This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2020.2996784, IEEE Internet of
Things Journal
IEEE INTERNET OF THINGS JOURNAL, VOL. , NO. , 2020 1

Collaborate Edge and Cloud Computing with

Distributed Deep Learning for Smart City
Internet of Things
Huaming Wu, Member, IEEE, Ziru Zhang, Chang Guan, Katinka Wolter, and Minxian Xu

Abstract—City Internet of Things (IoT) applications are becoming increasingly complicated and thus require large amounts of
computational resources and strict latency requirements. Mobile Cloud Computing (MCC) is an effective way to alleviate the limitation
of computation capacity by offloading complex tasks from mobile devices to central clouds. Besides, Mobile Edge Computing (MEC) is
a promising technology to reduce latency during data transmission and save energy by providing services in a timely manner. However,
it is still difficult to solve the task offloading challenges in heterogeneous cloud computing environments, where edge clouds and
central clouds work collaboratively to satisfy requirements of city IoT applications. In this paper, we consider the heterogeneity of edge
and central cloud servers in the offloading destination selection. To jointly optimize the system utility and the bandwidth allocation for
each mobile device, we establish a hybrid offloading model including the collaboration of MCC and MEC. A Distributed Deep
learning-driven Task Offloading (DDTO) algorithm is proposed to generate near-optimal offloading decisions over the mobile devices,
edge cloud server and central cloud server. Experimental results demonstrate the accuracy of the DDTO algorithm, which can
effectively and efficiently generate near-optimal offloading decisions in edge and cloud computing environments. Furthermore, it
achieves high performance and greatly reduces the computational complexity when compared with other offloading schemes that
neglect the collaboration of heterogeneous clouds. More precisely, the DDTO scheme can improve the computational performance by
63%, compared with the local-only scheme.

Index Terms—Mobile cloud computing, Mobile edge computing, City internet of things, Distributed deep learning, Task offloading.

1 I NTRODUCTION

W ITH the fast development of mobile networks and

the widespread application of city Internet of Things
(IoT) in various fields (e.g., smart transportation, smart
to Mobile Cloud Computing (MCC), which has emerged
as a solution to offload task workloads to computation-
rich cloud data centers. However, due to the limitations
home, smart manufacturing), the demand for Mobile De- of centralized service mode and access bandwidth, this
vices (MDs) is increasing drastically. However, MDs such approach still faces many challenges such as high latency,
as smartphones, tablet computers, unmanned aerial vehi- low bandwidth, and network congestion.
cles, and wearable devices are usually constrained by lim- Always offloading the tasks to the central cloud server,
ited resources, e.g., CPU computing power, storage space, however, is not suitable, especially for those tasks that
energy capacity, and environmental awareness. Complex are data-concentrated and latency-sensitive [1]. Recently,
computing tasks, e.g., Optical Character Recognition (OCR), Mobile Edge Computing (MEC) has emerged as a novel
Face Recognition (FR) and Augmented Reality (AR), are computing paradigm that harnesses computing resources in
inefficient to be handled locally. Further, a diversity of city the proximity of IoT devices. It has attracted extensive inter-
IoT applications such as delay-sensitive and delay-tolerant est [2]. In MEC, IoT devices are connected to edge servers
applications can cause a variety of different computation instead of directly to cloud servers. Around 29 billion IoT
and communication costs. devices are estimated to be connected to the internet by
To alleviate the limitations of mobile computation capac- 2022 [3]. Due to their proximity to mobile users the com-
ity, one effective way is to offload complex compute tasks munication cost for task offloading will become very small,
from the MDs to a central cloud. By taking advantage of which can greatly reduce the latency of network operations
the rich virtual resources and the fast processing speed of and service delivery, and further meet the requirements of
the cloud servers, we can lower the pressure on MDs in the ultra-high bandwidth and ultra-low latency of future
handling tasks locally. Considerable attention has been paid networks [4], [5]. However, the computing power of edge
cloud servers is relatively low and cannot efficiently satisfy
• H. Wu is with Center for Applied Mathematics, Tianjin University,
the requirements of the city IoT applications while central
Tianjin 300072, China. E-mail: whming@tju.edu.cn cloud servers have sufficient computing power.
• Z. Zhang and G. Chao are with School of Mathematics, Tianjin Univer- To better serve IoT users with diverse requirements,
sity, Tianjin 300072, China. Contributed Equally. heterogeneous clouds composed of edge clouds and central
• K. Wolter is with Institut für Informatik, Freie Universität Berlin, Berlin
14195, Germany. E-mail: katinka.wolter@fu-berlin.de. clouds should be jointly exploited to meet stringent delay
• M. Xu is with Shenzhen Institutes of Advanced Technology, CAS, Shen- requirements. This will make task execution faster, cheaper,
zhen 518055, China. E-mail: mx.xu@siat.ac.cn. and more stable. However, if all computing tasks are only
(Corresponding author: Huaming Wu) offloaded to the edge or central cloud server, the wireless

2327-4662 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Waterloo. Downloaded on June 01,2020 at 12:31:41 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2020.2996784, IEEE Internet of
Things Journal
IEEE INTERNET OF THINGS JOURNAL, VOL. , NO. , 2020 2

