Professional Documents
Culture Documents
fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2020.3022802, IEEE Internet of
Things Journal
1
Abstract—Mobile edge computing (MEC) is considered as a requirement [1]–[4]. However, IoT devices, i.e., smartphones
promising technology to reduce the energy consumption (EC) and wearable devices, are in general only with limited battery
and task accomplishment latency of smart mobile user equipment energy and computation resources due to their small physical
(UEs) by offloading computation-intensive tasks to the nearby
MEC servers. However, the quality of experience (QoE) for sizes. To tackle this problem, mobile cloud computing (MCC)
computation highly depends on the wireless channel conditions has been put forward in the last decade. In MCC systems, IoT
when computation tasks are offloaded to MEC servers. In this devices offload their computation tasks via wireless links to
paper, by considering the imperfect channel state information the resource-rich remote cloud servers, which can save battery
(CSI), we study the joint offloading decision, transmit power, energy of IoT devices [5].
and computation resources to minimize the weighted sum of
EC of all UEs while guaranteeing the probabilistic constraint However, in MCC, the execution delay may be very large
in multiuser MEC-enabled Internet of Things (IoT) networks. because of long distance and heavy network congestion be-
This formulated optimization problem is a stochastic mixed- tween IoT devices and cloud servers. This is unacceptable
integer nonconvex problem and challenging to solve. To deal for delay-sensitive applications. To address this issue, mobile
with it, we develop a low-complexity two-stage algorithm. In edge computing (MEC) has been proposed to deliver the cloud
the first stage, we solve the relaxed version of the original
problem to obtain offloading priorities of all UEs. In the second functionalities, i.e., computing and storage, to the edge of
stage, we solve an iterative optimization problem to obtain a wireless networks [6] [7]. The existing works [8] [9] have
suboptimal offloading decision. As both stages include solving a shown that the computation experience of IoT devices can be
series of nonconvex stochastic problems, we present a constrained significantly improved when offloading the computation tasks
stochastic successive convex approximation based algorithm to to MEC servers.
obtain a near optimal solution with low-complexity. Numerical
results demonstrate that the proposed algorithm provides a Recently, the MEC paradigm has attracted significant at-
comparable performance to existing approaches. tention from both academia and industry. Earlier works on
MEC have been focused on offloading platforms, such as,
Index Terms—Computation offloading, resource allocation,
stochastic programming, optimization. MAUI [10], CloneCloud [11], ThinkAir [12], and CONCERT
[13]. Unlike traditional cloud servers with rich resources, MEC
servers may be resource-limited. Therefore, it is crucial to opti-
I. I NTRODUCTION mize computation offloading strategies and resource allocation
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: Heriot-Watt University. Downloaded on September 22,2020 at 17:20:16 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.2020.3022802, IEEE Internet of
Things Journal
2
For multi-user MEC, a large number of works have also tive approach to solve the robust transmit beamforming design
been devoted to computation offloading [5] [18]–[31]. In [18], problem under the probabilistic signal-to-interference-plus-
a joint optimization of partial offloading and resource alloca- noise ratio (SINR) constraint for multiuser wireless systems.
tion has been proposed to minimize the EC for multiuser MEC Their idea is to replace the difficult probabilistic constraint
systems with TDMA. Different from the partial offloading, by the worst-case constraint with spherically bounded channel
binary offloading designs in multi-user MEC are mixed-integer errors. However, this idea can not be applied to our formulated
optimization problems which are usually nonconvex. In [5] problem, where the objective function involves the expectation
[19], binary offloading mechanisms based on game theory have and has no closed-form expression. To deal with this issue, we
been proposed. In [20], the problem of the total EC minimiza- propose a low-complexity two-stage optimization framework
tion has been formulated by optimizing transmit power for based on constrained stochastic successive convex approxi-
multiuser MIMO MEC systems. Note that, in [5] [19] [20], mation (CSSCA) to solve the formulated problem. The main
either offloading decision or resource allocation was optimized contributions of this paper are summarized as follows:
separately. In [21], a joint optimization of offloading decision • By taking imperfect CSI into account, we formulate the
and resource allocation has been developed to minimize the average weighted sum of EC minimization problem under
total EC under latency requirements. In [22] [23] [24], a the probabilistic delay constraint, and jointly optimize the
joint optimization framework of binary offloading decision offloading decision, UEs’ transmit power, and computa-
and resource allocation in heterogeneous networks (HetNets) tion resource allocation.
with MEC has been investigated. These works have been • To deal with the mixed-integer nature of the formulated
further extended to the multi-task offloading in HetNets in problem, we develop a novel low-complexity optimiza-
[25], where the authors have considered the joint computation tion framework. It includes two stages. In the first stage,
offloading, user association and resource allocation problem we solve the relaxed version of the original problem
and proposed an alternating optimization framework to solve to obtain offloading priorities of all UEs. In the sec-
it. Reference [26] has considered a hierarchical fog-cloud ond stage, we iteratively solve a stochastic optimization
computing system and studied a min-max fairness problem problem to obtain a suboptimal offloading decision and
under binary offloading decision constraint. Meanwhile, some corresponding resource allocation.
works have also considered multiuser MEC systems with EH • Both stages include solving a series of nonconvex s-
[27] [28]. In [27], a joint optimization framework of offloading tochastic problems. One of the difficulties in the stochas-
ratio, energy beamforming, and computing resource has been tic optimization problem is that the objective function in-
developed to minimize the total EC. Different from [27], volves the expectation and has no closed-form expression.
reference [28] has studied the weighted sum of computation To tackle this challenge, we propose a CSSCA-based
rate maximization by jointly optimizing the binary offloading iterative algorithm, in which a quadratic convex problem
and the transmission time allocation. Besides, reference [29] is solved at each iteration by a low-complexity algorithm
has investigated stochastic resource allocation in multiuser based on the dual decomposition method. In addition, we
MEC systems and proposed a low-complexity Lyapunov opti- have also analyzed the convergence and complexity of the
mization based online algorithm. More recently, some works proposed algorithm.
