EC EC EC path loss model
1) Communication Model: In the UAV-enabled network,  UL −2 α0
i  α0 di
hUL =  (1)
the LoS links are much more dominant than other channel QMU − QUAV 2
i
impairments, such as shadowing or small-scale fading due to
the high altitude of the UAV. Therefore, the uplink channel where α0 represents the received power at the reference dis-
tance of 1 m for a transmission power of 1 W, diUL denotes the
1 Note that we use mobile users and IoT devices interchangeably in this uplink distance from MU i to the UAV, and · denotes the
article. Euclidean norm of a vector. Similarly, the downlink channel

Authorized licensed use limited to: National Chung Cheng University. Downloaded on September 11,2024 at 06:22:36 UTC from IEEE Xplore. Restrictions apply.
YU et al.: JOINT TASK OFFLOADING AND RESOURCE ALLOCATION IN UAV-ENABLED MEC 3151

gain from the UAV to EC j can be described as the UAV or further offloaded to the ground ECs for processing.
 −2 α0 Denote {βij ∈ [0, 1], i ∈ N , j ∈ J } and {βi0 ∈ [0, 1], i ∈ N }
j  α0 dj
hDL =
DL
2 (2) as the portion of received tasks from MU i to be processed at
 UAV 
Q − QEC
j  EC j and the UAV, respectively. Then, the computation delay
at the UAV side to process the offloaded tasks from MU i can
where djDL denotes the downlink distance from the UAV to be calculated as
the EC j.
βi0 Li Ci
We assume the FDMA protocol for bandwidth sharing tiUAV = UAV (6)
among MUs during the task offloading process. According to fi
Shannon’s capacity, the achievable uplink transmission data
where fiUAV (in CPU cycles/s) is the computation resource that
rate (in bps) from MU i to the UAV can be expressed as
  the UAV allocates to MU i. Note that when βi0 equals 0, it
hUL
i Pi
MU means that no computation will be executed at the UAV side
Ri = Bi log2 1 +
UL UL
(3) while when βi0 equals 1, it indicates that no further offloading
σ2
will occur from the UAV to ECs.
where BULi , Pi , and σ represent the assigned bandwidth to
MU 2 A2G Downlink Transmission Delay From the UAV to ECs:
MU i, transmit power of MU i, and the noise power at the The UAV may further offload the tasks to more powerful ECs
UAV, respectively. For simplicity, we assume the noise power on the ground to reduce the computation latency. Then, the
is the same at UAV and ECs [29]. However, it can be easily A2G transmission delay from the MU i to EC j via UAV is
extended to the case when they are different. Similarly, the described as the ratio of offloaded task input data size and the
downlink transmission data rate (in bps) from the UAV to EC associated downlink transmission data rate
j can be computed as βij Li
 UAV
 tijA2G = DL . (7)
hDL
j PTX Rj
Rj = Bj log2 1 +
DL DL
(4)
σ2 Computation Delay at ECs: After receiving the offloaded
UAV task data from the UAV, ECs can start the computation process.
where BDL j and PTX represent the per-device bandwidth pre-
2
Therefore, the computation delay at the EC side to process the
assigned to EC j and transmit power of the UAV, respectively.
offloaded task from the MU i to EC j via UAV is
2) Delay Analysis: In our model, we assume that MUs do
not perform local computing due to their limited computation βij Li Ci
tijEC = (8)
capacities. In contrast, tasks will be first offloaded to the UAV, fijEC
and then, the UAV will determine the portion of tasks that are
processed locally or further offloaded to ECs on the ground. where fijEC (in CPU cycles/s) is the computation resource that
Note that the decision time to split a task is very short com- EC j allocates to MU i.
pared to the entire communication and computation latency, 3) UAV Energy Consumption Analysis: To ensure service
and therefore, can be neglected. Besides, the output data size availability, it is important to manage the energy consumption
of the computation results is often very small compared to the of the UAV due to its limited battery size. In this article, we
input data size in many computation-intensive applications, focus on computation and transmission energy consumption of
such as face recognition and video analysis. Thus, the time UAV, and ignore the hovering power since it is independent
needed to send the computation results back to MUs can be of our decisions.
ignored as well. Computation Energy Consumption: Similar to [40], we
In what follows, we will describe the four key components model the power consumption of the CPU in UAV as
of the total delay for the offloading process: 1) G2A uplink κ(fiUAV )3 , where κ denotes the effective switched capacitance
transmission delay from MUs to the UAV; 2) computation depending on the CPU architecture. It follows that the corre-
delay at the UAV; 3) A2G downlink transmission delay from sponding energy consumption of UAV when processing tasks
the UAV to the ECs; and 4) computation delay at the ECs. offloaded from MU i is given by the product of the power
G2A Uplink Transmission Delay From MUs to the UAV: As level and computation time
mentioned before, all the tasks will be offloaded to the UAV  3  2
first via G2A links without any local computation. Therefore, EiCP = κ fiUAV tiUAV = κβi0 Li Ci fiUAV . (9)
the G2A transmission delay from MU i to the UAV is com-
puted as the ratio of task input data size and the associated Transmission Energy Consumption: The transmission
uplink transmission data rate energy consumption of the UAV when receiving the task input
data via the G2A uplink transmission channels from MU i is
Li
tiG2A = . (5) given by
RUL
i
Li PUAV
Computation Delay at the UAV: The UAV will decide the EiRX = PUAV G2A
RX ti = RX
(10)
portion of the received tasks that will be processed locally at RUL
i

2 We assume that each MU is assigned a certain bandwidth beforehand where PUAV

RX is the receiving power of UAV. Besides, the trans-
when they communicate with ECs via the UAV. mission energy consumption of the UAV when offloading the

