Download as docx, pdf, or txt
Download as docx, pdf, or txt
You are on page 1of 68

JING et al.

: LONG-TERM MAX-MIN FAIRNESS GUARANTEE MECHANISM FOR INTEGRATED MULTI-RAT AND MEC
NETWORKS
1

Long-Term Max-Min Fairness Guarantee


Mechanism for Integrated Multi-RAT and MEC
Networks
Zewei Jing , Qinghai Yang , Member, IEEE, Meng Qin , Member, IEEE, Jinglei Li ,
and Kyung Sup Kwak , Senior Member, IEEE

Abstract—Recently, mobile edge computing (MEC) has been


recognized as an emerging paradigm to meet the ever-increasing I. INTRODUCTION

I
computation demands. Besides, multiple radio access technology N RECENT years, the rapid growth of computation-
(multi-RAT) has been proposed to enhance the network throughput
and service reliability. However, the two technologies have been intensive and latency-sensitive mobile applications, such as
evolving independently. In this paper, we will develop an integrated facial recognition, augmented reality, highly-interactive online
multi-RAT and MEC network framework, which enables smart gam- ing, etc., has driven increasing computing demands for a
devices (SDs) to offload computation tasks over multiple RAT wide range of smart devices (SDs) from smartphones to
links in parallel. Considering the resource-limited nature as well various wearable devices [1], [2]. Since SDs are commonly
as the time-varying property of the proposed network, we then
focus on investigating long-term fairness guarantee mechanisms resource- constrained, directly computing these applications
to facilitate fair resource sharing/allocation between SDs. locally might be too time-consuming, and thereby fails to meet
Specifically, we formulate a max-min stochastic optimization their stringent latency requirements. Moreover, the task
problem with the objective of maximizing the minimum long- computing for these computation-intensive applications will
term time-average offloading utility. An adaptive task splitting and consume too much elec- tricity, which also makes it
resource allocation algorithm is proposed based on the Lyapunov
optimization tech- nology, which jointly optimizes the SD task unsuitable for the local comput- ing [3]. To address this
splitting and uplink transmit power allocation, the RAT challenge, mobile cloud computing (MCC) had been
subcarrier allocation, and the MEC computation frequency proposed and applied in industry in the lase decade [4].
allocation. The adaptive algorithm can accommodate to the time- Nevertheless, MCC generally provides applica- tion/task
varying network dynamics without requiring their distribution
information. Moreover, the proposed algorithm is shown to have
offloading services depending on the remote central- ized
polynomial computation complexity and be asymptotically clouds. With the exponentially increasing of task traffic, such
optimal by rigorous analysis. Simulation results verify the traditional computing paradigm could hardly meet the new
theoretical analysis and show that the Jain’s fairness index of the computing demands and exposes some inherent imperfections,
proposed algorithm converges to 1 as the number of time-slots such as severe core network congestion, unsatisfactory end-to-
grows, which outperforms the benchmark algorithms.
end service latency, and so on [5].
Index Terms—Mobile edge computing, multi-RAT, max-min To further address this issue, mobile edge computing
fairness, resource allocation. (MEC) was proposed as a supplementary of MCC by
European Telecommunications Standard Institute (ETSI) [6].
Comparing with MCC, MEC provides task computing service
Manuscript received October 5, 2020; revised December 9, 2020; accepted at the radio access network edge in close proximity to SDs by
February 3, 2021. Date of publication February 17, 2021; date of current version distributed small edge servers, which conduces to lower
April 2, 2021. This work was supported in part by the NSF of China under Grant
61801365, Grant 61971327, Grant 61701371, and Grant 61901319, in part by the service latency. In addition, MEC has awareness of the real-
China Postdoctoral Science Foundation under Grant 2019M663015 and Grant time radio context benefitting from the proximity to SDs. This
2019TQ0210, in part by the National Science Foundation for Young Scientists merit brings the adaptive task offloading as well as the joint
of Shaanxi under Grant 2020JQ-311, in part by the Natural Science Basic
Research Plan in Shaanxi Province of China under Grant 2020JQ-328, in part communication and computation resource allocation into
by the NRF of Korea-Grant under Grant NRF-2020R1A2B5B02002478, and practice in MEC net- works [7]. Although several works had
in part by the ITRC support program under Grant IITP-2019-2014-1-00729. studied the adaptive task offloading problems, they mostly
The review of this article was coordinated by Dr. Kaigui Bian.
(Corresponding author: Meng Qin.) assumed that the SD tasks were offloaded by a single certain
Zewei Jing, Qinghai Yang, and Jinglei Li are with the State Key Laboratory radio access technology (RAT) and overlooked the fairness
of Integrated Services Networks, School of the Telecommunication among SDs. Considering the limited re- sources at the network
Engineering, Collaborative Innovation Center of Information Sensing and
Understanding, Xidian University, Xi’an 710071, China, and also with the edge, fair task offloading and resource al- location mechanisms
Guangzhou In- stitute of Technology, Xidian University, Guangzhou 510555, are of paramount importance. To enhance the task offloading
China (e-mail: zwjing@stu.xidian.edu.cn; qhyang@xidian.edu.cn; efficiency, in this paper, we propose to offload SD tasks in
jlli68@xidian.edu.cn).
Meng Qin is with the School of Electronics and Computer Engineering, parallel by exploiting the multiple RAT (multi-RAT) diversity.
Peking University, and Shenzhen Pengcheng Laboratory, Shenzhen 518055, Based on that, we focus on investigating the fair task
China (e-mail: yaochnqm@gmail.com). offloading and resource allocation for the considered network.
Kyung Sup Kwak is with the Graduate School of Information Technology
and Telecommunications, Inha University, Incheon 402-751, South Korea (e- Multi-RAT convergence has been considered as one of the
mail: kskwak@inha.ac.kr). key technologies for enhancing network throughput and ser-
Digital Object Identifier 10.1109/TVT.2021.3059944 vice reliability in 5G and beyond networks [8]. In multi-RAT