[30] [31] have proposed offloading scheme based on machine • We perform extensive numerical simulations to evaluate
learning. A reinforcement learning based offloading policy for the convergence of the proposed iterative algorithm and
an IoT device has been proposed in [30], while an after- compare the performance of the proposed solution with
state learning based computation offloading scheme has been the existing approaches.
designed for MEC systems with EH in [31]. The remaining of this paper is organized as follows. In
However, the aforementioned works assume that the perfect Section II, the system model is firstly described and the opti-
CSI is known at the base station (BS). This assumption may mization problem is then formulated. In Section III, the low-
be unrealistic in practical systems due to the estimation errors, complexity framework to solve the mixed-integer nonconvex
limited CSI feedback quantization, delays, etc. On the other stochastic optimization problem is presented. Convergence and
hand, the QoE of the computation heavily relies on the wireless complexity analysis of the proposed algorithm are shown in
fading channel conditions since task offloading requires effec- Section IV. In Section V and Section VI, simulation results
tive wireless transmission. The authors in [32] have studied the and conclusion are, respectively, presented.
fairness problem for MEC systems with imperfect CSI with Notation: E [·] presents the mathematical expectation. Pr {·}
the goal to minimize the maximum weighted EC under the b ∆
denotes the probability operator. [x]a = min {b, max {a, x}}.
upper-bound of the average latency constraint. Different from
it, our goal is to minimize the weighted sum of EC of UEs II. S YSTEM M ODEL AND P ROBLEM T RANSFORMATION
while guaranteeing the probabilistic delay constraint for MEC In this section, we introduce the communication model and
systems with imperfect CSI, which seeks to provide “safe” computation model of a multiuser MEC system, followed by
performance guaranteeing for a certain probability (often high) the optimization problem formulation.
of satisfying the delay requirements. Our formulated problem
is mixed-integer nonconvex stochastic optimization problem A. Communication Model
under the probabilistic constraint which is more difficult to Consider a multiuser MEC system described in Fig. 1,
solve. Note that the authors in [33] have presented a conserva- where K UEs are served by a BS equipped with the MEC
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: Heriot-Watt University. Downloaded on September 22,2020 at 17:20:16 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.2020.3022802, IEEE Internet of
Things Journal
3
B. Computation Model
UE 1 The task of UE k to be processed can be characterized
Task 1
by a three-tuple (Lk , Ck , Tkmax ), where Lk is the size of
task input data (in bits), Ck is the number of CPU cycles to
accomplish one bit task (in CPU cycles/bit), and Tkmax presents
Task k
the maximum tolerable delay for the task. The parameters Lk ,
UE k
Ck , and Tkmax can be measured by task profilers [37]. Assume
Task K that each task is atomic, i.e., cannot be further partitioned. That
is, it can be either locally processed at the UE or transferred
MEC
Server BS to the MEC server for remote processing. Let xk be the binary
UE K offloading decision variable for UE k, where xk = 1 implies
that UE k chooses to offload its task to the MEC server, xk = 0
Fig. 1. Multiuser MEC system.
means that UE k chooses to locally process its task.
In what follows, we present the process delay and the EC
for both local and remote computing.
server. For simplicity, we will use the MEC server and BS 1) Local Computing: The processing time Tkloc for local
interchangeably. Denote the set of UEs as K = {1, 2, ..., K}. task executing can be written as
Without loss of generality, assume that each UE has only one Lk Ck
Tkloc = loc , (6)
computation-intensive task which is required to be accom- fk
plished within a certain time. where fkloc is the CPU computational capability of UE k (in
To avoid the mutual interference among UEs, we suppose CPU cycles/s).
that each UE occupies an orthogonal quasi-static wireless As in [14] [16], the EC Ekloc for local task executing can
channel when offloading its task. Let hk be the channel fading be expressed as
coefficient between UE k and the BS. However, in practical 3 2
wireless systems, the accurate CSI at BS is difficult to obtain Ekloc = βk fkloc Tkloc = βk fkloc Lk Ck , (7)
owing to the channel estimation error or the feedback delay. 3
where βk fkloc is the computational power of UE k, and βk
Thus, in this paper, the actual CSI hk is modeled as:
is a constant related to chip architecture [15] [16].
hk = b
hk + ek , (1) 2) Remote Computing: When UE k decides to offload its
task to the MEC server for processing, the offloading process
where b hk is a channel estimate and ek is the estimation includes the following three stages. Firstly, UE k sends task
error. Assume that ek obeys a circularly symmetric
complex input data to the MEC server over the wireless channel. Then,
Gaussian distribution, i.e., ek ∼ CN 0, σe2k . Then, the the MEC server processes task k by allocating part of the
received signal associated with UE k is given by computation resources to it. Finally, the MEC server sends
√ √ back the task output data to UE k after completing the task.
yk = pk b hk sk + pk ek sk + nk , (2) Accordingly, the uplink transmission time can be expressed
| {z }
n0k by
Lk
where pk is the transmit power hof UEik, sk is the transmitted Tktr = . (8)
Rk (pk , θk , ek )
2
symbol from UE k satisfying E |sk | = 1, nk is the additive The remote computation time can be given by
white Gaussian noise with power spectral density N0 , n0k is Lk Ck
the effective noise, which combines the additive noise and Tkexe = , (9)
wk Fsmax
residual channel estimation error and its variance is
h i where wk Fsmax denotes the computation resources (in CPU
2 2 2
E |n0 k | = pk |ek | + σk2 = pk |ek | + θk BN0 , (3) cycles/s) allocated to UE k, wk is the normalized portion of
the computation resource assigned to UE k, and Fsmax presents
where B is the total bandwidth, and θk is the normalized the maximum computation resource of the MEC server.
portion of the bandwidth to UE k. As in [18] [27] [28], the size of task output data is, in
As in [34] [35] [36], since the effective noise n0k is neither general, much smaller than that of task input data. Thus, we
independent nor Gaussian, the effective signal to interference neglect the time consumed by returning the task output data to
plus noise ratio (SINR) for UE k is bounded below by UEs. Therefore, the total delay for remote processing consists
2 of two parts:
pk bhk
Lk Lk Ck
SINRk = . (4) Tkmec = Tktr + Tkexe = + . (10)
2
pk |ek | + θk BN0 Rk (pk , θk , ek ) wk Fsmax
The EC of UE k for remote processing is mainly the
Thus, the transmission rate for UE k in (2) is bounded below transmit energy [28], and thus can be written as
by
pk Lk
Rk (pk , θk , ek ) = θk Blog2 (1 + SINRk ) . (5) Ekmec = pk Tktr = . (11)
Rk (pk , θk , ek )
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: Heriot-Watt University. Downloaded on September 22,2020 at 17:20:16 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.2020.3022802, IEEE Internet of
Things Journal
4
C. Problem Formulation either executed locally or offloaded to the MEC server. (17c)
As introduced above, each task is either processed locally at describes the transmit power constraint of UE k. (17d) and
the UE or offloaded to the MEC server. Thus, the total delay (17e) are the bandwidth constraints. (17f) and (17g) describe
for completing task k can be expressed as the computation resource constraints. It should be pointed out
that the objective function is a function of x, p, and θ which
Tk (xk , pk , θk , wk , ek ) = (1 − xk ) Tkloc +xk Tkmec . (12) depend on the random state given by channel estimation error
ek .
