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
Joint Computation Offloading and Resource
Allocation for MEC-enabled IoT Systems with
Imperfect CSI
Jun Wang, Daquan Feng, Shengli Zhang, An Liu, Senior Member, IEEE, and Xiang-Gen Xia, Fellow, IEEE
Abstract—Mobile edge computing (MEC) is considered as a
promising technology to reduce the energy consumption (EC)
and task accomplishment latency of smart mobile user equipment
(UEs) by offloading computation-intensive tasks to the nearby
MEC servers. However, the quality of experience (QoE) for
computation highly depends on the wireless channel conditions
when computation tasks are offloaded to MEC servers. In this
paper, by considering the imperfect channel state information
(CSI), we study the joint offloading decision, transmit power,
and computation resources to minimize the weighted sum of
EC of all UEs while guaranteeing the probabilistic constraint
in multiuser MEC-enabled Internet of Things (IoT) networks.
This formulated optimization problem is a stochastic mixed-
integer nonconvex problem and challenging to solve. To deal
with it, we develop a low-complexity two-stage algorithm. In
the first stage, we solve the relaxed version of the original
problem to obtain offloading priorities of all UEs. In the second
stage, we solve an iterative optimization problem to obtain a
suboptimal offloading decision. As both stages include solving a
series of nonconvex stochastic problems, we present a constrained
stochastic successive convex approximation based algorithm to
obtain a near optimal solution with low-complexity. Numerical
results demonstrate that the proposed algorithm provides a
comparable performance to existing approaches.
Index Terms—Computation offloading, resource allocation,
stochastic programming, optimization.
I. I NTRODUCTION
W ITH the rapid growth of the Internet of Things (IoT)
devices, various new mobile applications, such as,
online gaming, augmented reality (AR), and face recognition,
are emerging. These applications pose new challenges to the
IoT systems since they not only consume a great number
of computation resources but also have a stringent delay
This work was supported in part by the National Natural Science Foundation
of China (NSFC) under Grant 61701317, Young Elite Scientists Sponsor-
ship Program by CAST under Grant 2018QNRC001, Guangdong Basic
and Applied Basic Research Foundation under Grant 2019B1515130003,
the Innovation Project of Guangdong Educational Department under Grant
2019KTSCX147, the Tencent “Rhinoceros Birds” Scientific Research Foun-
dation for Young Teachers of Shenzhen University, and The Start-up Fund of
Peacock Project. (Corresponding author: Daquan Feng.)
J. Wang, D. Feng, S. Zhang are with The Guangdong Key Laboratory of
Intelligent Information Processing, Shenzhen University, Shenzhen 518060,
China (email: johnwangqc@gmail.com; fdquan@gmail.com; zsl@szu.edu.cn)
A. Liu is with the College of Information Science and Electronic Engineer-
ing, Zhejiang University, Hangzhou 310027, China (email: anliu@zju.edu.cn)
X.-G. Xia is with the Department of Electrical and Computer Engi-
neering, University of Delaware, Newark, DE 19716 USA (e-mail: xi-
anggen@udel.edu)
Copyright (c) 20xx IEEE. Personal use of this material is permitted.
However, permission to use this material for any other purposes must be
obtained from the IEEE by sending a request to pubs-permissions@ieee.org.
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.
1
requirement [1]–[4]. However, IoT devices, i.e., smartphones
and wearable devices, are in general only with limited battery
energy and computation resources due to their small physical
sizes. To tackle this problem, mobile cloud computing (MCC)
has been put forward in the last decade. In MCC systems, IoT
devices offload their computation tasks via wireless links to
the resource-rich remote cloud servers, which can save battery
energy of IoT devices [5].
However, in MCC, the execution delay may be very large
because of long distance and heavy network congestion be-
tween IoT devices and cloud servers. This is unacceptable
for delay-sensitive applications. To address this issue, mobile
edge computing (MEC) has been proposed to deliver the cloud
functionalities, i.e., computing and storage, to the edge of
wireless networks [6] [7]. The existing works [8] [9] have
shown that the computation experience of IoT devices can be
significantly improved when offloading the computation tasks
to MEC servers.
Recently, the MEC paradigm has attracted significant at-
tention from both academia and industry. Earlier works on
MEC have been focused on offloading platforms, such as,
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-
mize computation offloading strategies and resource allocation
for better quality of experience (QoE) of IoT devices. Thus,
computation offloading in MEC systems has recently attracted
increasing attention.
For a single-user MEC, computation offloading strategy has
been widely considered [14]–[17]. In [14], a binary offloading
framework has been proposed for minimizing the energy
consumption (EC) under a stochastic wireless channel by
dynamically adjusting the local CPU frequency for mobile
execution and scheduling the data transmission rate for cloud
execution. Considering the divisibility of computation tasks,
partial offloading decision design has been studied in [15],
where the offloading ratio, local CPU frequency, and transmit
power of smart mobile user equipment (UE) are jointly opti-
mized. In [16], the authors have proposed a binary offloading
scheme to minimize the weighted sum of EC and execution
latency by jointly optimizing offloading decision and UE’s
CPU frequency, where an UE has multiple computation tasks
to offload simultaneously to multiple MEC servers. In [17],
a dynamic Lyapunov optimization based binary offloading
algorithm has been proposed for MEC systems with an energy
harvesting (EH) device.
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
For multi-user MEC, a large number of works have also
been devoted to computation offloading [5] [18]–[31]. In [18],
a joint optimization of partial offloading and resource alloca-
tion has been proposed to minimize the EC for multiuser MEC
systems with TDMA. Different from the partial offloading,
binary offloading designs in multi-user MEC are mixed-integer
optimization problems which are usually nonconvex. In [5]
[19], binary offloading mechanisms based on game theory have
been proposed. In [20], the problem of the total EC minimiza-
tion has been formulated by optimizing transmit power for
multiuser MIMO MEC systems. Note that, in [5] [19] [20],
either offloading decision or resource allocation was optimized