link between the IoT devices and the MCC or MEC servers • We conduct experiments in distinct situations to
can be congested and the latency of the computation can evaluate the effectiveness of DDTO. When compared
be unacceptable. In fact, MCC and MEC can cooperate in with several offloading schemes without the coop-
terms of computing, storage, and communication facilities eration of MEC and MCC, our proposed DDTO
since they are complementary to each other. To reduce the algorithm can achieve superior performance.
overall cost of delay and energy consumption, how to make
The remainder of this paper is organized as follows. In
real-time offloading decisions becomes the most significant
Section 2, we review the related work. The system model
challenge.
and problem formulation are described in Section 3. The
Due to the rapid changes of channel conditions, the num- proposed algorithm based on deep learning to generate
ber of users and other system parameters, offloading deci- the optimal binary offloading decisions is presented in Sec-
sions and resource allocation need to be completed within a tion 4. The numerical and comparison results are shown in
few milliseconds. In practical scenarios, however, large-scale Section 5. Finally, the paper is concluded in Section 6.
offloading decision-making is often involved since the total
number of possible decisions increases exponentially with
the number of users and tasks and it is very difficult to enu- 2 R ELATED WORK
merate all possible decisions. Conventional task offloading MCC and MEC have become important solutions to satisfy
techniques usually apply some heuristic algorithms, which the requirements of applications running on IoT devices, es-
involve difficult to solve and complex problems that require pecially for latency-sensitive applications and those running
a large amount of computation, and additional computation on energy-constrained IoT devices. A significant number of
is also needed to execute the offloading decisions. In addi- offloading decision schemes in MCC and MEC are provided
tion, the best solution can usually not be calculated in the in the literature, which can be classified as follows:
face of complex workflows with correlations and only local
optimal solutions can be given. Considering the absence of 2.1 Markov-based Offloading Decisions
the optimal decisions for each task workload of the users,
The Markov decision process is a well-known discrete-
deep learning becomes a promising method due to its ability
time mathematical framework applied for modeling deci-
to provide solutions based on labeled data. Deep learning-
sion making with uncertainty. It models a system based
driven approaches can facilitate offloading decision making,
on Markov chains during the time which experiences the
dynamic resource allocation and content caching as they
transition from one state to another according to certain
benefit from the growth in volumes of communication and
probabilistic rules.
computation for emerging city IoT applications. However,
Numerous stochastic offloading schemes via modeling
how to customize deep learning techniques for task offload-
the task offloading procedure as Markov decision processes
ing in IoTs is still unknown.
have been proposed in the literature to help them in making
To minimize the weighted sum of the task completion better offloading decisions [6]. Several queueing models
delay and energy consumption while maintaining the Qual- were applied in [7], [8] to mitigate the weighted sum of
ity of Service (QoS) for MDs, more intelligent technologies power usage and performance expressed in different met-
and effective parallel algorithms are required to address rics. Various offloading decision policies have been taken
such complicated offloading scenarios limited by high di- into account, where arriving tasks are either processed
mensionality. In this paper, the motivation of designing a locally in the MDs or offloaded to the remote cloud via
Distributed Deep learning-driven Task Offloading (DDTO) a WLAN or cellular network. Moreover, a Markov-based
algorithm is to find a way to proceed with optimal learn- offloading strategy was developed, which solved the prob-
ing in MEC and MCC heterogeneous environments, and lem of where to offload the tasks based on an M/G/1-
to further solve computationally expensive problems in FCFS queue model [7]. The offloading approach proposed
offloading decision-making. Considering the characteristics in [8] supports two delayed offloading policies, a partial
of the abundant computing resources in MCC and the low offloading model where jobs can leave the slow offloading to
transmission delay in MEC comprehensively, we integrate be executed locally, and a full offloading model, where jobs
MCC and MEC for task offloading. can be offloaded directly via the cellular network. Alasmari
The contributions of this paper can be summarized as et al. [6] proposed a mobile edge offloading method based
follows: on a Markov decision process to generate offloading deci-
sions, which used a numbering scheme (1, 2, 3) to denote
• We formalize the MCC and MEC hybrid task place-
executing the tasks in local devices, at the edge and cloud,
ment problem as a multi-objective optimization
respectively. Considering the clock frequency configuration,
problem. To jointly minimize the system utility and
transmission power allocation, channel rate scheduling and
the bandwidth allocation for each MD, we propose
offloading strategy selection, a distributed algorithm was
an effective and efficient offloading framework with
derived in [11], where an M/M/n queue model was also
intelligent decision-making capabilities.
used to optimize the offloading decision.
• We design a Distributed Deep learning-driven Task
Offloading (DDTO) algorithm, where multiple par-
allel Deep Neural Networks (DNNs) are adopted to 2.2 Graph-based Offloading Decisions
effectively and efficiently generate offloading deci- It is important to note that city IoT applications can be
sions over the MDs, edge cloud server and central viewed as heterogeneous workflows with a different num-
cloud server. ber of tasks and data flows. To make offloading decisions

2327-4662 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Waterloo. Downloaded on June 01,2020 at 12:31:41 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2020.2996784, IEEE Internet of
Things Journal
IEEE INTERNET OF THINGS JOURNAL, VOL. , NO. , 2020 3

TABLE 1: Comparison of different offloading decision schemes

Offloading Offloading Task Architectural Properties Decision Objectives

Categories
Schemes Mode Number MCC MEC MCC & MEC Time Energy Weighted
Scheme in [6] Partial Multiple × X × X X ×
Scheme in [7] Partial Multiple X × × X X X
Markov-based Scheme in [8] Partial Multiple X × × X X X
offloading decisions MCOWA [9] Partial Multiple × X × X X ×
Unit-slot [10] Partial Single X × × X × ×
Scheme in [11] Partial Multiple × X × X X ×
K-M-LARAC [12] Partial Multiple X × × X X ×
Scheme in [13] Partial Multiple × X × × X ×
Graph-based
Scheme in [14] Partial Multiple X × × X × ×
offloading decisions
One-climb policy [15] Partial Multiple X × × × X ×
MCOP [16] Partial Multiple X × × X X X
MAPCloud [17] Partial Multiple X × × X X ×
Scheme in [18] Partial Single × X × × X ×
SDTO [19] Full Multiple × X × X X ×
Optimization-based COM [20] Partial Multiple X × × X X ×
offloading decisions F-SGA&C-SGA [21] Partial Multiple × X × X × ×
EEDOA [22] Full Multiple × X × × X ×
Scheme in [23] Partial Multiple X × × X X ×
HGPCA [24] Full Multiple × X × × X ×
Scheme in [25] Partial Multiple × X × × X ×
DDLO [26] Partial Multiple × X × × × X
DROO [27] Partial Multiple × X × × X ×
Deep learning-based DRLO [28] Partial Multiple × X × × X ×
offloading decisions Scheme in [29] Partial Multiple × X × × X ×
Scheme in [30] Partial Multiple × X × X X X
MCCG [31] Partial Multiple × × X X X ×
Proposed DDTO scheme Partial Multiple X X X X X X

based on optimizing the response time or energy consump- 2.3 Optimization-based Offloading Decisions
tion, many research efforts have been devoted to computa- A diversity of platforms and algorithms have been proposed
tion partitioning in mobile computing. Automatic applica- to solve the problems of offloading binary decisions for
tion partitioning has attracted more and more attention. MCC and MEC.
An offloading platform named MAPCloud was pro-
The offloading operation can be modeled via a cost posed in [17], which consists of a local cloud and a com-
graph, where finding the optimal solution for offloading is mon cloud. MAPCloud determined the optimal location
equivalent to finding the constrained shortest path in this of tasks according to multiple QoS factors of users. Haber
graph [12], [32], [33]. Zhang et al. [15] modeled a mobile et al. [18] proposed a successive convex approximation
application as a general topology, which consists of a set of method, which approximately optimizes the computational
fine-grained tasks. Each task within the application can be cost and figures out the energy-efficient task offloading
either executed on the mobile device or on the cloud. By strategy mathematically. A computational offloading algo-
using arbitrary topographical consumption graphs, Wu et rithm based on the NSGA-III is presented in [20], where
al. [16] proposed a graph-cut-based partitioning algorithm, big data methods have been used for IoT-enabled cloud-
which determines whether the parts of the tasks run locally edge computing. In addition, computation offloading game
or offload to the cloud server. The decision engine in this theory has been discussed in [21], which proposed C-SGA
proposal is placed at the mobile device aiming at finding a and F-SGA algorithms to solve the problem.
group of tasks for offloading, by which the execution time Energy-efficient task offloading algorithms in MEC or
of a mobile application and the energy consumption of a MCC based on Lyapunov optimization theory have been
mobile device are reduced. widely investigated [22], [23]. The authors derived adaptive
offloading decision algorithms when taking advantage of
Preferably the graph partitioning between IoT devices Lyapunov optimization techniques. The algorithm deter-
and cloud/edge servers should be dynamic and the offload- mined when and on which network, and where to perform
ing decisions should be made adaptively at runtime. How- each application task (i.e., IoT device, edge server or cloud
ever, only homogeneous resources are considered in these server) such that the overall energy consumption is mini-
studies, and unlike them, we consider the edge and cloud mized while guaranteeing the average queue length.
computing to be heterogeneous environments to support Many optimization-based algorithms, e.g., traversal or
the IoT applications running on diverse devices in a better linear programming, can only obtain results after mul-
manner. tiple iterations, which often involve too many complex