The total EC for completing a task can be expressed as
Problem P0 is challenging to solve due to the following
Ek (xk , pk , θk , ek ) = (1 − xk ) Ekloc +xk Ekmec . (13) three reasons:
• Firstly, the objective function cannot be expressed in
Each UE requires its task to be completed during the
closed-form.
maximum tolerable delay Tkmax . However, in the presence of
• Secondly, there exists the coupling between the optimiza-
random CSI errors, the traditional deterministic delay require-
tion variables x, p, and θ, leading to the nonconvex
ments can no longer be guaranteed. Therefore, we model the
objective function and constraint (17a).
delay requirement in the form of a chance constraint as follows
• Thirdly, the binary variable x makes problem P0 a mixed-
Prek {Tk (xk , pk , θk , wk , ek ) ≤ Tkmax } ≥ 1 − ξkmax , (14) integer nonconvex problem.
Therefore, problem P0 is a stochastic mixed-integer nonconvex
where ξkmax is the maximum tolerable outage probability for
optimization problem which is generally difficult to solve. In
task k. The inequality can be recognized as a soft delay
the next section, we will propose a low-complexity algorithm
requirement. To facilitate the analysis, we have the following
to obtain a suboptimal solution for problem P0 .
lemma.
Lemma 1. The delay probability constraint in (14) is equiv- III. L OW- COMPLEXITY F RAMEWORK FOR P ROBLEM P0
alently expressed by In this section, we first develop a CSSCA framework to
Θk (xk , pk , θk , wk ) ≤ 0, (15) solve the relaxed version of problem P0 . Specifically, in the
CSSCA framework, the problem at each iteration is solved by
where Θk (xk , pk , θk , wk ) is given in (16) at the top of next using dual decomposition method [38]. Then, we propose a
page. ranking-based algorithm to obtain a suboptimal solution for
Proof: See Appendix A. the original problem.
For an MEC system, the QoE of UE k is mainly determined
by the EC and its task execution delay. Considering that UEs A. Problem Transformation
are usually energy-limited, our goal is to design an energy-
efficient joint offloading decision and resource allocation To handle the binary constraint (17b), we relax the binary xk
scheme to minimize the average weighted sum of EC of UEs as a continuous variable, i.e., xk ∈ [0, 1]. As a result, problem
under the probabilistic delay requirement. The optimization P0 can be reformulated as
problem can be stated as K
X
K
P1 : min Φk (zk ) (18)
z
k=1
X
P0 : min αk Eek [Ek (xk , pk , θk , ek )] (17)
x,p,θ,w
k=1
s.t. Θk (zk ) ≤ 0, ∀k, (18a)
s.t. Θk (xk , pk , θk , wk ) ≤ 0, ∀k, (17a) 0 ≤ xk ≤ 1, ∀k, (18b)
xk ∈ {0, 1} , ∀k, (17b) (17c) − (17g),
0 ≤ pk ≤ Pkmax , ∀k, (17c) where z = [x , p , θk , wk ]
T
∈ R4×1 and z =
T T k T T k k 4K×1
0 ≤ θk ≤ 1, ∀k, (17d) z1 , z2 , ..., zK ∈ R . The functions Φk (zk ) and
K
X ϕk (zk , ek ) are defined as
θk ≤ 1, (17e)
Φk (zk ) = E [ϕk (zk , ek )] , (19)
k=1
0 ≤ wk ≤ 1, ∀k, (17f) ϕk (zk , ek ) = αk Ek (xk , pk , θk , ek ). (20)
K
X It is noteworthy that problem P1 is still a nonconvex
wk ≤ 1, (17g) constrained stochastic optimization problem. There exist a
k=1 few algorithms that can deal with nonconvex stochastic opti-
where x, p, θ, and w denote the vectors of the offloading mization problems, i.e., sample average approximation (SAA)
decision xk , transmit power pk , the normalized portion θk [39]–[41] and online primal-dual algorithm [42]. However,
of bandwidth, and the normalized portion wk of computation they either have a higher computational complexity or cannot
resource, respectively. αk is the weight for UE k, where the guarantee the convergence. So it is necessary to develop
larger the value of αk , the higher the priority for energy a new effective method to handle it. In what follows, we
saving of user k. In problem P0 , (17a) gives the delay will develop a CSSCA-based algorithm to solve problem P1 ,
constraint for each UE. (17b) implies that each task can be which is more efficient and provably convergent [43] [44].
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: Heriot-Watt University. Downloaded on September 22,2020 at 17:20:16 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.2020.3022802, IEEE Internet of
Things Journal
5
2
hk
b
Lk Ck 1 θk BN0
Θk (xk , pk , θk , wk ) = (1 − xk ) − Tkmax +xk 2 − max
− ln ξk (16)
loc Lk
fk σ ek pk
max − Lk Ck
θk B Tk
wk Fsmax
−1
2
wk Fsmax
−1
2 2
∆
(22)
and
are its convex and nonconvex parts, respectively.