Authorized licensed use limited to: National Chung Cheng University. Downloaded on September 11,2024 at 06:22:36 UTC from IEEE Xplore. Restrictions apply.
3152 IEEE INTERNET OF THINGS JOURNAL, VOL. 7, NO. 4, APRIL 2020

task input data of MU i via the A2G downlink transmission that the allocated resources for uplink bandwidth, UAV and EC
channels to EC j is given by CPU frequencies are nonnegative and no more than their limits
while (14c), (14f), and (14g) constrain that the offloading tasks
βij Li PUAV
EijTX = PUAV A2G
TX tij = TX
. (11) of MUs are completely processed by UAV and ECs, and the
RDL
j values of partition variables are between 0 and 1.
Therefore, the total energy consumption of the UAV when
serving the task offloading and computation of MU i is IV. S OLUTION M ETHODOLOGY
given by Problem (14) is hard to solve due to the nonconvexity of
⎛ ⎞ the objective function and constraints. In what follows, we
EiUAV = λi ⎝EiCP + EiRX + EijTX ⎠. (12) will first linearize the maximum term in (13) by leveraging
auxiliary variables and reformulate the original optimization
j∈J
problem into a tractable one. Then, we develop an SCA-based
B. Problem Formulation algorithm to transform the nonconvex objective function and
constraints into suitable convex approximants to iteratively
In this article, we are interested in minimizing the total solve the resulting optimization problem.
energy consumption of the UAV when serving the computa-
tion and communication needs of the MUs and the total service
A. Problem Reformulation
delay of all MUs. To define the service delay of each MU, we
make the following assumptions: 1) the UAV cannot partition We first define an auxiliary variable for each MU i as zi 
a task until receiving its entire input data to ensure the accu- maxj∈J {tiUAV , tijA2G +tijEC }. Then, we linearize the service delay
racy of task splitting; 2) the UAV and ECs cannot start the term in (14a) using zi and reformulate the original optimization
processing of tasks until the end of the transmission between problem into the following:
MUs and the UAV or the UAV and ECs to ensure the reliabil-
min EiUAV + ρ (tiG2A + zi ) (15a)
ity of the computation results; and 3) the computation at the zi ,QUAV ,BUL i , i∈N i∈N
UAV can proceed simultaneously with the transmission of the βi0 ,βij ,fiUAV ,fijEC
tasks to each EC since the communication and computation
s.t. zi ≥ tiUAV ∀i (15b)
modules are often separated at the UAV. Based on the above
assumptions, the service delay of MU i can be represented as zi ≥ tijA2G + tijEC ∀i, j (15c)
  (14b)−(14i). (15d)
Ti = tiG2A + max tiUAV , tijA2G + tijEC . (13)
j∈J However, the reformulated optimization problem is still dif-
Our problem becomes jointly optimizing the UAV position ficult to solve due to the nonconvex objective function (15a)
QUAV , G2A uplink communication resource allocation BUL and nonconvex constraints (15b) and (15c). Note that both
i ,
task partition variables βi0 and βij , and computation resource the uplink and downlink transmission data rate functions (3)
allocation of the UAV fiUAV and ECs fijEC with the goal of and (4) are nonconvex with respect to the UAV position QUAV .
minimizing the weighted sum of total energy consumption of
UAV and total service delay of all MUs. It can be formulated B. Successive Convex Approximation
as the following optimization problem: In this section, we will show how to build the convex
approximation for the nonconvex objective function and non-
min EiUAV +ρ Ti (14a)
QUAV ,BUL i ,βi0 , i∈N
convex constraints in the reformulated problem (15) while
i∈N
βij ,fiUAV ,fijEC preserving the local first-order behavior of the original non-
convex problem and solve the resulting problem iteratively
i ≤B
BUL UL
s.t. (14b)
to obtain suboptimal solutions by means of SCA. Before
i∈N
we develop the SCA-based algorithm, we first present the
βi0 + βij = 1 ∀i (14c) background of SCA.
j∈J 1) Background of SCA: Consider the following
fiUAV ≤ F UAV (14d) optimization problem:
i∈N P: min U(x) (16a)
x
fijEC ≤ FjEC ∀j (14e)
s.t. gl (x) ≤ 0 ∀l = 1, . . . , m (16b)
i∈N
0 ≤ βij ≤ 1 ∀i, j (14f) x∈K (16c)
0 ≤ βi0 ≤ 1 ∀i (14g) where the objective function U : K → R is smooth (possibly
BULi , fi
UAV
≥0 ∀i (14h) nonconvex) and gl : K → R is smooth (possibly nonconvex),
fij ≥ 0 ∀i, j
EC
(14i) for all l = 1, . . . , m; the feasible set is denoted as X . A widely
used method for solving this specific problem is SCA (also
where ρ > 0 is a parameter defining the relative weight of known as majorization minimization) where at each iteration,
energy and delay, (14b), (14d), (14e), (14h), and (14i) ensure a convex approximation of the original problem is solved via

Authorized licensed use limited to: National Chung Cheng University. Downloaded on September 11,2024 at 06:22:36 UTC from IEEE Xplore. Restrictions apply.
YU et al.: JOINT TASK OFFLOADING AND RESOURCE ALLOCATION IN UAV-ENABLED MEC 3153

Algorithm 1 SCA Algorithm for Problem P respect to the argument x evaluated at (u; v). Note that
Find a feasible solution x ∈ X in P, choose a step size θ ∈ A1) and B1) make the problem Pxk strongly convex
(0, 1] and set k = 0. while A2) and A3) guarantee the iterate feasibility that
Repeat xk ∈ X (xk ) ⊆ X .
1) Compute x̂(xk ), the solution of Pxk ; Example 1 (Approximation of gl (x) [38, Example 3]):
2) Set xk+1 = xk + θ (x̂(xk ) − xk ); Suppose that gl has a difference of convex (DC) structure,
3) Set k ← k + 1. i.e., gl (x) = g+ − + −
j (x) − gj (x) with both gl and gl being con-
Until some convergence criterion is met. vex and continuously differentiable. By linearizing the concave
part g− l , we obtain the convex upper approximation of gl as
follows: for all x ∈ K and y ∈ X ,
replacing the nonconvex objective function and constraints by
suitable convex approximants. The convex approximation of g̃l (x; y)  g+ − −
l (x) − gl (y) − ∇x gl (y) (x − y) ≥ gj (x). (18)
the original problem can be stated as follows: given xk ∈ X Example 2 (Approximation of gl (x) [38, Example 4]):
 