2327-4662 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Waterloo. Downloaded on June 01,2020 at 12:31:41 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2020.2996784, IEEE Internet of
Things Journal
IEEE INTERNET OF THINGS JOURNAL, VOL. , NO. , 2020 4

calculation operations, e.g., matrix inversion and singular amount of computation within the same time interval. It
value decomposition, resulting in high running time cost in may incur different communication costs due to the specific
offloading decision-making. Moreover, these optimization connectivity and cloud availability [35].
methods that only take advantage of MEC or MCC struggle As shown in Table 2, compared to cloud servers, edge
to balance the complexity and optimality. Thus, it is neces- servers are closer to the mobile devices and thus have lower
sary to develop an algorithm that can be used for real-time latency. However, the edge server has low computing power
offloading decisions with MEC and MCC collaboration. as compared to the cloud server, which has relatively suffi-
cient computing power [36]. Therefore, MEC can be treated
2.4 Deep Learning-based Offloading Decisions as an extension of traditional MCC, but not an alternative to
MCC.
Deep learning is very promising for solving complicated
real-world scenarios, e.g., Internet of Vehicles (IoVs) [29], TABLE 2: Comparison of MCC and MEC [36]
Unmanned Aerial Vehicles (UAVs) [30] and industrial IoTs
[31]. Recently, deep learning-driven offloading schemes play Technical Aspect MCC MEC
an increasingly important role in dealing with task of- Deployment centralized distributed
floading decisions for MEC and/or MCC, i.e., intelligent Distance to IoT device high low
offloading [27]. Latency high low
Delay jitter high low
A model-free reinforcement learning offloading mecha- Computational power ample limited
nism was proposed in [25], which uses a gaming framework Storage capacity ample limited
and reaches 87.87% payoff compared to the optimal condi- Mobility support ample limited
tion. In order to solve the offloading decision problem in Privacy protect low high
the MEC environment, a distributed deep learning-based
offloading algorithm has been proposed in [26], where par- Few recent studies have focused on identifying and ad-
allel computing is utilized to speed up the computation. dressing important challenges of task offloading in hetero-
Apart from that, Min et al. [28] proposed a reinforcement geneous edge and cloud computing environments, where
learning-based solution to solve the task offload decision edge clouds and central clouds work collaboratively to sat-
of IoT devices with energy-harvesting functions, which en- isfy the city IoT application requirements. Here, we consider
ables IoT devices to optimize the offloading strategy without the heterogeneity of different edge and cloud servers in
knowing the MEC model, energy consumption model and the offloading destination selection. To jointly optimize the
delay model. system utility and the bandwidth allocation for each MD,
Many existing deep learning-based offloading schemes, we establish a hybrid offloading model with the collabo-
however, optimize all system parameters simultaneously, ration of MCC and MEC. In addition, a distributed deep
which will eventually identify infeasible solutions as the learning-driven algorithm is proposed to generate optimal
optimal offloading decision. Moreover, the heterogeneity offloading decisions for heterogeneous clouds. To the best of
of the servers is still ignored in the selection of the of- knowledge, our work is the first that adopts deep learning
floading destination and the definition of the convergence with the collaboration of MCC and MEC for heterogeneous
in these works is not clear. Inspired by recent advantages servers.
of deep learning in handling offloading decision problems
with large search spaces, we take advantage of parallel
computing of DNNs, meanwhile, the convergence of the
3 S YSTEM M ODEL AND P ROBLEM F ORMULATION
deep learning-based decision algorithm is clearly defined In this section, we consider a framework of hybrid
and improved. In addition, the heterogeneity of servers task offloading with heterogeneous clouds, in which mo-
and devices is also considered in the MEC and MCC en- bile devices can execute their workflows locally or com-
vironment. Once the IoT environment changes, the use of pletely/partially offload them to the central cloud and/or
deep learning methods requires new labeled data, and the to the edge cloud for execution.
offloading decision for complex tasks required by different
services should have long-term programming and contin- 3.1 System Model
uous learning capabilities to meet the requirements of city
Figure 1 presents an overview of our system model. We con-
IoT applications.
sider one edge cloud, one central cloud and multiple MDs,
where each MD can choose to offload its computation tasks
2.5 A Qualitative Comparison either to the edge cloud server or to the central cloud server.
As listed in Table 1, we identify and compare key elements We aim at effectively integrating heterogeneous computing
of related work with ours in terms of their offloading modes, resources in the MEC and MCC collaborative computing en-
architectural properties, and decision objectives. To summa- vironment, where the edge cloud and the central cloud can
rize, the literature above only concentrates on local devices be interconnected. The mobile application is divided into
and edge clouds, or ignores the possibly high dimensions of multiple tasks by the application partitioning algorithm. The
the problem. tasks can then be offloaded to cloud servers. Offloading is
In fact, there can be multiple offloading destinations performed according to the complexity of the tasks and the
and targets for task placement [34]. Due to the different present network environment, i.e., offloading the compute-
speeds of heterogeneous cloud servers, offloading the same intensive task to the central cloud server and offloading
application to different places may complete a different the data-intensive task to the edge cloud server, thereby

2327-4662 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Waterloo. Downloaded on June 01,2020 at 12:31:41 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2020.2996784, IEEE Internet of
Things Journal
IEEE INTERNET OF THINGS JOURNAL, VOL. , NO. , 2020 5

alleviating the difficulties of load bottlenecks, delays, and resources such as battery capacity, MDs can only perform
ensuring fault tolerance. fundamental tasks.
Let c(n,m) denote the total CPU cycles of computing the
mth task of the nth MD. Considering that the CPU cycles
are proportional to the workloads, which is given by:
c(n,m) = δw(n,m) , (3)
where δ denotes the positive coefficient of proportionality.
The energy used while executing the mth task at the nth
local device can be expressed as:
El(n,m) = θl c(n,m) = θl δw(n,m) , (4)
where θl denotes the energy consumption on the local
device per unit of workloads.
The execution time of the local device can be calculated
as:
c(n,m) δw(n,m)
T l(n,m) = = , (5)
fl fl
Fig. 1: System model of task offloading with heterogeneous
clouds where fl denotes the task processing rate of the MDs.
Therefore, the total computation time of the nth MD can
The system model consists of one central cloud server, be derived as:
one edge cloud server, one wireless Access Point (AP), M
X
multiple Mobile Devices (MDs), denoted by a set N = T l(n) = x1(n,m) T l(n,m) . (6)
{1, 2, · · · , N }, and some independent computational tasks, m=1
denoted by a set M = {1, 2, · · · , M }. We denote the size
of the mth task of the nth MD by w(n,m) , since each MD 3.1.2 Edge Cloud Computing Model
has several tasks to cope with. In addition, each MD can Edge cloud servers are close to the MDs and communicate
either execute their tasks locally or offload them to the cloud with them via different wireless communication technolo-
servers for further execution. Once the decision is taken gies such as Bluetooth or WiFi. Edge cloud servers provide
that a task will be offloaded to the cloud server, it can be a low-latency computing service to MDs because they form
offloaded either to the central cloud server or to the edge a local area network (LAN) with the MDs.
cloud server. We define two binary variables named x1(n,m) The transmission time for offloading the workload to the
and x2(n,m) to represent the offloading decisions. edge cloud server via the access point can be given by:
On one hand, x1(n,m) ∈ {0, 1} stands for the offloading w(n,m)
T t(n,m) = , (7)
decision for the mth task, which is measured as: bn
1, if task is executed on the nth MD,

1
x(n,m) = (1) where bn denotes the bandwidth of the nth MD.
0, if task is offloaded to the cloud server, The energy consumption for the transmission can be
expressed as:
where x1(n,m) = 1 denotes the mth task is processed locally
Et(n,m) = σw(n,m) , (8)
on the nth MD, and x1(n,m) = 0 indicates the mth task is
offloaded to the cloud server. where σ denotes the positive coefficient of proportionality.
On the other hand, once the mth task is decided to be After transmitting the tasks to the edge cloud, they will
offloaded to the cloud server, we further define x2(n,m) ∈ be executed by the edge cloud server. The completion delay
{0, 1} to represent the offloading destination selection for of the whole progress can be formulated as:
the mth task, which is measured as: c(n,m)