∆ p L
ϕck (zk , ek ) = αk xk k k 2
. (23) Let Θ b (i) (zk ) denote the convex surrogate function of
pk |b
hk | k
θk Blog2 1 + p |e |2 +θ BN Θk (zk ) at the i-th iteration of problem P1 . Similar to
k k k 0
b (i) (zk ), Θ
Φ b (i) (zk ) can be given as
k k
Let Φ b (i) (zk ) denote the convex surrogate function of
k
Φk (zk ) at the i-th iteration of problem P1 . The same as in
b (i) (zk ) is expressed as b (i) (zk ) = (1 − ρ(i) )Θ(i−1) + ρ(i) Θc (zk )
Θ
[43] [45], Φ k k
k
k
(i) c (i) (i) T c (i) (i)
b (i) (zk ) = (1 − ρ(i) )Φ(i−1) + ρ(i) ϕc (zk ) + ρ Θk zk + ρ ∇ Θk zk zk − zk (33)
Φ k k k T
2
(i−1) (i) (i)
(i) (i) (i) (i)
+ ρ(i) ϕck (zk , ek ) + ρ(i) ∇T ϕck (zk , ek ) zk − zk
(i) + (1 − ρ(i) ) vk zk − zk + τk1
zk − zk
,
(24)
T
2
(i−1) (i) (i)
+ (1 − ρ(i) ) uk zk − zk + τk0
zk − zk
,
(i) (i)
(i) (i) where τk1 is any positive number, Θk = Θk zk , and
where τk0 ishany positive inumber, Φhk and uk are i approxima- (i)
(i) (i)
(i)
tions for E ϕk (zk , ek ) and ∇E
(i)
ϕk (zk , ek ) , respectively, vk = ∇Θk zk . Similar to the procedure in (27), vk
(i) (i) (i)
which are updated recursively according to can be expressed as vk = ∇Θck zk + ∇Θck zk , where
(i) (i)
(i)
Φk = (1−ρ(i) )Φk
(i−1) (i)
+ρ(i) ϕk (zk , ek ),
(i)
(25) ∇Θck zk and ∇Θck zk are, respectively, given by (34a)
(i) (i−1) (i) (i) and (34b) at the top of next page. The components in (34b)
uk = (1−ρ(i) )uk +ρ(i) ∇ϕk (zk , ek ), (26) Lk
(i) Lk C k
max −
θk B Tk
(i)
(i)
wk Fsmax
is a given constant sequence. As a result, the optimization problem of the i-th iteration
(i) (i)
In (26), ∇ϕk (zk , ek ) is expressed as can be expressed as
(i) (i) (i) (i) (i)
∇ϕk zk , ek = ∇ϕck zk + ∇ϕck zk , ek , (27)
K
(i) b (i) (zk )
X
P2 : min Φ (36)
(i) (i) (i)
where ∇ϕck zk
and ∇ϕck zk , ek
are, respectively, giv- z k
k=1
en by (28a) and (28b) at the top of next page. The first b k (zk ) ≤ 0, ∀k.
(i) s.t. Θ (36a)
three components in (28b)! are given by (29) with mk =
(i)
p |b hk |
2 (17c) − (17g), (18b).
ln 1 + (i) (i)k 2 (i) .
pk ek +θk BN0
Similarly, Θk (zk ) is also divided into two parts as (i)
z(i) denote the
Note that P2 is not necessarily feasible. Let b
(i) (i)
Θk (zk ) = Θck (zk ) + Θck (zk ) (30) optimal solution to P2 . If P2 is infeasible, the following
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: Heriot-Watt University. Downloaded on September 22,2020 at 17:20:16 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.2020.3022802, IEEE Internet of
Things Journal
6
h iT
(i) 2
∇ϕck zk = −αk Lk βk fkloc Ck 0 0 0 (28a)
h iT
(i) (i) (i) (i) (i) (i) (i) (i)
∇ϕck zk , ek = ∂x∂ k ϕck zk , ek ∂ c
∂pk ϕk zk , ek
∂ c
∂θk ϕk zk , ek 0 (28b)
(i)
∂ c (i) (i) pk αk Lk ln 2
ϕ z ,e =
∂xk k k k (i) (i)
θk mk B
2
(i) (i)
(i) BN0 pk θk b hk
∂ c (i) (i) x αk Lk ln 2 (i)
ϕk zk , ek = k 2 mk −
2
2 2
∂pk (i) (i)
(i) (i) (i)
(i)
(i) (i) (i)
(29)
θk mk B pk ek + θk BN0 pk b hk + pk ek + θk BN0
2
(i) (i) b
∂ c (i) (i)
(i) (i)
x p αk Lk ln 2 p θ
k k k h BN 0
m(i) −
ϕk zk , ek = − k k 2 k 2
2 2
∂θk (i) (i) (i) (i) (i) (i) (i) (i)
(i)
θk mk B pk ek + θk BN0 pk ek + θk BN0 + pk b hk
h iT
(i) L C
∇Θck zk = Tkmax − fklock − ln ξkmax 0 0 0 (34a)
k
h iT
(i) (i) (i) (i) (i)
∇Θck zk = ∂x∂ k Θck zk ∂
Θ
∂pk k
c
zk
∂ c
∂θk k zk
Θ ∂ c
∂wk Θk zk (34b)
2
hk
b
∂ c (i) 1 θk BN0
Θ z = 2 −
∂xk k k σ ek pk nk − 1
(i) (i)
∂ c (i) x θ BN0
Θ z = − k k 2
∂pk k k (i)
σe2k pk
2
(35)
(i) hk Lk ln 2 (i)
b
∂ c (i) xk BN0 n
Θ z = 2 − 2 × k
∂θk k k σek p(i) (i) (i)
2
Lk Ck
k B θk nk − 1 Tkmax − (i)
wk Fsmax
2
(i)
xk L2k Ck bhk ln 2 (i)
∂ c
(i) nk
Θk zk = − 2 × 2
∂wk
2
(i) (i)
BFsmax σe2k θk wk (i)
nk − 1 max Lk Ck
Tk − (i) max
w k Fs
(i) (i)
problem will be solved to minimize the constraints Obviously, both problems P2 and P3 are convex, since the
(i)
(i)
P3 : min δ (37) surrogate functions in (24) and (33) are convex. Thus, P2
(i)
z,δ and P3 can be efficiently solved by interior-point methods
b k (zk ) ≤ δ, ∀k.
s.t. Θ (37a) [46].
(i)
− δ ≤ xk ≤ 1 + δ, ∀k, (37b) z(i) for problems P2
After obtaining the optimal solution b
(i)
or P3 , z can be updated as follows,
− δ ≤ pk ≤ Pkmax + δ, ∀k, (37c)
− δ ≤ θk ≤ 1 + δ, ∀k, (37d) z(i+1) = 1 − γ (i) z(i) +γ (i)bz(i) , (38)
K
where γ (i) ∈ (0, 1] is a given constant sequence.
X
θk ≤ 1 + δ, (37e)
k=1 The procedure to solve P1 is described in Algorithm 1.