Authorized licensed use limited to: Balochistan University of IT Engineering and MS. Downloaded on June 19,2021 at 11:09:32 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 70, NO. 3, MARCH
2 2021
0018-9545 © 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: Balochistan University of IT Engineering and MS. Downloaded on June 19,2021 at 11:09:32 UTC from IEEE Xplore. Restrictions apply.
SD offloading utility is defined as the long-term time-average
amount of offloaded tasks, which we refer to as the long-term
max-min fairness. In addition, we aim to design a long-term
max-min fairness guarantee mechanism for the integrated multi-
RAT and MEC network, which can achieve an
undifferentiated quality of service in terms of offloading utility
among SDs by an adaptive task splitting and resource allocation
algorithm design. The main contributions of this paper are
summarized as follows:
1) To guarantee the long-term max-min fairness for the
integrated multi-RAT and MEC network, we formulate
a max-min stochastic optimization problem, where the
Fig. 1. The architecture of integrated multi-RAT and MEC network. minimum SD offloading utility is maximized under the
mean rate stability of task queues at MECSs (the defini-
networks, SDs can be equipped with multi-RAT capabilities tion of mean rate stability will be given later) by jointly
(e.g., WiFi, LTE, 5NR, etc.) and integrate these heterogeneous optimizing the task splitting as well as the communication
RATs in different protocol layers. By doing so, SDs can and computation resource allocations.
maintain multiple RAT links simultaneously and split their 2) To make the problem easy to solve, we first transform
task/traffic flow into multiple subflows and transmit these the formulated max-min problem into an equivalent max
subflows over different RAT links, so that the network version by introducing an auxiliary variable for each
throughput and service reliability can be improved by such time- slot. Then, an adaptive task splitting and resource
parallel transmission [9]. In fact, new network frameworks for alloca- tion algorithm is proposed to handle the
supporting inter-operability of multi-RATs have been transformed prob- lem by using the Lyapunov
discussed in the literature and standard- ization body. For optimization technology [14], where several decoupled
example, dual-connectivity has been proposed in Release 12 of subproblems are solved at each time-slot without
LTE-A to support mobility-robust communica- tion by requiring priori network information.
connecting SDs to different types of base stations (e.g., macro 3) The proposed algorithm is analyzed to have polyno-
and pico base stations) [10]. Furthermore, to promote the mial computational complexity. Moreover, it achieves
coexistence of LTE and WiFi, LTE assisted access (LAA) an [O(1/V ), O(V )]-tradeoff between the minimum SD
[11] and the LTE-WiFi link aggregation (LWA) [12] have been of- floading utility optimality and task queue backlog.
spec- ified by the 3 rd Generation Partnership Project (3GPP). Sim- ulation results show that the Jain’s fairness index
Even so, multi-RAT has been evolving independently from of the proposed algorithm converges to 1
the MEC context. To merge the two advanced technologies asymptotically, which outperforms some other
together, a new integrated network is desired. benchmark algorithms greatly. Therefore, the long-term
In this paper, we investigate an integrated multi-RAT and max-min fairness can be guar- anteed.
MEC network, which is firstly introduced in our previous The organization of the rest of this paper is given as
work [13], as shown in Fig. 1. In this integrated network, the follows. The related works are given in Section II. We present
user plane protocol stack of each SD comprises a common the integrated multi-RAT and MEC network model and
packet data convergence protocol (PDCP) which is shared by formulate the max-min stochastic optimization problem in
multiple groups of lower-layer RAT protocols. Each group of Section III. In Section IV, we propose the adaptive algorithm
RAT protocols is composed of a radio link control (RLC), a and give the algo- rithm performance. In Section V, we present
medium access control (MAC), and a physical (PHY) simulation results to evaluate the proposed algorithm. Finally,
protocol. Such protocol design enables SDs to execute the task the conclusion is drawn in Section VI.
splitting operation at the PDCP layer and map the subtask
flows onto the lower protocol layers corresponding to each
RAT. As such, the SD tasks can be offloaded to multiple RAT II. RELATED WORKS
base stations and computed by the attached MEC servers Recently, there have appeared great efforts towards task of-
(MECSs) concurrently, and thus the task offloading and floading and resource allocation for MEC networks. The related
computing efficiency can be enhanced. works in the literature can be roughly divided into two cate-
However, the wireless RAT links are commonly spatially- gories: static offloading and adaptive/dynamic offloading.
temporally varying due to the large-scale and small-scale In the static offloading case, network randomness is neglected
chan- nel fading. The SD tasks also arrive randomly. Therefore, and thus deterministic optimization models are commonly
the task splitting for the integrated multi-RAT and MEC adopted. In addition, energy consumption or latency
network should be adaptive to the stochastic network dynamics. minimiza- tion is wildly utilized as the optimization objective.
Besides, to guar- antee the SD fairness under the dynamical For example, in [15], Wang et al. attempted to minimize
network environment with limited resources, the fairness energy consump- tion and task computation latency
guarantee mechanism should be carefully studied. To this end, respectively for single-user task offloading by using convex
we propose to maximize the minimum/worst SD offloading optimization methods. In [16], Cao et al. proposed an user
utility among all SDs, where the cooperation approach for latency- constrained task offloading
to improve energy efficiency under
given channel conditions. Problems with similar optimization TABLE I
objectives were also extended to multi-MECS and multi-user SUMMARY OF NOTATIONS
scenarios. For example, in [17], a low-complexity successive
convex approximation based algorithm was proposed for partial
task offloading in a macro-cell assisted heterogeneous MEC
network. In [18], Tran et al. formulated an energy
consumption and task computation latency tradeoff problem
for multi-MECS and multi-user task offloading and solved it
by a centralized heuristic algorithm.
In the adaptive/dynamic offloading case, the task offloading
and resource allocation strategies should accommodate to the
time-varying network dynamics. Several works have been
con- ducted in this context. In [19], the multi-user task
offloading process was constructed as a non-cooperative exact
potential game without priori network information, which was
proved to converge to the Nash equilibrium via best-response
dynamics. To break the curse of high dimensionality, deep
reinforcement learning based algorithms were proposed in [20]
and [21] respec- tively to learn the optimal stationary task
offloading policies. In [22], the Lyapunov optimization
technology was used for de- signing adaptive task offloading
algorithms for multi-user MEC systems. Furthermore, a two
time-scale Lyapunov optimization based task offloading
algorithm was proposed for ultra-reliable and low-latency
multi-MECS networks in [23].
Although the afore-mentioned works have investigated the
static and adaptive/dynamic task offloading extensively, most
of them assumed that user tasks can only be offloaded by a III. SYSTEM MODEL
single RAT and overlooked the fairness among users. To ensure
fair communication resource allocation, in [24], Zhu et al. As shown in Fig. 1, we consider an integrated multi-RAT
formulated the problem as a fair Nash bargaining game, where and MEC network where a set N of SDs and a set M of base
the existence and uniqueness of the solution were analyzed. stations equipped with different RATs are distributed within
In [25], Chen et al. studied the min-max fairness by an service area. Each RAT base station is attached with a MECS
minimizing the maximum weighted energy consumption and for computing tasks offloaded from SDs. As introduced before,
task computa- tion latency among users. However, the above SDs are endowed with multi-RAT capabilities so that they can
works neglected network randomness. To cooperate network keep connections with multiple RAT base stations
randomness, Lyu et al. [26] derived a proportional fair and simultaneously and offload tasks concurrently to the attached
adaptive algorithm by using partial network information for a MECSs. In addition, we consider a synchronized and time-
single-RAT MEC network. In [27], Kim et al. studied the slotted system protocol model, where the time horizon is divided
application throughput fairness and energy efficiency for a into back-to-back time- slots with equivalent duration of τ . The
multi-mode user device, where the user device is allowed to ∈ T {is denoted} by t = 0, 1, 2 For better
index of time-slots
make the optimal network interface selection according to the understanding, we
channel conditions. To en- hance the task offloading summarize the major notations in Table I.
efficiency, our previous conference work [13] proposed an
integrated multi-RAT and MEC network framework and A. Task Splitting Model
investigated the long-term max-min fairness for the first
At the beginning of each time-slot t, we assume that an
attempt. However, the previous work did not give the detailed
description of the proposed network framework and left out amount of∈ An(t)n [0, Amax] tasks are generated from the
some necessary technical details due to the limited space. This application layer of each SD n,n where Amax is the maximum
current paper is an extended version and serves as a value ∀of ∈ATn(t), t . Note that An(t) is an independently
supplement to the previous one. Specifically, we give a more identically distributed (i.i.d.) stochastic process and this
specific description and explanation of the protocol design of assumption has been wildly adopted and experimentally
our integrated multi-RAT and MEC network. We propose an validated. For instance, the number of poisson task arrivals
per time-slot is i.i.d. as time-slots are uniform in the
adaptive task splitting and resource allocation algorithm with
considered protocol model. We
polynomial computational complexity by using the continuity
relaxation and Lagrange dual decomposition. The algorithm collect all An(t) by vector A(t) = [An(t)]n∈N and represent
the expectation as Aav = E { A(t)} . Similar with [22][26],
performance is rigorously analyzed theoretically. Moreover, we
we provide more simulation results to examine the algorithm suppose that tasks are fine-grained and data-partitioned so that
performance. they can be split into several proportions/segments arbitrarily
and computed independently and in parallel. To exploit
benefits
of the integrated multi-RAT and MEC network, each SD splits rate from SD n to RAT m as
Σ
the arrived tasks at the PDCP layer and then forwards the k
rn,m (t) = n,m (t). (3)
resulted subtasks over different RAT links to distributed
k∈Km
MECSs for par- allel computing. Let an,m(t) denote SD n’s r
subtasks determined
to be forwarded over RAT m at time-slot t based on the In order to ensure the error-free transmission of subtasks,
task splitting operation and matrix a(t) = [an,m(t)]m∈M,n∈N the task splitting should be adaptive to the RAT link capacities
collect the subtasks for all SDs. Then, the following causality at every time-slot. In other words, we need to guarantee
constraint
should be satisfied τ rn,m(t) ≥ an,m(t), ∀m ∈ M,n ∈ N . (4)
Σ Constraint (4) indicates that the amount of subtasks split to
an,m(t) ≤ An(t), ∀n ∈ N . (1) offload over any RAT m for any SD n should be no greater
m∈M than the instantaneous channel capacity between them. This
constraint implies that the optimization of task splitting and
The causality constraint (1) implies that the total amount of tasks
communication resource allocation should be jointly carried out
offloaded by SD n through all RATs at a time-slot should not
to improve the offloading efficiency.
exceed the current task arrival.
C. Task Computation Model
B. Uplink Transmission Model We assume that each MECS is equipped with one single-core
Because of the efficient transmission robustness, orthogonal CPU processor, whose computation frequency can be adjusted
frequency division multiple access (OFDMA) technology has adaptively by using the dynamic voltage scaling technology [31].
been successfully adopted in various wireless standards, includ- After receiving the offloaded tasks, the MECSs will allocate
ing LTE, WiMax, and WiFi [28], and has been determined to be CPU computation frequencies to SDs for task computing. Let
used in 5G era [29]. Therefore, we assume that the OFDMA tech- fn,m(t) represent the CPU computation frequency allocated by
nology is adopted by all RATs in the considered network without MECS m to SD n at time-slot t and f (t) = [fn,m(t)]m∈M,n∈N
loss of generality. The available system spectrum is divided be the computation frequency allocation matrix for all
into MECSs. Then, the amount of tasks that can be computed by
a setK = [ Km]m∈M of subcarriers, where subset K m, ∀m ∈ M MECS m for SD n at time-slot t can be given by
contains the subcarriers occupied by RAT m. Since data fn,m(t)τ
signals
over different RATs are generally modulated on different fre-
quency spectrum, we assume that there is no overlap between
Km, ∀m ∈ M [9]. Denote x(t) = [xn, k (t)]m∈M,n∈N ,k∈Km the bn,m(t) = , (5)
βn
mk
subcarrier allocation matrix, where xn, (t) is a binary indicator where β is the computation workload (CPU cycles per bit)
n
and xkn, (t) = 1 if subcarrier k of RAT m
m is allocated to SD for SD n’s task, which can be acquired by the offline measure-
n at time-slot
m t, otherwise, xn, k
(t) = 0. Since any subcar-
rier
Σ can only be exclusively allocated
m to one SD in OFDMA ments [32].
systems, we have
n∈N xkn,m (t) ≤ 1, ∀m ∈ M,k ∈ Km. We D. Task Queue Model
consider the flat block fading channel model for the uplink
RAT links, namely, the channel power gain (CPG) gkn, (t) The unprocessed tasks of different SDs at each MECS will
from any SD n to any RAT base station m on any subcar- m be buffered in their task queues. We denote Qn,m(t) the task
rier k remains unchanged during one time-slot and has the queue backlog for SD n at MECS m at time-slot t, and let
potential to change in the next one in an i.i.d. manner (more matrix Q(t) = [Qn,m(t)]m∈M,n∈N collect the task queue
general channel models like Markov channel model can also backlogs for all SDs and MECSs. With an,m(t) as the task
be assumed, please see [30] and references therein). We use arrival and bn,m(t) as the task departure for SD n at MECS
m, the dynamics of
matrix g(t) = [gn, k
(t)]m∈M,n∈N ,k∈Km to collect all CPGs. Let Qn,m(t) can be given by
pkn,m (t) be the uplink
m
transmit power allocated by SD n to Qn,m(t + 1) = [Qn,m(t) − bn,m(t)] + an,m(t), (6)
RAT +
m on subcarrier k and p(t) = n, (t)]m∈M,n∈N ,k∈Km be the where [x]+ = max{x, 0}.
[pk m
uplink transmit power matrix, respectively. Then, the achievable
uplink transmit rate from SD n to RAT m on subcarrier k at According to the queue theory [33], the long-term time-
time-slot t can be written as average departure should be no less than the long-term time-
average arrival to guarantee the stability of a queue. Hence,
we
rkk (t) = k
x (t)Bk log2 pk (t)g (t) , (2)
1+
n,m
have
n,m n,m
Bk n, T −1 T −1
m Σ 1 E Σ 1 E
N0 li n, (t)}≥ lim n, (t)} , (7)
m {b m
{a m

where T →∞ T →∞
Bk is the spectrum bandwidth of subcarrier k ∈ Km and T
t=0
T
t=0