Pxk : min Ũ x; xk (17a) Suppose that gl (x) has a product of functions (PF) structure,
x i.e., gl (x) = f1 (x)f2 (x) with both f1 and f2 being convex and
s.t. g̃l (x; xk ) ≤ 0 ∀l = 1, . . . , m (17b) nonnegative. Observe that gl (x) can be rewritten as a function
x∈K (17c) with the DC structure
1 1 2 
where Ũ(x; xk ) and g̃l (x; xk ) represent the approximants of gl (x) = (f1 (x) + f2 (x))2 − f1 (x) + f22 (x) . (19)
2 2
U(x) and gl (x) at current iterate xk , respectively; the feasi-
ble set is denoted as X (xk ). More specifically, we consider Then, the convex upper approximation of gl can be obtained
the SCA method presented in Algorithm 1. It is assumed that by linearizing the concave part in (19): for any y ∈ X
at each iteration, some original functions U(x) and gl (x) are 1 1 2 
approximated by their upper bounds where the same first-order g̃l (x; y)  (f1 (x) + f2 (x))2 − f1 (y) + f22 (y)
2 2
behavior is preserved [41]. − f1 (y)f1 (y)(x − y) − f2 (y)f2 (y)(x − y) ≥ gl (x).
2) SCA-Based Algorithm: Scutari et al. [38] proposed a
(20)
framework that unifies several existing SCA-based algorithms
to solve the problem P in a parallel and distributed fashion. It Example 3 (Approximation of U(x) [38, Example 8]):
also offers much flexibility in the choice of the convex approx- Suppose that U(x) has a PF structure, i.e., U(x) =
imation functions, and the objective function U need not be an h1 (x)h2 (x) with both h1 and h2 being convex and nonneg-
upper bound of itself at any feasible point. Multiple examples ative. For any y ∈ X , a convex approximation of U(x) is
are summarized to find the candidate approximants g̃l (x) and given by
Ũ(x) while necessary assumptions are satisfied to develop the
SCA-based algorithm. We first present the assumptions and Ũ(x; y) = h1 (x)h2 (y) + h1 (y)h2 (x)
τ
examples that we will utilize to approximate the nonconvex + (x − y) H(y)(x − y) (21)
2
terms in our problem as follows.
Assumption 1: The key assumptions on the choice of the where τ > 0 is a positive constant, and H(y) is a uniformly
approximated function g̃l : K × X → R are given as follows. positive-definite matrix.
A1) g̃l (•; y) is convex on K for all y ∈ X . Then, we transform the nonconvex constraints and noncon-
A2) Upper Bound: gl (x) ≤ g̃l (x; y) ∀x ∈ K, y ∈ X . vex objective function in the reformulated problem (15) into
A3) Function Value Consistency: g̃l (y; y) = gl (y), for all suitable approximants by following the above examples. For
y ∈ X. constraint (15b), we observe that the nonconvex term tiUAV
A4) g̃l (•; •) is continuous on K × X . can be written as the product of convex and nonnegative
A5) ∇x g̃l (•; •) is continuous on K × X . functions3
   
A6) Gradient Consistency: ∇x g̃l (y; y) = ∇x gl (y), for all tiUAV = Li Ci gl βi0 , fiUAV = Li Ci f1 (βi0 )f2 fiUAV (22)
y ∈ X , where ∇x g̃l (y; y) denotes the partial gradient of
the function g̃l with respect to the argument x evaluated where f1 (βi0 ) = βi0 and f2 (fiUAV ) = 1/fiUAV . Then, given
at (y; y). a feasible solution βi0 (k) and fiUAV (k) for the kth iteration
Assumption 2: The key assumptions on the choice of the of the SCA-based algorithm, we derive a convex upper
approximated function Ũ : K × X → R are given as follows. approximation of tiUAV by using Example 2 as
B1) Ũ(•; y) is uniformly strongly convex on K with con-  
stant μ > 0, i.e., for all x, z ∈ K and y ∈ X : tiUAV ≤ t̃iUAV βi0 , fiUAV ; βi0 (k), fiUAV (k)
  ⎡ ⎛ 2  2 ⎞
(x − z) ∇x Ũ(x; y) − ∇x Ũ(z; y) ≥ μx − z2 . 1 1 1
 Li Ci ⎣ ⎝ βi0 + UAV −(βi0 (k))2 − UAV ⎠
B2) Gradient Consistency: ∇x Ũ(y; y) = ∇x U(y), for all 2 fi fi (k)
y ∈ X.
B3) ∇x Ũ(•; •) is continuous on K × X , where ∇x Ũ(u; v) 3 Without loss of generality, we factorize the constants (L , C , etc.) out of
i i
denotes the partial gradient of the function Ũ with the term since they will not affect the convexity.

Authorized licensed use limited to: National Chung Cheng University. Downloaded on September 11,2024 at 06:22:36 UTC from IEEE Xplore. Restrictions apply.
3154 IEEE INTERNET OF THINGS JOURNAL, VOL. 7, NO. 4, APRIL 2020

− (βi0 (k)(βi0 − βi0 (k))) where h1 (βi0 ) = βi0 and h2 (fiUAV ) = (fiUAV )2 in (28), and
 3  ⎤ h1 (βij ) = βij and h3 (φj ) = 1/φj in (29), while h3 (BUL
1 1 1 i ) =
+ − UAV ⎦. (23) 1/BUL i and h 3 (γ i ) = 1/γ i in (30). Then, given a feasible solu-
fiUAV (k) fiUAV fi (k) tion βi0 (k), βij (k), φj (k), γi (k), BUL (k), and f UAV (k) for the kth
i i
iteration of SCA-based algorithm, we derive convex approx-
For constraint (15c), tijA2G can be written as the product imation of EiCP , EijTX , tiG2A , and EiRX by using Example 3
of Li , βij , and 1/RDLj . However, 1/Rj
DL is a nonconvex func-
as
tion with respect to the UAV location QUAV , and therefore,  
Example 2 cannot be directly applied to derive a convex upper ẼiCP βi0 , fiUAV ; βi0 (k), fiUAV (k)
  2  2 
approximation. To tackle the nonconvexity, we replace it by
nonnegative auxiliary variables {φj }j∈J . Then, the noncon-  κLi Ci βi0 fiUAV (k) + βi0 (k) fiUAV
vex term tijA2G can be written as the product of convex and τβi0 τf UAV  2
nonnegative functions + (βi0 − βi0 (k))2 + i fiUAV − fiUAV (k) (32)
2 2  
        βij βij (k)
tijA2G = Li gl βij , φj = Li f1 βij f2 φj (24) Ẽij βij , φj ; βij (k), φj (k)  Li PTX
TX UAV
+
φj (k) φj
where f1 (βij ) = βij and f2 (φj ) = 1/φj in (24). Similarly, the τβij  2 τφj  2
nonconvex term tijEC in (15c) can be written as the product of + βij − βij (k) + φj − φj (k) (33)
2 2  
convex and nonnegative functions
  1 1
      G2A UL
t̃i Bi , γi ; Bi (k), γi (k)  Li UL
UL
+
tijEC = Li Ci gl βij , fijEC = Li Ci f1 βij f2 fijEC (25) Bi γi (k) BUL i (k)γi
τBUL  2 τγi
where f1 (βij ) = βij and f2 (fijEC ) = 1/fijEC in (25). Then, + i BUL i − Bi (k) +
UL
(γi − γi (k))2 (34)
2 2
given a feasible solution βij (k), φj (k), and fijEC (k) for the kth and
iteration of the SCA-based algorithm, we derive convex upper  
ẼiRX BUL i , γi ; Bi (k), γi (k)
UL
approximation of tijA2G and tijEC by using Example 2 as  
 PUAV G2A UL
RX t̃i Bi , γi ; BUL
i (k), γi (k) (35)
 