− δ ≤ wk ≤ 1 + δ, ∀k, (37f) However, it is not necessary to guarantee that the optimal
K
X
2327-4662 (c) 2020 IEEE. Personal use is permitted,wbut ≤ 1 + δ.
k republication/redistribution (37g)
requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Heriot-Watt University. Downloaded on September 22,2020 at 17:20:16 UTC from IEEE Xplore. Restrictions apply.
k=1
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.2020.3022802, IEEE Internet of
Things Journal
7
Algorithm 1 Relaxed Offloading Decision and Resource Therefore, the dual problem is expressed by
Allocation Algorithm (RODRA) for solving P1
1: Initialization: max D(i) (µ, η, λ) (41)
µ,η,λ
Choose proper sequences γ (i) and ρ(i) ; s.t. µ0, η ≥ 0, λ ≥ 0.
Choose an initial point z(0) meeting all constraints in P1 ;
Set iteration index i = 0. (i)
Since problem P2 is convex and the Slater’s condition is
2: repeat satisfied, the strong duality holds between it and its dual
(i) (i)
3: Generate a channel error vector e(i) = [e1 , ..., eK ]T . (i)
problem. Thus, the optimal solution to problem P2 can be
(i) (i)
4: Calculate the surrogate functions Φ b (zk ) , Θ
k
b (zk ),
k obtained through solving problem (41).
∀k using (24) and (33). After some algebraic manipulations, (39) can be rewritten
(i)
5: if P2 is feasible then by
(i)
6: Solve problem P2 to obtain b z(i) .
K
7: else X 2 T
8:
(i)
Solve problem P3 to obtain b z(i) . L(i) (z, µ, η, λ) = ak kzk k + (bk ) zk + ck , (42)
k=1
9: end if
10: Update z(i+1) according to (38). where ak , bk , and ck are given by (43) at the top of next page.
11: Set i = i + 1. Clearly, for a given (µ, η, λ), the dual function D(i) (µ, η, λ)
12: until the stopping criterion is met. in (40) can be decomposed into K independent subproblems
13: Output: The solution zRel = z(i+1) for problem P1 . as
2 T
min ak kzk k + (bk ) zk + ck (44)
zk
value of the offloading decision is in the integral form since s.t. (18b), (17c), (17d), and (17f).
we have relaxed the binary constraint. Thus, Algorithm 1 can
be referred to as Relaxed Offloading Decision and Resource for k = 1, ..., K.
Allocation Algorithm (RODRA). For the optimal solution of (44), we have the following
Remark 1: Algorithm 1 is a centralized algorithm and Lemma.
executed at the BS, which has strong computing power and
communication capability. Lemma 2. For given (µ, η, λ), the optimal solution z∗k =
T
[x∗k , p∗k , θk∗ , wk∗ ] for subproblem (44) is
1
B. Low-complexity Solution for Problem P2
(i) x∗k = [xk ]0 , (45)
P max
(i)
Although problems P2 and P3 are convex and can be
(i) p∗k = [pk ]0 k , (46)
1
solved by adopting general interior-point methods [46], this θk∗ = θk 0 , (47)
method suffers from relatively high complexity since it does 1
wk∗ = [wk ]0 , (48)
not utilize their special structures. Hence, we develop a low-
complexity algorithm based on the dual decomposition theory k,1 b k,2 k,3 b k,4 b b
(i) (i) where xk = − 2a , pk = − 2a , θk = − 2a , and wk = − 2a .
[38] to obtain the optimal solution for problem P2 (or P3 ) k k k k
where µ = [µ1 , µ2 , . . . , µK ], η, and λ are Lagrange multipliers where l is the iteration index of subgradient method;
T
associated to constraints (36a), (17e), and (17g), respectively. z∗k = [x∗k , p∗k , θk∗ , wk∗ ] is the optimal solution to (44) with
+
(µ(l), η(l), λ(l)); δµ,l , δη,l , δλ,l are positive step-sizes; [·] =
The dual function is then written as
max {0, ·}.
D(i) (µ, η, λ) = min L(i) (z, µ, η, λ) (40) Since problem (40) can be decomposed as multiple indepen-
z
dent subproblems, it is convenient to implement at the MEC
s.t. (18b), (17c), (17d), and (17f). server in a parallel way.
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: Heriot-Watt University. Downloaded on September 22,2020 at 17:20:16 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.2020.3022802, IEEE Internet of
Things Journal
8
ak = τk0 + µk τk1
(i) (i) (i) (i−1) (i−1) (i)
bk = ρ(i) ∇ϕck zk , ek + µk ∇Θck zk + 1 − ρ(i) uk + µk vk − 2 τk0 + µk τk1 zk
(i) max loc 2
Lk Ck max T T
−ρ ln ξk + αk βk fk Lk Ck + µk − Tk [1, 0, 0, 0] + [0, 0, η, λ]
fkloc
1
(i−1) (i−1)
(i) (i)
(i) (43)
ck = − (η + λ) + 1 − ρ(i) Φk + µk Θk + ρ(i) ϕck zk , ek + µk Θck zk
K
T T
(i−1) (i−1) (i) (i) (i) (i) (i)
− 1 − ρ(i) uk + µk vk zk − ρ(i) ∇ϕck zk , ek + µk ∇Θck zk zk
(i)
2 Lk Ck
2
+ τk0 + µk τk1
zk
+ ρ(i) αk βk fkloc Lk Ck + µk − T max
k
fkloc
C. Ranking-based Recovery of Binary Offloading Decision Algorithm 2 Ranking-based Binary Offloading and Resource
Variables Allocation Algorithm (RBORA) for solving P0
So far, we can solve problem P1 successfully and efficiently. Stage 1:
However, the solution xRel obtained from solving problem P1 1: Obtain the solution zRel for problem P1 by using Algo-
is not necessarily a vector with integer elements. If xRel is a rithm 1.
vector with binary elements, z = zRel . If xRel is not a vector 2: Order xRel from largest to smallest, i.e., xRel Rel
k1 ≥ xk2 ≥
Rel
with integer elements, which is infeasible for problem P0 . . . . ≥ xk K .