N0 represents the single-sided spectral density of the Additive where the expectation operator E{·} is related to the dy-
White Gaussian Noise (AWGN). By aggregating the transmit namic network state information (NSI) s(t) = {A(t), g(t) }
rates on all subcarriers k ∈ Km, we obtain the total transmit and the task splitting and resource allocation policy α(t) =
a(t), x(t), p(t), f (t) . Next, we give the concept of a the current decision actions will impact those in the future; (c)
{ }
tighter queue stability, called the mean rate stability by The objective is designed to achieve the long-term max-min
Definition 1. fairness, which further increases the difficulty of solving the
Definition 1: A queue is defined as mean rate stable if problem. To make it solvable, we first transform the max-min
E {|Qn,m(T )|}
lim = 0. (8) problem into a max problem equivalently by introducing an
T →∞
auxiliary variable for each time-slot, and then we propose an
T
From [14], the queue stability condition in (7) is actually a adaptive algorithm by exploiting the Lyapunov optimization
necessary condition of the mean rate stability (8). In other words, technology, which makes decisions only responding to the
if we can guarantee the mean rate stability of all task queues cur- rent NSI and task queue backlog.
in the network, (7) can be naturally satisfied.
IV. PROBLEM SOLUTION
E. Problem Formulation
A. Problem Transformation
In this subsection, we focus on formulating the max-min
stochastic optimization problem to investigate the long-term To solve problem P1, we show that it can be equivalently
max-min fairness for the integrated multi-RAT and MEC transformed into a stochastic maximization problem by intro-
network. The optimization target is to maximize the mini- ducing an auxiliary variable φ(t) for each time-slot. The
mum/worst long-term time-average offloading utility among details are given in the following lemma.
all SDs subject to the mean rate stability of their task Lemma 1: Problem P1 can be equivalently transformed into
queues. The long-term time-average offloading utility of each problem P2 as follows
ΣT −1 ¯
SD n is
defined
= as U¯ o = limT →∞ 1 E{U o (t)}, where U o (t)

∈M

0 n n P2 : max φ

Σ n T t=
n.
m an,m(t) is the instantaneous offloading utility of SD
As a result, the max-min stochastic optimization problem can
be formulated as follows
P1 : max min U¯ o a(t),x(t),p(t),f (t),φ(t)

n
s.t. C1 − C7,
C8:U¯ o − φ¯ ≥ 0, ∀n ∈ N ,
Σ

n, n
m C9:0 ≤ φ(t) ≤ φmax, (10)
T −1
n
where ¯
φ = limT →∞T t=0 E{φ(t)}
1
and φ ≥
max

max
ma ∈N
n Amax
n. Without
ma
loss of generality, we take
x x

a(t),x(t),p(t),f (t) n∈N

ΣC4:
x
s.t. C1: Σ
All task queues are mean rate stable,
C2: a (t) ≤ A (t), ∀n ∈ N ,
Σ
C5: f (t) ≤ f
Σ , ∀m ∈ M,n,m
C6:
x
n,
m

n, (t), ∀m ∈ M,n ∈ N
m∈M φ = maxn∈N An in this paper.

n, n,
m m

P P
m ,
C3: τ rn,m(t) ≥ P
Proof: Please see Appendix A.

P P

a
Note that problem P2 is easier to handle than P1 even
though
P
P

k
n,
m
k∈Km
(t)pk
(t) ≤ pmax , ∀m ∈
M,n ∈ N ,
it introduces two
extra constraints
C8 and C9.
Constraint C9 is
a simple
instantaneous
constraint. The
difficulty comes
from
constraint C8.
We use the
virtual queue
technology to
tackle
3: max φ¯
a(t),x(t),p(t),f (t),φ(t) splitting and resource allocation (ATSRA) algorithm to solve
s.t. C1 − C7, C9, problem P3 based on the Lyapunov optimization technology.
C10: All virtual queues are mean rate stable. (12) B. ATSRA Algorithm
So far, we have transformed problem 1 into problem 3. In In order to bypass directly optimizing the long-term time-
the next subsection, we will propose an adaptive task average objective of problem P3, the Lyapunov optimization
technology suggests to minimize the upper bound of one-time-
slot Lyapunov drift plus the objective penalty term. Denote Algorithm 1: The ATSRA Algorithm.
Y(t) = [Q(t), Z(t)] the queue state information (QSI) of both 1: At the beginning of each time-slot t ∈ {0, 1, 2, · · · }
task queues and virtual queues, a quadratic Lyapunov function , observe
can be defined as the current NSI s(t) and QSI Y(t).
1 Σ Σ 2 1Σ 2 2: Determine the task splitting and resource
L(Y(t)) = Qn,m(t)+ Zn(t). (13) allocation decision α(t) = {a(t), x(t), p(t), f (t)}
2 m∈M n∈N 2 n∈N and the
auxiliary variable φ(t) by solving the following
problem:
Then, the one-time-slot conditional Lyapunov drift is given by P4 :
Δ(Y(t)) E {L(Y(t + 1)) − L(Y(t))|Y(t)} . (14) Σ
min n (t) (φ(t) − Uno(t))
Z
a(t),x(t),p(t),f (t),φ(t)
Adding the expected objective function of problem P3 into n∈N
(14),
we obtain the conditional Lyapunov drift plus penalty term, as
given by
Σ Σ n, (t) n, (t) − n, (t))−V φ(t)
+ m (a m m
Q b

ΔV (Y(t)) = Δ(Y(t)) − V E {φ(t)|Y(t)} , (15) m∈M n∈N

s.t. C2 − C7, C9. (18)


where V is a positive control parameter, which makes a
tradeoff between the two terms in (15). 3: Update task queues Q(t) and virtual queues Z(t)
according to (6) and (11), respectively.
We then give the upper bound of ΔV (Y(t)) by the following
lemma.
Lemma 2: For any feasible decision policy α(t) =
x(t), p(t), f (t) }, auxiliary variable φ(t), QSI Y(t) and
variable decision subproblem SP 4 −1, computation frequency
{
a(t),
allocation subproblem SP
control parameter V > 0, we have −
4 2, and joint task splitting and
communication resource allocation subproblem SP 4 3. −
ΔV (Y(t)) ≤ B − V E {φ(t)|Y(t)} 1) Auxiliary Variable Decision: Decoupled from as
auxiliary variable decision subproblem can be written P4, the
Σ Σ
+ Qn,m (t)E {an,m (t) − bn,m (t)|Y(t)}
Σ Zn(t) − φ(t)
V
SP −
m∈M n∈N
Σ o
4 1 : min
φ(t)
n∈N
+ Zn (t)E {φ(t) − Un (t)|Y(t)} , (16)
s.t. C9:0 ≤ φ(t) n∈N
≤ φmax. (19)
where Σ Σ } Clearly, problem SP4 − 1 is a single-variable linear program-
1 (t)2 + ming problem, where the optimal φ(t) can be obtained by
B≥ E n, b n, (t)|Y(t)
a2 m m
2 m∈M n∈N ⎧ Σ
1 Σ } ⎪⎨ φmax , if Zn (t) − V ≤ 0,
+ E φ2 (t) + (U o (t))2 |Y(t) . (17)
φ(t) = n∈N (20)
2n
n∈N ⎪0, otherwise.
Proof: Please see Appendix B. ⎩
2) Computation Frequency Allocation: The optimal compu-
The main idea of our approach is to minimize an estimate tation frequency allocation can be obtained by minimizing the
of the right hand of (16) under all instantaneous constraints in fourth term in the objective of problemP4 subject to constraint
problem P 3 at each time-slot, which is helpful for maximizing C5. Moreover, this problem is separable. For each m ∈ , it
the objective while stabilizing all task queues asymptotically can be written as
in the long term. The procedures of the proposed ATSRA M
Σ
algorithm
are summarized in Algorithm 1. Note that the ATSRA algorithm SP 4− 2 : max ρn,m(t)fn,m(t)
fm(t)
makes decisions at each time-slot only based on the current n∈N
NSI and QSI and does not require the statistical distribution of Σ
s.t. C5: n,m(t) ≤ fm
max
,
any random parameter. Therefore, it suffices to solve the
determin-
f
istic problem P4. It can be seen that the optimization variables
n∈N
f (t), { a(t), x(t), p(t) }and φ(t) of problem 4 are independent solving these subproblems are given in the following subsection.
of each other in the both P objective and constraints. Consequently,
it can be naturally decomposed into several decoupled C. Optimal Solutions to Problem P4
subprob- lems which can be solved in parallel. The details of
In this subsection, we give the optimal solutions to problem C11:0 ≤ fn,m(t) ≤ αn,m(t), ∀n ∈ N , (21)
P4 by solving three decoupled subproblems, i.e., auxiliary
where fm(t) = [fn,m(t)]n∈N , ρn,m(t) = Qn,m(t)τ /βn, n∀ ∈
N and αn,m(t) = Qn,m(t)βn/τ, n ∀ ∈ N . Constraint C11 is
added for constricting the allowed computation frequency
SP − 4
allo- cation according to the task queue backlogs. Problem
2 can be generally classified into a linear programming
problem, which can be solved by the typical simplex method.
However, we find SPthat− problem 4 2 has good structural
properties so that the optimal closed-form solution can be
achieved. Specifically,
by interpreting ρn,m(t) as the unit profit of SD n and f max as
m M n(t), ∀n ∈ N are time-varying according to the dynamics
the capacity of a knapsack, problem 4 2 can be treated as
SP −
the linear relaxation of a knapsack problem. According to
of the task queues and virtue queues. Then, at each time-slot t,
we can simply set an,m(t) = 0, xn, k
(t) = 0 and pn,
k
(t) = 0
[34], we can obtain the optimal computation frequency
for any n ∈ N , m ∈ M − M m
n(t), k
∈ m
m. As a result,
allocation by selecting SD with higher unit profit to fulfill
we only needK to optimize an,m(t), n, xk (t) and n,
pk (t)
the knapsack capacity f max. for any n ∈ N ,m ∈ M n(t), k
∈ K m
m in problem
SP 4 − 3.
m
For the msake of expression, we sort SDs in the
We collect the reduced task splitting, subcarrier allocation
descending order of unit profit, i.e., ρi,m(t) ≥ ρj,m(t) for i
and transmit power allocation variables by matrices a˘(t)
> j. Define =
the breaking item as the SD that is firstly not allocated its full [an,m(t)]m∈M,n∈Nm (t), x˘(t) = (t)]m∈M,n∈Nm (t),k∈Km
k
n,
[x
computation frequency. The index of the breaking term w can and p˘(t) = m (t)]m∈M,n∈Nm (t),k∈Km , respectively, for
[pk
be identified, as given by each time-slot t. However, the resulted problem is still non-
iΣ convex and hard to solve. To this end, we will take advantage
w = arg min αj,m(t) > fmax
m . (22) of the continuity relaxation and Lagrange dual
i
j=1
decomposition method [35], [36] to handle this obstacle. For