1, if offloaded to edge cloud & x1(n,m) = 0 T e(n,m) = T t(n,m) + , (9)
2
x(n,m) = (2) fe
1
0, if offloaded to central cloud & x(n,m) = 0
where fe indicates the task processing rate of the edge cloud
where x2(n,m) = 1 indicates that the mth task is offloaded to server. Then the total time delay of the nth MD can be
the edge cloud server, and x2(n,m) = 0 denotes that the mth formulated as:
task is offloaded to the central cloud server. XM  
For convenience, all parameters used in this paper are T e(n) = 1 − x1(n,m) x2(n,m) T e(n,m) . (10)
listed in Table 3. The detailed operations of local computing, m=1
edge cloud computing and central cloud computing models The energy consumption during all steps can be com-
are illustrated as follows, respectively. puted as:
Ee(n,m) = Et(n,m) + θe c(n,m) , (11)
3.1.1 Local Computing Model
We first introduce the local computing model when the where θe denotes the energy consumption per unit of work-
MDs decide to execute their tasks locally. Due to limited load at the edge cloud server.

2327-4662 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Waterloo. Downloaded on June 01,2020 at 12:31:41 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2020.2996784, IEEE Internet of
Things Journal
IEEE INTERNET OF THINGS JOURNAL, VOL. , NO. , 2020 6

TABLE 3: Summary of notations

Notation Description
w(n,m) The mth task workload of the nth MD
δ The positive proportion coefficient
c(n,m) The total CPU cycles of computing the mth task of the nth MD
x1(n,m) x1(n,m) = 1 if computing the mth task at local, x1(n,m) = 0 if computing the mth task at the cloud
x2(n,m) When x1(n,m) = 0, x2(n,m) = 0 if computing the task at the central cloud, x2(n,m) = 1 if computing the task at the edge cloud
El(n,m) The energy consumption of the mth task of the nth MD
θl The local device energy consumption per unit of workloads
T l(n,m) The execution time of the nth MD
fl The task processing rate of the MD
T l(n) The total execution time of the nth MD
T t(n,m) The transmission time of offloading the mth task to the edge cloud server via the access point
bn The bandwidth of the nth MD
B The total available bandwidth of all users
T e(n,m) The time delay of offloading the mth task of the nth MD to the edge cloud server
fe The task processing rate of the edge cloud server
T e(n) The total time delay of offloading all the tasks of the nth MD to the edge cloud server
Ee(n,m) The energy consumption of offloading the mth tasks of the nth MD to the edge cloud server
θe The edge cloud energy consumption per unit of workloads
T c(n,m) The time delay of offloading the mth task of the nth MD to the central cloud server
fc The task processing rate of the central cloud server
Ec(n,m) The energy consumption of offloading the mth task of the nth MD to the central cloud server
θc The central cloud energy consumption per unit of workloads
T c(n) The total time delay of offloading all the tasks of the nth MD to the central cloud server

3.1.3 Central Cloud Computing Model The total energy consumption consumed in the whole
Central cloud servers can provide the most powerful com- hybrid offloading model can be expressed by:
puting capacity and can be a private cloud or public cloud
offered by cloud service providers. N X
X M  
Similarly to the edge cloud computing model, we as- E= x1(n,m) El(n,m) + (1 − x1(n,m) ) x2(n,m) Ee(n,m)
sume that the transmission time and the energy consump- n=1 m=1
tion from the local device to the central cloud server are 
+(1 − x2(n,m) )Ec(n,m) . (15)
approximately equal to T t(n,m) and Et(n,m) , respectively.
Then, the total execution time and the energy consumption
can be given by:
Meanwhile, the total execution time required to execute
c(n,m) all the tasks can be given by:
T c(n,m) = T t(n,m) + , (12)
fc
Ec(n,m) = Et(n,m) + θc c(n,m) , (13) N
X 
where θc denotes the energy consumption per unit of work- T = max T l(n,m) , T e(n,m) , T c(n,m) . (16)
n=1
load of the central cloud. fc denotes the task processing rate
of the central cloud server.
In general, the computing power of MDs, edge cloud Then, the function Q(w, x, b) can be calculated as:
server and cloud server satisfy the following: fl < fe < fc ,
which means that the central cloud server has the strongest Q(w, x, b) = ψ × E + (1 − ψ) × T, (17)
computing power, followed by the edge cloud server and
then the MDs [37]. 
Therefore, the total execution time of the nth MD can be where w n= w(n,m) |n ∈ N , m ∈ M , bo = {bn |n ∈ N },
derived as: and x = x1(n,m) , x2(n,m) |n ∈ N , m ∈ M . The parameter
M 
X   ψ with 0 ≤ ψ ≤ 1 is a weighting parameter that represents
T c(n) = 1 − x1(n,m) 1 − x2(n,m) T c(n,m) . (14) the relative significance of the energy consumption and
m=1 the execution time, by which the weighted cost model can
be adjusted according to the users’ requirements. To focus
3.2 Problem Formulation more on improving the performance, ψ should be less than
In order to minimize both, the execution time of all the 0.5; to focus more on reducing the energy consumption, ψ
tasks and the energy consumption of MDs, we introduce should be greater than 0.5. We only consider the execution
a function Q (w, x, b), which is the weighted sum of the time in the case ψ = 0, and we only consider the energy
execution time and the energy consumption. The weighted consumption at MDs when ψ = 1.
sum is related to the workload, the offloading decision and Next, we formulate an optimization problem (P ) to
the bandwidth allocated to the task. minimize the Q(w, x, b) by jointly optimizing offloading

2327-4662 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Waterloo. Downloaded on June 01,2020 at 12:31:41 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2020.2996784, IEEE Internet of
Things Journal
IEEE INTERNET OF THINGS JOURNAL, VOL. , NO. , 2020 7

decisions and bandwidth allocation, which is expressed as:

(P ) : Q(w) = min Q(w, x, b), (18a)

x,b
s.t. : bn ≥ 0, ∀n ∈ N , (18b)
XN
bn ≤ B, (18c)
n=1
x1(n,m) , x2(n,m) ∈ {0, 1} , (18d)

where B denotes the total available bandwidth of N users.