tijA2G ≤ t̃ijA2G βij , φj ; βij (k), φj (k)
    2  where τβi0 , τβij , τφj , τγi , τBUL , τf UAV > 0. Therefore, the convex
1 1 2 1 i i
surrogate objective function of (15a) can be denoted as the
 Li βij + − (βij (k)) −
2
2 φj φj (k) nonnegative weighted sum of convex functions
    ⎛ ⎞
1 3 1 1  
− (βij (k)(βij − βij (k)))+ −
φj (k) φj φj (k) λi ⎝ẼiCP + ẼiRX + ẼijTX ⎠ + ρ t̃iG2A + zi (36)
(26) i∈N j∈J i∈N

and where the convexity is preserved.

  Moreover, as we replace the nonconvex data rate functions
tijEC ≤ t̃ijEC βij , fijEC ; βij (k), fijEC (k) in both objective function and constraints by the auxil-
⎡ ⎛ 2  2 ⎞ iary variables {φj }j∈J and {γi }i∈N , we obtain equality con-
1 1 1 UL
 Li Ci ⎣ ⎝ βij + EC − (βij (k))2 − EC ⎠ straints {φj }j∈J = 1/RDL
j and {γi }i∈N = 1/Ri . To further
2 fij fij (k) address the nonconvexity, we first relax them as the following
   inequalities:
− βij (k) βij − βij (k)
 3  ⎤ 0 ≤ φj ≤ RDL ∀j (37)
1 1 1 j
+ − EC ⎦. (27) UL
fijEC (k) fijEC fij (k) 0 ≤ γi ≤ Ri ∀i (38)

UL where the optimality is preserved since at optimal solutions the

By defining Ri  log2 (1 + [(hUL i Pi )/(σ )]), we replace
MU 2
auxiliary variables will equate their upper bounds. The key
UL
1/Ri in tiG2A by nonnegative auxiliary variables {γi }i∈N since observation is that in (37) and (38), although RDL and Ri
UL
j
it is a nonconvex function with respect to the UAV location are not concave with respect to QUAV
QUAV . Then, the nonconvex terms in objective function (15a) 2, theyare convex func-
2
 UAV 
can be written as the product of convex and nonnegative tions with respect to Q − Q  and QMU − QUAV  ,
EC
j i
functions respectively. Recall that any convex function is globally
  lower bounded by its first-order Taylor expansion at any
EiCP = κLi Ci h1 (βi0 )h2 fiUAV (28) point [42]. Therefore, by taking the first-order Taylor expan-
  
UL  UAV 2
Eij = Li PTX h1 βij h3 (φj )
TX UAV
(29) sion of Rj and Ri with respect to Q
DL − QEC
j  and
 MU 
tiG2A = Li h3 (BUL Q − QUAV 2 , respectively, we obtain lower bounds of RDL
i )h3 (γi ) (30) i j
UL
EiRX = PUAV G2A
RX ti = RX Li h3 (Bi )h3 (γi )
PUAV UL
(31) and Ri at local point QUAV (k) for the kth iteration of the

Authorized licensed use limited to: National Chung Cheng University. Downloaded on September 11,2024 at 06:22:36 UTC from IEEE Xplore. Restrictions apply.
YU et al.: JOINT TASK OFFLOADING AND RESOURCE ALLOCATION IN UAV-ENABLED MEC 3155

SCA-based algorithm as follows: Algorithm 2 SCA-Based Algorithm for Problem (43)

    Input: ψ(0) = (zi , QUAV (0), BUL i (0), βi0 (0), βij (0),
RDL
j ≥ RDL
j,LB Q
UAV
; QUAV (k)  RDL j QUAV (k) fiUAV (0), fijEC (0), φj (0), γi (0)), and τβi0 , τβij , τφj , τγi , τBUL ,
 2  2  i
BDL η
 UAV
−Q EC  −  UAV (k)−QEC  τf UAV > 0 for i ∈ N and j ∈ J , θ (k) ∈ (0, 1]. Set k = 0,
j Q j  Q j  i
  α = 0.5.
− 2  2  ∀j
   UAV EC  Repeat
ln 2 QUAV (k) − QEC j  η + Q (k)−Q j 
1) Compute ψ̂(ψ(k)), the solution of (43);
and
   
(39) 2) Set ψ(k + 1) = ψ(k) + θ (k)(ψ̂(ψ(k)) − ψ(k)), with
UL
Ri
UL
≥ Ri,LB Q UAV
;Q UAV
(k)  Ri Q
UL UAV
(k) θ (k) = θ (k − 1)(1 − αθ (k));
     3) Set k ← k + 1.
εi QMU − QUAV  − QMU − QUAV (k)
2 2


i i Until ψ(k) is a stationary solution of (14).
− 2   MU 2  ∀i
QUAV , BUL i , βi0 , βij , fi
UAV and f EC .

ln 2 Qi − Q
MU UAV (k) 
εi + Qi − Q UAV (k) Output: ij
(40)

TX /σ and εi  α0 Pi /σ . Note that both

where η  α0 PUAV 2 MU 2
DL UL
Rj,LB and Ri,LB are concave functions with respect to QUAV .
UL
Then, by replacing RDL
j and Ri with their lower bounds, we
obtain the approximated convex constraints as
 
0 ≤ φj ≤ RDL
j,LB Q
UAV
; QUAV (k) ∀j (41)
 
UL
0 ≤ γi ≤ Ri,LB QUAV ; QUAV (k) ∀i. (42)

Finally, we denote the set of decision vari-

ables for our optimization problem as ψ = Fig. 2. Locations of 10 MUs and 4 ECs in the MEC system.
(zi , QUAV , BUL
i , βi0 , βij , fi
UAV , f EC , φ , γ ).
ij j i The convex
approximation of the reformulated problem (15) with a TABLE II
feasible solution ψ(k) for the kth iteration of the SCA-based S IMULATION PARAMETERS
algorithm is given by
⎛

min λi ⎝ẼiCP (ψ; ψ(k)) + ẼiRX (ψ; ψ(k))