Thus, we need to recover the binary x. In this subsection, 3: Set S off = ∅.
we propose a ranking-based algorithm to obtain a suboptimal Stage 2:
solution of P0 where x is a binary vector. This algorithm is 4: for j = 1, ..., K do
composed of two stages summarized in Algorithm 2. 5: Set S temp = S off ∪ {kj } and calculate U (z|S temp ) by
In the first stage, the MEC server computes the compu- using Algorithm 1.
tation offloading priorities of all UEs, denoted by xRel . The 6: if U (z|S temp ) ≤ U z|S off then
computation offloading priority of UE k can be defined as xRel
k
7: S off = S temp .
obtained by Algorithm 1. Furthermore, we sort the elements 8: else
of xRel in the decreasing order. Let S off = {k ∈ K|xk = 1} 9: Break.
denote the set of offloading UEs. 10: end if
In the second stage, the UE is added to S off in descending 11: end for
order of offloading priorities until the weighted sum of EC can 12: Output: S off and corresponding p, θ, and w.
not be further
decreased. For example, if U z|S off ∪ {kj } ≥
off off
U z|S , S is the final set of offloading UEs. Here,
U z|S off denotes the optimal value of P1 under given the second stage needs to solve P4 (namely, P1 under given
offloading variable x and can be obtained by solving the x according to offloading priorities obtained from the first
following problem stage) by using Algorithm 1. Thus, we only need to check
K the convergence of Algorithm 1.
Note that Algorithm 1 includes a nested loop. In the inner
X
P4 : min Φk (zk ) (52) (i) (i)
z loop (lines 5-10), convex problem P2 or P3 is solved by
k=1
s.t. xk = 1, ∀k ∈ S off , (52a) using Lagrange dual decomposition method. For the outer loop
of Algorithm 1, with the properly chosen sequences γ (i)
(17a), (17c) − (17g). (i)
and ρ , it almost surely converges to a stationary point of
Note that P4 is a special form of P1 , and thus can be solved problem P1 . In other words, it almost surely converges to a
by Algorithm 1 by setting xk = 1, ∀k ∈ S off . local optimal point. For a rigorous proof of this convergence,
The whole procedure to solve P0 is summarized in Algo- interested readers are referred to [43].
rithm 2 which is nested by Algorithm 1.
B. Complexity Analysis
IV. C ONVERGENCE AND C OMPLEXITY A NALYSIS
As mentioned above, Algorithm 2 for solving P0 is com-
A. Convergence Analysis of Algorithm 1 posed of two stages. For the first stage, it needs to run Algo-
The whole procedure to solve P0 is summarized in Algorith- rithm 1 once. For the second stage, it needs to run Algorithm
m 2, which includes two stages. In the first stage, P1 is solved 1 at most K times. Note that Algorithm 1 includes a nested
(i) (i)
by using Algorithm 1. In the second stage, the ranking-based loop. At its each iteration, a convex problem P2 or P3
method is adopted to search S off and compute corresponding is solved by using the Lagrange dual decomposition method.
resource allocation variables. Specifically, each iteration of Thus, the complexity of Algorithm 1 is O(K) [47]. Supposing
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: Heriot-Watt University. Downloaded on September 22,2020 at 17:20:16 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.2020.3022802, IEEE Internet of
Things Journal
9
K=12, F max
k
=10
1.6
Notation Parameter Value
K=6, F max
k
=5
r Cell radius 500 m 1.4
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: Heriot-Watt University. Downloaded on September 22,2020 at 17:20:16 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.2020.3022802, IEEE Internet of
Things Journal
10
0.9 1
0.8 0.9
RBORA RBORA
ES 0.8 ES
0.7 NonR
NonR
NonR-bisec
Weighted sum of EC (J)
NonR-bisec 0.7
0.6 Robust
Robust
Local-only
Local-only 0.6
0.5
CDF
0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0 0
0.9 1.1 1.3 1.5 1.7 1.9 2.1 2.3 2.5 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2
Maximum tolerable delay T max
k
(s) Task execution delay T k (s)
(a) σe2k = 0.002 Fig. 4. CDF of realistic task execution delay Tk for Tkmax =1.5 s
0.9
0.9
0.8
RBORA
0.8
ES
0.7 RBORA
NonR
0.7 ES
Weighted sum of EC (J)
NonR-bisec
0.6 NonR
Robust
0.8
RBORA
NonR-bisec
can observe that for all schemes except the NonR, all UEs 0.6 Robust
Local-only
obtain the prescribed task execution delay with a very high 0.5
weighted sum of EC of the NonR, as shown in Fig. 3, is at Maximum computation resource F max
s
(GHz)
the cost of severe violation of the delay constraint. Moreover, (b) Tkmax = 2 s
the proposed RBORA algorithm has almost the same delay
outage performance as the ES. Fig. 5. Weighted sum of EC versus maximum computation resource of MEC
Figs. 5a and 5b show the impact of the maximum computa-
tion resource Fsmax of MEC on the weighted sum of EC of the
proposed RBORA and benchmarks for the case K = 6. It can corresponds to Figs. 5a and 5b for the case Fsmax = 5 GHz.
be seen that the larger Fsmax , the less the weighted sum of EC This phenomenon is similar to that of Fig. 4, that is, all UEs
for all schemes except the Local-only. Moreover, it can been for all schemes except the NonR meet the prescribed task
observed that the weighted sum of EC of the RBORA scheme execution delay with a very high probability.
is fairly close to that of the ES. In addition, when the delay Fig. 7 shows the impact of the task input data size Lk on
requirement is looser, i.e., Tkmax = 2 s, the weighted sum of the weighted sum of EC for the case K = 6, Fsmax = 5 GHz,
EC obtained by all schemes except the Local-only and the and Tkmax = 1.6 s. As Lk becomes large, the weighted sum
NonR will sooner stabilize. This means that under the looser of EC of all schemes increase. Besides, we also note that the
Tkmax , almost all UEs prefer to offload their tasks even when weighted sum of EC of the proposed RBORA is almost the
Fsmax is small, i.e., Fsmax = 7 GHz. same as that of the ES and higher than that of the NonR by at
Similarly, we also plot the CDF of the realistic task execu- most 0.1J, but lower than that of the NonR-bisec, the Robust
tion delay of all UEs in Figs. 6a and 6b, which respectively and the Local-only.