Then, the optimal computation frequency allocation for simplicity of expression, we omit the time-slot index t in the
MECS following.
1) Continuity Relaxation and Convexity: By relaxing xk
m can be determined
⎧ by
n,m
αn,m (t), if n < into a continuous interval [0,1] and introducing a new variable
w,
s = p n,
x m
Σn−1 (t),
w, if n = (23) , and
k letting ˘s = ]m∈M,n∈Nm ,k∈Km , the
f max
− k
n,
m k k
n,
m
[s
n,
m
f (t) = α

⎪⎨
n,m i=1 i,m relaxed problem of subproblem SP4 − 3 can be written as
m Σ Σ
⎪⎪that, RSP4 − 3 : min ( n, — n ) n,
It is remarkable ⎩ to alleviate the task traffic congestion, a˘,x˘,˘s Qm Z a m
the MECS gives more 0, priority to the SD which
otherwise.
has larger task m∈M n∈Nm

queue backlog and smaller task computation workload when Σ


s.t. C12 : an,m ≤ An, ∀n ∈ N ,
determining the computation frequency allocation. m∈Mn
3) Joint Task Splitting and Communication Resource
Alloca-
tion: Combining the second and third terms in the objective of Σ : τ rn,m (x˘, ˘s) ≥ an,m , ∀m ∈ M, n ∈ Nm ,
C13
problem P4 subject to constraints C2-C4, C6 and C7, the joint C14 :
task splitting and communication resource allocation subprob- ≤ pmax
n,m, ∀m ∈ M,n ∈ Nm,
k
lem can be written as k∈Ksm n,m
Σ Σ
SP4 − 3 : min (Qn,m(t) − Zn(t)) an,m(t) Σ
a(t),x(t),p(t) k
m∈M n∈N C15 : x
n,m ≤ 1, ∀m ∈ M,k ∈ Km,
n∈Nm
Σ
an,m(t) ≤ An(t), ∀n ∈ N , C16 01 (29)
k
s.t. C2: (24)
m∈M :, xn,m ∈ [ ], ∀m ∈ M,n ∈ N m ,k ∈ Km,
C3: τ rn,m (x, p) (t) ≥ an,m(t), ∀m ∈ M,n ∈ N , (25)
where
Σ
C4: n, (t)pn, (t) ≤ pmax
n, , ∀m ∈ M,n ∈ N , rn, (x˘, ˘s)
xk m k m
(26) m m
k∈Km sn,m(t)gn,m(t)
Σ Σ k
k k

C6: ≤ 1, ∀m ∈ M,k ∈ Km, (27)


= xn,m(t)Bk log2 1 xkn,m (t)N0B k
. (30)
n,(t)
m k∈K
n∈N
xk m

It can be noted that the relaxed rate function rn,m (x˘, ˘s) has a
C7: xn,
k
(t) ∈ {0, 1}, ∀m ∈ M,n ∈ N ,k ∈ Km. (28)
m form of classic perspective concave function. Hence, constraint
Problem SP 4− 3 is a non-convex mixed integer nonlinear C13 is a convex constraint. Furthermore, since the objective and
programming (MINLP) problem which is hard to solve. the remainder constraints C12 and C14-C16 are affine, problem
However, we find that the optimal solutions to an,m(t), xk n, (t) RSP4 − 3 is substantially a convex problem and the Slater’s
and pn,k (t) can be achieved at zero if Qn,m(t)— Zn(t) m
0. condition can be satisfied. Hence, there exists a zero dual gap
This inspires us to modify the problem by removing the
m between problem RSP 4− 3 and its dual problem [36].
redundant variables. Specifically, we define a SD set Nm(t) = 2) Lagrange Dual Decomposition Method: By introducing
{n|Qn,m (t) − Zn (t) < 0, n ∈ N } for each RAT m ∈ M, two groups of Lagrange multipliers ω˘ = [ωn,m ]m∈M,n∈Nm and
and a RAT set Mn (t) = {m|Qn,m (t) − Zn (t) < 0, m ∈ M} γ˘ = [γn,m ]m∈M,n∈Nm for relaxing constraints C13 and
for each SD n ∈ N . Note that sets Nm(t), ∀m ∈ M and C14, respectively, we can obtain the Lagrangian of problem
RSP4 −
+
3 as follows ωn,mτ NB
n, 0 k
gk
Bk
Σ Σ −m n,
L(x˘, ˘s, a˘, ω˘ , γ˘) (Q
= n, — ) an,m × ln 2 g m
m Zn log2 1+ k n,
m N0 B k ,
γn,
m
m∈M n∈Nm ∀m ∈ M,n ∈ N ,k ∈ K , (36)
an,m − Σ
m m
+ Σ Σ ωn,m
k
τ n,mBk we rewrite the dual function as
m∈M n∈Nm k∈Km x

k k
log2 1 + sn,m n,m g(ω˘ , γ˘)
k Σ
xn,m N0Bk = min Σ k xn,m
Σ n,
m
Γ
k
Σ Σ Σ k a˘,x˘
max (31) m∈M n∈Nm k∈Km
γ sn,m − pn,m . Λ
mn,
an,m −Σ
Σ γn,m
+ m∈M mn, k∈K max
m
Σ n∈N
Σm n,m p
n∈Nm
Then, the dual function is given by +
m∈M m∈M n∈Nm

g(ω˘ , γ˘) = min (x˘, ˘s, a˘, ω˘ s.t. C12, C15, C16. (37)
, γ˘) L
a˘,x
˘,˘s It can be seen that problem (37) is a classical linear
Σ program- ming problem that is separable with respect ton,
s.t. C12 : an,m ≤ An, ∀n ∈ N an,m and xk m .

