IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 70, NO.

10, OCTOBER 2021 10925

Energy-Aware Task Offloading and Resource

Allocation for Time-Sensitive Services in
Mobile Edge Computing Systems
Mingxiong Zhao , Member, IEEE, Jun-Jie Yu, Wen-Tao Li, Di Liu , Shaowen Yao , Wei Feng ,
Changyang She , Member, IEEE, and Tony Q. S. Quek , Fellow, IEEE

Abstract—Mobile Edge Computing (MEC) is a promising archi- I. INTRODUCTION

tecture to reduce the energy consumption of mobile devices and
ITH the rapid development of Internet of Things (IoT),
provide satisfactory quality-of-service to time-sensitive services.
How to jointly optimize task offloading and resource allocation
to minimize the energy consumption subject to the latency re-
W the demand for high-speed, ultra-low latency, and dy-
namically configurable computing resources at the edge of net-
quirement remains an open problem, which motivates this paper. works is extensively growing [2]. The increasingly high compu-
When the latency constraint is taken into account, the optimization
variables, including offloading ratio, transmission power, and sub-
tation loads and critical latency requirements of time-sensitive
carrier and computing resource allocation, are strongly coupled. services (e.g., virtual reality/augmented reality (AR/VR), online
To address this issue, we first decompose the original problem into games, and drones) bring significant challenges to the IoT de-
three subproblems named as offloading ratio selection, transmis- vices with limited computational capability and battery lifetime.
sion power optimization, and subcarrier and computing resource To deal with such a challenging problem, the new architecture
allocation. Then, we propose an iterative algorithm to deal with
them in a sequence. To be specific, we derive the closed-form
named as Mobile Edge Computing (MEC) has been envisioned
solution of offloading ratios, employ the equivalent parametric as a potential approach to push the cloud services down to the
convex programming to obtain the optimal power allocation policy, edge of the networks, and process the offloaded computation-
and deal with subcarrier and computing resource allocation by intensive tasks from IoT devices [3]–[5]. In addition, the devel-
the primal-dual method. Simulation results demonstrate that the opment of MEC-related applications also requires higher service
proposed algorithm can save 20%–40% energy compared with the
reference schemes, and can converge to local optimal solutions.
capacity provided by optical networks, which can further support
the elastic deployment of multi-dimensional resources in MEC
Index Terms—Mobile edge computing (MEC), resource systems [6], [7].
allocation, time-sensitive services, Internet of Things (IoT). As an effective method to liberate IoT devices from
computation-intensive workloads, MEC can also efficiently re-
duce the energy consumption of IoT devices for task processing,
Manuscript received October 13, 2020; revised January 31, 2021, March and thus has been considered as a promising architecture for
24, 2021, and June 30, 2021; accepted August 23, 2021. Date of publication the scenarios with energy-constrained IoT devices. Although
August 30, 2021; date of current version October 15, 2021. This work was MEC helps cut down the energy consumption caused by local
supported in part by the National Natural Science Foundation of China under
Grants 61801418 and 61902341, in part by Yunnan Applied Basic Research computation at IoT devices, offloading computation-intensive
Projects2019FD-129 and 202101AT070182, in part by the Open Foundation or energy-consuming tasks to MEC servers also invokes extra
of Key Laboratory in Software Engineering of Yunnan Province under Grant energy for data transmission. In addition, communication la-
2020SE316, and in part by the National Research Foundation, Singapore and
Infocomm Media Development Authority under its Future Communications tency could be the dominant component of end-to-end (E2E)
Research & Development Programme, and MOE ARF Tier 2 under Grant latency. As one of the important metrics to measure the quality
T2EP20120-0006. This paper was accepted in part at IEEE GLOBECOM of service of time-sensitive services, the E2E latency, including
2020 [1]. The review of this article was coordinated by Dr. Kaigui Bian.
(Corresponding author: Di Liu.) both communication delay and processing delay, should be
Mingxiong Zhao, Jun-Jie Yu, Wen-Tao Li, Di Liu, and Shaowen Yao carefully addressed in MEC systems.
are with the Engineering Research Center of Cyberspace, National Pilot
School of Software, Yunnan University, Kunming 650106, China, and also
with the Key Laboratory in Software Engineering of Yunnan Province, A. Related Works
Yunnan University, Kunming 650106, China (e-mail: [email protected];
[email protected]; [email protected]; [email protected]; Depending on the problem formulations, we can classify
[email protected]). the existing studies into three branches: 1) minimizing energy
Wei Feng is with the School of Communication Engineering, Hangzhou consumption subject to the latency requirement; 2) minimizing
Dianzi University, Hangzhou 310014, China (e-mail: [email protected]).
Changyang She is with the School of Electrical and Information En- E2E latency; 3) improving the trade-off between latency and
gineering, University of Sydney, Sydney, NSW 2006, Australia (e-mail: energy consumption.
[email protected]). 1) Minimizing Energy Consumption: The results in the first
Tony Q. S. Quek is with the Singapore University of Technology and Design,
Singapore 487372, Singapore (e-mail: [email protected]). branch of studies indicated that MEC can efficiently reduce the
Digital Object Identifier 10.1109/TVT.2021.3108508 energy consumption of mobile devices while ensuring different
0018-9545 © 2021 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: University of Ottawa. Downloaded on February 05,2024 at 15:43:44 UTC from IEEE Xplore. Restrictions apply.
10926 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 70, NO. 10, OCTOBER 2021

kinds of latency requirements [8]–[11]. To be specific, task first branch of studies. To further reduce energy consumption
offloading in an MEC system was formulated as a stochastic under latency constraints, there are two issues to be addressed:
optimization problem in [8], where the energy consumption for 1) Without optimal power allocation; 2) Without partial offload-
transmitting and processing the tasks was minimized subject ing mechanism.
to an average queue length requirement. Since the average 1) Without Optimal Power Allocation: As a significant com-
queue length requirement is not suitable for latency-critic tasks, ponent of energy consumption of MEC networks, the transmis-
a statistical constraint on task queue lengths was considered sion power allocation has not been investigated in the above
in [11]. Since a shorter queue length does not necessarily mean existing works. This is due to the fact that deriving the closed-
a lower latency, the authors of [9], [10], [12]–[16] analyzed the form expression of the optimal transmission power is very chal-
delay components in MEC systems, and minimized the energy lenging, especially with latency constraints. In the literature, the
consumption under the E2E latency requirement. Under the user transmission power was averaged among the associated subcar-
delay requirement, the energy consumption was minimized in a riers of users [1], or set at a fixed level [1], [12], [13], [15], [16] for
single-cell MEC system [13], or in a multi-cell MEC system [14] simplicity, or obtained resorting to the arithmetic solution [14].
by jointly optimizing resource allocation and offloading deci- Apparently, transmission power optimization has a good impact
sions. The authors in [15] investigated hierarchical multilevel on system performance, and facilitates the efficiency of energy
computation offloading among users, edge and cloud to mini- resource utilizing. For instance, it can help reduce the energy
mize the energy consumed by all users via bandwidth allocation. consumption of the network. Besides, due to the limited battery
Moreover, the energy-efficient joint offloading and wireless capacity of IoT devices, it has growing importance to design
resource allocation strategies for delay-critical applications were the power allocation strategy for these devices, in order to
proposed in [12] and [16] to minimize the weighted sum (or total) prolong the lifetime of the energy-constrained wireless nodes.
energy consumption of mobile devices, respectively. Therefore, transmission power allocation was optimized in some
2) Minimizing E2E Latency: In the second branch of studies, literature [26], [27]. However, the authors in [26] considered
the authors aimed to minimize the E2E latency, which consists of the binary offloading, and averagely allocated the subcarriers
the communication delay and the task processing delay. The au- for mobile devices. Meanwhile, the authors in [27] did not
thors investigated the latency minimization problem in a multi- investigate the latency requirement of time-sensitive services.
user time-division multiple access MEC offloading system [17], 2) Without Partial Offloading Mechanism: The first brunch
and then minimized the weighted sum of latency experienced of studies ONLY investigated binary offloading, i.e., each task is
by all mobile devices via the collaboration between cloud com- either offloaded to the MEC or processed locally. Nevertheless,
puting and edge computing [18]. Meanwhile, the authors in [19] partial offloading is more suitable for the application with more
considered a partial offloading scheme and developed a heuristic stringent latency requirements, since the tasks are processed in
algorithm to minimize the maximum delay of mobile devices via parallel at both the mobile devices and the MEC servers. As
jointly optimizing the subcarrier and power allocation, while the indicated in [27]–[29], partial offloading can achieve higher
average latency of users was minimized in [20]. Furthermore, computation capacity than binary offloading via utilizing the
the authors in [21] studied how to minimize the total delay for distributed computation resources at different nodes.
processing a series of interdependent tasks in an MEC system Motivated by the aforementioned issues, in our considered
under an energy constraint. MEC system, we devote to tackle these problems in the first
3) Improving Energy-Latency Trade-Off: In the third branch branch to further minimize the energy consumption with the
of studies, the fundamental trade-off between energy consump- consideration of the latency requirements of time-sensitive com-
tion and latency was investigated. To satisfy diverse latency putation tasks.
requirements of various IoT applications, the authors in [22]
found a trade-off between the energy consumption and latency,
and formulated the problem into a constrained multi-objective C. Novelty and Contributions
optimization problem. Meanwhile, the authors in [23] mini- Comparing to these existing works, our novelty lies in two
mized the weighted sum of the execution delay and energy con- aspects: First, we optimized transmission power allocation pol-
sumption while guaranteeing the transmission power constraint icy together with subcarrier and computing resource allocation,
of IoT devices based on partial offloading. By jointly optimizing where the subcarrier allocation in orthogonal frequency division
task offloading and resource allocation, the weighted sum of multiple access (OFDMA) systems results in more complicated
task completion time and energy consumption reduction was mixed integer nonlinear programming (MINLP) problems. Sec-
maximized [24]. Moreover, the trade-off between the energy ond, we adopted partial offloading in our framework, and hence
consumption of smart mobile devices and latency of their tasks our problem is more challenging than the problems with binary
was considered to design the energy-aware offloading scheme offloading.
in [25]. 1) Differences From Our Previous Conference Paper: With-
out consideration of optimizing the transmission power in our
conference version [1], in this paper, the transmission power
B. Motivations
of users at their associated subcarriers has been optimized as
In this paper, we aim to minimize the energy consumption another variable to further minimize the total energy consump-
under the latency requirement. Thus, our work belongs to the tion. However, the optimization problem related to transmission