separately. In [21], a joint optimization of offloading decision
and resource allocation has been developed to minimize the
total EC under latency requirements. In [22] [23] [24], a
joint optimization framework of binary offloading decision
and resource allocation in heterogeneous networks (HetNets)
with MEC has been investigated. These works have been
further extended to the multi-task offloading in HetNets in
[25], where the authors have considered the joint computation
offloading, user association and resource allocation problem
and proposed an alternating optimization framework to solve
it. Reference [26] has considered a hierarchical fog-cloud
computing system and studied a min-max fairness problem
under binary offloading decision constraint. Meanwhile, some
works have also considered multiuser MEC systems with EH
[27] [28]. In [27], a joint optimization framework of offloading
ratio, energy beamforming, and computing resource has been
developed to minimize the total EC. Different from [27],
reference [28] has studied the weighted sum of computation
rate maximization by jointly optimizing the binary offloading
and the transmission time allocation. Besides, reference [29]
has investigated stochastic resource allocation in multiuser
MEC systems and proposed a low-complexity Lyapunov opti-
mization based online algorithm. More recently, some works
[30] [31] have proposed offloading scheme based on machine
learning. A reinforcement learning based offloading policy for
an IoT device has been proposed in [30], while an after-
state learning based computation offloading scheme has been
designed for MEC systems with EH in [31].
However, the aforementioned works assume that the perfect
CSI is known at the base station (BS). This assumption may
be unrealistic in practical systems due to the estimation errors,
limited CSI feedback quantization, delays, etc. On the other
hand, the QoE of the computation heavily relies on the wireless
fading channel conditions since task offloading requires effec-
tive wireless transmission. The authors in [32] have studied the
fairness problem for MEC systems with imperfect CSI with
the goal to minimize the maximum weighted EC under the
upper-bound of the average latency constraint. Different from
it, our goal is to minimize the weighted sum of EC of UEs
while guaranteeing the probabilistic delay constraint for MEC
systems with imperfect CSI, which seeks to provide “safe”
performance guaranteeing for a certain probability (often high)
of satisfying the delay requirements. Our formulated problem
is mixed-integer nonconvex stochastic optimization problem
under the probabilistic constraint which is more difficult to
solve. Note that the authors in [33] have presented a conserva-
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.
2
tive approach to solve the robust transmit beamforming design
problem under the probabilistic signal-to-interference-plus-
noise ratio (SINR) constraint for multiuser wireless systems.
Their idea is to replace the difficult probabilistic constraint
by the worst-case constraint with spherically bounded channel
errors. However, this idea can not be applied to our formulated
problem, where the objective function involves the expectation
and has no closed-form expression. To deal with this issue, we
propose a low-complexity two-stage optimization framework
based on constrained stochastic successive convex approxi-
mation (CSSCA) to solve the formulated problem. The main
contributions of this paper are summarized as follows:
• By taking imperfect CSI into account, we formulate the
average weighted sum of EC minimization problem under
the probabilistic delay constraint, and jointly optimize the
offloading decision, UEs’ transmit power, and computa-
tion resource allocation.
• To deal with the mixed-integer nature of the formulated
problem, we develop a novel low-complexity optimiza-
tion framework. It includes two stages. In the first stage,
we solve the relaxed version of the original problem
to obtain offloading priorities of all UEs. In the sec-
ond stage, we iteratively solve a stochastic optimization
problem to obtain a suboptimal offloading decision and
corresponding resource allocation.
• Both stages include solving a series of nonconvex s-
tochastic problems. One of the difficulties in the stochas-
tic optimization problem is that the objective function in-
volves the expectation and has no closed-form expression.
To tackle this challenge, we propose a CSSCA-based
iterative algorithm, in which a quadratic convex problem
is solved at each iteration by a low-complexity algorithm
based on the dual decomposition method. In addition, we
have also analyzed the convergence and complexity of the
proposed algorithm.
• We perform extensive numerical simulations to evaluate
the convergence of the proposed iterative algorithm and
compare the performance of the proposed solution with
the existing approaches.
The remaining of this paper is organized as follows. In
Section II, the system model is firstly described and the opti-
mization problem is then formulated. In Section III, the low-
complexity framework to solve the mixed-integer nonconvex
stochastic optimization problem is presented. Convergence and
complexity analysis of the proposed algorithm are shown in
Section IV. In Section V and Section VI, simulation results
and conclusion are, respectively, presented.
Notation: E [·] presents the mathematical expectation. Pr {·}
b ∆
denotes the probability operator. [x]a = min {b, max {a, x}}.
II. S YSTEM M ODEL AND P ROBLEM T RANSFORMATION
In this section, we introduce the communication model and
computation model of a multiuser MEC system, followed by
the optimization problem formulation.
A. Communication Model
Consider a multiuser MEC system described in Fig. 1,
where K UEs are served by a BS equipped with the MEC
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
Task 1
Task k
Task K
MEC
Server BS
Fig. 1. Multiuser MEC system.
server. For simplicity, we will use the MEC server and BS
interchangeably. Denote the set of UEs as K = {1, 2, ..., K}.
Without loss of generality, assume that each UE has only one
computation-intensive task which is required to be accom-
plished within a certain time.
To avoid the mutual interference among UEs, we suppose
that each UE occupies an orthogonal quasi-static wireless
channel when offloading its task. Let hk be the channel fading
coefficient between UE k and the BS. However, in practical
wireless systems, the accurate CSI at BS is difficult to obtain
owing to the channel estimation error or the feedback delay.
Thus, in this paper, the actual CSI hk is modeled as:
hk = b
hk + ek ,
where b hk is a channel estimate and ek is the estimation
error. Assume that ek obeys a circularly symmetric