,
m∈Mn Combining the first term in the objective of problem (37) and
Σ constraints C15 and C16, the optimal subcarrier allocation can
C15 : xkn,m ≤ 1, ∀m ∈ M,k ∈ Km,
n∈Nm
be determined by

C16 :
∈ [0, 1], ∀m ∈ M,n ∈ N m ,k ∈ Km, (32) ⎪1, if < 0 and
xk n,
m k
Γ
n, n,
m m
}
, ∀m ∈ M,k ∈ Km ,
n,
RSP4 g(ω˘
− 3 is
k
and the dual problem of max
γ˘). ,
⎪⎨
x m= n = arg min Γk
⎪ n∈Nm
ω˘ , (33) ⎪ ⎩
γ˘≥0 0,
otherwise. (38)
It is well known the minimum of the primal problem is
equivalent to the maximum of its dual problem. Therefore, we Since sn,
k
= kn,m pkn,m , the optimal transmit power pn,k
can
can solve problem RSP4 − 3 in the dual domain. m x
be recovered after computing the optimal x n, k
according
m to
By using the Karush-Kuhn-Tucker (KKT) conditions, we can (34). m

achieve the relationship between the optimal n, and xn,


k
with The optimal task splitting can be obtained by combining the
s k m m second term in the objective of problem (37) and constraint C12,
given Lagrange multipliers, as given by as given by
ωn,mτ N0Bk +
skn,m = kn,m − k . (34) ⎧⎪⎨ n
Bk γn,m gn,m
x A , if n, < 0 and
Substituting ln 2 (32), the dual function can be
into problem a Λ m
n, n, m = arg min { n, m
, ∀n ∈ Nm} (39)
m
= ⎩ Λ ,
rewritten as problem
m
(35), shown at the bottom of this page. m∈M

Letting Λn,m = ωn,m + Qn,m − Zn and 0, otherwise.
Γkn,m Given the optimal a , xk and pk , the Lagrange
multipliers
method [36],can be updated
which is given by using the subgradient decent
n,m n, n,m
+
= ωn,
m
τB k − N0 B k — ωn,mτ
m

γn,m γn,m ln 2 g Bk
k
n,m

g(ω˘ , γ˘) = min


Σ
a˘,x˘ Σ (Qn,m − Zn) an,m
m∈M n∈Nm

Σ Σ Σ
+ ωn, an,m − Bk log2 ωn,mτ +
m 1 k+
B N0Bk gk
n,

k
τx g
m∈M n∈Nm n,m γn,m ln 2 k Nm
0Bk
k∈Km n,
m
Σ Σ γn,m +
Σ k ωn,mτ N0Bk max
+ xBk n,m γ — — n,m
n,m gk
ln 2 p
m∈M n∈Nm k∈Km n,m

s.t. C12, C15, C16. (35)


individual decision results at each iteration with each other, until
Algorithm 2: The DJTSCRA Algorithm. the Lagrange multipliers satisfy the convergence threshold ϵ
Input: The iteration number l = 1 and its maximum value
max or the iteration index number reaches its maximum Lmax.
L , the convergence factor ϵ, the step size κ, the
current QSI Y(t) and NSI s(t), the AWGN
D. Computational Complexity Analysis
spectral density N0 and the CPG g(t);
Output: The optimal a˘∗ , x˘ ∗ and p˘ ∗ ; In this subsection, we give the computational complexity
1: for all RAT base stations m ∈ M do of the proposed ATSRA algorithm at each time-slot. The
2: Initialize its Lagrange multipliers computation complexity of the auxiliary variable decision
1 1 1
ω˘ m = n, ]n∈Nm and m = n, ]n∈Nm , then is O(1) as it is solved in closed form. The computation
m 1 m
[ω γ˘ [γ
broadcast them to all complexity of the computation frequency allocation mainly
SDs; 3: end for depends on the sorting of2 { ρn,m(t) }n∈N , ∀m ∈ , which
4: repeat can be given by O(|M||N M ) when the bubble sort
5: for all SDs n ∈ N do |
algorithm is adopted. The joint task splitting and
6: Compute the uplink transmit power communication resource allocation subproblem is solved
by using the
p˘ n = n, ]m∈M ,k∈K and the task DJTSCRA algorithm. According to [37], in order to
[pk m
n m
splitting
a˘n = [an,m ]m∈Mn according to (34) and (39), achieve a δ-optimality, i.e., g(ω˘ , γ˘) g(ω˘ ∗ , γ˘ ∗ )
respectively, then broadcast them to RAT δ, the subgradient method| takes −the cost of | ≤ O(1δ2)
base stations
∈ m n; iterations, which does not depend on the number of
7: end for variables. For each
M SD n, the complexity for computing the transmit power
Σ
8: for all RAT base stations m ∈ M do p˘ n = n, ]m∈Mn,k∈K m is O( m∈Mn |Km|) and that for
9: Compute the subcarrier computing
[p k m the task splitting a˘ n = [an,m ]m∈Mn is O(|
allocation Mn |)
x˘km =
[x ]n∈N ,k∈K according to (38) in each iteration. As a consequence, the total computation
and
n,m m m Σ
update the Lagrange complexity for SD is 1 2 m∈M |Km| + 1)). For
n O( / n
multipliers (
ω˘ m = [ωn,m ]n∈Nm and γ˘ m =
each RAT base station m,δ the complexity for computing
according to (40), then broadcast them to the subcarrier allocation x˘ m = ]n∈Nm ,k∈Km is

SDs ]n∈Nm
n,m
[xk n,
n ∈ N m; O(|Nm||Km|) and that for updating the m Lagrange multipliers
10: end for ω˘ m = [ωn,m ]n∈Nm and γ˘ m = [γn,m ]n∈Nm is O(2|Nm|).
11: until (ǁ ω˘ l+1− ω˘ǁl +ǁ γ˘ l+1
− γ˘ǁ )/2 ϵ or
l
As a result, the total computation complexity for RAT base
l=L ; ≤
max
station m is O(1/δ2 |Nm |( |K m |+ 2)). Therefore, the total
computational
Σ complexity of the Σ DJTSCRA
Σ algorithm is
2 |Nm |(|Km | + 2) + (|K | +
O(1/δ 2 [ m∈M Σ Σ
1 1 n ∈N
3 . m∈M n m

ωl+1 = ω
l + an,m − τ x
k B M, ∀n ∈ N , the computational complexity of the
κ DJTSCRA m∈Mn (|Km| +
k

)]) = O( /δ n∈N )) Σ Since Mn ∈


algorithm is at worst O(1/δ2
Σ (|Km | + with
n,m n,m
k∈K
n,m
Σ
n∈N

M
m
p gm 3))
Σ
, 2
a|M||N
given δ. In summary, the computational complexity of the
log2 |

Σ p —
n,m ATSR A
k kp
+

1+
n,m
Σ
K
M
(|m| ∈+
3)m+

algorithm is O
2
n∈N

γ = n,m +
κ

l+1 max
, (40)
+
N0Bk ) for each time-slot. Obviously, it is a polynomial

γl

n,m n,

k
complexity and thereby can be accepted in practice.
n

n,m
E. Performance Results

k∈Km
In this subsection, we provide the performance results of the
where κ is the sufficiently small step size and l is the
iteration
index. Since the subgradient satisfies the Lipchitz continuity
condition, the convergence of multipliers towards the optimal
3) Distributed Implementation:By the virtue of the Lagrange
ω˘ ∗ and γ˘ ∗ is guaranteed.
dual decomposition method, the joint task splitting and commu-
nication resource allocation can be implemented in a distributed
manner. The procedure details of distributed joint task splitting
and communication resource allocation (DJTSCRA) algorithm
rithm, the SDs are responsible for making the uplink transmit
is given in Algorithm 2. In the distributed DJTSCRA algo-
base stations take charge of making the subcarrier allocation
power allocation and task splitting decisions, while the RAT
of the both partners are done alternatively by exchanging their
decision and updating the Lagrange multipliers. The decisions
proposed ATSRA algorithm by Theorem 1 under the

Slater’s condition assumption.


following
and φ(ε) (where 0 ≤ φ(ε) ≤ φmax ) and a s-only policy αˆ(t) =
Assumption 1.(Slater’s condition):There are values ε > 0

{aˆ(t), xˆ(t), pˆ(t), ˆf (t)} that satisfies:


E{φˆ(t)} = φ(ε),
E{φˆ(t)} ≤ E{Uˆ o (t)} − ε, ∀n ∈ N , sta- tionary and randomized policies which independently
chooses a control action αˆ(t), t as a pure (possibly
E{aˆn,m (t)} ≤ E{ˆbn,m (t)} − ε, ∀m ∈ M, n ∈ N , (41) randomized) function of the observed NSI s(t) and regardless
where the s-only policies are defined as the special class of of the QSI Y(t).
TABLE II
SIMULATION PARAMETERS computation utility for each SD, as given by
T −1
Σ
U¯ p = lim 1 E {U np(t)} , ∀n ∈ N , (44)
n T →∞
T t=0
Σ
where Unp(t) = m∈M min{ n, (t), (t) } is the instanta-
m Q
b n,m
neous computation utility which is equivalent to the total amount
of tasks computed for SD n at time-slot t.
For algorithm comparison, we provide three benchmark al-
gorithms as follows:
1) Static Max-Min Fairness (SMMF): The SMMF algorithm
maximizes the minimum instantaneous offloading
utility,
i.e., max minn∈N U o(t), and the minimum instantaneous
n
p
computation utility, i.e., max minn∈N Un(t), among all
SDs, respectively, at each time-slot t. The optimizations
over time-slots are decoupled and independent of the QSI
Q(t). Hence, it does not preserve the task queue stability.
2) Sum Rate Maximization (SRM): The SRM algorithm
maximizes the sum of the instantaneous offloading util-
Σ
ities, i.e., max U o(t), and the sum of the computa-
Theorem 1: Let = min , tion utilities, i.e., U¯Σo Unp(t)
be the
n n∈N
max
respectively, at
each time-slot t. Similar with SMMF, it does not preserve the
o pt
offloading utility ws SDs, n∈ N nthe optimum of task queue stability, neither.
1, and
among Uws be P
3) Adaptive Sum Rate Maximization (ASRM): Compared
φopt be the optimum ofP3. Suppose that the NSI s(t) is i.i.d. with the SRM algorithm, the ASRM algorithm maxi-
over time-slots, E { L(Y(0))} < and that Assumption 1 is mizes the sum of the time-average offloading utilities, i.e.,
satisfied, then for
following performance: ∞the ATSRA algorithm has the
any V > 0,
no
a) The performance bound of the minimum/worst offloading max
Σ n∈N U¯ , while preserving the task queue stabil-