The constraint in (18b) indicates that the allocation of the
bandwidth should not be negative. In addition, the sum
of bn cannot exceed the maximum bandwidth B , which is
given in (18c). The binary offloading decisions x1(n,m) and
x2(n,m) are defined in (1) and (2), respectively. Studies on
efficiently solving the bandwidth allocation problem have
been shown in [38], [39], where the bandwidth allocation is
a convex problem that can be solved by an optimizer. Here,
we just consider the given workloads w and the offloading
decision x to optimize the function Q(w, x).
This is a Mixed-Integer Programming (MIP) problem
with high-dimensional state space. In order to tackle such
a complex problem, one needs to find an optimal offloading
decision in MEC and MCC heterogeneous environments. In
this problem, there are a total of 3N M possible offloading
decisions to select from. Due to the exponentially large
search space, the optimization problem is difficult to be Fig. 2: The procedure of the DDTO algorithm
solved in conventional ways such as with heuristic search
algorithms. To solve the problem (P ) in an effective and
efficient way, we will in the next section introduce a deep 4.1 Offloading Decision Generation
learning-driven algorithm to generate offloading decisions. In this section, we propose a method to obtain the approx-
imate optimal offloading decisions. The mean square error
(MSE) function is applied to obtain the optimal offloading
4 DDTO A LGORITHM decision by minimizing the loss function in deep learning.
The M SE function is formulated by:
In this section, we propose a distributed deep learning-
n
driven task offloading (DDTO) algorithm for the MCC and 1X
M SE = |logitst − outputst |2 , (19)
MEC hybrid offloading model, which is based on multiple n t=1
parallel deep neural networks. The architecture of DDTO
algorithm is as depicted in Fig. 2. where logits and outputs denote the label and the predicted
value, respectively. As each element of the decision is binary,
When all users’ task workloads are given by
the logits can only be 0 or 1. It is straight forward to prove
w = [w(1,1) , w(1,2) , · · · , w(N,M ) ], our target is to
that if the output is larger than 1/2, the logits will be 1,
figure out the optimal offloading decision x =
otherwise 0. In this way, the M SE function is minimized,
[x1(1,1) , x2(1,2) , x1(1,2) , · · · , x1(N,M ) , x2(N,M ) ]. We assume that
which means the precision of the model is highest.
all users have the same number of tasks because the appli-
As depicted in Fig. 3, the generation process of the
cation can be divided into multiple tasks and the workloads
offloading decisions can be expressed as follows: when the
of extra tasks can be treated as zeros. Furthermore, since we
inputs w are given, we first use the DNNs to get the outputs.
cannot get the optimal decisions directly, it is an unsuper-
Then we use the method described above to generate the
vised learning problem which is difficult to solve. Therefore,
offloading decision as our logits. This is how we create the
we propose a method to obtain the offloading decisions and
labeled database and then the problem can be solved by
turn it into a supervised learning problem.
deep learning.
We regard the workloads w as the input to the neural
networks and the optimal offloading decisions x as our
output. Importantly, we store w, the best decisions x and 4.2 Training
the minimum value of the function Q(w, x) into a database We decide to train S parallel DNNs to solve the optimization
together. We then use these labeled data to train our mul- problem. Each of the DNNs consists of one input layer, two
tiple parallel deep neural networks (DNNs) and generate hidden layers and one output layer.
new data to replace the old data in the database. Thus we Once the best offloading decision x∗ is obtained, we save
can update the database and train the DNNs to solve the the workloads w, the best offloading decision x∗ , and the
NP-hard problem well. value of the function Q(w, x∗ ) together in a database. The

2327-4662 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Waterloo. Downloaded on June 01,2020 at 12:31:41 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2020.2996784, IEEE Internet of
Things Journal
IEEE INTERNET OF THINGS JOURNAL, VOL. , NO. , 2020 8

We find Q∗1 by using a time-consuming greedy algorithm,

where we enumerate all offloading decision combinations
and identify the true optimal one. Then we compute a ratio
of the minimum value to the optimal value. We defined it
as R2 , which can be given by:
Q∗1
R2 = , (22)
Q∗2
where 0 < R2 ≤ 1 indicates how close the solution found by
our algorithm comes to the true optimal solution achieved
by the greedy algorithm. When R2 = 1, it means that we
have found the true optimal solution and we call it relative
optimality.
The whole progress of the DDTO algorithm for the MCC
and MEC hybrid offloading model is displayed in Algo-
rithm 1. The database structure is initially empty and mul-
Fig. 3: The process of generating the offloading decisions tiple DNNs are initialized with random parameter values
λk . The proposed DDTO framework learns from the past
offloading experiences in MEC and MCC heterogeneous en-
size of the database is limited and can be set to an arbitrary vironments and then automatically adjusts the parameters
value. The database works in a round-robin fashion, i.e. to generate near-optimal offloading decisions. In this way,
when the database is full, the oldest data will be abandoned it eliminates the need for solving complex MIP problems
and new data will replace the old one. In addition, the and then avoids the curse of dimensionality with a high-
labeled data from the structure can be used to train all dimensional search space. The DDTO algorithm only needs
DNNs. One issue we considered here is that it will take to choose from a few candidate offloading decisions each
too long if all DNNs are trained by all labeled data from the time and thus the computational complexity will not in-
database. Therefore, the relay technology is used in this part crease dramatically with the growth in the numbers of users
[40]. More specifically, the database is shared by all DNNs and tasks. Good convergence performance can be achieved
and each of them can extract a batch of data randomly from because of the high diversity in the generated offloading
the database to train the neural network. As the database decisions.
is constantly updated, and the newly generated data will
be more precise than the older one, the efficiency will be ALGORITHM 1: Distributed Deep learning-driven
improved by this database. Task Offloading (DDTO) algorithm
This is a classification problem, we determine to perform Input: Workloads w of local MDs
the cross-entropy as the loss function, which is given by: Output: Optimal offloading decisions
L(λk ) = −xT log fλk (w) − (1 − x)T log(1 − fλk (w)), (20) 1 Initialization:
2 Initialize S DNNs with random parameter λj
where λk is the parameter value of the DNN. We employ 3 Empty the database
the gradient descent method to minimize the cross-entropy 4 for i = 1, 2, · · · , N do
loss, and then update the parameters of all DNNs. 5 Replicate ith offloading decision candidate xi
from the ith DNN
4.3 Testing 6 Select the optimal offloading decision x∗ by
We generate offloading decisions as our logits, and up- minimizing Q(w, x) and calculate Q(w, x∗ ) as
date the database structure continuously, thus we define Q∗
the convergence as the process of approaching towards 7 if database is not full then
a defined value, i.e., the extremum. More specifically, we 8 Store (wi , x∗ , Q∗ ) into the database
denote the minimum value of the function Q(w, x) during 9 else
the process of generating offloading decisions as Q1 . When 10 Discard the oldest data and save the new one
we randomly select a batch of data from the database and 11 end
repeat the previous procedure, we obtain another optimal 12 end
value of Q(w, x), denoted as Q2 . Then the ratio R1 can be 13 for j = 1, 2, · · · , S do
formulated as: 14 Randomly choose a batch of data from database
min(Q1 , Q2 ) 15 Train the DNNs and update the parameter λj
R1 = , (21) 16 end
max(Q1 , Q2 )
where the ratio R1 can be interpreted as the convergence of
the DDTO algorithm.
We cannot treat the extremum as the true minimum
or maximum value, so we decide to enumerate all cases 5 P ERFORMANCE EVALUATION
to find the true minimum value of Q(w, x) expressed as In this section, we demonstrate the experimental results
Q∗1 and compare the result with our optimal value Q∗2 . of our proposed DDTO algorithm for solving the problem

2327-4662 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Waterloo. Downloaded on June 01,2020 at 12:31:41 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2020.2996784, IEEE Internet of
Things Journal
IEEE INTERNET OF THINGS JOURNAL, VOL. , NO. , 2020 9