complex
Gaussian distribution, i.e., ek ∼ CN 0, σe2k . Then, the
received signal associated with UE k is given by
√ √
yk = pk b hk sk + pk ek sk + nk , (2)
| {z }
n0k
where pk is the transmit power hof UEik, sk is the transmitted
2
symbol from UE k satisfying E |sk | = 1, nk is the additive
white Gaussian noise with power spectral density N0 , n0k is
the effective noise, which combines the additive noise and
residual channel estimation error and its variance is
h i
2 2 2
E |n0 k | = pk |ek | + σk2 = pk |ek | + θk BN0 , (3)
where B is the total bandwidth, and θk is the normalized
portion of the bandwidth to UE k.
As in [34] [35] [36], since the effective noise n0k is neither
independent nor Gaussian, the effective signal to interference
plus noise ratio (SINR) for UE k is bounded below by
Things Journal
3
B. Computation Model
UE 1 The task of UE k to be processed can be characterized
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
the maximum tolerable delay for the task. The parameters Lk ,
UE k
Ck , and Tkmax can be measured by task profilers [37]. Assume
 that each task is atomic, i.e., cannot be further partitioned. That
is, it can be either locally processed at the UE or transferred


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
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.
1) Local Computing: The processing time Tkloc for local
task executing can be written as
Lk Ck
Tkloc = loc , (6)
fk
where fkloc is the CPU computational capability of UE k (in
CPU cycles/s).
As in [14] [16], the EC Ekloc for local task executing can
be expressed as

3
2
Ekloc = βk fkloc Tkloc = βk fkloc Lk Ck , (7)

3
where βk fkloc is the computational power of UE k, and βk
is a constant related to chip architecture [15] [16].
(1) 2) Remote Computing: When UE k decides to offload its
task to the MEC server for processing, the offloading process
includes the following three stages. Firstly, UE k sends task
input data to the MEC server over the wireless channel. Then,
the MEC server processes task k by allocating part of the
computation resources to it. Finally, the MEC server sends
back the task output data to UE k after completing the task.
Accordingly, the uplink transmission time can be expressed
by
Lk
Tktr = . (8)
Rk (pk , θk , ek )
The remote computation time can be given by
Lk Ck
Tkexe = , (9)
wk Fsmax
where wk Fsmax denotes the computation resources (in CPU
cycles/s) allocated to UE k, wk is the normalized portion of
the computation resource assigned to UE k, and Fsmax presents
the maximum computation resource of the MEC server.
As in [18] [27] [28], the size of task output data is, in
general, much smaller than that of task input data. Thus, we
neglect the time consumed by returning the task output data to
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
 