Authorized licensed use limited to: University of Ottawa. Downloaded on February 05,2024 at 15:43:44 UTC from IEEE Xplore. Restrictions apply.
ZHAO et al.: ENERGY-AWARE TASK OFFLOADING AND RESOURCE ALLOCATION 10927

power is non-convex since the objective function involves the

sum-of-ratio minimization, and the corresponding constraint
makes the optimization problem even more difficult to tackle.
Thus, we resort to the equivalent parametric convex program-
ming to derive the optimal transmission power, and analyze
the corresponding computation complexity. Moreover, to fur-
ther reduce the complexity of the aforementioned algorithm
and make it favorable for practical application, we proposed
another algorithm based on equal power allocation to facilitate
the procedure of achieving the transmission power.
2) Contributions: The main contributions of this paper are
summarized as follows:
r We establish a framework to minimize the total energy
consumption under the latency constraint in MEC systems,
Fig. 1. Illustration of our system model.
where offloading ratio, transmission power, computation
capability assignment, and subcarrier allocation are jointly
optimized. execute the time-sensitive computation tasks. All nodes are
r To solve the original non-convex problem, we decompose equipped with a single antenna as shown in Fig. 1. Denote K 
it into three subproblems with respect to (w.r.t.) offload- {1, 2, . . . , K} as the set of users, and let N  {1, 2, . . . , N } be
ing ratio selection PO , transmission power optimization the index for multiple orthogonal subcarriers, each of which has
PT , subcarriers and computing resource allocation PS , bandwidth B. A task generated by the k-th user is described by a
respectively. The corresponding optimal solution of each tuple of four parameters {Rk , ck , λk , T }, where Rk indicates the
subproblem can be obtained when the solutions of the other amount of input data to be processed, ck represents the number
subproblems are given. Then, we employ the Block Coor- of CPU cycles required to process 1-bit of input data, λk ∈ [0, 1]
dinate Descent (BCD) method to obtain a local optimal is the proportion of data to be offloaded to the MEC, and T is the
solution of the original problem. required delay bound.1 We assume full granularity in data par-
r To solve each of the subproblems, we first derive the tition [31]. Thus, the amount of data processed by the MEC and
closed-form solution of the optimal offloading ratios by by the mobile device are λk Rk and (1 − λk )Rk , respectively.
solving PO . Given the offloading ratios, we derive the Since time-sensitive IoT services have rigorous requirements
optimal transmit power in PT according to the equivalent on latency [16], we assume that the latency experienced by the
parametric convex programming, which turns out to be a k-th user, denoted by tk , should not exceed the required delay
kind of “water-filling” policy. Since PS is an MINLP prob- bound T , which is shorter than the channel coherence time. All
lem, we optimize the computation capability assignment the notations used in the paper are summarized in Table I.
via the bisection method, and allocate the subcarriers by
using the primal-dual method, with which a local optimal A. Computing Resources
solution can be obtained [30]. The local CPU frequency of the k-th user is denoted by fkL . If
r We analyze the complexity of the proposed iterative algo- the k-th user offloads part of its data to the MEC, the computing
rithm and validate the convergence of the algorithm via resource allocated to the user from the MEC server is denoted
simulation. Our simulation results show that compared by fkM . Both fkL and fkM are measured by the number of CPU
with equal power allocation, our optimal power allocation cycles per second. Since the computing capability of an MEC is
can save 25% energy when the transmit power is small and limited, the computing resources allocated to all the associated
10% energy when the transmit power is large. This obser- usersshould not exceed the computing capacity, denoted by F ,
vation implies that optimizing transmit power can achieve i.e., k∈K fkM ≤ F .
a larger performance gain in a lower Signal-to-Noise-Ratio
(SNR) regime. Besides, by optimizing computing resource B. Latency
allocation, the system can save another 20% − 40% energy
1) Local Processing Delay: Given the amount of data pro-
compared with the equal computing resource allocation.
cessed at the k-th user, (1 − λk )Rk , and the computing capa-
The rest of paper is organized as follows. The system model
bility, fkL , the local processing delay can be easily obtained as
and problem formulation are introduced in Section II. In Sec-
follows
tion III, offloading and resource allocation strategy is proposed.
ck (1 − λk ) Rk
We further assume equal power allocation to optimize the trans- tLk = . (1)
mission power in Section IV. Numerical results are presented in fkL
Section V, and conclusions are drawn in Section VI. 2) Communication and Processing Delays in MEC: We as-
sume that orthogonal subcarriers are assigned to different users
II. SYSTEM MODEL AND PROBLEM FORMULATION in the OFDMA system. Thus, there is no intra-cell interference.

We consider an OFDMA-based MEC system with K users 1 Our proposed algorithm can also be applied to the scenarios that the required
and one base station (BS) integrated with an MEC server to delay bound is different for different users.
Authorized licensed use limited to: University of Ottawa. Downloaded on February 05,2024 at 15:43:44 UTC from IEEE Xplore. Restrictions apply.
10928 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 70, NO. 10, OCTOBER 2021

TABLE I
SUMMARY OF NOTATIONS

Denote xk,n as the indicator of subcarrier allocation, i.e., if the For the k-th user, the delay experienced by the task offloaded
n-th subcarrier is allocated to the k-th user, xk,n = 1. Otherwise, to the MEC is denoted by toff 2
k , which consists of two parts: the
U
xk,n = 0. Therefore, the aggregated data rate from the k-th user uplink transmission delay, tk , and the corresponding processing
to the BS can be expressed as delay at MEC server, tM k . Therefore, the communication and
processing delays in the MEC is given by
  pk,n gk,n  λk R k λk R k c k
rk = B xk,n log2 1 + , (2) toff U M
k = tk + tk = + . (3)
n∈N
σ2 rk fkM

Due to the parallel computing at users and MEC server, the

where pk,n is the transmission power that the k-th user allocates total latency for the k-th user depends on the larger one between
on the n-th subcarrier, gk,n is the channel gain between the k-th tLk and toff L off
k , and can be expressed as tk = max{tk , tk }.
user and the BS on the n-th subcarrier, σ 2 is the variance of the
additive white Gaussian noise on each subcarrier. For notational
p g
simplicity, we define g̃k,n  k,nσ2 k,n .
2 In practice, the MEC-integrated BS will provide sufficient transmission
Moreover, given the maximum transmit power of the k-
max power, while the amount of output data from MEC server to the k-th user
 user pk , the max
th power allocation policy should satisfy is usually much less than that of the input data, the time consumed and the
n∈N x k,n p k,n ≤ p k . transmission energy for delivering the computed results are negligible [32].

Authorized licensed use limited to: University of Ottawa. Downloaded on February 05,2024 at 15:43:44 UTC from IEEE Xplore. Restrictions apply.
ZHAO et al.: ENERGY-AWARE TASK OFFLOADING AND RESOURCE ALLOCATION 10929

C. Energy Consumption of MEC server. Constraints (7g) and (7h) enforce that each
subcarrier can only be used by one user to avoid interference.
The total energy consumption of the system consists of two
parts: the energy for local computing and that for transmitting
and processing the data offloaded to the MEC. III. OFFLOADING AND RESOURCE ALLOCATION STRATEGY
1) Energy Consumption for Local Computing: Given the In this section, we provide offloading and resource allocation
processor’s computing capability, fkL , the power consumption of strategy for the considered optimization problem P, which is
the processor is modeled as κk (fkL )3 (joule per second), where intractable to deal with due to the coupled variants in both the
κk represents the computation energy efficiency coefficient re- constraints and the objective function based on our observation.
lated to the processor’s chip equipped at the k-th user [31], [33], To decouple these variants, we will divide the original problem
[34]. Taking the processing delay (1) into consideration, the P into three subproblems: 1) PO , offloading ratio selection;
energy consumption for local computing is given by 2) PT , transmission power optimization; 3) PS , subcarriers and
computing resource allocation.
EkL = κk (fkL )3 tLk = κk ck (1 − λk ) Rk (fkL )2 . (4)
Firstly, with given transmission power p, computation ca-
2) Energy Consumption for Offloaded Data: For the data pability assignment f and subcarrier allocation strategy, we
offloaded to the MEC, the energy is consumed by the uplink can obtain the optimal offloading ratio λ in the outer loop.3
transmission and the data processing in the MEC server, i.e., Secondly, with the newly obtained offloading ratio λ , we can
optimize the transmission power p according to the first-order
Ekoff = EkU + EkM
Taylor expansion and the transformation of the sum of fractional
 λk R k functions, while updating its dual variable indicated by (31) and
= xk,n pk,n + κM λk ck Rk (fkM )2 , (5)
rk (32), and then renewing auxiliary variable presented in (35) and
n∈N
(36) in its inner loop. Thirdly, we can optimize (f , X) at one
where κM is the computation energy efficiency coefficient re- iteration with the newly achieved (λ , p ) in its inner loop, and
lated to the processor’s chip of the MEC server. renew the corresponding auxiliary variable according to (52).
Therefore, the total energy consumption of the k-th user is Then, with the newly optimized (f  , X  ), we further update the
given by corresponding dual variables via the subgradient method at the
next iteration in its inner loop. Finally, we will iteratively update
Ek = EkL + EkU + EkM . (6)
the derived (λ, p, f , X) in the outer loop, and the procedures
In this paper, we minimize the overall energy consumption are known as the BCD method [35], [36]. In this section, the
of the considered system by jointly optimizing offloading ratio, joint optimization on offloading ratio, transmission power, and
transmission power, and subcarriers and computing resource al- subcarriers and computing resource allocation will be proposed
location. Mathematically, the energy consumption minimization in accordance with the iterative approach based on the BCD
problem can be formulated as follows method as follows.