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: Heriot-Watt University. Downloaded on September 22,2020 at 17:20:16 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.2020.3022802, IEEE Internet of
Things Journal
11
1 1
0.9 0.9
RBORA RBORA
0.8 ES 0.8 ES
NonR NonR
NonR-bisec NonR-bisec
0.7 0.7
Robust Robust
Local-only Local-only
0.6 0.6
CDF
CDF
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Task execution delay T k (s) Task execution delay T k (s)
(a) Tkmax = 1 s Fig. 8. CDF of realistic task execution delay Tk for Tkmax =1.6 s
1
2.5
RBORA
0.9
RBORA NonR
ES NonR-bisec
0.8
NonR 2 Robust
NonR-bisec Local-only
0.7
Robust
0.5
0.4
1
0.3
0.2
0.5
0.1
0
0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 0
Task execution delay T k (s) 5 7 9 11 13 15 17
Number of UEs
ES 10
1.6
NonR
NonR-bisec
1.4 Robust 8
Local-only
Weighted sum of EC (J)
1.2
6
1
4
0.8
0.6 2
0.4
0
5 7 9 11 13 15 17
0.2 Number of UEs
0
0.2 0.3 0.4 0.5 0.6 0.7 0.8
(b) Number of offloading UEs
Task input size L k (MB)
Fig. 9. Offloading performance versus the number of UEs
Fig. 7. Weighted sum of EC versus task input size
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: Heriot-Watt University. Downloaded on September 22,2020 at 17:20:16 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.2020.3022802, IEEE Internet of
Things Journal
12
0.9 2
RBORA
hk
NonR
b
0.8
NonR-bisec 1 θk BN0
= 1 − exp 2 − .
Robust
0.7 Lk
Local-only
σek pk B max −
Tk
1
Lk Ck
wk Fsmax
0.6
2 −1
(55)
CDF
0.5
0.4 The second equality above is due to eek = σ1e ek . The last
k
2
0.3 equality holds since eek ∼ CN (0, 1) and thus |e ek | follows
0.2 an exponential distribution with mean one. So, in this case,
0.1 (14) becomes
0 2
0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6
hk
b
Task execution delay T k (s)
1 θk BN0 max
− ≤ ln ξk (56)
Fig. 10. CDF of realistic task execution delay Tk for Tkmax =1.6 s 2
σ ek pk Lk
max −
L C
k k
θk B Tk
max
−1
2 wk F s
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: Heriot-Watt University. Downloaded on September 22,2020 at 17:20:16 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.2020.3022802, IEEE Internet of
Things Journal
13
[8] K. Kumar, J. Liu, Y. Lu, and B. Bhargava, “A survey of computation [30] Z. Wei, B. Zhao, J. Su, and X. Lu, “Dynamic edge computation
offloading for mobile systems,” Mobile Netw. Appl., vol. 18, no. 1, pp. offloading for internet of things with energy harvesting: A learning
129–140, Feb. 2013. method,” IEEE Internet Things J., vol. 6, no. 3, pp. 4436–4447, Jun.
[9] W. Shi, J. Cao, Q. Zhang, Y. Li, and L. Xu, “Edge computing: Vision 2019.
and challenges,” IEEE Internet Things J., vol. 3, no. 5, pp. 637–646, [31] M. Min, L. Xiao, Y. Chen, P. Cheng, D. Wu, and W. Zhuang, “Learning-
Oct. 2016. based computation offloading for IoT devices with energy harvesting,”
[10] E. Cuervo, A. Balasubramanian, D. Cho, A. Wolman, S. Saroiu, IEEE Trans. Veh. Technol., vol. 68, no. 2, pp. 1930–1941, Feb. 2019.
R. Chandra, and P. Bahl, “Maui: Making smartphones last longer with [32] T. T. Nguyen, L. Le, and Q. Le-Trung, “Computation offloading in MI-
code offload,” in Proc. IEEE Int. Conf. Mobile Syst. Appl. Services MO based mobile edge computing systems under perfect and imperfect
(MobiSys), San Francisco, CA, USA, Jun. 2010, pp. 49–62. CSI estimation,” IEEE Trans. Services Comput., pp. 1–1, 2019.
[11] B.-G. Chun, S. Ihm, P. Maniatis, M. Naik, and A. Patti, “Clonecloud: [33] K. Wang, T. Chang, W. Ma, and C. Chi, “A semidefinite relaxation based
Elastic execution between mobile device and cloud,” in Proc. EuroSys, conservative approach to robust transmit beamforming with probabilistic
Salzburg, Austria, Apr. 2011, pp. 301–314. SINR constraints,” in Proc. European Signal Process. Conf., Aug. 2010,
[12] S. Kosta, A. Aucinas, P. Hui, R. Mortier, and X. Zhang, “Thinkair: pp. 407–411.
Dynamic resource allocation and parallel execution in the cloud for [34] B. Hassibi and B. M. Hochwald, “How much training is needed in
mobile code offloading,” in Proc. IEEE Int. Conf. Comput. Commun. multiple-antenna wireless links?” IEEE Trans. Inf. Theory, vol. 49, no. 4,
(INFOCOM), Orlando, FL, USA, Mar. 2012, pp. 945–953. pp. 951–963, Apr. 2003.
[13] J. Liu, T. Zhao, S. Zhou, Y. Cheng, and Z. Niu, “Concert: a cloud- [35] T. Van Chien, E. Bjornson, and E. G. Larsson, “Joint power allocation
based architecture for next-generation cellular systems,” IEEE Wireless and user association optimization for massive MIMO systems,” IEEE
Commun., vol. 21, no. 6, pp. 14–22, Dec. 2014. Trans. Wireless Commun., vol. 15, no. 9, pp. 6384–6399, Sep. 2016.
[14] W. Zhang, Y. Wen, K. Guan, D. Kilper, H. Luo, and D. O. Wu, “Energy- [36] C. Pan, H. Ren, M. Elkashlan, A. Nallanathan, and L. Hanzo, “Weighted
optimal mobile cloud computing under stochastic wireless channel,” sum-rate maximization for the ultra-dense user-centric TDD C-RAN
IEEE Trans. Wireless Commun., vol. 12, no. 9, pp. 4569–4581, Sep. downlink relying on imperfect CSI,” IEEE Trans. Wireless Commun.,
2013. vol. 18, no. 2, pp. 1182–1198, Feb. 2019.
[15] Y. Wang, M. Sheng, X. Wang, L. Wang, and J. Li, “Mobile-edge com- [37] S. Melendez and M. P. McGarry, “Computation offloading decisions
puting: Partial computation offloading using dynamic voltage scaling,” for reducing completion time,” in Proc. IEEE Annu. Consum. Commun.