 
 
 
 
 

The key idea of the CSSCA is to convert the complicated where

nonconvex stochastic optimization problem into an iterative
optimization problem. At each iteration, variable z is updated  
by solving a simple convex problem which is obtained by ∆ Lk Ck
Θck (zk ) = (1 − xk ) − Tkmax − xk ln ξkmax (31)
replacing the nonconvex functions Φk (zk ) and Θk (zk ) with fkloc
their corresponding convex surrogate functions.
Let ϕck (zk ) and ϕck (zk , ek ) denote convex and nonconvex
parts of ϕk (zk , ek ), respectively. Then, ϕk (zk , ek ) can be and
expressed as
  2 
ϕk (zk , ek ) = ϕck (zk ) + ϕck (zk , ek ) , (21) hk 
b 
∆ xk  θk BN0
Θck (zk ) = −  (32)

Lk
σe2k pk

where 



max −
θk B Tk


Lk Ck







wk Fsmax
 

−1
 

2 2
 
 

∆
 
 

ϕck (zk ) = αk (1 − xk ) βk fkloc Lk Ck ,

 
 

(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
 
 

(−1) (−1) are given by (35) with nk =2 .

 
 
 

= 0, and ρ(i) ∈ (0, 1]

 
 

with the initial values = 0, Φk uk

 

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

with closed-form update at each iteration. Proof: See Appendix B. 

(i) (i)
Due to the similar structures of problems P2 and P3 , we On the other hand, the dual variables (µ, η, λ) can be
(i) updated by adopting the subgradient method, which are given
only focus on solving P2 in the remainder of this subsection
for brevity. In what follows, we illustrate how to exploit the as follows:
(i)
Lagrange dual decomposition method to solve P2 . The partial h i+
(i) b (i) (z∗k ) ,
µk (l + 1) = µk (l) + δµ,l Θ (49)
Lagrangian of P2 is written as k
" K
!#+
K K X
X
b (i) (zk )+
X
b (i) (zk ) η (l + 1) = η (l) + δη,l θk∗ − 1 , (50)
L(i) (z, µ, η, λ) = Φ k µk Θ k k=1
k=1 k=1 !#+
K
! K
! (39) " K
X
λ (l + 1) = λ (l) + δλ,l wk∗ −1 , (51)
X X
+η θk − 1 +λ wk − 1 ,
k=1
k=1 k=1

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

TABLE I: SIMULATION PARAMETERS 1.8

K=12, F max
k
=10
1.6
Notation Parameter Value
K=6, F max
k
=5
r Cell radius 500 m 1.4

Weighted sum of EC (J)

B Total bandwidth 10 MHz 1.2

K Number of UEs {6 − 20} 1

Pkmax Maximum transmit power 23 dBm

0.8
N0 Noise power density −174 dBm/Hz
0.6
σe2 Variance of estimation error 0.002
k
0.4
Lk Size of task input data 0.42 MB
Ck Density of task computation 297.6 cycles/bit 0.2

Tkmax Maximum tolerable delay {0.9 − 2.5} s 0

5 10 15 20 25 30 35 40 45
F max Maximum computation resource of MEC {5 − 20} GHz Number of Iterations

fkloc Local computation resource 1.2 GHz Fig. 2. Convergence of Algorithm 1

βk Energy consumption coefficient
max
ξk Maximum delay outage probability
max
that Algorithm 1 requires Iinner iterations to converge, the
max
complexity of Algorithm 1 is at most O(Iinner K) . Thus, the
max
total complexity of Algorithm 2 is at most O(Iinner K 2 ).
As one of the important benchmarks, the complexity of the
robust method in [32] is O(N2OA 2K + N1bis K 3.5 N3feas ),

weighted EC under the upper bound of the average latency
where N1bis , N2OA , and N3feas are the number of bisection
search, the number of alternating optimization of outer loop,
and the number of iteration of feasibility verification, respec-
tively. Obviously, the complexity of this algorithm is much
higher than that of the proposed algorithm.
V. N UMERICAL S IMULATION
In this section, numerical simulations are presented to
evaluate the performance of the proposed ranking-based binary
offloading and resource allocation algorithm (RBORA). We
consider a multiuser MEC system consisting of a BS and
K UEs. The BS is in the center of a circle where UEs
are randomly distributed. The large-scale fading model is
128.1 + 37.5 log10 r[Km] (dB), and the log-normal shadowing
standard deviation 10 dB. For simplicity, we assume that
the estimated channel