P : min Ek (7a) A. Offloading Ratio Selection
λ,p,f ,X
k∈K
Given transmission power p, computation capability assign-
s.t. 0 ≤ λk ≤ 1, ∀k, (7b) ment f and subcarrier allocation strategy X, the optimal offload-
max{tLk , toff ing ratio λ can be obtained by solving the following problem,
k } ≤ T, ∀k, (7c)
 
xk,n pk,n ≤ pmax , ∀k, (7d) PO : min Ek
k λ
k∈K
n∈N

fkM ≥ 0, ∀k, (7e) s.t. (7b)(7c), (8)

 which can be decoupled into K subproblems related to each
fkM ≤ F, (7f)
user, given by
k∈K
 PO 1 : min Ek (9a)
xk,n ≤ 1, ∀n, (7g) λk
k∈K
s.t. 0 ≤ λk ≤ 1, (9b)
xk,n ∈ {0, 1}, ∀k, n, (7h)
T fkL T rk fkM
where λ  {λk }, p  {pk,n }, f  {fkM } and X  1− ≤ λk ≤ , (9c)
ck Rk Rk fkM + rk Rk ck
{xk,n }, ∀k, n. Constraint (7c) guarantees the latency
requirements of time-sensitive services. Constraint (7d) where (9c) can be derived with the help of (1) and (3). Ap-
ensures that the total transmission power allocated on different parently, PO 1 is a convex problem w.r.t. λk , and the optimal
subcarriers does not exceed the maximum transmit power of
each user. Constraint (7f) indicates that the computing resources 3 The outer loop is designed for the three subproblems, i.e., P , P and P ,
O T S
allocated to users cannot exceed the total computation capability and to update the achieved variants iteratively.

Authorized licensed use limited to: University of Ottawa. Downloaded on February 05,2024 at 15:43:44 UTC from IEEE Xplore. Restrictions apply.
10930 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 70, NO. 10, OCTOBER 2021

offloading ratio λk at the k-th user can be achieved according to the growing fkM is over the threshold which makes ∂Ek
< 0 into
∂λk
the following theorem. ∂Ek T fL
≥ 0 according to (10), the k-th user will choose (1 − ck Rkk )
Theorem 1: With given (p, f , X), the optimal λk for PO 1 ∂λk
as its offloading ratio. That is because users will have additional
is given by4
⎧ energy consumption and communication delay while uploading
L
⎨ max 1 − T fk , 0 , if ∂E
∂λk ≥ 0,
k data to the MEC-integrated BS, even though MEC server has
 c k Rk
λk = M
T rk f k
(10) more efficient computation. Taking EkL and Ekoff into consid-
⎩ min , 1 , otherwise.
Rk f M +rk Rk ck
k
eration, it is more beneficial for users to compute partial data
Proof: It can be obtained resorting to the first-order condi- locally to reduce the energy consumption while satisfying the
tion and comparing with the boundary points provided by (9b) latency requirement of time-sensitive computation tasks at users,
and (9c). in despite of the assigned large enough computation capability.
Remark 1: Based on the first-order derivative, we can obtain
the optimal λk for the k-th user, ∀k. When ∂E ∂λk ≥ 0, the k-th
k B. Transmission Power Optimization
user is willing to compute its data locally with the increasing fkL , In this subsection, with the newly obtained offloading ratio
but tends to offload more data to the MEC-integrated BS when it λ , the optimal transmission power p can be obtained by solv-
has very limited computation capability, i.e., low CPU frequency ing the following problem with given computation capability
at local. When ∂E k
∂λk < 0, the k-th user prefers offloading to the assignment f and subcarrier allocation strategy X,
MEC-integrated BS if the MEC server assigns more and more 
computation resource, fkM to the user, or the user has a gradual PT : min EkU (12a)
p
increasing transmission rate rk , and vice versa. k∈K

To get more insight into the above theorem, we can obtain s.t. pk,n ≤ pmax , ∀k, (12b)
k
the following corollary based on some asymptotic assumptions, n∈Nk
and further get some meaningful conclusions.
Corollary 1: When the CPU of the k-th user is very powerful tU M
k + tk ≤ T, ∀k, (12c)
such that fkL → +∞, or MEC server assigns a large enough  λk R k
computation capability for the k-th user such that fkM → +∞, where EkU = n∈Nk pk,n  B log2 (1+pk,n g̃k,n ) , Nk repre-
n∈Nk

the optimal λk that minimizes the energy consumption in PO 1 sents the set of subcarriers allocated to the k-th user, and |Nk |
is given as represents the size of the related set. In PT , constraint (12c) can
⎧ be deduced from constraint (7c) with the newly obtained λ .
⎪ T fL
⎨ 1 − ck Rkk , if fkM → +∞, Based on the observation of PT , it is clear that PT is non-convex

λk = (11) since the objective function involves the sum-of-ratio mini-
⎪
⎩ T rk fkM
, if f L
→ +∞.
M
Rk f +rk Rk ck
k
k mization, and constraint (12c) further makes the optimization
problem even more difficult to tackle.
Proof: The first-order condition of Ek is given by
To address the above issues, we first deal with the non-convex
∂Ek  Rk constraint (12c). Based on (2) and (3), the constraint for the
= −κk ck Rk (fkL )2 + pk,n + κM ck Rk (fkM )2 .
∂λk rk latency requirement of time-sensitive computation tasks (12c)
n∈Nk
can be recast as
When the CPU frequency of the k-th user approaches to infinity,  λk R k
i.e., fkL → +∞, it is straightforward that ∂E
∂λk  0, and thus we
k
log2 (1 + pk,n g̃k,n ) ≥ , ∀k, (13)
T rk fkM n∈N
BTkU
have λk = Rk fkM +rk Rk ck
referring to (10). Similarly, we can k

obtain the optimal λk when fkM → +∞.


where TkU  T − λkfRMk ck . However, the transformed constraint
k
Remark 2: From Theorem 1, we observe that the k-th user (13) is still non-convex since the summation of concave func-
will compute data locally by offloading a part of data, i.e., tions is greater than or equal to the constant. In order to deal with
T fL
(1 − ck Rkk ) to MEC server as its CPU frequency fkL is growing the non-convexity of (13), we obtain the following inequalities
when ∂E L
∂λk ≥ 0. Notice that when the increasing fk exceeds a
k by applying the first-order Taylor expansion at the given points
prk,n , ∀k, n in the r-th iteration,
certain value and thus makes ∂E k
∂λk < 0, the k-th user prefers to
T r fM
k k

offload its data based on ( Rk f M +r k Rk c k
), rather than computes log2 (1 + pk,n g̃k,n ) ≥
k
its whole data locally even though fkL → +∞. The reason is that n∈Nk
⎧  ⎫
MEC server has more efficient computation energy coefficient
 ⎨ g̃k,n pk,n − prk,n ⎬
than the users, i.e., κM < κk , ∀k, and thus users are inclined to log2 1 + prk,n g̃k,n +  , ∀k,
offload partial data to reduce the energy consumption even if ⎩ 1 + prk,n g̃k,n ln 2 ⎭
n∈Nk
they have sufficiently large CPU frequencies. Moreover, when
(14)
4 The value of E is independent of λ when ∂Ek where the right-hand side of (14) is now a linear function w.r.t.
k k ∂λk = 0, thus we can choose
any offloading ratio as the optimal ratio. pk,n . With some mathematic manipulations, constraint (13) can

Authorized licensed use limited to: University of Ottawa. Downloaded on February 05,2024 at 15:43:44 UTC from IEEE Xplore. Restrictions apply.
ZHAO et al.: ENERGY-AWARE TASK OFFLOADING AND RESOURCE ALLOCATION 10931

be further recast as where ϕ  {ϕk }, ϑ  {ϑk }, ∀k are the non-negative Lagrange

 g̃k,n pk,n multipliers for the corresponding constraints. It can be readily
  ≥ ok , ∀k, (15) proved that PT 2 is convex for given bk and ak , ∀k, and satisfies
n∈Nk 1 + prk,n g̃k,n ln 2
Slater’s condition. Thus, strong duality holds between the primal
λk R k  g̃k,n prk,n and dual problems, which means solving PT 2 is equivalent to
where we define ok  + n∈Nk { (1+prk,n g̃k,n ) ln 2 } −
BTkU solving the dual problem. Define P as all sets of possible p that
log2 (1 + prk,n g̃k,n ). satisfy constraint (12b). Notice that the Lagrange dual function
Next, we start to tackle the non-convex objective function in is then defined as
(12), which is in the sum-of-ratios form. For notational brevity,
we denote g(ϕ, ϑ) = min L(p, ϕ, ϑ). (22)
p∈P

dk (pk ) = pk,n λk Rk , (16) Moreover, the Lagrange dual problem becomes
n∈Nk
 max g(ϕ, ϑ)
hk (pk ) = B log2 (1 + pk,n g̃k,n ), (17) ϕ,ϑ
n∈Nk
s.t. ϕ 0, ϑ 0. (23)
where pk  {pk,n }, ∀n ∈ Nk . By introducing auxiliary vari- In the following, we obtain the optimal transmission power p for
ables a, where a  {ak }, ∀k, the original problem with some given auxiliary variables (b, a) and Lagrange multipliers (ϕ, ϑ)
simple transformations can be rewritten as at first, then the Lagrange multipliers are updated via gradient

