You are on page 1of 8

Optimal Online Resource Allocation for

SWIPT-Based Mobile Edge Computing Systems


Hamed Mirghasemi, Luc Vandendorpe and Mateen Ashraf
Université Catholique de Louvain, Louvain-la-Neuve, Belgium.
Emails: {seyed.mirghasemi, luc.vandendorpe, mateen.ashraf@uclouvain.be}

Abstract—The integration of simultaneous wireless information the statistics of the system state is very difficult and a highly
and power transfer (SWIPT) and mobile-edge computing (MEC) resource consuming process. To address the issue of system
technologies is emerging as a promising technique to overcome state information, Lyapunov optimization technique has been
the performance limits of ultra-low power devices (ULPD) due
to their low battery capacities and their limited computation recently used in the literature to jointly optimize the computa-
capabilities in the Internet of Things (IoT) era. In this paper, tional and radio resources of MEC systems in a multi-slot man-
we propose an online resource allocation algorithm for multi- ner and with limited requirement of system state information
user SWIPT-based MEC systems with the aim of maximizing [1], [5]–[7]. In [5], a Lyapunov-based resource management
the proportional fairness computational utility function subject has been developed to minimize the long-term time-average
to the stability of task and energy queues. Lyapunov optimization
framework is used to jointly optimize the amounts of time allo- energy consumption of a regularly-powered MEC system. This
cated for energy harvesting, information decoding and offloading, work has been extended to WPT-based MEC in [1], [6],
the transmission power for offloading and CPU-cycles frequencies [7]. For WPT enabled MEC systems, in [1], the radio and
for local computing. Moreover, rigorous performance analysis has computational resource management problem with the aim of
been done to prove the asymptotic optimality of our proposed maximizing the computational throughput has been proposed.
algorithm. Simulation results are also presented to demonstrate
the gains of our proposed algorithms over alternative online In [7], an online resource management has been proposed to
approaches and the impact of different network parameters on maximize the system utility for heterogeneous IoT systems in
the performance of our algorithm. which wireless devices are divided into two types according
to their energy storage capability. Differently from the wire-
I. I NTRODUCTION less power transfer (WPT)-based MEC systems considered
Alongside the rapid development of the Internet of Things in the works mentioned before, in this work, we study the
and expanding popularity of intensive computational applica- performance of SWIPT-based MEC system where the AP
tions, new technologies are emerging to mitigate the com- simultaneously transmits command messages and power to
puting and energy bottlenecks of wireless devices. To relieve low power devices in the downlink channel. Using the time
resource-limited wireless devices of the burden of intensive switching (TS) SWIPT, devices harvest energy and decode the
computations, the mobile edge computing paradigm has been command messages where each command message requires
introduced to bring computational resources to the proximity a specific number of computational tasks to be performed.
of wireless devices. On the other hand, wireless power transfer Different from the random generation traffic model used in
(WPT) is envisioned as another promising technology to existing works on WPT-based MEC systems, in our model, we
prolong the lifetime of energy-constrained wireless devices assume that the number of arriving computational tasks at each
and to guarantee the self-sustainability of the network. The device depends on the number of bits decoded from the signal
integration of WPT and MEC technologies to achieve the transmitted by the AP, and therefore on the time duration
gains of both these paradigms has recently attracted grow- allocated for information decoding, which implies a trade-off
ing interest [1, and references therein]. In order to achieve in time allocation between command message decoding and
the potential gains of these two technologies, the problem harvesting energy during the downlink phase. More specifi-
of computational and communication resource management cally, to increase the computational throughput of any device,
for WPT-based MEC systems has been recently addressed more amount of time is needed to be allocated for command
in offline and online manners. The applicability of offline message decoding which decreases the amount of harvested
optimal resource allocation approaches, proposed e.g., in [2], is energy and consequently the computational and offloading
conditioned on knowing beforehand the system state, including capabilities of that device, which leads to a larger queueing
task traffic load, the amount of harvested energy and channel delay. Using Lyapunov optimization technique, in this work,
gains, which is difficult to meet in the practice. Some existing we propose an online resource management algorithm that
works have also developed online resource allocation strategies determines the amount of admitted tasks and the harvested
where either the resource allocation problem has been modeled energy as well as the amount of offloaded or locally computed
as a single-slot optimization [3], or a multi-slot optimization tasks at each time slot. Similar to [7], we set our objective
has been addressed under the assumption that the statistics to maximize the proportional-fair computational throughput
of system state is known [4]. In reality, accurately obtaining while stabilizing task/energy queues at the devices and the