hk ∼ CN (0, 1) and estimation error
b maximum tolerable delay Tkmax for two cases σe2k = 0.002
2
ek ∼ CN 0, σek . For comparison, the weight of each UE
is αk = 1. Other default simulation parameters are shown in
Table I, if not specified.
We compare our proposed RBORA algorithm with the
following benchmarks.
1) Exhaustive search (ES): The optimal offloading decision
variable is obtained by exhaustive search. Given the offloading
decision, we optimize other variables by using Algorithm
1. This provides a performance upper bound for practical
schemes.
2) Nonrobust scheme (NonR): When the estimated chan-
nels are used as the perfect CSI, the offloading decision
and resource allocation are jointly optimized to minimize
the weighted sum of EC, which is a combination of works
[21] and [26]. Note that the nonrobust method is designed
without considering the channel errors, this will result in the
dissatisfaction of delay constraint (15).
3) Nonrobust scheme with bisection search (NonR-bisec):
Due to the dissatisfaction of the delay constraint in the NonR,
1 × 10−28
10%
the bisection search procedure [33] is applied to the nonro-
bust method, which will make the delay constraint satisfied.
Readers are referred to [33] for the details.
4) Robust scheme (Robust): The offloading decision and
resource allocation (including transmit power and computation
resource) are jointly optimized to minimize the maximum
and resource constraints, as in [32].
5) Local execution only (Local-only): All tasks are always
locally executed without any optimization.
A. Convergence Behavior of Algorithm 1
To verify the convergence of the proposed algorithm 1 (RO-
DRA), Fig. 2 plots the objective function in problem P1 versus
the number of iterations for two cases K = 6, Fsmax = 5 GHz
, and Tkmax =1.5 s and K = 12, Fsmax = 10 GHz, and Tkmax
=1.5 s, respectively. Observe that the algorithm 1 converges
to a stationary point at most 30 iterations.
B. Performance Comparison
Figs. 3a and 3b show the weighted sum of EC comparison
of the proposed RBORA and benchmarks versus different
and σe2k = 0.005, respectively. From the figures, we can see
that the weighted sum of EC of the RBORA is close to that
of the ES, especially when Tkmax is smaller. This is because
a tighter delay requirement leads to almost the same set of
offloading UEs and so the weighted sum of EC is almost
the same. But compared to the ES, the RBORA has a lower
complexity especially when K is very large. In addition, com-
pared to the NonR, the RBORA consumes more energy, since
it accounts for the channel estimation error and guarantees
the delay outage requirement. However, the RBORA scheme
outperforms the NonR-bisec and the Robust in terms of the
weighted sum of EC. For the Robust, it is because that this
method aims to minimize the maximum weighted EC instead
of the overall EC. Furthermore, comparing Figs. 3a and 3b, it
can be observed that when σe2k increases, the RBORA takes
more energy to meet the delay outage requirement.
To highlight the advantages of the proposed RBORA al-
gorithm, Fig. 4 depicts the cumulative distribution function
(CDF) of the realistic task execution delay of all UEs for

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

Weighted sum of EC (J)

NonR-bisec
Local-only 0.6
Robust
0.5
Local-only
0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0
0.9 1.1 1.3 1.5 1.7 1.9 2.1 2.3 2.5
0
Maximum tolerable delay T max
k
(s) 5 7 9 11 13 15 17 19
Maximum computation resource F max
s
(GHz)
(b) σe2k = 0.005
(a) Tkmax = 1 s
Fig. 3. Weighted sum of EC versus maximum tolerable delay
0.9

0.8
RBORA

the case K = 6, = 5 GHz, and Fsmax

= 1.5 s. We Tkmax 0.7
ES
NonR
Weighted sum of EC (J)

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

probability. In particular, for the Robust, the reason is that 0.4

it is constrained by the upper bound of the average delay, 0.3

which leads to that the realistic task execution delay of all
0.2
UEs is almost smaller than Tkmax . However, for the NonR,
0.1
only a few UEs meet the prescribed task execution delay
with a very high probability. This implies that the lower the 0
5 7 9 11 13 15 17 19

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

Weighted sum of EC (J)

Local-only
0.6
1.5
CDF

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

(b) Tkmax = 2 s (a) Weighted sum of EC

Fig. 6. CDFs of realistic task execution delay Tk 14

RBORA
NonR
12 NonR-bisec
1.8 Robust
RBORA Local-only
Number of offloading 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

stringent delay requirement, only some UEs prefer to offload

Fig. 8 depicts the CDF of the realistic task execution delay their tasks to the MEC server with the limited computation
of all UEs, which corresponds to Fig. 7 for the case L = 0.5 resource, which implies the rest of UEs can only process their
MB. We can observe the same result as in Fig. 4 and Fig. 6. tasks locally and thus cause a rapid increase in terms of the
Figs. 9a and 9b show the influence of the number of UEs weighted sum of EC.
on the offloading performance for the case Fsmax = 10 GHz Fig. 10 depicts the CDF of the realistic task execution delay
and Tkmax = 1.6 s. Since the ES has high computational of all UEs for the case K = 11, which corresponds to Fig. 9.
complexity in MEC systems with a large number of users, The phenomenon is consistent with the observations in Fig. 4,
we ignore it here. From Fig. 9a, we can see that the weighted Fig. 6, and Fig. 8.
sum of EC increases with the number of UEs for all schemes.
Besides, the weighted sum of EC of all schemes except the VI. C ONCLUSION
Local-only first increases slowly and then increase rapidly. The In this paper, we have studied the joint offloading de-
reason can be explained by Fig. 9b. As shown in Fig. 9b, due to cision, transmit power, and computation 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
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
 