PT 1 : min ak (18a) descent method. At last, the auxiliary variables are updated based
p,a on the iterative method as well.
k∈K
1) Update pk,n : To minimize the Lagrange dual function in
s.t. (12b)(15), (22), g(ϕ, ϑ) can be decomposed into K independent subprob-
dk (pk ) ≤ ak hk (pk ), ∀k, (18b) lems. To be specific, the subproblem for the k-th user is given
by
which is still difficult to solve due to the coupled variables in
the constraints. Therefore, we employ the equivalent parametric g(ϕk , ϑk ) = min L(pk , ϕk , ϑk ), (24)
pk ∈P
convex programming described in the following lemma to cope
with PT 1. where
Lemma 1: [37, Lemma 2.1] For ∀k, n, if ({pk,n }, {ak }) is 
L(pk , ϕk , ϑk ) = bk [dk (pk )−ak hk (pk )]+ϕk pk,n
the the optimal solution of PT 1, there must exist {bk } such that
n∈Nk
{pk,n } satisfies the Karush-Kuhn-Tucker (KKT) condition of

the following problem for bk = bk and ak = ak , − ϑk 
g̃k,n pk,n
 . (25)
 n∈Nk 1 + prk,n g̃k,n ln 2
PT 2 : min bk [dk (pk ) − ak hk (pk )]
p
k∈K On the observation of (25), it can be further decomposed into
s.t. (12b)(15). (19) |Nk | subproblems w.r.t. the subcarriers allocated to the k-th user,
Furthermore, {pk,n } satisfies the following equations for bk = given as
bk and ak = ak : gn (ϕk , ϑk ) = min L(pk,n , ϕk , ϑk ), (26)
pk,n ∈P
1 dk (pk )
bk = , ak = , ∀k, (20) where
hk (pk ) hk (pk )
L(pk,n , ϕk , ϑk ) = bk [pk,n λk Rk − ak B log2 (1 + pk,n g̃k,n )]
Where {bk } is the non-negative multiplier of constraint (18b),
and we define b  {bk }, ∀k for simplicity. ϑk g̃k,n pk,n
+ ϕk pk,n −   , (27)
Proof: It can be proved by taking the derivative of the La-
1 + prk,n g̃k,n ln 2
grange function of PT 1, the details of which can be referred
to [37], and thus is omitted for brevity.  combined with (16) and (17).
Lemma 1 implies that the optimal solution of PT 1 can be Theorem 2: The optimal transmission power for the k-th user
obtained by solving the equations of (20) among the solution of at the n-th subcarrier is pk,n , where n ∈ Nk , given by
PT 2. The Lagrangian of PT 2 is ⎡ ⎤+
L(p, ϕ, ϑ) pk,n = ⎣
a k b k B
−
1 ⎦ . (28)
  ϑ g̃
(bk λk Rk + ϕk ) ln 2− kr k,n g̃k,n
   1+pk,n g̃k,n
= bk [dk (pk )−ak hk (pk )]+ ϕk pk,n −pmax
k
k∈K k∈K n∈Nk
Proof: The proof is presented in Appendix A. 
⎡ ⎤ Remark 3: To minimize the energy consumption of offloading

  g̃k,n pk,n from users to MEC, k∈K EkU , the k-th user will distribute
+ ϑk ⎣o k −   ⎦, (21) more power to its allocated subcarrier with a higher channel gain
k∈K n∈Nk 1 + prk,n g̃k,n ln 2
g̃k,n , and the assigned power is proportional to the bandwidth
Authorized licensed use limited to: University of Ottawa. Downloaded on February 05,2024 at 15:43:44 UTC from IEEE Xplore. Restrictions apply.
10932 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 70, NO. 10, OCTOBER 2021

of subcarrier, B. These can be seen from the optimal expression TABLE II

SIMULATION PARAMETERS
of pk,n . On the other hand, the k-th user is inclined to reduce
its transmission power when it plans to offload more data to the
MEC-integrated BS, in order to reduce the energy consumption
of offloading, which can be deduced from the formulation of
EkU in (12a).
Interestingly, an asymptotic solution can be obtained in a
strong channel gain scenario, presented by the following corol-
lary.
Corollary 2: When the channel gain for the k-th user at the
n-th subcarrier is very strong such that g̃k,n → +∞, the optimal
transmission power pk,n is given by
⎡ ⎤+
a k bk B
pk,n = ⎣ ⎦ , (29)
(bk λk Rk + ϕk ) ln 2 − pϑr k
k,n

where pk,nis irrelevant to the channel gain.

Proof: It can be directly deduced from (28), and thus we omit method to update auxiliary variables
it. 
τi
2) Lagrange Multipliers Update: With the achieved p , we bi+1
k = (1 − τ i )bik + , ∀k, (35)
start to update the Lagrange multiplier (ϕ, ϑ). With known p , hk (pk )
the Lagrange dual problem (23) is always convex, which can be dk (pk )
ai+1
k = (1 − τ i )aik + τ i , ∀k, (36)
decomposed into K independent subproblems w.r.t. each user, hk (pk )
and the one related to the k-th user is given as
where τ i is the largest τ that satisfies
max g(ϕk , ϑk )
ϕk ,ϑk
W bik + τ l qiK+1:2 K , aik +τ l qi1:K
s.t. ϕk ≥ 0, ϑk ≥ 0, (30)
≤ (1 − zτ l ) W (bik , aik ) , (37)
which is an affine function w.r.t dual variables. Thus, we can
apply the subgradient method for the variable update, and the where q is the Jacobian matrix of W , l ∈ {1, 2, · · · }, τl ∈ (0, 1)
dual variables (ϕk , ϑk ) can be updated according to the follow- and z ∈ (0, 1). Note that when τ i = 1, it becomes the standard
ing formulations. For example, (ϕj+1 j+1 Newton method, and the detailed process of achieving p is
k , ϑk ) can be obtained at
the (j + 1)-th iteration, respectively, as summarized in Algorithm 1.
  +
j+1 j
 C. Subcarriers and Computing Resource Allocation
ϕk = ϕk − ιk pk,n −pmax
k , (31)
n∈Nk With the newly achieved offloading ratio λ and transmission
⎧ ⎡ ⎤⎫+ power p , the subcarriers and computing resource allocation will
⎨  g̃k,n pk,n ⎬ be designed to assign computation capability of MEC server,
ϑj+1 = ϑ j
−ν k
⎣ o k −   ⎦ , (32) and allocate subcarriers for each user, to further reduce the
k ⎩ k ⎭
r
n∈Nk 1 + pk,n g̃k,n ln 2 energy consumption. In this subsection, we aim to minimize the
energy consumption via the following
 optimization problem, the
where ιk and νk are the stepsizes corresponding to the related
objective function of which is k∈K Ekoff indeed,
dual variable during iterations. Notice that all the Lagrange
variables must be non-negative, the details of which are given  λk R k
in Table II. PS : min xk,n pk.n +κM λk ck Rk (fkM )2
f ,X rk
k∈K n∈N
3) Auxiliary Variables Update: Then, we can update auxil-
iary variables (a, b) in this subsection. According to Lemma 1, s.t. (7d) − (7h)(12c). (38)
the optimal pk also satisfies the following conditions:
The main challenge in solving PS is that the considered op-
ak hk (pk ) − dk (pk ) = 0, ∀k, (33) timization problem is a mixed-integer programming and thus
bk hk (pk ) − 1 = 0, ∀k. (34) is NP-hard and non-convex, finding the optimal solution is
generally prohibitively due to the computation complexity. For-
Similarly, referring to [37], we define functions for notational tunately, the duality gap becomes zero in multi-carrier systems
brevity. Specifically, let Wn (an ) = an hk (pk )−dk (pk ) and as the number of subcarriers goes to infinity according to the
Wn+K (bn ) = bn hk (pk )−1, n ∈ {1, 2, . . ., K}. The optimal so- time-sharing condition [38], [39]. Thus, the optimal solution
lutions of (bk , ak ) can be obtained by solving W (bk , ak )  for a non-convex resource allocation problem in a multi-carrier
[W1 , W2 , . . .W2 K ] = 0. Then, we can apply the iterative system can be achieved in the dual domain.

Authorized licensed use limited to: University of Ottawa. Downloaded on February 05,2024 at 15:43:44 UTC from IEEE Xplore. Restrictions apply.
ZHAO et al.: ENERGY-AWARE TASK OFFLOADING AND RESOURCE ALLOCATION 10933

Algorithm 1: Transmission Power Algorithm. Algorithm 2: Proposed Bisection Method.

Input: Given offloading ratio λ , computation capability Input: given offloading ratio λ , transmission power p ,
assignment f  , and subcarrier allocation strategy X  . and the algorithm accuracy indicator 3 .
Output: The optimal p . 1: for k ∈ K do
1: Initialization: give iteration r = 0, and the algorithm 2: Initialize: fkM,UB = F and fkM,LB = 0.
accuracy indicators ( 1 , 2 ), where 1 and 2 are a very 3: repeat
small constant for controlling accuracy. 4: Set fkM = 21 (fkM,UB + fkM,LB ).
2: repeat (from 2 to 18) ∂L
5: Compute ∂f M according to (43).
3: With given local points prk,n . ∂L
k
6: if ∂fkM
> 0 then
4: Initialization: auxiliary variables b and a.
5: repeat (from 5 to 14) 7: set fkM,UB = fkM ;
6: Initialization: Lagrange variables ϕ and ϑ. 8: else
7: repeat (from 7 to 12) 9: set fkM,LB = fkM .
∂L
8: repeat (from 8 to 11) 10: until | ∂f M | ≤ 3.
k
9: Obtain transmission power pk,n . 11: Obtain the optimal fkM, .
10: until Lagrangian function converges.
11: Update Lagrange variables based on gradient
descent. method in (31) and (32).
12: until ϕ and ϑ converge. where α  {αk }, β  {βk }, δ  {δk }, ∀k and γ are the non-
13: update auxiliary variables from (35) and (36). negative Lagrange multipliers corresponding to the related con-
14: until W(bk , ak ) ≤ 1 . straints. Define F as all sets of possible f that satisfy constraint
15: Output the optimal pk,n . (7e), X as all sets of possible X that satisfy constraints (7g), (7h)
r+1 and (12c), and Q as all sets of possible φ that satisfy constraint
16: Update local point pk,n = pk,n .
(39b). The Lagrange dual function is then defined as
17: r = r + 1.
r+1
18: until |pk,n − prk,n | ≤ 2 . g(α, β, δ, γ) = min L(f , X, φ, α, β, δ, γ). (41)
f ∈F,X∈X ,φ∈Q