ity. It is only different from ATSRA in the optimization


utility satisfies
objective and thus can be addressed by the Lyapunov
U opto − ≤ φopt − φ¯ ≤ B/V. (42) optimization technology directly.
¯
Uws ws
Note that since the SMMF and SRM algorithms do not
b) All task queues and virtual queues are mean rate stable, preserve the task queue stability, the time-average
and thus (7) and constraint C8 are satisfied. computation utilities might be smaller than the time-average
c) The time-average task queue backlog has the performance offloading utili- ties under the two algorithms, which means
bound that the task queue backlogs under the two algorithms might
increase continuously.
In contrast, the proposed ATSRA and the ASRM algorithms are
lim 1 Σ Σ Σ B + V (φmax −φ(ε))
E n,
m
}≤ self-adaptive to the time-varying wireless environments and the
T −1
.
T →∞ T ε random task arrivals. Therefore, the task queues of them will re-
t=0 m∈Mn∈N
(43) main stable after the algorithm converge. The simulation results
Proof: Please see Appendix C. of the convergence performances under different algorithms are
Theorem 1 asserts that the minimum offloading utility U¯ o given in the following subsection.
is
within 1
of the optimum opt. Hence, ¯ o ws
O( /V ) Uws Uws can be pushed
arbitrarily close to the optimum by using a sufficiently large otherwise stated, the simulation parameters are given in Table II.
value of V . The tradeoff is that an O(V ) queue backlog length Recall that the time-average offloading utility U¯ o is equivalent
is necessary when an O(1/V ) close-to-optimum is achieved. to the time-average enqueue rate of the admitted tasks for each
SD n at the MECSs. However, the task computing/service rate is
V. SIMULATION RESULTS characterized by the dequeue of task queues, which depends on
the computation frequency allocation of MECSs. For this rea-
In this section, we provide simulation results to evaluate
son, we define another performance metric, called time-average
the performance of the proposed ATSRA algorithm. Unless

n
A. Convergence Performance
Fig. 2 shows the average task queue backlogs of different
algorithms varying over time. We can see that the task
queue backlogs of the proposed ATSRA and the ASRM
algorithms converge to stable states as the number of time-
slots increases, while those of the SMMF and SRM
algorithms increase con- tinuously. This verifies that the
ATSRA and ASRM algorithms show better performances in
preserving task queue stability than the SMMF and SRM
algorithms. Furthermore, we observe that our proposed
ATSRA algorithm starts to stabilize nearly 6 times earlier and
has a 7.5 times smaller task queue backlog than the ASRM
algorithm under the same parameter setting. It shows that the
proposed algorithm can achieve low queue latency compared
with other algorithms.
Fig. 4. Time-average computation utility versus number of time-slots.
Fig. 2. Average task queue backlog versus number of time-slots.

Fig. 5. LTJFI w.r.t offloading utility versus number of time-slots.


Fig. 3. Time-average offloading utility versus number of time-slots.

task computation throughput compared with the SMMF algo-


rithm from the long-term time-average perspective. However,
Fig. 3 depicts the time-average offloading utilities of different combining Figs. 3 and 4, we can observe that the max-min
algorithms as the number of time-slots increase. Combining fair algorithms (SMMF and ATSRA) achieve smaller offloading
Figs. 2 and 3, we can observe that the offloading utilities of and computation utilities compared with the sum utility max-
the ATSRA and ASRM algorithms first decrease as their task imization algorithms (SRM and ASRM). This is because the
queue backlogs increase, and then converge to the constant max-min fair algorithms need to allocate more resources to
values when the task queues reach stable states. However, the the network-edge SDs to improve fairness.
offloading utilities of the SMMF and SRM algorithms
consistently remain unchanged even when their task queue
backlogs are very large. This is because the ATSRA and B. Long-Term Jain’s Fairness Index
ASRM algorithms can control the task offloading rate (or Figs. 5 and 6 plot the long-term Jain’s fairness indexs
enqueue rate of the task queue) adaptively according to the (LTJFIs) of different algorithms with respect to (w.r.t) the
task queue backlog, which helps to bound the task queue offloading and computation utilities as the number of time-
backlog under a finite value, but the SMMF and SRM slots increases, respectively. Different from [24], the LTJFI
algorithms do not have this property. measures fairness from the long-term time-average
Fig. 4 shows the time-average computation utilities of perspective. Specifically, the LTJFIs w.r.t the offloading and
differ- ent algorithms as the number of time-slots increases. computation utilities are defined as follows, respectively,
Similar with Fig. 3, the computation utilities of the SMMF
Σ 2 Σ 2
and SRM algorithms remain unchanged basically from the n∈N U¯n o n∈N U¯n
p
¯ ¯ (45)
beginning to o n p n
the end, while those of the ATSRA and ASRM algorithms
vary p 2,
adaptively with the task queue dynamics. Moreover, we see n∈N n∈N (U¯ )
that the computation utilities of the proposed ATSRA and the J(U ) = )n n∈N
N (U¯ o 2 ; J(U =
Σ n Σ ) N
ASRM algorithms converge to those of the SMMF and SRM where ¯ o = [U¯ U¯ p = , J(U¯ o ) ∈ (0,
¯p Uo
] n∈N , p 1],
o∈ [U¯ ]
algorithms asymptotically, respectively. This illustrates that the
proposed algorithm does not degrade the performance of the J(U ) (0, 1]. A larger LTJFI stands for a higher fairness. If
J(U¯ ) = J(U¯ p ) = 1, the offloading and computation
utilities
Fig. 6. LTJFI w.r.t computation utility versus number of time-slots.
Fig. 8. Time-average offloading utility versus control parameter V.

be observed that the time-average task queue backlogs of the


ATSRA and ASRM algorithms increase monotonously as V
rises. This behaviour is coincident with the theoretical
analysis of part (c) in Theorem 1 that the performance
bound of the time-average task queue backlog is proportional
to the control parameter V . That means that a larger V will
result in a higher task queue backlog. In contrast, the time-
average task queue backlogs of the SMMF and SRM
algorithms remain unchanged since they are irrelevant to V .
Based on the Little’s law that the time-average task queue
backlog is directly proportional to the average service latency,
we draw the conclusion that the service latency under the
proposed algorithm increases as V increases. Fig. 8 compares
Fig. 7. Time-average task queue backlog versus control parameter V. the impact of control parameter V on the time-average
offloading utilities of different algorithms. We can see that
of different SDs are the same, which implies that the resource the time-average offloading utility of the proposed ATSRA
allocation is perfectly fair. algorithm increases as V increases. This implies that the gap
From Figs. 5 and 6, we see that the LTJFIs of our proposed of the time-average offloading utility and its optimum will
ATSRA algorithm converge to the value of 1 in a very few vanish gradually as V increases, which verifies part (a) in
time-slots. This states that the resource allocation decisions Theorem 1. Moreover, we can observe from Fig. 8 that
Σ
the average 1 offloading nutility of the proposed ATSRA algo-
made under the proposed algorithm are perfectly fair from the rithm, i.e., n∈N U
¯ o , is almost identical to the minimum
|N |
long-term time-average perspective. Moreover, we see that the can
LTJFIs of the SMMF and SRM algorithms do not change with
the increase of the number of time-slots. However, the LTJFIs of
the ASRM algorithm degrades as time grows. This is because the
SDs with worse CPG information might have few
opportunities to offload their tasks under the ASRM
algorithm, which results in lower offloading utilities for these
SDs as well as unfairness. The accumulated unfairness will
become more and more severe as time grows, and thus causes
the degradation of the LTJFIs. From Fig. 6, even though the
LTJFI w.r.t. the computation utility of the SMMF algorithm is
roughly equal to that of the proposed algorithm, the SMMF
algorithm does not guarantee the task queue stability.
Therefore, we can conclude that our proposed algorithm can
achieve the best fairness while preserving the task queue
stability compared with other benchmark algorithms.

C. The Impact of Control Parameter V


Fig. 7 compares the impact of control parameter V on the
time-average task queue backlogs of different algorithms. It
offloading utility among all SDs, i.e., minn∈N U¯ o .
However, the performance gaps between the average and n
minimum offloading
utilities under the benchmark algorithms are much larger.
This further demonstrates that our proposed algorithm
shows better performance in guaranteeing the long-term
fairness compared with other algorithms.