(P ) and evaluate the performance of different offloading definition of R1 , when S = 1 the neural network cannot be
strategies. trained very well and the result will not converge. However,
when S ≥ 2 the convergence performs very well after
10, 000 learning steps. It illustrates that as the number of
5.1 Parameter Setting
DNNs increases, the function of the DDTO algorithm will
In our experiments, the proposed DDTO algorithm and be improved. However, the computation time will also in-
other offloading decision algorithms are implemented and crease. Therefore, we set the number of DNNs to S = 6 as a
evaluated in Python using the ML library Tensorflow. An compromise between the best configurations to optimize the
edge and cloud computing heterogeneous environment is time consumption of running the code and the performance
built. We set the number of users or MDs N = 3, and each of the convergence.
of them has M = 3 independent computation tasks at the As depicted in Fig. 4(b), the ratio R1 increases with the
same time. learning step. Importantly, when the size of the database
Apart from that, we set the parameter θc = 1, the central equals 1400, the convergence performs best. This is because
cloud energy consumption per unit of workloads, since the the data will be updated at a low rate when the size of the
energy consumption at the central cloud server is the lowest. database is too large. Meanwhile, the data that is randomly
Then we set the edge cloud energy consumption per unit of selected from the database will not be acceptable when
workloads θe = 1.5 J/MB and the local energy consumption it is too small. Therefore, the size of the database has a
per unit of workloads θl = 3 J/MB, respectively [41]. considerable influence on the gain ratio, because it alters
There are two kinds of energy profilers that can be used the speed of updates.
to estimate the energy consumption of MDs, namely, soft- As shown in Fig. 4(c), the best performance is achieved
ware and hardware monitors. Although the measurement when α0 =0.01. We analyze that when α0 is too small, the
results provided by the former are not as accurate as those convergence rate is low. Simultaneously, when α0 is too
provided by the latter, they are more convenient to use and large, it will converge to another extremum, thus the ratio
the result is still reasonable [16]. Similarly, we set the task R1 will be very low. Therefore, the learning rate α0 also has
processing rate fl = 1 and the processing rates of the edge an influence on the convergence.
cloud server and the central cloud server are fe = 800
MHz and fc = 1200 MHz, respectively. The parameters
satisfy: fl < fe < fc . In addition, we suppose that the input 5.3 The Performance of the relative optimality R2
workloads of all tasks are randomly distributed between 0 We demonstrate that the relative optimality, the ratio R2 , is
and 30 MB. In all simulations, the set value of the bandwidth affected by the number of DNNs, the learning rate and the
limit bn of each user is 50 Mbps. The weighting parameter ψ size of the database, respectively.
is set to 0.5, indicating that our focus is on both, balancing Figure 5(a) depicts the impact of the number of DNNs
performance and reducing power consumption. We train on the relative optimality R2 . It can be seen that the value
DNNs using batches of size 500 of the labeled data from of R2 is increasing when the DNN number increases and
the database. The summary of our evaluation parameters the value of R2 reaches its maximum value approximately
and their respective values are presented in Table 4. when the number of DNNs is larger than 5. We selected
S = 6 before, thus the figure indicates that this was a good
TABLE 4: Evaluation Parameters choice.
Figure 5(b) shows the performance of the relative opti-
Evaluation Parameters Values
The number of users or mobile devices N =3
mality R2 under different sizes of the database. The value
The number of independent computational tasks M =3 of R2 is fluctuating when the size is increasing. It is visible
The local energy consumption per unit θl = 3 J/MB that when the size of database reaches 1400, the R2 gets
The edge cloud energy consumption per unit θe = 1.5 J/MB the vertex, which is greater than 0.925. We previously chose
The central cloud energy consumption per unit θc = 1 J/MB 1400 as the size of the database, and this figure supports this
The processing rate of the local device fl = 100 MHz selection.
The processing rate of the edge cloud server fe = 800 MHz
The processing rate of the central cloud server fc = 1200 MHz
Figure 5(c) shows the performance of the relative opti-
The input workloads of all tasks [0 − 30] MB mality R2 with different learning rates. The value of R2 is
The bandwidth limit bn = 50 Mbps up to maximum when the learning rate is 0.01. It indicates
The weighting parameter of time and energy ψ = 0.5 that the previous choice is the optimal one since when the
learning rate is larger or smaller, R2 is decreasing.
Figures 5(a)-5(c) demonstrate that the former selection of
the DNN number, the size of the database and the learning
5.2 Convergence Performance rate are all reasonable. In addition, the value of R2 exceeds
We demonstrate the convergence of the DDTO algorithm 0.92 under our choice.
in distinct situations, where it converges to the optimal
solution under a wide range of parameter settings. The con-
vergence performance of the DDTO algorithm is analyzed 5.4 Comparison Analysis
on the basis of the number of DNNs, the size of the database To gain some insights and analyze the efficiency of the
and the learning rate, respectively. proposed DDTO algorithm, the following state-of-the-art
We observe from Fig. 4(a) that R1 converges to 1 as offloading decision methods are implemented for compar-
the learning step increases when S ≥ 2. According to the isons:

2327-4662 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Waterloo. Downloaded on June 01,2020 at 12:31:41 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2020.2996784, IEEE Internet of
Things Journal
IEEE INTERNET OF THINGS JOURNAL, VOL. , NO. , 2020 10

(a) The number of DNNs (b) The size of database (c) The learning rate

Fig. 4: Impact of the number of DNNs, the size of the database and the learning rate on R1

(a) The number of DNNs (b) The size of database (c) The learning rate

Fig. 5: Impact of the number of DNNs, the size of the database and the learning rate on R2

• Local-only Scheme: i.e., the zero offloading scheme. In

this method, all tasks of workflows are executed lo-
cally on their respective MDs, and hence, no parallel
execution of tasks can be performed for workflows.
Here, the offloading decisions x1(n,m) will be 1. The
results of this method can be used as a benchmark to
analyze the gain of different types of task offloading
techniques.
• Edge-only Scheme: i.e., the edge cloud-only offloading
scheme. In this method, all tasks of workflows are
fully offloaded to the edge cloud server for execution
[42]. The offloading decisions x1(n,m) and x2(n,m) will
be 0 and 1, respectively.
• Central-only Scheme: i.e., the central cloud-only of- Fig. 6: Comparison with several offloading schemes
floading scheme. In this method, all tasks of work-
flows are fully offloaded to the central cloud server
for further processing [23]. The offloading decisions the other four schemes. For example, the R2 −value of the
x1(n,m) and x2(n,m) will be 0 and 0, respectively. local-only scheme is only about 0.3, and the local & central
• Local & Central Scheme: i.e., the local execution scheme is approximately 0.75. This is because unlike the
and central cloud partial offloading scheme. In this edge-only and the central-only schemes, the DDTO scheme
method, some tasks of workflows are processed lo- dynamically offloads tasks according to the heterogeneous
cally on the MDs, while some of them are offloaded computing environment such as task workloads, communi-
to the central cloud server for further processing [26]. cation data and network conditions. Especially when the
• Our Algorithm: i.e., the proposed DDTO scheme. In network bandwidth is very low offloading tasks to the
this method, we adopt the proposed DDTO algo- edge/cloud server may not be beneficial. Therefore, the
rithm to generate optimal offloading decisions over proposed DDTO scheme can achieve near-optimal offload-
the MDs, the edge cloud server and the central cloud ing decisions in edge and cloud computing heterogeneous
server. environments.

The comparison of the results of different offloading

schemes are shown in Fig. 6. It can be seen that the optimal 6 C ONCLUSIONS AND F UTURE W ORK
decisions of the proposed DDTO algorithm perform very In contrast to conventional distributed deep learning ap-
well since the relative optimality, the ratio R2 , of the DDTO proaches, we have proposed a DDTO algorithm with het-
scheme exceeds 0.93, which is much higher than for any of erogeneous clouds, i.e., the central cloud and the edge