Furthermore, the Lagrange dual problem is given by

However, we cannot transform the primal domain of PS into
the dual domain directly since rk is in the denominator and max g(α, β, δ, γ)
its form in (2) makes the problem more intractable. Therefore, s.t. α 0, β 0, δ 0, γ ≥ 0. (42)
a new non-negative auxiliary variable φ  {φk }, ∀k will be
introduced to transform PS into the following problem, With the achieved offloading ratio and transmission power,
we apply the primal-dual method to obtain the optimal solutions
 λk R k for computing resource and subcarrier allocation [30].
PS 1 : min xk,n pk,n + κM λk ck Rk (fkM )2
f ,X,φ φk 1) Computation Capability Assignment: Employing the
k∈K n∈N
KKT conditions, the following condition is both necessary and
(39a)
sufficient for computation capability assignment’s optimality:
s.t. (7d) − (7h)(12c),
∂L(f , X, φ, α, β, δ, γ)
0 ≤ φk ≤ rk , ∀k, (39b) ∂fkM

where we can update φ to help minimize the energy consump- αk c k R k λk

= 2fkM κM λk Rk ck − + γ = 0. (43)
tion. The Lagrangian for the above problem with the achieved (fkM )2
λ and p can be written as
However, it is difficult to find a closed-form expression for
the optimal solution, fkM, . Fortunately, we can resort to the
L(f , X, φ, α, β, δ, γ)
following proposition to obtain fkM, .
 λk R k
= xk,n pk,n + κM λk ck Rk (fkM )2 Proposition 1: L is a convex function of fkM .
k∈K n∈N
φk Proof: The proof is given in Appendix B. 
∂L
Since L is a convex function of fkM , and ∂f
  λk R k λk R k c k
  M increases mono-
k
+ αk + − T + δk (φk − rk ) tonically along with fkM , we can adopt the bisection method
φk fkM
k∈K k∈K to obtain fkM, within 0 ≤ fkM ≤ F . The detailed process of
   
   achieving fkM, is summarized in Algorithm 2.
+ βk xk,n pk,n − pmaxk +γ fkM − F , 2) Subcarrier Allocation Strategy: When the optimal com-
k∈K n∈N k∈K puting capability assignment is achieved, the optimal subcarrier
(40) allocation can be obtained through the following procedures.

Authorized licensed use limited to: University of Ottawa. Downloaded on February 05,2024 at 15:43:44 UTC from IEEE Xplore. Restrictions apply.
10934 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 70, NO. 10, OCTOBER 2021

With some mathematic manipulations, we can rewrite (40) as

Algorithm 3: Proposed Algorithm.
L(X, φ, α, β, δ, γ) 1: Initialization: given {p, f , X, φ, α, β, δ, γ}, we set
  z = 0, and denote zmax as the maximum number of
= xk,n Lk,n + γfkM + δk φk − βk pmax
k iterations.
k∈Kn∈N k∈K
  2: repeat (from 2 to 14)
 λk R k λk R k c k 3: Determine offloading ratio λ according to (10).
+ αk + −T +κM λk ck Rk (fkM )2 −γF,
φk fkM 4: Determine transmission power p by Algorithm 1.
k∈K
(44) 5: repeat (from 5 to 12)
6: repeat (from 6 to 10)
where 7: Allocate computing resource f by Algorithm 2.
pk,n λk Rk 8: Determine subcarrier allocation X according to
Lk,n = +pk,n βk −Bδk log2 (1+pk,n g̃k,n ) . (45)
φk (49).
On the observation of (44), we further suppose that the n-th 9: Update φ .
subcarrier is assigned to the k-th user, and thus have the La- 10: until Lagrangian function converges.
grangian for the k-th user, 11: Update α, β, δ and γ.
 12: until α, β, δ, γ converge.
Lk = Lk,n + ωk − γF, (46) 13: z = z + 1.
n∈Nk 14: until the difference successive values of the objective
where function is less than 3 or z > zmax .
ωk = γfkM + δk φk − βk pmax
k + κM λk ck Rk (fkM )2
 
λk R k λk R k c k Theorem 3: Given the optimal computation capability assign-
+ αk + − T . (47) ment f  , and the optimal subcarrier allocation strategy X  , the
φk fkM
optimal φk is given by
Thus, the subproblem is given by ⎧
⎨ φk,1 , if φok < φk,1 ,
min Lk,n (xn , φ, β, δ) (48) 
φk = φok , if φk,1 ≤ φok ≤ r̃k , (52)
xn ∈X
⎩
r̃k , otherwise,
where xn  {xk,n }, ∀k, and this problem can be solved in-
dependently. To minimize each Lk,n , the optimal xn can be where φk,1 and φok are given by
obtained as λk Rk fkM
 φk,1 = , (53)
1, if k = k  = argmink Lk,n , M
T fk − λk R k c k
xk,n = (49)
0, otherwise.  

 
3) Auxiliary Variable Update: In this part, with the newly o
φk =  αk + pk,n λ̃k , (54)
obtained {f  , X  } in the above subsections, we now try to n∈Nk
find the optimal φ , which can be solved via the following
optimization problem, where λ̃k = λkδR k
.
  k
Proof: The proof is presented in Appendix C. 
  pk,n λk Rk αk λk Rk
PS 2 : min + +δk φk (50a) 4) Lagrange Multipliers Update: In this subsection, since
φ
k∈K
φk
n∈Nk
φk f  , X  and φ are obtained, we can deal with the dual problem
(42), which is a convex function, by updating α, β, δ and γ
s.t. (12c), using the subgradient method.
0 ≤ φk ≤ r̃k , ∀k, (50b) According to III-A, III-B and III-C, the details to solve P are
 summarized in Algorithm 3 as follows.
where r̃k  n∈Nk B log2 (1 + pk,n g̃k,n ). It is obvious that
PS 2 can be further decoupled into K subproblems w.r.t. users, D. Complexity Analysis
given by
  In this part, the complexity of our proposed algorithm to solve
λk R k  α k λk R k P is discussed. We decompose the original problem into three
PS 3 : min pk,n + + δk φ k (51a)
φk φk φk subproblems: PO , PT and PS , then solve them iteratively, and
n∈Nk
the maximum number of iterations is zmax .
λk R k λk R k c k Firstly, we compute the offloading ratio λ for PO , and the
s.t. + ≤ T, (51b)
φk fkM complexity of solving PO is O(K), a linear complexity with
the number of users K.
0 ≤ φk ≤ r̃k , (51c)
Secondly, we have a three-layer iteration to solve PT : 1)
where the optimal auxiliary variable φk , ∀k can be obtained by The first-layer (innermost) iteration includes p and Lagrangian
the following theorem. variables (ϕ, ϑ), where the computing complexity to solve p

Authorized licensed use limited to: University of Ottawa. Downloaded on February 05,2024 at 15:43:44 UTC from IEEE Xplore. Restrictions apply.
ZHAO et al.: ENERGY-AWARE TASK OFFLOADING AND RESOURCE ALLOCATION 10935

according to (28) is O(KN ) and the updating complexity of La- user will transmit its data to the MEC-integrated BS using the
grangian variables is O(K 2 ). Hence, the calculation complexity same transmission power at each subcarrier belonging to it. With
of this iterative process is O(K 2 + KN ). 2) The second-layer some simple transformations, (56) can be rewritten into a linear
iteration is the updating iteration of auxiliary variables (a, b), constraint, given by
which has the complexity independent of K, thus the complexity
of this layer is O(K). 3) The third-layer (outermost) iteration is p̄k ≥ 2|N̄k | , ∀k, (57)
related to the fixed point prk,n , and the computation complexity where p̄k is the equal power allocation for subcarriers assigned
of this layer is O(KN ). Therefore, the complexity of solving to the k-th user, and |N̄k | is given by
PT is O(K 3 N 2 + K 4 N ). Since the number of subcarriers N is  
greater than the number of users K, the complexity is O(K 3 N 2 ) 1 λk R k 
for PT . |N̄k | = − log2 (g̃k,n ) . (58)
|Nk | BTkU
Thirdly, we have a two-layer iteration to compute X and f for n∈N k

PS . In the inner iteration, the method of solving f is dichotomy, Then, we can reformulate PT with the latest constraint (57)
and thus the computation complexity is O(K log2 (F )). Mean- as follows,
while, we get the subcarrier allocation strategy according to (49), 
the computation complexity of which is O(KN ), and the com- PET : min ĒkU (59a)
p̄
putation complexity related with φ is O(K). Then, we optimize k∈K

(f , X, φ) iteratively until Lagrange function converges, and the pmax

computation complexity is O(K). In the outer iteration, the com- s.t. 2|N̄k | ≤ p̄k ≤ k
, ∀k, (59b)
|Nk |
plexity of updating Lagrangian variables (α, β, δ, γ) is O(K 2 ).
Therefore, the complexity of solving PS is O(K 4 log2 (F ) + where p̄ = {p̄k }, ∀k and
K 4 N + K 4 ), which can be simplified as O(K 4 N ). |Nk |p̄k λk Rk
Finally, the computation complexity to solve the above ĒkU =  . (60)
n∈Nk B log2 (1 + p̄k g̃k,n )
three subproblems iteratively is O(zmax ). In summary, our
proposed algorithm PA has a total computation complexity According to the procedures described in III-B, some mod-
O(Kzmax + K 3 N 2 zmax + K 4 N zmax ), which can be abbrevi- erate changes should be made to adapt to the current situation
ated as O(K 3 N 2 zmax ). before directly applied. Similar to (18)-(20), the Lagrangian is
given as
IV. EQUAL-POWER-ALLOCATION BASED ALGORITHM 
L(p̄, ϕ, ϑ) = bk [dk (p̄k ) − ak hk (p̄k )]
Instead of utilizing the first-order Taylor expansion to cope k∈K
with the non-convex constraint (13) of III-B, which further
     