VI. CONCLUSION
In this paper, we developed an integrated multi-RAT and
MEC network framework, in which the SDs are allowed to
offload their computation tasks over multiple RAT links in
parallel by the task splitting operation. To enable fair
resource sharing/allocation between SDs under dynamical
network settings, the long-term max-min fairness guarantee
mechanism was investigated for the proposed network. An
adaptive task splitting and resource allocation algorithm was
proposed, which makes decisions only depending on the
current NSI and QSI without requiring their distribution
information. We showed that the proposed algorithm achieves
an [O(1/V ), O(V )]-tradeoff between the objective
optimality and the task queue backlog by rigorous analysis. APPENDIX C
Simulation results corroborated the theoretical analysis and indi- PROOF OF THEOREM 1
cated that the long-term Jain’s fairness index under the proposed
mechanism approaches to 1 asymptotically, which implies that Because the ATSRA algorithm minimizes the right hand of
the resource allocation is perfectly fair. Our study provides an (16), we have
effective approach to design fairness-based resource allocation
Δ(Y(t)) − V E {φ(t)|Y(t)} ≤ B − V E ,φδ (t)|Y(t),
algorithms for stochastic and time-varying MEC systems. The
proposed mechanism can be extended to more general MEC sys- Σ
tems with other advanced RAT technologies like non-orthogonal + Σ Qn,m (t)E{δan,m (t) − δbn,m (t)|Y(t)}
multiple access and massive MIMO. m∈M n∈N
Σ
Zn (t)E{φδ (t) − Unδ o (t)|Y(t)}, (49)
APPENDIX A n∈N
PROOF OF LEMMA 1
where φδ (t), Unδ o (t), δan,m (t) and δbn,m (t) are resulted from
Proof: Since problem P2 involves all constraints of prob- any other alternative policies.
lem P 1, we only need to prove that the optimum of prob- Suppose that the assumptions in Theorem 1 are met.
lem P2 is no less than that of problem P1 in the presence Accord- ing to [14], for any η > 0, there at least exists a s-
of C8 and C9. Denote the minimum/worst offloading utility only policy
among SDs as wsU¯ o = minnn∈N U¯ o . Let φopt be the α˜(t) = {a˜(t), x˜(t), p˜(t), ˜f (t)}, such that
maximum
E{φ˜(t)} = φopt + η,
value of problemP 2, and α(t){ = a(t), x(t), p(t), }f (t)
be E{φ˜(t) − Un˜ o (t)} ≤ η, ∀n ∈ N ,
the corresponding decision action at time-slot t. According to
C8, we always have U¯ o (α(t)) ≥ φopt (α(t)). Then, let U opt
=
¯ opt ws opt ws E{a˜n,m (t) − ˜bn,m (t)} ≤ η, ∀m ∈ M, n ∈ N . (50)
minn∈N Un be the optimum of problem P1 and α (t) =
opt
{a (t),
opt
(t), opt (t), opt (t)} be its optimal-solution at time Plugging (50) into (49) and taking a limit as η → 0 yields
x p f
slot t. Clearly, under C8 and C9, if we take the strategy φ(t) =
U¯ws (α (t)), ∀t in problem P2, it achieves φopt = Uwsopt . Hence,
o opt Δ(Y(t)) − V E {φ(t)|Y(t)} ≤ B − V φ opt . (51)
¯o opt
we have Uws ≥ Uws , i.e., problem P1 is equivalent to problem Taking expectations of both sides of (51) and using the telescop-
P2. ing sums over t ∈ {0, 1,... ,T − 1}, we have
T −1
APPENDIX B } −)) { E }L(Y(0))
EΣ{L(Y(T − V { E φ(t)}
PROOF OF LEMMA 2 t=0

Proof: Given Q ≥ 0, b ≥ 0, and a ≥ 0, the inequality ([Q− ≤ T (B − V φopt ) (52)


b]+ + a)2 ≤ Q2 + b2 + a2 + 2Q(a − b) always holds. There-
fore, according to (6) and (11), we have Rearranging terms, dividing (52) by TV and taking a limit
as T → ∞, we get φopt − φ¯ ≤ VB based on the
condition
Q2 (t + 1) − Q2 (t) ≤ a2 (t)+ b2 (t) E{L(Y(0))} < ∞. Since problems P1 and P3 have the same¯
n,m n,m n, n,m optimum,
¯ o i.e., U opt = φopt . Based on C8, we have ≥φ .
m U
ws ws
+
n, (t) n, (t) − b n, (t)), (46) opt
Therefore, we finally have U ws − U¯wso ≤ opt
− φ¯ ≤ B/V . We
2Q m m m
(a φ
Z2 (t + 1) − Z2 (t) ≤ φ2(t)+ (U o(t))2 have completed the proof of part (a) in Theorem
n n n 1.
Similarly, plugging the s-only policy in Assumption 1 into
+ 2Zn(t)(φ(t) − U no(t)). (47) (49), taking expectations of both sides and using the law of
iterated expectations yields
Taking sum over all m ∈ M and n ∈ N and dividing by 2 at
the both sides of (46) and (47), respectively, and adding them Δ(Y(t)) − V E {φ(t)|Y(t)} ≤ B − V φ(ε)
Σ Σ Σ (53)
together, we have —ε E{Qn,m(t)}− ε E{Zn(t)}.
2 2 m∈M n∈N n∈N
Σ an,m(t)+ bn,m(t) Taking telescoping sums over 0 1 1 , we have
L(Y(t + 1)) − L(Y(t)) ≤
Σ

2 t ∈ { , , . . . ,T − }
Σ m∈MΣ
n∈N
2 o 2 T −1
φ ( t )+ (Un(t)) Σ
+ 2 + (t)(φ(t) − Uno(t)), E{L(Y(T ))}− E{L(Y(0))}− V E{φ(t)}
n∈N

Zn
n∈N
T −1 Σ
Σ Σ n, (t) n, (t) − b n, (t)) (48) ≤ T (B − V φ(ε)) − ε Σ E{Qn,m(t)}
+
Q
m (a m m
Σ
m∈M n∈N t=0 m∈M n∈N
Σ
Taking expectation and adding the penalty term −1

—V E{ φ(t) |Y(t) at} the both sides of (48), we obtain the —ε E{Zn(t)} (54)
upper bound expression (16) in Lemma 2. t=0 n∈N
Rearranging (54) and using E{φ(t)}≤ φmax, ∀t as well as the [14] M. J. Neely, Stochastic Network Optimization With Application to Com-
fact that Qn,m(t) ≥ 0, Zn(t) ≥ 0, ∀m ∈ M,n ∈ N , we obtain munication and Queueing Systems. San Rafael, CA, USA: Morgan and
Claypool, 2010.
E{Z2n(T )}≤ 2TB + 2VT (φmax − φ(ε)) + 2E{L(Y(0))}. [15] Y. Wang, M. Sheng, X. Wang, L. Wang, and J. Li, “Mobile-edge comput-
(55) ing: Partial computation offloading using dynamic voltage scaling,”
Because the variance of Zn(t), ∀t is non-negative, i.e., IEEE Trans. Commun., vol. 64, no. 10, pp. 4268–4282, Oct. 2016.
D{Zn(t)} = E{Z (t)}−
2 E{|Zn(t)|} 2≥ 0, we have [16] X. Cao, F. Wang, J. Xu, R. Zhang, and S. Cui, “Joint computation and
communication cooperation for energy-efficient mobile edge
n computing,”
2 2 IEEE Internet Things J., vol. 6, no. 3, pp. 4188–4200, Jun. 2019.
E{Zn(t)}≥ E{|Zn(t)|} . Hence, the following formula [17] E. El Haber, T. M. Nguyen, C. Assi, and W. Ajib, “Macro-cell assisted
holds task offloading in MEC-based heterogeneous networks with wireless back-
√ haul,” IEEE Trans. Netw. Service Manag., vol. 16, no. 4, pp. 1754–
E{|Zn(T )|} ≤ 2TB + 2VT (φmax − φ(ε)) + 1767, Dec. 2019.
[18] T. X. Tran and D. Pompili, “Joint task offloading and resource
(56)
2E{L(Y(0))}. allocation for multi-server mobile-edge computing networks,” IEEE
Trans. Veh. Technol., vol. 68, no. 1, pp. 856–868, Jan. 2019.
[19] T. Q. Dinh, Q. D. La, T. Q. S. Quek, and H. Shin, “Learning for
Dividing by T and taking a limit as T → ∞, we obtain computation
E{|Zn(T )|}
lim = 0. (57) offloading in mobile edge computing,” IEEE Trans. Commun., vol. 66,
T →∞ no. 12, pp. 6353–6367, Dec. 2018.
[20] X. Chen, H. Zhang, C. Wu, S. Mao, Y. Ji, and M. Bennis, “Optimized