 
 
 
 
 

Combining the above two cases, we can obtain the equivalent

problem to minimize the average weighted sum of EC while constraint in (15).
meeting the probabilistic delay constraint in the multiuser
MEC-enabled IoT network with imperfect CSI. To solve the
formulated mixed-integer nonconvex stochastic problem, we A PPENDIX B
develop a novel low-complexity framework. This framework P ROOF OF L EMMA 2
includes two stages. In the first stage, we solve the relaxed Proof: Problem (44) can be further decomposed into three
version of the original problem to obtain offloading prior- quadratic subproblems
ities of all UEs. In the second stage, we iteratively solve
a stochastic optimization problem to obtain a suboptimal min ak x2k + bk,1 xk (57)
xk
offloading decision and corresponding resource allocation. We s.t. (18b),
have also analyzed the convergence and complexity of the
proposed algorithms. Our numerical results have confirmed the min ak p2k + bk,2 pk (58)
pk
advantages of the proposed algorithms over existing methods s.t. (17c),
in terms of the weighted sum of EC and the violation of the
min ak θk2 + bk,3 θk (59)
delay constraint. θk
s.t. (17d),
A PPENDIX A min ak wk2 + bk,4 wk (60)
wk
P ROOF OF L EMMA 1
s.t. (17f).
Lk Ck
When xk =0, Tk (xk , pk , θk , wk , ek ) = Obviously, it fkloc
.
Because (57)–(60) are quadratic problems with one-
does not contain the random variable ek and thus the left dimensional variable, the optimal solutions only appear at
hand side of (14) can only be equal to 0 or 1. Therefore, three points, i.e., two endpoints of a feasible set and the
if the inequality (14) holds, the left hand side of it must be stationary point of objective function. So we can express their
equal to 1, that is optimal solutions as in (45)–(48), respectively. 
Prek {Tk (xk , pk , θk , wk , ek ) ≤ Tkmax } = 1 (53)
R EFERENCES
In this case, (14) is reduced as [1] M. Chiang and T. Zhang, “Fog and IoT: An overview of research
opportunities,” IEEE Internet Things J., vol. 3, no. 6, pp. 854–864, Dec.
Lk Ck
≤ Tkmax . (54) 2016.
fkloc [2] J. Pan and J. McElhannon, “Future edge cloud and edge computing for
internet of things applications,” IEEE Internet Things J., vol. 5, no. 1,
pp. 439–449, Feb. 2018.
When xk =1, Tk (xk , pk , θk , wk , ek ) = Rk (pkL,θk k ,ek ) + Lk Ck
wk F max , [3] J. Wang, D. Feng, S. Zhang, J. Tang, and T. Q. S. Quek, “Computation
thus, the probability in (14) is calculated as offloading for mobile edge computing enabled vehicular networks,”
IEEE Access, vol. 7, pp. 62 624–62 632, 2019.
Prek {Tk (xk , pk , θk , wk , ek ) ≤ Tkmax } [4] B. Cao, L. Zhang, Y. Li, D. Feng, and W. Cao, “Intelligent offloading in
  2  multi-access edge computing: A state-of-the-art review and framework,”
IEEE Commun. Mag., vol. 57, no. 3, pp. 56–62, Mar. 2019.
hk 
 b  

2
 θk BN0  [5] X. Chen, “Decentralized computation offloading game for mobile cloud
=Prek |ek | ≤ Lk − computing,” IEEE Trans. Parallel Distrib. Syst., vol. 26, no. 4, pp. 974–
 B
1
max − Lk Ck
pk 
 Tk
wk Fsmax
 983, Apr. 2015.
2 − 1 [6] E. T. S. I. (ETSI), “Mobile-edge computing-introductory technical white
   2 
paper,” Sep. 2014.
h
 b  
 1   k  θk BN0  [7] P. Yang, N. Zhang, S. Zhang, L. Yu, J. Zhang, and X. Shen, “Content
2
=Preek |e
e k | ≤ 2  Lk − popularity prediction towards location-aware mobile edge caching,”
σ ek 1
pk