ψ
i∈N
⎞

+ ẼijTX (ψ; ψ(k))⎠

j∈J

+ρ (t̃iG2A (ψ; ψ(k)) + zi ) (43a)

i∈N
s.t. zi ≥ t̃iUAV (ψ; ψ(k)) ∀i (43b)
zi ≥ t̃ijA2G (ψ; ψ(k)) + t̃ijEC (ψ; ψ(k)) ∀i, j (43c) V. N UMERICAL E XPERIMENTS
0 ≤ φj ≤ RDL j,LB (ψ; ψ(k)) ∀j (43d) In this section, we validate the effectiveness of our proposed
UL SCA-based algorithm via extensive numerical experiments.
0 ≤ γi ≤ Ri,LB (ψ; ψ(k)) ∀i (43e) All the experiments are implemented in MATLAB R2018a
(14b)−(14i) (43f) using CVX on a desktop computer with an Intel Core i7-4790
3.60-GHz CPU and 16-GB RAM. The convergence tolerance
which has a unique solution denoted by ψ̂(ψ(k)). The above threshold ζ for the proposed algorithm is set to be 10−2 .
optimization problem (43) is convex, and the SCA-based
algorithm is summarized in Algorithm 2.
Note that a diminishing step-size rule is applied in step 2), A. Simulation Setup
which is numerically more efficient than a constant one. The We consider a UAV-enabled MEC system with 4 ground
convergence of Algorithm 1 is guaranteed if the step size θ (k) ECs placed at each vertex and 10 ground MUs that are ran-
is chosen so that θ (k) ∈ (0, 1], θ (k) → 0, and v θ (k) = ∞, domly distributed within a 2-D area of 1000 × 1000 m2 , as
then ψ(k) is bounded and at least one of its limit points illustrated in Fig. 2. The UAV is deployed and operated to
is stationary [38]. For the termination criterion, it is very facilitate the MEC service provisioning, and the optimal 3-D
convenient to use ψ̂(ψ(k)) − ψ(k), which is a measure location of UAV can be found using our proposed SCA-based
of stationarity. Thus, a reliable termination rule is to check algorithm. The simulation parameter settings are summarized
ψ̂(ψ(k)) − ψ(k) ≤ ζ , where ζ is the desired accuracy. in Table II unless otherwise stated.

Authorized licensed use limited to: National Chung Cheng University. Downloaded on September 11,2024 at 06:22:36 UTC from IEEE Xplore. Restrictions apply.
3156 IEEE INTERNET OF THINGS JOURNAL, VOL. 7, NO. 4, APRIL 2020

TABLE III
As mentioned in Section I, our system settings involving the S YSTEM C OST C OMPARISON FOR O PTIMIZED UAV LOCATION AND
interactions among IoT devices, UAV, and ECs are different R ANDOM UAV L OCATION S CHEMES
from prior works. The approaches proposed in their studies are
not directly applicable to our settings. Therefore, we consider
the following intuitive methods as baselines.
1) Random UAV Location Scheme: The task splitting and
resource allocation decisions are optimized while the
UAV location is randomly selected without optimization.
2) UAV-Only Scheme: All tasks are offloaded and processed
at the UAV without further offloading to any ECs.
3) EC-Only Scheme: All tasks are first offloaded to the
UAV without any computations and further offloaded to
ECs for processing.
4) Fixed UAV-EC Scheme: Half of the tasks is processed at
the UAV while the other half is processed at ECs.
Note that the UAV is deployed at the optimal position for
the last three baselines similar to our proposed method. To
investigate the importance of UAV location optimization, we
name our proposed method as optimized UAV location scheme
and compare it with the random UAV location scheme. To
study the benefits of utilizing computing capacity at both UAV
and ECs, we rename our proposed method as collaborative
UAV-EC scheme and compare it with the UAV-only, EC-only, Fig. 3. Optimal task splitting ratios of the UAV βi0 (i = 1, 2, . . . , 10) and
and fixed UAV-EC schemes. ECs βij (j = 1, 2, 3, 4) for MUs.

B. Experimental Results average are first processed at the UAV (i.e., βi0 ) to reduce the
In this section, we first simulate and analyze how UAV data size, and then the remaining tasks are distributed to ECs
position and per-device bandwidth in the downlink communi- for further processing.
cation will affect the system cost of the studied UAV-enabled 2) Impact of the Per-Device Bandwidth: In this part, we
MEC system. Then, we compare the performances of our first study how the per-device bandwidth will affect the optimal
proposed collaborative UAV-EC scheme with UAV-only, EC- task splitting ratios at ECs. As mentioned before, per-device
only, and fixed UAV-EC schemes to verify the effectiveness bandwidth is assigned to each MU beforehand, and it plays an
of our method in reducing the overall system cost as well important role in affecting the optimal task splitting ratios at
as the benefits of UAV-EC collaboration. We set the simula- ECs and system cost. To proceed, we increase the per-device
tion parameters FjEC (j = 1, 2, 3, 4) and Li (i = 1, 2, . . . , 10) bandwidth BDL 1 assigned to EC1 from 0.5 to 5 MHz while the
to be [8, 9, 6, 7] GHz and [3, 5, 2, 3, 5, 1, 1, 5, 4, 5] Mbits, other three BDL DL DL
2 , B3 , and B4 remain unchanged. In Fig. 4,
respectively. we observe that the optimal task splitting ratio β11 for MU1
1) Importance of Optimizing the UAV Position: In this part, at EC1 is increasing while the other three β12 , β13 , and β14
we compare the performances of our proposed optimized UAV are decreasing. The reason is that as BDL1 increases, the A2G
downlink transmission delay t11 A2G from the UAV to the EC1
location scheme with the random UAV location scheme where
the location of UAV is randomly assigned without optimization can be reduced, and then, more tasks will be offloaded to EC1
in terms of reducing the system cost. The results are summa- for further processing and therefore, the corresponding optimal
rized in Table III. It is shown that under the optimized UAV task splitting ratio grows.
location scheme, the optimal 3-D position (x∗UAV , yUAV ∗ , H) Next, we investigate how the per-device bandwidth will
found for the UAV is at (558.11, 724.52, 100) m. The system affect the system cost. To proceed, we increase the per-device
cost of the optimized UAV location scheme is 20.83, which bandwidth BDL j assigned to the jth EC (j = 1, 2, 3, 4) from
is the best compared with randomly selected UAV locations 0.5 to 5 MHz while the other three remain at 0.5 MHz. In
(at the center or near each EC), and our proposed scheme can Fig. 5, we observe that the system cost reduces as the per-
achieve high-cost saving as 13.39%. The rationale behind the device bandwidth assigned to the jth EC increases. The reason
system cost difference is that for our proposed approach, the is that as more bandwidth assigned to each MU when tasks
UAV location is optimized to obtain better channel condition are offloaded from the UAV to the jth EC, higher downlink
when providing the offloading opportunities for ground MUs transmission data rates can be achieved, and thus, the down-
while for the random UAV location scheme, the UAV loca- link transmission delay and downlink transmission energy
tion is randomly assigned beforehand without optimization. consumption of the UAV can be reduced accordingly.
Besides, the optimal task splitting ratios of MUs for UAV and 3) Benefits of UAV-EC Collaboration: In this part, we com-
ECs are shown in Fig. 3. We observe that for MUs 2, 5, 8, 9, pare the performances of our proposed collaborative UAV-EC
and 10 with large amount of input data size, 32.12% tasks in schemes with UAV-only, EC-only and fixed UAV-EC schemes