pmax
makes the complexity of the algorithm in III-B unfavorable for + ϕk p̄k − k + ϑk 2|N̄k | − p̄k , (61)
practical application with the increase of K and N , another |Nk |
k∈K k∈K
algorithm is proposed to reduce complexity based on the equal
power allocation (EPA) in this section. where dk (p̄k ) and hk (p̄k ) are given as follows, respectively,
Firstly, we make the following approximation at the high SNR dk (p̄k ) = |Nk |p̄k λk Rk , (62)
regime at the right-hand side of (13): 
  hk (p̄k ) = B log2 (1 + p̄k g̃k,n ) . (63)
log2 (pk,n g̃k,n ) ≈ log2 (1 + pk,n g̃k,n ), (55) n∈Nk
n∈Nk n∈Nk
Referring to Theorem 2, the optimal transmission power for
where the high SNR approximation made by (55) is commonly those subcarriers assigned to the k-th user could be obtained
used to simplify the derivations and get more insight of the through solving the following equation
problem [40]. The difference between the actual value and the
approximate value becomes negligible as the SNR increases, ∂L(p̄, ϕ, ϑ)
= bk |Nk |λk Rk + ϕk − ϑk
and hence the constraint (13) can be recast as ∂ p̄k
 λk R k  ak bk Bg̃k,n
log2 (pk,n g̃k,n ) ≥ , ∀k, (56) − = 0, (64)
n∈N
BTkU (1 + p̄k g̃k,n ) ln 2
n∈Nk
k

where
 the left-hand side of (56) serves as a tight lower-bound which can be solved by Matlab. Since the closed-form solution
for n∈Nk log2 (1 + pk,n g̃k,n ) at the high SNR regime, and of p̄k is difficult to obtain, we have the following propositions
the approximation is also accurate even at moderate-low SNR to achieve the closed-form solution of p̄k , ∀k based on different
regime [41]. However, constraint (56) is still non-convex w.r.t. assumptions: 1) equivalent channel gains; 2) high SNR approxi-
pk,n , ∀k, n. In order to further deal with constraint (56) and make mation, to get further knowledge about the relationship between
the optimization problem tractable, we assume that the k-th user the transmission power and other variables.
will equivalently distribute its transmission power pk among its Proposition 2: If the channel gains g̃k,n , n ∈ Nk are equiva-
allocated subcarrier n, where n ∈ Nk . That is to say, the k-th lent for all the subcarriers allocated to the k-th user, ∀k, redefined

Authorized licensed use limited to: University of Ottawa. Downloaded on February 05,2024 at 15:43:44 UTC from IEEE Xplore. Restrictions apply.
10936 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 70, NO. 10, OCTOBER 2021

as g̃k , the optimal transmission power p̄k can be directly deduced

from (64), given by
 +
 ak bk B|Nk | 1
p̄k = − . (65)
(bk λk Rk |Nk | + ϕk − ϑk ) ln 2 g̃k

Remark 4: It is straightforward to know that the k-th user

will assign more power p̄k to offload data with a better channel
gain, and also allocate power according to the assigned number
of subcarriers |Nk |, which is different from Theorem 2. These
can be seen from the optimal formulation of p̄k in (65).
Proposition 3: In the high SNR regime, the optimal transmis-
sion power p̄k can be directly deduced from (64) by assuming
(1 + p̄k g̃k,n ) ≈ p̄k g̃k,n , given by
 +
ak bk B|Nk |
p̄k = . (66)
(bk λk Rk |Nk | + ϕk − ϑk ) ln 2
Fig. 2. The comparisons of the total energy consumption versus different
numbers of users where N = 128 for different algorithms.
Remark 5: The high SNR approximation is commonly used
in the literature [40], [41], as mentioned above. Different from
Theorem 2 and Proposition 2, the transmission power p̄k has
no relationship with the channel gain under the high SNR
hypothesis similar to Corollary 2.

V. SIMULATION AND NUMERICAL RESULTS

This section presents the numerical results to demonstrate the
better performance of our Proposed Algorithm (PA) compared
with the reference schemes,5 named as
r Local computation algorithm (LC), where all user de-
vices process their tasks by local CPU.
r Equal power allocation (EPA), where the user equiva-
lently distributes its transmission power among its allo-
cated subcarriers.
r Equal computing resource allocation (ECRA), where
MEC computation resource is equivalently assigned to its
accessed users.
In the simulations, the path loss model is Rayleigh distributed Fig. 3. The total energy consumption versus the latency requirements of time-
and denoted by |β|d−2 , where |β| and d represent the short-term sensitive computation tasks of users for different algorithms where K = 5 and
N = 128.
channel fading and the distance between two nodes, respectively.
User devices have the same maximum transmit power, which are
evenly and independently distributed in a circular area around
the MEC server with a radius of 50 meters. Moreover, we set and ECRA schemes, respectively. Besides, it can help save 40%-
(κk , κM ) as (10−24 , 10−26 ) referring to [27], [33], [42], and 70% energy compared with LC.
other parameters employed in the simulations are summarized In Fig. 3, we study the influence of the latency requirements of
in Table II, unless otherwise mentioned. time-sensitive computation tasks at users on the total energy con-
In Fig. 2, for the proposed and reference schemes, we plot the sumption for different algorithms where K = 5 and N = 128.
total energy consumption w.r.t. different numbers of users where On the observation of Fig. 3, the total energy consumption is
N = 128 for different algorithms. As the number of users grows, reducing along with the gradually decreasing latency require-
the total energy consumption of different algorithms increases. ment of time-sensitive computation tasks. This is because users
The results show that PA can achieve superior performance than can ask for more help from MEC server via uploading more
the other schemes. Specifically, PA can reduce around 5% − data since the latency requirement is not so strict, and users
30% and 10% − 50% energy consumption compared with EPA can save the energy on processing data. Moreover, the reducing
extent of energy consumption w.r.t. the latency will gradually
decrease when the latency requirement approaches to a certain
5 Since there is no similar scenario in the literature, which investigated the
threshold, that is to say, the growing latency has a diminishing
energy-aware offloading in time-sensitive MEC network based on OFDMA, we influence on the total energy consumption. Therefore, the results
propose some reference schemes for comparison herein. in Fig. 3 reveal the trade-off between the energy consumption

Authorized licensed use limited to: University of Ottawa. Downloaded on February 05,2024 at 15:43:44 UTC from IEEE Xplore. Restrictions apply.
ZHAO et al.: ENERGY-AWARE TASK OFFLOADING AND RESOURCE ALLOCATION 10937

Fig. 5. The comparisons of the total energy consumption versus different CPU
Fig. 4. The total energy consumption versus transmission power of users for
frequencies of MEC server for different algorithms (and the average offloading
different algorithms (and the average offloading ratio for PA) where K = 5 and
ratio for PA) where K = 5 and N = 128.
N = 128.

and the latency. In addition, the results show that PA can save
20% − 40% energy compared with EPA and ECRA with diverse
latency requirements.
In Fig. 4, we investigate the relation between transmission
power of users and the total energy consumption, where K = 5
and N = 128. The results show that the total energy consump-
tion decreases as the transmission power of users increases.
The reason is that the uplink transmission delay tU k decreases
with the transmit power. With higher transmit power, users
can offload more data to the MEC within the same period of
time. As a result, the energy consumption for processing the
tasks can be reduced remarkably via more efficient computation
capability at the MEC server. In addition, the results demonstrate
that the performance gain achieved by PA decreases with the
Fig. 6. The comparisons of the total energy consumption versus different
transmit power. When the transmit power is less than 400 mW, numbers of subcarriers for different algorithms where K = 5.
PA can save 25% and 40% energy compared with EPA and
ECRA, respectively. The performance gains reduce to 10% and
20%, when the transmit power is greater than 400 mW. This increasing number of subcarriers. The reason is that the user
observation implies that PA can achieve a high performance will have a better chance to select the subcarrier with preferable
gain in the low SNR regime. channel gain to get the reduction of transmission power. Alterna-
In Fig. 5, we show the total energy consumption versus dif- tively, the user can offload more data with the same amount of
ferent CPU frequencies of MEC for different algorithms, where transmission power, and thus reduce the energy consumption
K = 5 and N = 128. The results show that the total energy with the aid of MEC server. The results also show that PA
consumption decreases with the CPU frequency of the MEC can save 10% − 25% (or 20% − 45%) energy consumption
server. The reason is when the computation capability of MEC compared with EPA (or ECRA).
is strengthened, users are more inclined to offload data to save In Fig. 7, we present the performances of the total energy
energy, instead of computing locally according to Theorem 1. consumption versus different CPU frequencies of users for
However, the total energy consumption of ECRA will increase K = 5. It is noted that the total energy consumption increases
when the CPU frequency is higher than 9 GHz. This is because with the CPU frequencies of users. The reason is that users
the computation resource assigned to the users is more than prefer to compute data locally along with stronger computation
enough. Comparing PA with EPA (or ECRA), we can see that capabilities according to Remark 1, and thus the proportion of
the percentage of energy saving increases from 5% to 15% (or data offloaded to the MEC server is decreasing, which is more
15% to 60%) as the CPU frequency grows. Additionally, it is energy-consuming. What is noteworthy is that the increase of
observed that offloading ratio increases with the growth of CPU the total energy consumption has been slowing or saturated
frequency at MEC due to its highly efficient computation. when the CPU frequencies of users grow in the regime of fkL ≥
In Fig. 6, we evaluate the effect of the number of subcarriers 0.5GHz, ∀k. This is because users are energy hogs compared
on the total energy consumption, where K = 5. It can be seen with MEC server as κk > κM , and users are willing to offload
that the total energy consumption is reducing along with the more data to reduce their energy consumption even if they have

Authorized licensed use limited to: University of Ottawa. Downloaded on February 05,2024 at 15:43:44 UTC from IEEE Xplore. Restrictions apply.
10938 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 70, NO. 10, OCTOBER 2021

cope with, we apply the equivalent parametric convex program-

ming to optimize the transmission power of user on its assigned
subcarriers with the help of auxiliary variables. Moreover, as
PS is a challenging non-convex MINLP problem with cou-
pled parameters, we distribute computation capability of MEC
server, and allocate subcarriers to each user in an alternating
manner in the dual domain. Simulation results demonstrate that
the proposed algorithm can effectively reduce the total energy
consumption of computation tasks, and outperform the reference
schemes with a better performance.
Although our iterative algorithm can obtain the local optimal
solution of the non-convex optimization problem, the com-
putational complexity remains as the bottleneck for real-time
implementation. To address this issue, one promising future
direction is to integrate the theoretical knowledge into deep
Fig. 7. The comparisons of the total energy consumption versus different CPU
frequencies of users where N = 128.
learning [43]. Given the optimization algorithm in our work,
the optimal solutions can serve as the labeled training samples
in supervised deep learning. For other problems without labeled
samples, unsupervised deep learning can be applied to find the
optimal policy. If the problem turns out to be a sequential-
decision making problem, we can model it as a Markov decision
process and solve it by deep reinforcement learning.