T
Therefore, the virtual queue Zn(t) is mean rate stable, and computation offloading performance in virtual edge computing systems
C8 via deep reinforcement learning,” IEEE Internet Things J., vol. 6, no. 3,
pp. 4005–4018, Jun. 2019.
is guaranteed. A similar proof can be given for the task queue [21] Y. Liu, H. Yu, S. Xie, and Y. Zhang, “Deep reinforcement learning
Qn,m(t), and thus (7) is also satisfied. We have completed for offloading and resource allocation in vehicle edge computing and
the proof of part (b) in Theorem 1. networks,” IEEE Trans. Veh. Technol., vol. 68, no. 11, pp. 11158–
11168, Nov. 2019.
Rearranging (54) and dividing it by εT , and then taking a [22] Y. Mao, J. Zhang, S. H. Song, and K. B. Letaief, “Stochastic joint
limit as T → ∞, we can prove part (c) in Theorem 1.
Σ max radio and computational resource management for multi-user mobile-edge
li 1 TΣ−1 E (t)}≤ B + V (φ − φ(ε)). computing systems,” IEEE Trans. Wireless Commun., vol. 16, no. 9,
m Σ
{Q
T →∞ n,m pp. 5994–6009, Sep. 2017.
ε
t=0 m∈M n∈N [23] C. Liu, M. Bennis, M. Debbah, and H. V. Poor, “Dynamic task of-
T (58) floading and resource allocation for ultra-reliable low-latency edge
We finish the proof of Theorem 1. trum, Rel. 13,” Tech. Rep. 36.889, Jul. 2015.
[12] 3GPP, “Evolved universal terrestrial radio access network (E-UTRAN) and
wireless LAN (WLAN); LTE-WLAN aggregation adaptation proto- col,
REFERENCES Rel. 13,” Tech. Rep. 36.360, Mar. 2016.
[13] Z. Jing, Q. Yang, M. Qin, and K. Kwak, “Long term max-min fairness
[1] M. Qin et al., “Service-oriented energy-latency tradeoff for IoT task guarantee mechanism: Adaptive task splitting and resource allocation in
partial offloading in MEC-enhanced multi-RAT networks,” IEEE MEC-enabled networks,” in Proc. IEEE Int. Symp. Personal, Indoor
Internet Things J., vol. 8, no. 3, pp. 1896–1907, Feb. 2021. Mobile Radio Commun., 2019, pp. 1–6.
[2] J. Ahn, J. Lee, S. Yoon, and J. K. Choi, “A novel resolution and power
control scheme for energy-efficient mobile augmented reality applications
in mobile edge computing,” IEEE Wireless Commun. Lett., vol. 9, no. 6,
pp. 750–754, Jun. 2020.
[3] E. El Haber, T. M. Nguyen, and C. Assi, “Joint optimization of compu-
tational cost and devices energy for task offloading in multi-tier edge-
clouds,” IEEE Trans. Commun., vol. 67, no. 5, pp. 3407–3421, May
2019.
[4] H. Wu, Y. Sun, and K. Wolter, “Energy-efficient decision making for
mobile cloud offloading,” IEEE Trans. Cloud Comput., vol. 8, no. 2,
pp. 570–584, Jun. 2020.
[5] L. Lin, X. Liao, H. Jin, and P. Li, “Computation offloading toward edge
computing,” Proc. IEEE, vol. 107, no. 8, pp. 1584–1607, Aug. 2019.
[6] Y. C. Hu, M. Patel, D. Sabella, N. Sprecher, and V. Young, “Mobile edge
computing—A key technology towards 5G,” ETSI White Paper, vol. 11,
pp. 1–16, 2015.
[7] Q. Zhang, L. Gui, F. Hou, J. Chen, S. Zhu, and F. Tian, “Dynamic task
offloading and resource allocation for mobile-edge computing in dense
cloud RAN,” IEEE Internet Things J., vol. 7, no. 4, pp. 3282–3299,
Apr. 2020.
[8] M. Qin et al., “Green-oriented dynamic resource-on-demand strategy
for multi-RAT wireless networks powered by heterogeneous energy
sources,” IEEE Trans. Wireless Commun., vol. 19, no. 8, pp. 5547–5560,
Aug. 2020.
[9] W. Wu, Q. Yang, P. Gong, and K. S. Kwak, “Energy-efficient traffic
splitting for time-varying multi-RAT wireless networks,” IEEE Trans.
Veh. Technol., vol. 66, no. 7, pp. 6523–6535, Jul. 2017.
[10] 3GPP, “Study on small cell enhancements forE-UTRA and E-UTRAN:
Higher Layer Aspects, Rel. 12,” Tech. Rep. 36.842, Jul. 2014.
[11] 3GPP, “Feasibility study on licensed-assisted access to unlicensed spec-
computing,” IEEE Trans. Commun., vol. 67, no. 6, pp. 4132–4150,
Jun. 2019.
[24] Z. Zhu et al., “Fair resource allocation for system throughput
maximiza- tion in mobile edge computing,” IEEE Access, vol. 6, pp.
5332–5340, 2018.
[25] X. Chen, Y. Cai, M. Zhao, and M. Zhao, “Joint computation offloading
and resource allocation for min-max fairness in MEC systems,” in
Proc. IEEE Wireless Commun. Netw. Conf., Marrakesh, Morocco,
2019, pp. 1–6.
[26] X. Lyu et al., “Optimal schedule of mobile edge computing for
internet of things using partial information,” IEEE J. Sel. Areas
Commun., vol. 35, no. 11, pp. 2606–2615, Nov. 2017.
[27] Y. Kim, H. Lee, and S. Chong, “Mobile computation offloading for
appli- cation throughput fairness and energy efficiency,” IEEE Trans.
Wireless Commun., vol. 18, no. 1, pp. 3–19, Jan. 2019.
[28] L. Diez, A. Garcia-Saavedra, V. Valls, X. Li, X. Costa-Perez, and R.
Agero, “LaSR: A supple multi-connectivity scheduler for multi-RAT
OFDMA systems,” IEEE Trans. Mobile Comput., vol. 19, no. 3, pp.
624–639, Mar. 2020.
[29] 3GPP, “Release Description; Summary of Rel-15 Work Items, Rel.
15,” Tech. Rep. 21.915, Sep. 2019.
[30] L. Huang and M. J. Neely, “Max-weight achieves the exact
[O(1/V ), O(V )] utility-delay tradeoff under markov dynamics,”
Aug. 2010, arXiv:1008.0200.
[31] S. Kiamehr, M. Ebrahimi, M. S. Golanbari, and M. B. Tahoori,
“Temperature-aware dynamic voltage scaling to improve energy
efficiency of near-threshold computing,” IEEE Trans. Very Large
Scale Integr. Syst., vol. 25, no. 7, pp. 2017–2026, Jul. 2017.
[32] A. P. Miettinen and J. K. Nurminen, “Energy efficiency of mobile
clients in cloud computing,” in Proc. USENIX Conf. Hot Top. Cloud
Comput., Boston, MA, USA, Jun. 2010, pp. 1–7.
[33] A. O. Allen, Probability, Statistics, and Queueing Theory. San
Diego, USA: Academic Press, INC., 2014.
[34] D. Pisinger, “Algorithms for Knapsack Problems,” Ph.D.
dissertation, Dept. Comput. Sci., Copenhagen, Denmark: Univ.
Copenhagen, 1995.
[35] M. Qin et al., “Learning-aided multiple time-scale SON function
coordina- tion in ultra-dense small-cell networks,” IEEE Trans.
Wireless Commun., vol. 18, no. 4, pp. 2080–2092, Apr. 2019.
[36] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge,
U.K.: Cambridge Univ. Press, 2004.
[37] N. Mokari, M. R. Javan, and K. Navaie, “Cross-layer resource al-
location in OFDMA systems for heterogeneous traffic with
imper- fect CSI,” IEEE Trans. Veh. Technol., vol. 59, no. 2, pp.
1011–1017, Feb. 2010.
Zewei Jing received the B.S. degree from Inner
Mongolia University, Hohhot, China, in 2016. He Jinglei Li received the B.S. degree in electronic in-
is currently working toward the Ph.D. degree with formation engineering from PLA Information Engi-
the School of Telecommunication Engineering, Xi- neering University, Zhengzhou, China, in 2008, and
dian University, Xi’an, China. His research interests the M.S. and Ph.D. degrees in communication and
include stochastic network optimization, and radio information systems from Xidian University, Xi’an,
resource allocation and their applications in mobile China, in 2011 and 2016, respectively. He is
edge computing networks. currently with Xidian University. His research
interests include wireless network connectivity and
node-selfishness management.

Qinghai Yang (Member, IEEE) received the B.S.


degree in communication engineering from the
Shan- dong University of Technology, Zibo, China, Kyung Sup Kwak (Senior member, IEEE)
in 1998, the M.S. degree in information and received the Ph.D. degree from the University of
communication systems from Xidian University, California San Diego, San Diego, CA, USA. He
Xi’an, China, in 2001, and the Ph.D. degree in was with Hughes Network Systems, Germantown,
communication engi- neering from Inha University, MD, USA, and the IBM Network Analysis Center,
Incheon, South Korea, in 2007. From 2007 to 2008, Armonk, NY, USA. Since then, he has been with
he was a Research Fel- low with UWB-ITRC, the School of Information and Communication
Incheon, South Korea. Since 2008, he has been with Engineering, Inha University, Incheon, South
Xidian University. His current research interests Korea, as a Professor, and was the Dean with the
include autonomic communication, Graduate School of Informa- tion Technology and
content delivery networks, LTE-A techniques, and mobile edge computing. Telecommunications, and since 2003, he has been
He was the recipient of University-President Award from Inha University. the Director with UWB Wireless
Communications Research Center, IT Research Center, Incheon, South Korea. In
2006 and 2009, he was the President with the Korean Institute of Communication
Sciences (KICS), Seoul, South Korea, and the Korea Institute of Intelligent
Transport Systems, Seoul, South Korea. He has authored or coauthored more
Meng Qin (Member, IEEE) received the B.S. than 200 peer-reviewed journal papers. His research interests include multiple
degree in communication engineering from the access communication systems, mobile and UWB radio systems, the future
Taiyuan Uni- versity of Technology, Taiyuan, IoT, and wireless body area network, which include nano networks and
China, in 2012, the molecular communications. He is a Member of the IEICE, KICS, and KIEE.
M.S. and Ph.D degrees in information and He was the TPC and the Track Chair or Organizing Chair for several IEEE
communi- cation systems from Xidian University, related confer- ences. In 1993, he was the recipient of the Engineering
Xi’an, China, in 2015 and 2018, respectively. He is College Achievement Award from Inha University, a Service Award from the
currently a joint Postdoctoral Fellow with the Institute of Electronics Engineers of Korea, Distinguished Service Awards
School of Electronics and Computer Engineering from the KICS in 1996 and 1999, the LG Paper Award in 1998, the Motorola
Peking University, Bei- jing, China, and Peng Paper Award in 2000, official commendations for UWB Radio Technology
Cheng Laboratory, Shenzhen, China. His research Research and Development from the Ministry of Information and
interests include AI-aided wire- less network Communication, Prime Minister, and President of Korea, in 2005, 2006, and
operation and management, machine 2009, respectively, the Haedong Paper Award in 2007, and the Haedong
learning, self-organized network, statistical quality of service provisioning, Scientific Award of Research Achievement in 2009. In 2008, he was elected
and applications of stochastic optimization in intelligent wireless networks, for Inha Fellow Professor and is currently an Inha Hanlim Fellow Professor.
edge intelligence, and green cloud storage.

You might also like