Authorized licensed use limited to: National Chung Cheng University. Downloaded on September 11,2024 at 06:22:36 UTC from IEEE Xplore. Restrictions apply.
YU et al.: JOINT TASK OFFLOADING AND RESOURCE ALLOCATION IN UAV-ENABLED MEC 3157

Fig. 4. Optimal task splitting ratios at each EC for MU1 as a function of Fig. 7. System cost as a function of the UAV transmission power PUAV
TX
per-device bandwidth BDL
1 assigned to EC1. under four different offloading schemes.

their tasks to the UAV for execution without further offloading

to any ECs. We further observe that for the other three offload-
ing schemes, system cost increases as PUAV
TX increases since the
downlink transmission energy consumption of the UAV is an
increasing function of PUAV
TX . Under the above two scenarios,
we observe that our proposed approach largely outperforms
baseline schemes, such as UAV-only, EC-only, and fixed UAV-
EC offloading schemes in terms of reducing the system cost,
which verifies the benefits of UAV-EC collaboration in the task
offloading processes.
Fig. 5. System cost as a function of per-device bandwidth BDL
j assigned to
the jth EC (j = 1, 2, 3, 4) while fixing others at 0.5 MHz. VI. C ONCLUSION
In this article, we have studied an innovative UAV-enabled
MEC system involving the interactions among IoT devices,
UAV, and ECs. We have proposed to deploy a UAV prop-
erly to facilitate the MEC service provisioning to a set of
stationary IoT devices in regions where the existing ECs
cannot be accessible to IoT devices due to terrestrial signal
blockage and shadowing. The UAV and ECs in our system
collaboratively provide MEC services to the IoT devices using
aerial-to-ground communications. We have formulated a non-
convex optimization problem with the goal of minimizing the
weighted sum of the service delay of all IoT devices and
UAV energy consumption by jointly optimizing UAV posi-
Fig. 6. System cost as a function of the UAV computation capacity F UAV tion, communication and computing resource allocation, and
under four different offloading schemes.
task splitting decisions. We have developed an SCA-based
algorithm to tackle the nonconvexity of the original problem
by first transforming the original nonconvex problem into its
in terms of reducing the system cost. Meanwhile, we inves-
approximated convex form and then solve it efficiently. We
tigate how system cost behaves as the UAV computation
have also conducted numerical experiments to verify that our
capacity and UAV transmission power change, respectively.
proposed collaborative UAV-EC offloading scheme largely out-
First, we study how system cost behaves as the UAV compu-
performs baseline schemes that solely rely on UAV or ECs for
tation capacity F UAV increases from 3 to 30 GHz. As described
MEC in IoT. In the future, we will extend our work to the set-
in Fig. 6, the system cost of EC-only scheme does not change
ting of multiple UAVs and investigate the task offloading and
as F UAV varies since this scheme prescribes that all MUs must
UAV swarm placement in the multihop MEC scenario.
offload their tasks to ECs without any computations at the
UAV side. We further observe that for the other three offload-
R EFERENCES
ing schemes, all system cost decreases as F UAV increases since
more computation resources are available to reduce the task [1] Internet of Things Data to Top 1.6 Zettabytes by 2020. Accessed:
Jan. 6, 2019. [Online]. Available: https://campustechnology.com/articles/
computation delay at the UAV side. Then, we investigate how 2015/04/15/internet-of-things-data-to-top-1-6-zettabytes-by-2020.aspx
system cost will be affected as the UAV transmission power [2] GSMA Intelligence, the Mobile Economy 2018. Accessed: Jan. 6, 2019.
PUAV
TX increases from 1 to 10 W. As illustrated in Fig. 7, the
[Online]. Available: https://www.gsma.com/mobileeconomy/
wp-content/uploads/2018/02/The-Mobile-Economy-Global-2018.pdf
system cost of the UAV-only scheme remains constant as PUAVTX [3] G. Ananthanarayanan et al., “Real-time video analytics: The killer app
varies since this scheme indicates that all MUs must offload for edge computing,” Computer, vol. 50, no. 10, pp. 58–67, 2017.

Authorized licensed use limited to: National Chung Cheng University. Downloaded on September 11,2024 at 06:22:36 UTC from IEEE Xplore. Restrictions apply.
3158 IEEE INTERNET OF THINGS JOURNAL, VOL. 7, NO. 4, APRIL 2020