2327-4662 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Waterloo. Downloaded on June 01,2020 at 12:31:41 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2020.2996784, IEEE Internet of
Things Journal
IEEE INTERNET OF THINGS JOURNAL, VOL. , NO. , 2020 11

cloud, which optimizes the weighted sum of the energy [5] K. Yu, Y. Wang, J. Yu, D. Yu, X. Cheng, and Z. Shan, “Localized
consumption and the execution time in the MCC and MEC and distributed link scheduling algorithms in IoT under rayleigh
fading,” Computer Networks, vol. 151, pp. 232–244, 2019.
hybrid offloading model. This is achieved by generating [6] K. R. Alasmari, R. C. Green, and M. Alam, “Mobile edge offloading
and storing the offloading decisions with the workloads using markov decision processes,” in International Conference on
and system consumption together in a database and then Edge Computing. Springer, 2018, pp. 80–90.
training and updating multiple parallel DNNs with a batch [7] H. Wu, W. Knottenbelt, and K. Wolter, “Analysis of the energy-
response time tradeoff for mobile cloud offloading using com-
of labeled data. bined metrics,” in Teletraffic Congress (ITC 27), 2015 27th Interna-
The DDTO algorithm determines whether a task should tional. IEEE, 2015, pp. 134–142.
be executed at a local device or whether it should be [8] H. Wu and K. Wolter, “Stochastic analysis of delayed mobile of-
offloaded to the clouds, and if it should be offloaded to floading in heterogeneous networks,” IEEE Transactions on Mobile
Computing, vol. 17, no. 2, pp. 461–474, 2018.
the clouds the algorithm determines, whether to offload the [9] K. Peng, M. Zhu, Y. Zhang, L. Liu, J. Zhang, V. C. Leung, and
task to the central cloud or to the edge cloud. The numerical L. Zheng, “An energy-and cost-aware computation offloading
results demonstrate the accuracy of DDTO algorithm and method for workflow applications in mobile edge computing,”
EURASIP Journal on Wireless Communications and Networking, vol.
in comparison with several previously known schemes our 2019, no. 1, p. 207, 2019.
results are significantly better. In future studies, we will [10] Y. Nan, W. Li, W. Bao, F. C. Delicato, P. F. Pires, and A. Y.
consider more factors in the hybrid offloading model to Zomaya, “A dynamic tradeoff data processing framework for
further improve the capability of our algorithm in handling delay-sensitive applications in cloud of things systems,” Journal
of Parallel and Distributed Computing, vol. 112, pp. 53–66, 2018.
realistic mobile offloading scenarios. Moreover, we will [11] Q. Wang, S. Guo, J. Liu, and Y. Yang, “Energy-efficient computa-
build a platform that can evaluate the performance of the tion offloading and resource allocation for delay-sensitive mobile
DDTO algorithm during the actual task offloading progress. edge computing,” Sustainable Computing: Informatics and Systems,
Offloading decisions in MEC and MCC are becoming vol. 21, pp. 154–164, 2019.
[12] V. Haghighi and N. S. Moayedian, “An offloading strategy in mo-
more intelligent with the emergence of innovative technolo- bile cloud computing considering energy and delay constraints,”
gies and paradigms such as fog-aided wireless networks, IEEE Access, vol. 6, pp. 11 849–11 861, 2018.
blockchain and artificial intelligence [43], [44]. To meet [13] L. Dong, M. N. Satpute, J. Shan, B. Liu, Y. Yu, and T. Yan, “Com-
more stringent requirements for security and environmental putation offloading for mobile-edge computing with multi-user,”
in 2019 IEEE 39th International Conference on Distributed Computing
adaptability, we plan to use blockchain and meta-learning Systems (ICDCS). IEEE, 2019, pp. 841–850.
techniques for intelligent offloading in the future. [14] J. Barrameda and N. Samaan, “A novel statistical cost model and
In view of the single point of failure, data privacy an algorithm for efficient application offloading to clouds,” IEEE
Transactions on Cloud Computing, vol. 6, no. 3, pp. 598–611, 2018.
and security problems faced by the current centralized IoT
[15] W. Zhang and Y. Wen, “Energy-efficient task execution for appli-
systems, we will develop a blockchain-based decentralized cation as a general topology in mobile cloud computing,” IEEE
offloading scheme, to address the challenge of data loss Transactions on cloud Computing, vol. 6, no. 3, pp. 708–719, 2018.
or privacy disclosure that may occur in the process of [16] H. Wu, W. Knottenbelt, and K. Wolter, “An efficient application
partitioning algorithm in mobile environments,” IEEE Transactions
task offloading, effectively promote data intelligence across on Parallel and Distributed Systems, vol. 30, no. 7, pp. 1464–1480,
devices, and ensure data integrity. When the environment 2019.
of the IoT system changes, such as the performance of the [17] M. R. Rahimi, N. Venkatasubramanian, S. Mehrotra, and A. V.
edge server or the bandwidth, deep learning-based methods Vasilakos, “MAPCloud: Mobile applications on an elastic and scal-
able 2-tier cloud architecture,” in Proceedings of the 2012 IEEE/ACM
have to train from scratch. To solve the problem of poor fifth international conference on utility and cloud computing. IEEE
portability, we also introduce meta-learning to ensure that Computer Society, 2012, pp. 83–90.
the offloading decision model can quickly adapt to the new [18] E. El Haber, T. M. Nguyen, D. Ebrahimi, and C. Assi, “Compu-
environment by learning the initial parameters of the neural tational cost and energy efficient task offloading in hierarchical
edge-clouds,” in 2018 IEEE 29th Annual International Symposium on
network in a different environment. Personal, Indoor and Mobile Radio Communications (PIMRC). IEEE,
2018, pp. 1–6.
[19] M. Chen and Y. Hao, “Task offloading for mobile edge computing
ACKNOWLEDGMENT in software defined ultra-dense network,” IEEE Journal on Selected
This work is partially supported by the National Key R & Areas in Communications, vol. 36, no. 3, pp. 587–597, 2018.
[20] X. Xu, Q. Liu, Y. Luo, K. Peng, X. Zhang, S. Meng, and
D Program of China (2018YFC0809800), the National Nat- L. Qi, “A computation offloading method over big data for IoT-
ural Science Foundation of China (61801325), the Natural enabled cloud-edge computing,” Future Generation Computer Sys-
Science Foundation of Tianjin City (18JCQNJC00600), and tems, vol. 95, pp. 522–533, 2019.
the Huawei Innovation Research Program (HO2018085138). [21] M. Li, Q. Wu, J. Zhu, R. Zheng, and M. Zhang, “A computing
offloading game for mobile devices and edge cloud servers,”
Wireless Communications and Mobile Computing, vol. 2018, pp. 1–10,
2018.
R EFERENCES [22] Y. Chen, N. Zhang, Y. Zhang, X. Chen, W. Wu, and X. S. Shen,
[1] K. Kumar, J. Liu, Y.-H. Lu, and B. Bhargava, “A survey of “Energy efficient dynamic offloading in mobile edge computing
computation offloading for mobile systems,” Mobile Networks and for internet of things,” IEEE Transactions on Cloud Computing, 2019.
Applications, vol. 18, no. 1, pp. 129–140, 2013. [23] H. Wu, Y. Sun, and K. Wolter, “Energy-efficient decision making
[2] R. Mahmud, R. Kotagiri, and R. Buyya, “Fog computing: A for mobile cloud offloading,” IEEE Transactions on Cloud Comput-
taxonomy, survey and future directions,” in Internet of everything. ing, 2018.
Springer, 2018, pp. 103–130. [24] F. Guo, H. Zhang, H. Ji, X. Li, and V. C. Leung, “An efficient com-
[3] B. Varghese and R. Buyya, “Next generation cloud computing: putation offloading management scheme in the densely deployed
New trends and research directions,” Future Generation Computer small cell networks with mobile edge computing,” IEEE/ACM
Systems, vol. 79, pp. 849–861, 2018. Transactions on Networking, vol. 26, no. 6, pp. 2651–2664, 2018.
[4] D. Yu, Y. Zou, J. Yu, Y. Zhang, F. Li, X. Cheng, F. Dressler, and F. C. [25] T. Q. Dinh, Q. D. La, T. Q. Quek, and H. Shin, “Learning for com-
Lau, “Implementing abstract MAC layer in dynamic networks,” putation offloading in mobile edge computing,” IEEE Transactions
IEEE Transactions on Mobile Computing, pp. 1–1, 2020. on Communications, vol. 66, no. 12, pp. 6353–6367, 2018.