B max − Lk Ck
 
 Tk
wk Fsmax
 IEEE Trans. Multimedia, vol. 21, no. 4, pp. 915–929, Apr. 2019.
2 −1

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] K. Kumar, J. Liu, Y. Lu, and B. Bhargava, “A survey of computation
offloading for mobile systems,” Mobile Netw. Appl., vol. 18, no. 1, pp.
129–140, Feb. 2013.
[9] W. Shi, J. Cao, Q. Zhang, Y. Li, and L. Xu, “Edge computing: Vision
and challenges,” IEEE Internet Things J., vol. 3, no. 5, pp. 637–646,
Oct. 2016.
[10] E. Cuervo, A. Balasubramanian, D. Cho, A. Wolman, S. Saroiu,
R. Chandra, and P. Bahl, “Maui: Making smartphones last longer with
code offload,” in Proc. IEEE Int. Conf. Mobile Syst. Appl. Services
(MobiSys), San Francisco, CA, USA, Jun. 2010, pp. 49–62.
[11] B.-G. Chun, S. Ihm, P. Maniatis, M. Naik, and A. Patti, “Clonecloud:
Elastic execution between mobile device and cloud,” in Proc. EuroSys,
Salzburg, Austria, Apr. 2011, pp. 301–314.
[12] S. Kosta, A. Aucinas, P. Hui, R. Mortier, and X. Zhang, “Thinkair:
Dynamic resource allocation and parallel execution in the cloud for
mobile code offloading,” in Proc. IEEE Int. Conf. Comput. Commun.
(INFOCOM), Orlando, FL, USA, Mar. 2012, pp. 945–953.
[13] J. Liu, T. Zhao, S. Zhou, Y. Cheng, and Z. Niu, “Concert: a cloud-
based architecture for next-generation cellular systems,” IEEE Wireless
Commun., vol. 21, no. 6, pp. 14–22, Dec. 2014.
[14] W. Zhang, Y. Wen, K. Guan, D. Kilper, H. Luo, and D. O. Wu, “Energy-
optimal mobile cloud computing under stochastic wireless channel,”
IEEE Trans. Wireless Commun., vol. 12, no. 9, pp. 4569–4581, Sep.
2013.
[15] Y. Wang, M. Sheng, X. Wang, L. Wang, and J. Li, “Mobile-edge com-
puting: Partial computation offloading using dynamic voltage scaling,”
IEEE Trans. Commun., vol. 64, no. 10, pp. 4268–4282, Oct. 2016.
[16] T. Q. Dinh, J. Tang, Q. D. La, and T. Q. S. Quek, “Offloading in mobile
edge computing: Task allocation and computational frequency scaling,”
IEEE Trans. Commun., vol. 65, no. 8, pp. 3571–3584, Aug. 2017.
[17] Y. Mao, J. Zhang, and K. B. Letaief, “Dynamic computation offloading
for mobile-edge computing with energy harvesting devices,” IEEE J.
Sel. Areas in Commun., vol. 34, no. 12, pp. 3590–3605, Dec. 2016.
[18] C. You, K. Huang, and H. Chae, “Energy efficient mobile cloud
computing powered by wireless energy transfer,” IEEE J. Sel. Areas
Commun., vol. 34, no. 5, pp. 1757–1771, May 2016.
[19] X. Chen, L. Jiao, W. Li, and X. Fu, “Efficient multi-user computation
offloading for mobile-edge cloud computing,” IEEE/ACM Trans. Netw.,
vol. 24, no. 5, pp. 2795–2808, Oct. 2016.
[20] S. Sardellitti, G. Scutari, and S. Barbarossa, “Joint optimization of radio
and computational resources for multicell mobile-edge computing,”
IEEE Trans. Signal Inf. Process. Netw., vol. 1, no. 2, pp. 89–103, Jun.
2015.
[21] X. Lyu, H. Tian, W. Ni, Y. Zhang, P. Zhang, and R. P. Liu, “Energy-
efficient admission of delay-sensitive tasks for mobile edge computing,”
IEEE Trans. Commun., vol. 66, no. 6, pp. 2603–2616, Jun. 2018.
[22] C. Wang, F. R. Yu, C. Liang, Q. Chen, and L. Tang, “Joint computation
offloading and interference management in wireless cellular networks
with mobile edge computing,” IEEE Trans. Veh. Technol., vol. 66, no. 8,
pp. 7432–7445, Aug. 2017.
[23] C. Wang, C. Liang, F. R. Yu, Q. Chen, and L. Tang, “Computation
offloading and resource allocation in wireless cellular networks with
mobile edge computing,” IEEE Trans. Wireless Commun., vol. 16, no. 8,
pp. 4924–4938, Aug. 2017.
[24] Q. Pham, T. Leanh, N. H. Tran, and C. S. Hong, “Decentralized compu-
tation offloading and resource allocation in heterogeneous networks with
mobile edge computing.” arXiv: Networking and Internet Architecture,