APPENDIX A
PROOF OF THEOREM 2
Taking the derivative of the Lagrangian L(pk,n , ϕk , ϑk ) w.r.t.
pk,n yields

∂L(pk,n , ϕk , ϑk )
∂pk,n
ak bk Bg̃k,n ϑk g̃k,n
Fig. 8. The comparisons of the total energy consumption versus number of
= bk λ k R k − +ϕk −   .
(1 + pk,n g̃k,n ) ln 2 1 + prk,n g̃k,n ln 2
iteration for PA with different numbers of users and subcarriers.
(A.1)
growing computation capabilities, which has been explained in ∂L(p ,ϕ ,ϑ )
Let k,n
∂pk,n
k k
= 0, the optimal pk,n can be obtained, and
details in Remark 2, where PA can reduce around 10% − 30%
thus we have Theorem 1.
and 30% − 40% energy consumption compared with EPA and
ECRA, respectively.
Fig. 8 is presented to demonstrate the convergence result of APPENDIX B
PA, where we illustrate the total energy consumption versus PROOF OF PROPOSITION 1
iteration times. Since the energy consumption remains constant The Hessian of L in (40) w.r.t. fkM is given by
after a few number of iterations, the total energy consumption
converges to a local optimal solution. ∂ 2 L(f , X, φ, α, β, δ, γ) α k λk R k c k
= 2κM λk Rk ck + ,
2
fkM
3
∂fk,m
VI. CONCLUSIONS AND FUTURE DIRECTIONS (B.1)
which is non-negative w.r.t. fkM , thus L is a convex function of
In the paper, we investigate the OFDMA-based MEC network
fkM [30].
with computation capability, and design PA method to minimize
the total energy consumption with the consideration of latency
requirements of time-sensitive computation tasks for users. Due APPENDIX C
to the coupled optimization variables, we decompose the for- PROOF OF THEOREM 3
mulated minimization problem into three subproblems named We can get the following inequality about φk from (51b) and
as PO , PT and PS , and optimize them sequentially instead of (51c)
solving the original problem directly. For PO , we can derive the
closed-form solution of offloading ratios. Since the objective λk Rk fkM
≤ φ ≤ r̃k , (C.1)
function of PT is in the sum-of-ratios form and thus difficult to M
T fk − λk R k c k

Authorized licensed use limited to: University of Ottawa. Downloaded on February 05,2024 at 15:43:44 UTC from IEEE Xplore. Restrictions apply.
ZHAO et al.: ENERGY-AWARE TASK OFFLOADING AND RESOURCE ALLOCATION 10939


where r̃k  n∈Nk B log2 (1 + pk,n g̃k,n ). Thus, we can rewrite [11] C.-F. Liu, M. Bennis, M. Debbah, and H. V. Poor, “Dynamic task offloading
and resource allocation for ultra-reliable low-latency edge computing,”
PR 3 as IEEE Trans. Commun., vol. 67, no. 6, pp. 4132–4150, Jun. 2019.
[12] C. You, K. Huang, H. Chae, and B.-H. Kim, “Energy-efficient resource
PR 3A : min T (φk ) allocation for mobile-edge computation offloading,” IEEE Trans. Wireless
φk
Commun., vol. 16, no. 3, pp. 1397–1411, Mar. 2017.
s.t. (C.1), (C.2) [13] W. Wen, Y. Fu, T. Q. Quek, F.-C. Zheng, and S. Jin, “Joint uplink/downlink
sub-channel, bit and time allocation for multi-access edge computing,”
 IEEE Commun. Lett., vol. 23, no. 10, pp. 1811–1815, Oct. 2019.
where T (φk ) = λkφRk k n∈Nk pk,n + αk φλkkRk + δk φk , and [14] A. Khalili, S. Zarandi, and M. Rasti, “Joint resource allocation and offload-
PR 3A is convex since the second derivative of T (φk ) with ing decision in mobile edge computing,” IEEE Commun. Lett., vol. 23,
no. 4, pp. 684–687, Apr. 2019.
respect to φk is positive. Therefore, resorting to the first-order [15] T. Mahn, H. Al-Shatri, and A. Klein, “Distributed algorithm for energy
condition efficient joint cloud and edge computing with splittable tasks,” in Proc.
 IEEE Wireless Commun. Netw. Conf., 2019, pp. 1–6.
∂T (φk ) λk Rk ( n∈Nk pk,n + αk ) [16] X. Yang, X. Yu, H. Huang, and H. Zhu, “Energy efficiency based joint
= δk − , (C.3)
∂φk φ2k computation offloading and resource allocation in multi-access MEC
systems,” IEEE Access, vol. 7, pp. 117054–117062, 2019.
∂T (φk ) [17] J. Ren, G. Yu, Y. Cai, and Y. He, “Latency optimization for resource
and letting ∂φk = 0, the optimal φok is given as allocation in mobile-edge computation offloading,” IEEE Trans. Wireless
 
Commun., vol. 17, no. 8, pp. 5506–5519, Aug. 2018.
 [18] J. Ren, G. Yu, Y. He, and G. Y. Li, “Collaborative cloud and edge computing
  λk R k
φ k =  αk +
o
pk,n . (C.4) for latency minimization,” IEEE Trans. Veh. Technol., vol. 68, no. 5,
pp. 5031–5044, May 2019.
δk
n∈Nk [19] M. Li, S. Yang, Z. Zhang, J. Ren, and G. Yu, “Joint subcarrier and power
allocation for OFDMA based mobile edge computing system,” in Proc.
Therefore, we can discuss the relationship between the bound- IEEE 28th Annu. Int. Symp. Pers., Indoor, Mobile Radio Commun., 2017,
aries of (C.1) and φok to get the optimal φk , and thus obtain pp. 1–6.
[20] W. Feng, H. Liu, Y. Yao, D. Cao, and M. Zhao, “Latency-aware offloading
Theorem 3. for mobile edge computing networks,” IEEE Commun. Lett., vol. 25, no. 8,
pp. 2673–2677, Aug. 2021.
[21] Y. Zhu, Y. Hu, and A. Schmeink, “Delay minimization offloading for
ACKNOWLEDGMENT interdependent tasks in energy-aware cooperative MEC networks,” in
Any opinions, findings and conclusions or recommendations Proc. IEEE Wireless Commun. Netw. Conf., 2019, pp. 1–6.
[22] L. Cui et al., “Joint optimization of energy consumption and latency in
expressed in this material are those of the author(s) and do not mobile edge computing for Internet of Things,” IEEE Internet Things J.,
reflect the views of National Research Foundation, Singapore vol. 6, no. 3, pp. 4791–4803, Jun. 2019.
and Infocomm Media Development Authority. [23] Z. Kuang, L. Li, J. Gao, L. Zhao, and A. Liu, “Partial offloading scheduling
and power allocation for mobile edge computing systems,” IEEE Internet
Things J., vol. 6, no. 4, pp. 6774–6785, Aug. 2019.
REFERENCES [24] T. X. Tran and D. Pompili, “Joint task offloading and resource allocation
for multi-server mobile-edge computing networks,” IEEE Trans. Veh.
[1] J. Yu et al., “Energy minimization for mobile edge computing networks Technol., vol. 68, pp. 856–868, Jan. 2019.
with time-sensitive constraints,” in Proc. IEEE Glob. Commun. Conf., [25] J. Zhang et al., “Energy-latency tradeoff for energy-aware offloading in
2020, pp. 1–6. mobile edge computing networks,” IEEE Internet Things J., vol. 5, no. 4,
[2] S. Mu, Z. Zhong, D. Zhao, and M. Ni, “Joint job partitioning and col- pp. 2633–2645, Aug. 2018.
laborative computation offloading for Internet of Things,” IEEE Internet [26] J. Zhang et al., “Joint offloading and resource allocation optimization
Things J., vol. 6, no. 1, pp. 1046–1059, Feb. 2019. for mobile edge computing,” in Proc. IEEE Glob. Commun. Conf., 2017,
[3] P. Porambage, J. Okwuibe, M. Liyanage, M. Ylianttila, and T. Taleb, “Sur- pp. 1–6.
vey on multi-access edge computing for Internet of Things realization,” [27] Y. Wu, Y. Wang, F. Zhou, and R. Q. Hu, “Computation efficiency max-
IEEE Commun. Surv. Tut., vol. 20, no. 4, pp. 2961–2991, Oct.-Dec. 2018. imization in OFDMA-Based mobile edge computing networks,” IEEE
[4] S.-R. Yang, Y.-J. Tseng, C.-C. Huang, and W.-C. Lin, “Multi-access edge Commun. Lett., vol. 24, no. 1, pp. 159–163, Jan. 2020.
computing enhanced video streaming: Proof-of-concept implementation [28] F. Zhou, Y. Wu, R. Q. Hu, and Y. Qian, “Computation rate maximization in
and Prediction/QoE models,” IEEE Trans. Veh. Technol., vol. 68, no. 2, UAV-Enabled wireless-powered mobile-edge computing systems,” IEEE
pp. 1888–1902, Feb. 2019. J. Sel. Areas Commun., vol. 36, no. 9, pp. 1927–1941, Sep. 2018.
[5] M. Zhao, W. Li, L. Bao, J. Luo, Z. He, and D. Liu, “Fairness-aware [29] X. Cao, F. Wang, J. Xu, R. Zhang, and S. Cui, “Joint computation and
task scheduling and resource allocation in UAV-enabled mobile edge communication cooperation for energy-efficient mobile edge computing,”
computing networks,” IEEE Trans. Green Commun. Netw., to be published, IEEE Internet Things J., vol. 6, no. 3, pp. 4188–4200, Jun. 2019.
doi: 10.1109/TGCN.2021.3095070. [30] S. Boyd, S. P. Boyd, and L. Vandenberghe, Convex Optimization. Cam-
[6] H. Yang, A. Yu, X. Zhao, Q. Yao, J. Zhang, and Y. Lee, “Multi-dimensional bridge, U.K.: Cambridge Univ. Press, 2004.
resources allocation based on reconfigurable radio-wavelength selec- [31] Y. Wang, S. Min, X. Wang, W. Liang, and J. Li, “Mobile-edge computing:
tive switch in cloud radio over fiber networks,” Opt. Exp., vol. 26, Partial computation offloading using dynamic voltage scaling,” IEEE
pp. 34719–34733, 2018. Trans. Commun., vol. 64, no. 10, pp. 4268–4282, Oct. 2016.
[7] H. Yang et al., “Bandwidth compression protection against collapse [32] X. Hu, K. K. Wong, and K. Yang, “Wireless powered cooperation-assisted
in fog-based wireless and optical networks,” IEEE Access, vol. 6, mobile edge computing,” IEEE Trans. Wireless Commun., vol. 17, no. 4,
pp. 54760–54768, 2018. pp. 2375–2388, Apr. 2018.
[8] Y. Chen, N. Zhang, Y. Zhang, X. Chen, W. Wu, and X. Shen, “Energy [33] W. Zhang, Y. Wen, K. Guan, K. Dan, and D. O. Wu, “Energy-optimal
efficient dynamic offloading in mobile edge computing for Internet of mobile cloud computing under stochastic wireless channel,” IEEE Trans.
Things,” IEEE Trans. Cloud Comput., vol. 9, no. 3, pp. 1050–1060, 2021. Wireless Commun., vol. 12, no. 9, pp. 4569–4581, Sep. 2013.
[9] K. Cheng, Y. Teng, W. Sun, A. Liu, and X. Wang, “Energy-efficient joint [34] S. Bi, J. Ying, and Zhang, “Computation rate maximization for wireless
offloading and wireless resource allocation strategy in Multi-MEC server powered mobile-edge computing with binary computation offloading,”
systems,” in Proc. IEEE Int. Conf. Commun., 2018, pp. 1–6. IEEE Trans. Wireless Commun., vol. 17, no. 6, pp. 4177–4190, Jun. 2018.
[10] Y. Dai, D. Xu, S. Maharjan, and Y. Zhang, “Joint computation offloading [35] P. Richtárik and M. Takáč, “Iteration complexity of randomized block-
and user association in multi-task mobile edge computing,” IEEE Trans. coordinate descent methods for minimizing a composite function,” Math.
Veh. Technol., vol. 67, no. 12, pp. 12 313–12 325, Dec. 2018. Prog., vol. 144, no. 1-2, pp. 1–38, 2014.