2327-4662 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Waterloo. Downloaded on June 01,2020 at 12:31:41 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2020.2996784, IEEE Internet of
Things Journal
IEEE INTERNET OF THINGS JOURNAL, VOL. , NO. , 2020 12

[26] L. Huang, X. Feng, A. Feng, Y. Huang, and L. P. Qian, “Distributed Huaming Wu received the B.E. and M.S. de-
deep learning-based offloading for mobile edge computing net- grees from Harbin Institute of Technology, China
works,” Mobile Networks and Applications, pp. 1–8, 2018. in 2009 and 2011, respectively, both in electrical
[27] L. Huang, S. Bi, and Y. J. Zhang, “Deep reinforcement learning for engineering. He received the Ph.D. degree with
online computation offloading in wireless powered mobile-edge the highest honor in computer science at Freie
computing networks,” IEEE Transactions on Mobile Computing, Universität Berlin, Germany in 2015. He is cur-
2019. rently an associate professor in the Center for
[28] M. Min, L. Xiao, Y. Chen, P. Cheng, D. Wu, and W. Zhuang, Applied Mathematics, Tianjin University, China.
“Learning-based computation offloading for IoT devices with en- His research interests include model-based eval-
ergy harvesting,” IEEE Transactions on Vehicular Technology, vol. 68, uation, wireless and mobile network systems,
no. 2, pp. 1930–1941, 2019. mobile cloud computing and deep learning.
[29] Z. Ning, P. Dong, X. Wang, L. Guo, J. J. P. C. Rodrigues, X. Kong,
J. Huang, and R. Y. K. Kwok, “Deep reinforcement learning for
intelligent internet of vehicles: An energy-efficient computational
offloading scheme,” IEEE Transactions on Cognitive Communications
and Networking, vol. 5, no. 4, pp. 1060–1072, 2019.
[30] M.-A. Messous, S.-M. Senouci, H. Sedjelmaci, and S. Cherkaoui,
“A game theory based efficient computation offloading in an UAV Ziru Zhang is currently working toward the
network,” IEEE Transactions on Vehicular Technology, vol. 68, no. 5, Bachelor degree at the School of Mathematics,
pp. 4964–4974, 2019. Tianjin University. His research interests include
[31] Z. Hong, W. Chen, H. Huang, S. Guo, and Z. Zheng, “Multi- distributed deep learning and edge computing.
hop cooperative computation offloading for industrial IoT–edge–
cloud computing environments,” IEEE Transactions on Parallel and
Distributed Systems, vol. 30, no. 12, pp. 2759–2774, 2019.
[32] D. Yu, L. Ning, Y. Zou, J. Yu, X. Cheng, and F. C. Lau, “Distributed
spanner construction with physical interference: constant stretch
and linear sparseness,” IEEE/ACM Transactions on Networking,
vol. 25, no. 4, pp. 2138–2151, 2017.
[33] Q. Hua, Y. Shi, D. Yu, H. Jin, J. Yu, Z. Cai, X. Cheng, and H. Chen,
“Faster parallel core maintenance algorithms in dynamic graphs,”
IEEE Transactions on Parallel and Distributed Systems, vol. 31, no. 6,
pp. 1287–1300, 2020.
[34] M. Goudarzi, H. Wu, M. S. Palaniswami, and R. Buyya, “An
application placement technique for concurrent iot applications Chang Guan is currently working toward the
in edge and fog computing environments,” IEEE Transactions on Bachelor degree at the School of Mathematics,
Mobile Computing, 2020. Tianjin University. His research interests include
[35] D. Yu, Y. Zou, J. Yu, X. Cheng, Q.-S. Hua, H. Jin, and F. C. distributed deep learning and edge computing.
Lau, “Stable local broadcast in multihop wireless networks under
SINR,” IEEE/ACM Transactions on Networking, vol. 26, no. 3, pp.
1278–1291, 2018.
[36] Y. Mao, C. You, J. Zhang, K. Huang, and K. B. Letaief, “A survey on
mobile edge computing: The communication perspective,” IEEE
Communications Surveys & Tutorials, vol. 19, no. 4, pp. 2322–2358,
2017.
[37] F. Wang, J. Xu, X. Wang, and S. Cui, “Joint offloading and com-
puting optimization in wireless powered mobile-edge computing
systems,” IEEE Transactions on Wireless Communications, vol. 17,
no. 3, pp. 1784–1797, 2017.
[38] H. Zhang, H. Liu, J. Cheng, and V. C. Leung, “Downlink energy
efficiency of power allocation and wireless backhaul bandwidth Katinka Wolter received her PhD degree from
allocation in heterogeneous small cell networks,” IEEE transactions Technische Universität Berlin in 1999. She has
on communications, vol. 66, no. 4, pp. 1705–1716, 2017. been Assistant professor at Humboldt-University
[39] C. You, K. Huang, H. Chae, and B.-H. Kim, “Energy-efficient Berlin and lecturer at Newcastle University be-
resource allocation for mobile-edge computation offloading,” IEEE fore joining Freie Universität Berlin as a profes-
Transactions on Wireless Communications, vol. 16, no. 3, pp. 1397– sor for dependable systems in 2012. Her re-
1411, 2016. search interests are model-based evaluation and
[40] D. Horgan, J. Quan, D. Budden, G. Barth-Maron, M. Hessel, improvement of dependability, security and per-
H. Van Hasselt, and D. Silver, “Distributed prioritized experience formance of distributed systems and networks.
replay,” arXiv preprint arXiv:1803.00933, 2018.
[41] Y. Wang, M. Sheng, X. Wang, L. Wang, and J. Li, “Mobile-edge
computing: Partial computation offloading using dynamic voltage
scaling,” IEEE Transactions on Communications, vol. 64, no. 10, pp.
4268–4282, 2016.
[42] Z. Zhang, Z. Hong, W. Chen, Z. Zheng, and X. Chen, “Joint com-
putation offloading and coin loaning for blockchain-empowered
mobile-edge computing,” IEEE Internet of Things Journal, vol. 6, Minxian Xu is currently an assistant professor
no. 6, pp. 9934–9950, 2019. at Shenzhen Institutes of Advanced Technology,
[43] S. Singh, I. Chana, and R. Buyya, “STAR: SLA-aware autonomic Chinese Academy of Sciences. He received the
management of cloud resources,” IEEE Transactions on Cloud Com- BSc degree in 2012 and the MSc degree in
puting, 2020. 2015, both in software engineering from Uni-
[44] S. S. Gill, S. Tuli, M. Xu, I. Singh, K. V. Singh, D. Lindsay, S. Tuli, versity of Electronic Science and Technology of
D. Smirnova, M. Singh, U. Jain, H. Pervaiz, B. Sehgal, S. S. Kaila, China. He obtained his PhD degree from the
S. Misra, M. S. Aslanpour, H. Mehta, V. Stankovski, and P. Gar- University of Melbourne in 2019. His research
raghan, “Transformative effects of IoT, blockchain and artificial interests include resource scheduling and opti-
intelligence on cloud computing: Evolution, vision, trends and mization in cloud computing.
open challenges,” Internet of Things, vol. 8, p. 100118, 2019.

2327-4662 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Waterloo. Downloaded on June 01,2020 at 12:31:41 UTC from IEEE Xplore. Restrictions apply.