2018.
[25] Y. Dai, D. Xu, S. Maharjan, and Y. Zhang, “Joint computation offloading
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.
13
[30] Z. Wei, B. Zhao, J. Su, and X. Lu, “Dynamic edge computation
offloading for internet of things with energy harvesting: A learning
method,” IEEE Internet Things J., vol. 6, no. 3, pp. 4436–4447, Jun.
2019.
[31] M. Min, L. Xiao, Y. Chen, P. Cheng, D. Wu, and W. Zhuang, “Learning-
based computation offloading for IoT devices with energy harvesting,”
IEEE Trans. Veh. Technol., vol. 68, no. 2, pp. 1930–1941, Feb. 2019.
[32] T. T. Nguyen, L. Le, and Q. Le-Trung, “Computation offloading in MI-
MO based mobile edge computing systems under perfect and imperfect
CSI estimation,” IEEE Trans. Services Comput., pp. 1–1, 2019.
[33] K. Wang, T. Chang, W. Ma, and C. Chi, “A semidefinite relaxation based
conservative approach to robust transmit beamforming with probabilistic
SINR constraints,” in Proc. European Signal Process. Conf., Aug. 2010,
pp. 407–411.
[34] B. Hassibi and B. M. Hochwald, “How much training is needed in
multiple-antenna wireless links?” IEEE Trans. Inf. Theory, vol. 49, no. 4,
pp. 951–963, Apr. 2003.
[35] T. Van Chien, E. Bjornson, and E. G. Larsson, “Joint power allocation
and user association optimization for massive MIMO systems,” IEEE
Trans. Wireless Commun., vol. 15, no. 9, pp. 6384–6399, Sep. 2016.
[36] C. Pan, H. Ren, M. Elkashlan, A. Nallanathan, and L. Hanzo, “Weighted
sum-rate maximization for the ultra-dense user-centric TDD C-RAN
downlink relying on imperfect CSI,” IEEE Trans. Wireless Commun.,
vol. 18, no. 2, pp. 1182–1198, Feb. 2019.
[37] S. Melendez and M. P. McGarry, “Computation offloading decisions
for reducing completion time,” in Proc. IEEE Annu. Consum. Commun.
Netw. Conf. (CCNC), Las Vegas, NV, USA, Jan. 2017, pp. 160–164.
[38] D. P. Palomar and M. Chiang, “A tutorial on decomposition methods
for network utility maximization,” IEEE J. Sel. Areas Commun., vol. 24,
no. 8, pp. 1439–1451, Aug. 2006.
[39] A. J. Kleywegt, A. Shapiro, and T. Homem-de Mello, “The sample
average approximation method for stochastic discrete optimization,”
SIAM Journal on Optimization, vol. 12, no. 2, pp. 479–502, 2002.
[40] S. Kim, R. Pasupathy, and S. G. Henderson, “A guide to sample average
approximation,” in Handbook of simulation optimization. Springer,
2015, pp. 207–243.
[41] J. Tang, B. Shim, and T. Q. S. Quek, “Service multiplexing and revenue
maximization in sliced C-RAN incorporated with urllc and multicast
embb,” IEEE J. Sel. Areas Commun., vol. 37, no. 4, pp. 881–895, Apr.
2019.
[42] M. Mahdavi, T. Yang, and R. Jin, “Online stochastic optimization
with multiple objectives,” Advances in Neural Information Processing
Systems, 2013.
[43] A. Liu, V. K. N. Lau, and B. Kananian, “Stochastic successive con-
vex approximation for non-convex constrained stochastic optimization,”
IEEE Trans. Signal Process., vol. 67, no. 16, pp. 4189–4203, Aug. 2019.
[44] A. Liu, V. K. N. Lau, and M. Zhao, “Online successive convex
approximation for two-stage stochastic nonconvex optimization,” IEEE
Trans. Signal Process., vol. 66, no. 22, pp. 5941–5955, Nov. 2018.
[45] Y. Yang, G. Scutari, D. P. Palomar, and M. Pesavento, “A parallel de-
composition method for nonconvex stochastic multi-agent optimization
problems,” IEEE Trans. Signal Process., vol. 64, no. 11, pp. 2949–2964,
June 2016.
[46] S. Boyd and L. Vandenberghe, Convex optimization. Cambridge, U.K.:
Cambridge Univ. press, 2004.
[47] Q. Wu, M. Tao, D. W. Kwan Ng, W. Chen, and R. Schober, “Energy-
efficient resource allocation for wireless powered communication net-
works,” IEEE Trans. Wireless Commun., vol. 15, no. 3, pp. 2312–2327,
Nov. 2016.