[4] J. Wang, B. Amos, A. Das, P. Pillai, N. Sadeh, and M. Satyanarayanan, [28] M. Mozaffari, W. Saad, M. Bennis, Y.-H. Nam, and M. Debbah, “A tuto-
“A scalable and privacy-aware IoT service for live video analytics,” in rial on UAVs for wireless networks: Applications, challenges, and open
Proc. ACM MMSys, 2017, pp. 38–49. problems,” IEEE Commun. Surveys Tuts., vol. 21, no. 3, pp. 2334–2360,
[5] M. Billinghurst et al., “A survey of augmented reality,” Found. TrendsR 3rd Quart., 2019.
Human–Comput. Interact., vol. 8, nos. 2–3, pp. 73–272, 2015. [29] Y. Zeng, R. Zhang, and T. J. Lim, “Throughput maximization for
[6] A. Zanella, N. Bui, A. Castellani, L. Vangelista, and M. Zorzi, “Internet UAV-enabled mobile relaying systems,” IEEE Trans. Commun., vol. 64,
of Things for smart cities,” IEEE Internet Things J., vol. 1, no. 1, no. 12, pp. 4983–4996, Dec. 2016.
pp. 22–32, Feb. 2014. [30] Q. Wu, Y. Zeng, and R. Zhang, “Joint trajectory and communication
[7] A. Al-Fuqaha, M. Guizani, M. Mohammadi, M. Aledhari, and design for multi-UAV enabled wireless networks,” IEEE Trans. Wireless
M. Ayyash, “Internet of Things: A survey on enabling technologies, pro- Commun., vol. 17, no. 3, pp. 2109–2121, Mar. 2018.
tocols, and applications,” IEEE Commun. Surveys Tuts., vol. 17, no. 4, [31] M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Mobile unmanned
pp. 2347–2376, 4th Quart., 2015. aerial vehicles (UAVs) for energy-efficient Internet of Things com-
[8] Y. C. Hu, M. Patel, D. Sabella, N. Sprecher, and V. Young, “Mobile munications,” IEEE Trans. Wireless Commun., vol. 16, no. 11,
edge computing—A key technology towards 5G,” vol. 11, ETSI, pp. 7574–7589, Nov. 2017.
Sophia Antipolis, France, White Paper, pp. 1–16, 2015. [32] C. Zhan and Y. Zeng, “Completion time minimization for multi-UAV-
[9] Y. Mao, C. You, J. Zhang, K. Huang, and K. B. Letaief, “A sur- enabled data collection,” IEEE Trans. Wireless Commun., vol. 18, no. 10,
vey on mobile edge computing: The communication perspective,” IEEE pp. 4859–4872, Oct. 2019.
Commun. Surveys Tuts., vol. 19, no. 4, pp. 2322–2358, 4th Quart., 2017. [33] S. Jeong, O. Simeone, and J. Kang, “Mobile edge computing via a UAV-
[10] N. Abbas, Y. Zhang, A. Taherkordi, and T. Skeie, “Mobile edge com- mounted cloudlet: Optimization of bit allocation and path planning,”
puting: A survey,” IEEE Internet Things J., vol. 5, no. 1, pp. 450–465, IEEE Trans. Veh. Technol., vol. 67, no. 3, pp. 2049–2063, Mar. 2018.
Feb. 2018. [34] F. Zhou, Y. Wu, R. Q. Hu, and Y. Qian, “Computation rate maximization
[11] Y. Xiao and M. Krunz, “QoE and power efficiency tradeoff for fog com- in UAV-enabled wireless-powered mobile-edge computing systems,”
puting networks with fog node cooperation,” in Proc. IEEE INFOCOM, IEEE J. Sel. Areas Commun., vol. 36, no. 9, pp. 1927–1941, Sep. 2018.
2017, pp. 1–9. [35] A. Asheralieva and D. Niyato, “Hierarchical game-theoretic and
[12] J. Xu, L. Chen, and P. Zhou, “Joint service caching and task offload- reinforcement learning framework for computational offloading in
ing for mobile edge computing in dense networks,” in Proc. IEEE UAV-enabled mobile edge computing networks with multiple service
INFOCOM, 2018, pp. 207–215. providers,” IEEE Internet Things J., vol. 6, no. 5, pp. 8753–8769,
[13] M. Chen and Y. Hao, “Task offloading for mobile edge computing in Oct. 2019.
software defined ultra-dense network,” IEEE J. Sel. Areas Commun., [36] X. Hu, K.-K. Wong, K. Yang, and Z. Zheng, “UAV-assisted relaying and
vol. 36, no. 3, pp. 587–597, Mar. 2018. edge computing: Scheduling and trajectory optimization,” IEEE Trans.
Wireless Commun., vol. 18, no. 10, pp. 4738–4752, Mar. 2019.
[14] J. Ren, G. Yu, Y. He, and G. Y. Li, “Collaborative cloud and edge
[37] G. Scutari, F. Facchinei, L. Lampariello, S. Sardellitti, and
computing for latency minimization,” IEEE Trans. Veh. Technol., vol. 68,
P. Song, “Parallel and distributed methods for constrained nonconvex
no. 5, pp. 5031–5044, May 2019.
optimization—Part II: Applications in communications and machine
[15] S. Sardellitti, G. Scutari, and S. Barbarossa, “Joint optimization of
learning,” IEEE Trans. Signal Process., vol. 65, no. 8, pp. 1945–1960,
radio and computational resources for multicell mobile-edge comput-
Apr. 2017.
ing,” IEEE Trans. Signal Inf. Process. Netw., vol. 1, no. 2, pp. 89–103,
[38] G. Scutari, F. Facchinei, and L. Lampariello, “Parallel and distributed
Jun. 2015.
methods for constrained nonconvex optimization—Part I: Theory,” IEEE
[16] K. Zhang et al., “Energy-efficient offloading for mobile edge computing
Trans. Signal Process., vol. 65, no. 8, pp. 1929–1944, Oct. 2017.
in 5G heterogeneous networks,” IEEE Access, vol. 4, pp. 5896–5907,
[39] Q. Wu, L. Liu, and R. Zhang, “Fundamental trade-offs in communication
2016.
and trajectory design for UAV-enabled wireless network,” IEEE Wireless
[17] C. You, K. Huang, H. Chae, and B.-H. Kim, “Energy-efficient Commun., vol. 26, no. 1, pp. 36–44, Feb. 2019.
resource allocation for mobile-edge computation offloading,” IEEE [40] Y. Wang, M. Sheng, X. Wang, L. Wang, and J. Li, “Mobile-edge com-
Trans. Wireless Commun., vol. 16, no. 3, pp. 1397–1411, Mar. 2017. puting: Partial computation offloading using dynamic voltage scaling,”
[18] X. Chen, L. Jiao, W. Li, and X. Fu, “Efficient multi-user computation IEEE Trans. Commun., vol. 64, no. 10, pp. 4268–4282, Oct. 2016.
offloading for mobile-edge cloud computing,” IEEE/ACM Trans. Netw., [41] M. Razaviyayn, “Successive convex approximation: Analysis and appli-
vol. 24, no. 5, pp. 2795–2808, Oct. 2016. cations,” Ph.D. dissertation, Faculty Grad. School, Univ. Minnesota,
[19] M.-H. Chen, B. Liang, and M. Dong, “Joint offloading and resource Minneapolis, MN, USA, 2014.
allocation for computation and communication in mobile cloud with [42] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, U.K.:
computing access point,” in Proc. IEEE INFOCOM, 2017, pp. 1–9. Cambridge Univ. Press, 2004.
[20] M.-H. Chen, M. Dong, and B. Liang, “Resource sharing of a computing [43] W. Yuan and K. Nahrstedt, “Energy-efficient soft real-time CPU schedul-
access point for multi-user mobile cloud offloading with delay con- ing for mobile multimedia systems,” ACM SIGOPS Oper. Syst. Rev.,
straints,” IEEE Trans. Mobile Comput., vol. 17, no. 12, pp. 2868–2881, vol. 37, no. 5, pp. 149–163, 2003.
Dec. 2018. [44] W. Yuan and K. Nahrstedt, “Energy-efficient CPU scheduling for
[21] J. Zhang, W. Xia, F. Yan, and L. Shen, “Joint computation offloading and multimedia applications,” ACM Trans. Comput. Syst., vol. 24, no. 3,
resource allocation optimization in heterogeneous networks with mobile pp. 292–331, 2006.
edge computing,” IEEE Access, vol. 6, pp. 19324–19337, 2018.
[22] Q. Hu, Y. Cai, G. Yu, Z. Qin, M. Zhao, and G. Y. Li, “Joint offloading
and trajectory design for UAV-enabled mobile edge computing systems,”
IEEE Internet Things J., vol. 6, no. 2, pp. 1879–1892, Apr. 2019.
[23] M. J. Neely, “Intelligent packet dropping for optimal energy-delay trade-
offs in wireless downlinks,” IEEE Trans. Autom. Control, vol. 54, no. 3,
pp. 565–579, Mar. 2009.
[24] H. Yu and M. J. Neely, “A new backpressure algorithm for joint rate
control and routing with vanishing utility optimality gaps and finite
queue lengths,” IEEE/ACM Trans. Netw., vol. 26, no. 4, pp. 1605–1618,
Aug. 2018.
[25] G. Sharma, R. Mazumdar, and N. B. Shroff, “Delay and capacity trade- Zhe Yu (Student Member, IEEE) received the
offs in mobile ad hoc networks: A global perspective,” IEEE/ACM Trans. B.Eng. degree in communication engineering from
Netw., vol. 15, no. 5, pp. 981–992, Oct. 2007. the University of Science and Technology Beijing,
[26] Z. Mao, C. E. Koksal, and N. B. Shroff, “Near optimal power and rate Beijing, China, in 2014, and the M.S. degree in
control of multi-hop sensor networks with energy replenishment: Basic electrical engineering from Vanderbilt University,
limitations with finite energy and data storage,” IEEE Trans. Autom. Nashville, TN, USA, in 2016. He is currently pursu-
Control, vol. 57, no. 4, pp. 815–829, Apr. 2012. ing the Ph.D. degree in electrical and computer engi-
[27] X. Cao, P. Yang, M. Alzenad, X. Xi, D. Wu, and H. Yanikomeroglu, neering with Oklahoma State University, Stillwater,
“Airborne communication networks: A survey,” IEEE J. Sel. Areas OK, USA, where his research is on resource man-
Commun., vol. 36, no. 9, pp. 1907–1926, Sep. 2018. agement for energy-efficient cloud and edge systems.