Authorized licensed use limited to: University of Ottawa. Downloaded on February 05,2024 at 15:43:44 UTC from IEEE Xplore. Restrictions apply.
10940 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 70, NO. 10, OCTOBER 2021

[36] M. Zhao, J. Y. Ryu, J. Lee, T. Q. Quek, and S. Feng, “Exploiting trust degree Shaowen Yao received the B.S. and M.S. degrees in
for multiple-antenna user cooperation,” IEEE Trans. Wireless Commun., telecommunication engineering from Yunnan Univer-
vol. 16, no. 8, pp. 4908–4923, Aug. 2017. sity, Kunming, China, in 1988 and 1991, respectively,
[37] Y. Jong, “An efficient global optimization algorithm for nonlinear sum- and the Ph.D. degree in computer application tech-
of-ratios problem,” Optimization Online, 2012. nology from the University of Electronic Science and
[38] K. Seong, M. Mohseni, and J. M. Cioffi, “Optimal resource allocation for Technology of China, Chengdu, China, in 2002. He
OFDMA downlink systems,” in Proc. IEEE Int. Symp. Inf. Theory, 2006, is currently a Professor with the School of Software,
pp. 1394–1398. Yunnan University. His current research interests in-
[39] Y. Wei and R. Lui, “Dual methods for nonconvex spectrum optimiza- clude neural network theory and applications, cloud
tion of multicarrier systems,” IEEE Trans. Commun., vol. 54, no. 7, computing, and big data computing.
pp. 1310–1322, Jul. 2006.
[40] D. Tse and P. Viswanath, Fundamentals of Wireless Communication.
Cambridge, U.K.: Cambridge Univ. Press, 2005. Wei Feng received the B.E. degree in electronics
[41] J. Tang and X. Zhang, “Cross-layer resource allocation over wireless relay and information engineering from Hubei Engineering
networks for quality of service provisioning,” IEEE J. Sel. Areas Commun., University, Xiaogan, China, in 2005, and the M.E.
vol. 25, no. 4, pp. 645–656, May 2007. and Ph.D. degrees in communication and informa-
[42] K. Kumar and Y. Lu, “Cloud computing for mobile users: Can offloading tion systems from the South China University of
computation save energy,” Computer, vol. 43, no. 4, pp. 51–56, 2010. Technology, Guangzhou, China, in 2009 and 2014,
[43] C. She et al., “A tutorial on ultra-reliable and low-latency communications respectively. From 2005 to 2006 and 2009 to 2011, she
in 6G: Integrating domain knowledge into deep learning,” in Proc. IEEE, was an FAE with LITE-ON Technology Corporation,
vol. 109, no. 3, pp. 204–246, Mar. 2021. Guangzhou, China, and a Network Engineer with
Huaxin Consulting Company Ltd., Hangzhou, China.
She is currently a Lecturer with Hangzhou Dianzi
University, Hangzhou, China. Her research interests include energy efficiency
Mingxiong Zhao (Member, IEEE) received the B.S. and physical layer security in future wireless communications.
degree in electrical engineering and the Ph.D. de-
gree in information and communication engineer-
ing from the South China University of Technology, Changyang She (Member, IEEE) received the
Guangzhou, China, in 2011 and 2016, respectively. B.Eng. degree from the Honours College (formerly
From 2012 to 2013 and from 2015 to 2016, he School of Advanced Engineering) of Beihang Uni-
was a Visiting Ph.D. Student with the University of versity (BUAA), Beijing, China, in 2012 and the
Minnesota, Twin Cities, MN, USA, and Singapore Ph.D. degree from the School of Electronics and
University of Technology and Design, Singapore. Information Engineering, BUAA, in 2017. From 2017
Since 2016, he has been with the School of Software, to 2018, he was a Postdoctoral Research Fellow with
Yunnan University, Kunming, China, where he is the Singapore University of Technology and Design,
currently an Associate Professor. His current research interests include physical Singapore. From 2018 to 2021, he was a Postdoctoral
layer security, mobile edge computing, and edge AI techniques. Research Associate with The University of Sydney,
Sydney, NSW, Australia. His research interests in-
clude ultra-reliable and low-latency communications, deep learning in wireless
networks, industrial Internet-of-Things, and tactile Internet. He was the recipient
Junjie Yu was born in 1995 in Xiaogan, China. of the Australian Research Council (ARC) Discovery Early Career Research
He received the B.S. degree in software engineering Award (DECRA). From 2021 to 2024, he will be an ARC DECRA Fellow with
from Southwest Minzu University, Chengdu, China, the University of Sydney.
in 2018 and the M.S. degree in software engineering
from Yunnan University, Kunming, China, in 2021.
His research interests include resource allocation and Tony Q.S. Quek (Fellow, IEEE) received the B.E.
mobile edge computing. and M.E. degrees in electrical and electronics
engineering from the Tokyo Institute of Technol-
ogy, Tokyo, Japan, in 1998 and 2000, respectively,
and the Ph.D. degree in electrical engineering and
computer science from the Massachusetts Institute
of Technology, Cambridge, MA, USA, in 2008. He
Wen-Tao Li received the B.S. degree in software is currently the Cheng Tsang Man Chair Professor
engineering and the M.S. degree in artificial intel- with the Singapore University of Technology and
ligent and machine learning from Yunnan University, Design (SUTD), Singapore. He is also the Director of
Kunming, China, in 2014 and 2021, respectively. His Future Communications R&D Programme, the Head
research interests include mobile edge computing and of ISTD Pillar, and the Deputy Director of the SUTD-ZJU IDEA. His current
convex optimization. research topics include wireless communications and networking, network
intelligence, Internet-of-Things, URLLC, and big data processing.
He is actively involved in organizing and chairing sessions, and was a Member
of the Technical Program Committee and Symposium Chair in a number of
international conferences. He is currently the Editor of the IEEE TRANSAC-
TIONS ON WIRELESS COMMUNICATIONS and an elected Member of the IEEE
Signal Processing Society SPCOM Technical Committee. He was an Executive
Di Liu received the B.Eng. and M.Eng. degrees from Editorial Committee Member for the IEEE TRANSACTIONS ON WIRELESS COM-
Northwestern Polytechnical University, Xi’an, China, MUNICATIONS, the Editor of IEEE TRANSACTIONS ON COMMUNICATIONS, and
in 2007 and 2011, respectively, and the Ph.D. degree the Editor of IEEE WIRELESS COMMUNICATIONS LETTERS.
from Leiden University, Leiden, The Netherlands, in He was honored with the 2008 Philip Yeo Prize for Outstanding Achievement
2017. He is currently an Assistant Professor with the in Research, 2012 IEEE William R. Bennett Prize, 2015 SUTD Outstanding
School of Software, Yunnan University, Kunming, Education awards – Excellence in Research, 2016 IEEE Signal Processing So-
China. He is a Research Fellow with HP-NTU Corp ciety Young Author Best Paper Award, 2017 CTTC Early Achievement Award,
Lab, Nanyang Technological University, Singapore. 2017 IEEE ComSoc AP Outstanding Paper Award, 2020 IEEE Communications
His research interests include edge systems, ma- Society Young Author Best Paper Award, 2020 IEEE Stephen O. Rice Prize,
chine learning on embedded systems, energy-efficient 2020 Nokia Visiting Professor, and 2016-2020 Clarivate Analytics Highly Cited
multicore/many core systems, and cyber-physical Researcher. He is a Distinguished Lecturer of the IEEE Communications Society.
systems.

Authorized licensed use limited to: University of Ottawa. Downloaded on February 05,2024 at 15:43:44 UTC from IEEE Xplore. Restrictions apply.