IEEE Trans. Commun., vol. 64, no. 10, pp. 4268–4282, Oct. 2016. Netw. Conf. (CCNC), Las Vegas, NV, USA, Jan. 2017, pp. 160–164.
[16] T. Q. Dinh, J. Tang, Q. D. La, and T. Q. S. Quek, “Offloading in mobile [38] D. P. Palomar and M. Chiang, “A tutorial on decomposition methods
edge computing: Task allocation and computational frequency scaling,” for network utility maximization,” IEEE J. Sel. Areas Commun., vol. 24,
IEEE Trans. Commun., vol. 65, no. 8, pp. 3571–3584, Aug. 2017. no. 8, pp. 1439–1451, Aug. 2006.
[17] Y. Mao, J. Zhang, and K. B. Letaief, “Dynamic computation offloading [39] A. J. Kleywegt, A. Shapiro, and T. Homem-de Mello, “The sample
for mobile-edge computing with energy harvesting devices,” IEEE J. average approximation method for stochastic discrete optimization,”
Sel. Areas in Commun., vol. 34, no. 12, pp. 3590–3605, Dec. 2016. SIAM Journal on Optimization, vol. 12, no. 2, pp. 479–502, 2002.
[18] C. You, K. Huang, and H. Chae, “Energy efficient mobile cloud [40] S. Kim, R. Pasupathy, and S. G. Henderson, “A guide to sample average
computing powered by wireless energy transfer,” IEEE J. Sel. Areas approximation,” in Handbook of simulation optimization. Springer,
Commun., vol. 34, no. 5, pp. 1757–1771, May 2016. 2015, pp. 207–243.
[19] X. Chen, L. Jiao, W. Li, and X. Fu, “Efficient multi-user computation [41] J. Tang, B. Shim, and T. Q. S. Quek, “Service multiplexing and revenue
offloading for mobile-edge cloud computing,” IEEE/ACM Trans. Netw., maximization in sliced C-RAN incorporated with urllc and multicast
vol. 24, no. 5, pp. 2795–2808, Oct. 2016. embb,” IEEE J. Sel. Areas Commun., vol. 37, no. 4, pp. 881–895, Apr.
[20] S. Sardellitti, G. Scutari, and S. Barbarossa, “Joint optimization of radio 2019.
and computational resources for multicell mobile-edge computing,” [42] M. Mahdavi, T. Yang, and R. Jin, “Online stochastic optimization
IEEE Trans. Signal Inf. Process. Netw., vol. 1, no. 2, pp. 89–103, Jun. with multiple objectives,” Advances in Neural Information Processing
2015. Systems, 2013.
[21] X. Lyu, H. Tian, W. Ni, Y. Zhang, P. Zhang, and R. P. Liu, “Energy- [43] A. Liu, V. K. N. Lau, and B. Kananian, “Stochastic successive con-
efficient admission of delay-sensitive tasks for mobile edge computing,” vex approximation for non-convex constrained stochastic optimization,”
IEEE Trans. Commun., vol. 66, no. 6, pp. 2603–2616, Jun. 2018. IEEE Trans. Signal Process., vol. 67, no. 16, pp. 4189–4203, Aug. 2019.
[22] C. Wang, F. R. Yu, C. Liang, Q. Chen, and L. Tang, “Joint computation [44] A. Liu, V. K. N. Lau, and M. Zhao, “Online successive convex
offloading and interference management in wireless cellular networks approximation for two-stage stochastic nonconvex optimization,” IEEE
with mobile edge computing,” IEEE Trans. Veh. Technol., vol. 66, no. 8, Trans. Signal Process., vol. 66, no. 22, pp. 5941–5955, Nov. 2018.
pp. 7432–7445, Aug. 2017. [45] Y. Yang, G. Scutari, D. P. Palomar, and M. Pesavento, “A parallel de-
[23] C. Wang, C. Liang, F. R. Yu, Q. Chen, and L. Tang, “Computation composition method for nonconvex stochastic multi-agent optimization
offloading and resource allocation in wireless cellular networks with problems,” IEEE Trans. Signal Process., vol. 64, no. 11, pp. 2949–2964,
mobile edge computing,” IEEE Trans. Wireless Commun., vol. 16, no. 8, June 2016.
pp. 4924–4938, Aug. 2017. [46] S. Boyd and L. Vandenberghe, Convex optimization. Cambridge, U.K.:
[24] Q. Pham, T. Leanh, N. H. Tran, and C. S. Hong, “Decentralized compu- Cambridge Univ. press, 2004.
tation offloading and resource allocation in heterogeneous networks with [47] Q. Wu, M. Tao, D. W. Kwan Ng, W. Chen, and R. Schober, “Energy-
mobile edge computing.” arXiv: Networking and Internet Architecture, efficient resource allocation for wireless powered communication net-
2018. works,” IEEE Trans. Wireless Commun., vol. 15, no. 3, pp. 2312–2327,
[25] Y. Dai, D. Xu, S. Maharjan, and Y. Zhang, “Joint computation offloading Nov. 2016.
and user association in multi-task mobile edge computing,” IEEE Trans.
Veh. Technol., vol. 67, no. 12, pp. 12 313–12 325, Oct. 2018.
[26] J. Du, L. Zhao, J. Feng, and X. Chu, “Computation offloading and
resource allocation in mixed fog/cloud computing systems with min-
max fairness guarantee,” IEEE Trans. Commun., vol. 66, no. 4, pp.
1594–1608, Apr. 2018.
[27] F. Wang, J. Xu, X. Wang, and S. Cui, “Joint offloading and computing
optimization in wireless powered mobile-edge computing systems,”
IEEE Trans. Wireless Commun., vol. 17, no. 3, pp. 1784–1797, Mar.
2018.
[28] S. Bi and Y. J. Zhang, “Computation rate maximization for wireless
powered mobile-edge computing with binary computation offloading,”
IEEE Trans. Wireless Commun., vol. 17, no. 6, pp. 4177–4190, Jun.
2018.
[29] Y. Mao, J. Zhang, S. H. Song, and K. B. Letaief, “Stochastic joint
radio and computational resource management for multi-user mobile-
edge computing systems,” IEEE Trans. Wireless Commun., vol. 16, no. 9,
pp. 5994–6009, Sep. 2017.
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: Heriot-Watt University. Downloaded on September 22,2020 at 17:20:16 UTC from IEEE Xplore. Restrictions apply.