Authorized licensed use limited to: National Chung Cheng University. Downloaded on September 11,2024 at 06:22:36 UTC from IEEE Xplore. Restrictions apply.
YU et al.: JOINT TASK OFFLOADING AND RESOURCE ALLOCATION IN UAV-ENABLED MEC 3159

Yanmin Gong (Member, IEEE) received the B.Eng. Yuanxiong Guo (Senior Member, IEEE) received
degree in electronics and information engineer- the B.Eng. degree in electronics and information
ing from the Huazhong University of Science and engineering from the Huazhong University of
Technology, Wuhan, China, in 2009, the M.S. degree Science and Technology, Wuhan, China, in 2009,
in electrical engineering from Tsinghua University, and the M.S. and Ph.D. degrees in electrical
Beijing, China, in 2012, and the Ph.D. degree and computer engineering from the University of
in electrical and computer engineering from the Florida, Gainesville, FL, USA, in 2012 and 2014,
University of Florida, Gainesville, FL, USA, in respectively.
2016. From 2014 to 2019, he was with the School
She is currently an Assistant Professor with the of Electrical and Computer Engineering, Oklahoma
Department of Electrical and Computer Engineering, State University, Stillwater, OK, USA. He has been
University of Texas at San Antonio, San Antonio, TX, USA. She was an an Assistant Professor with the Department of Information Systems and Cyber
Assistant Professor with the School of Electrical and Computer Engineering, Security, University of Texas at San Antonio, San Antonio, TX, USA, since
Oklahoma State University, Stillwater, OK, USA, from 2016 to 2018. Her August 2019. His current research interests include data analytics, security,
research interests include security and privacy in big data, machine learning, and privacy with applications to Internet of Things, cyber-physical systems,
Internet of Things, cyber-physical systems, mobile computing, and wireless and cloud/edge systems.
networks. Dr. Guo is a recipient of the Best Paper Award in the IEEE Global
Dr. Gong was a recipient of the Best Paper Award of GLOBECOM Communications Conference 2011. He is currently serving as an Editor for
2017. She is serving as an Associate Editor for IEEE W IRELESS the IEEE T RANSACTIONS ON V EHICULAR T ECHNOLOGY.
C OMMUNICATIONS.

Shimin Gong (Member, IEEE) received the B.E.

and M.E. degrees in electronics and information
engineering from the Huazhong University of
Science and Technology, Wuhan, China, in 2008 and
2012, respectively, and the Ph.D. degree in computer
engineering from Nanyang Technological University,
Singapore, in 2014.
He was an Associated Researcher with the
Shenzhen Institutes of Advanced Technology,
Chinese Academy of Sciences, Beijing, China. He is
currently an Associate Professor with the School of
Intelligent Systems Engineering, Sun Yat-sen University, Guangzhou, China.
He was a Visiting Scholar with the Chinese University of Hong Kong,
Hong Kong, in 2011, and the University of Waterloo, Waterloo, ON, Canada,
in 2012. His research interests include wireless powered IoT, deep reinforce-
ment learning, backscatter communications, and networking.
Dr. Gong is a recipient of the Best Paper Award on MAC and Cross-
layer Design in IEEE WCNC 2019. He has been the Lead Guest Editor
of the IEEE T RANSACTIONS ON C OGNITIVE C OMMUNICATIONS AND
N ETWORKING—Special Issue on Deep Reinforcement Learning on Future
Wireless Communication Networks.

Authorized licensed use limited to: National Chung Cheng University. Downloaded on September 11,2024 at 06:22:36 UTC from IEEE Xplore. Restrictions apply.