‹,(((

Authorized licensed use limited to: Khon Kaen University provided by UniNet. Downloaded on March 02,2021 at 07:40:05 UTC from IEEE Xplore. Restrictions apply.
of each time slot and shared with the AP through a feedback
 
channel [8].
A. SWIPT Phase
 In the first frame of each time slot with duration of
AP μ(t) ∈ [μmin , μmax ], ULPDs use the TS SWIPT scheme to
harvest energy and also decode command messages from the


   RF signal transmitted by the AP. At each time slot t, the
following inequalities must be satisfied:
Fig. 1. SWIPT-based MEC system
μi,H (t) + μi,ID (t) ≤ μ(t) , μi,H (t), μi,ID (t) ≥ 0 . (1)
Following [9], the number of decoded command bits at ith
device at time slot t can be expressed as
   
hi (t)H wi (t)2
Ãi (t) = μi,ID (t) W log 1 + N  2
 H  2
j=i hi (t) wj (t) + δi
Fig. 2. Time slot structure  
R̃i,DL (t)

MEC server. where hi (t) is the channel gain vector from AP to the ith
device, wi (t) is the MRT beamforming vector implemented
II. S YSTEM MODEL with total power budget PAP max
, W is the bandwidth, and δi2 is
We consider a SWIPT-based MEC system consisting of the channel decoding noise at ui . We also assume that channel
N single-antenna and single-core ULPDs, denoted by U = gains are bounded as: hi (t)2 ≤ h̄max i . Denoting by Li the
{u1 , · · · , uN }, a multiple-antenna access point integrated with length of each command message, the number of command
a multiple-core MEC server, as shown in Fig. 1. We denote messages, i.e., the number of computational tasks, received
by NAP the number of antennas at the AP and we assume that by ui is Mi (t) = Ãi (t)/Li . Additionally, we denote by Lloc i
time is slotted, i.e., T = {1, 2, · · · }, with each slot lasting T the number of local bits required to perform a computational
seconds. At each time slot, this SWIPT-based MEC system task in ui . Thus, each computational task at ui is equivalent to
has two working phases: the SWIPT Phase with duration of Li +Lloc
i bits, and therefore, the total number of computational
μ(t) in which the energy and control messages are transmitted tasks (in terms of bits) arriving at the end of time slot t is given
to ULPDs and the Computing Phase with duration of T −μ(t) by
in which devices execute some number of tasks stored in their Ai (t) = Mi (t)(Li + Lloc i ) = μi,ID (t)Ri,DL (t) , (2)
task queues, as shown in Fig. 2. In the first phase, devices def Lloc
use TS SWIPT architecture to harvest energy and decode where Ri,DL (t) = (1 + Lii )R̃i,DL (t). Similarly, the amount of
the command messages. These command messages specify harvested energy by the ith ULPD at the end of time slot t
a number of fine-grained computational tasks needed to be can be expressed as
performed by each device. Following TS SWIPT scheme, for 
 2 
 H 
any device ui , the first phase is further divided into two sub- Ei,H (t) = μi,H (t) φ hi (t)wj (t) + δi2 , (3)
slots: one for energy harvesting (μi,H (t)) and one for command 
j

message decoding (μi,ID (t)). In the second phase with duration Pi,DL (t)
of T − μ(t), the ULPDs can partially offload their task bits
and upload the outcome of locally processed tasks to the MEC where φ(.) characterizes the relationship between the received
server using the TDMA scheme. Considering the AP as the power and the harvested power. In this paper, we adopt the
endpoint of this computation process, the results of locally non-linear logistic EH model proposed in [10], given by:
ax

computed tasks should be uploaded to the AP. In this paper, φ(x) = Msat 1+e1−e
−a(x−b) , where Msat is the maximum harvested

we assume that the output size of the computation process is power at saturation and (a, b) are two constants depending on
relatively small with respect to the input size, and therefore the EH circuit design.
the uploading phase is neglected. B. Computing Phase
It is further assumed that the channel between the AP and
During the task execution phase, the ith device will offload
ULPDs as well as the channels between the MEC server and
part of its computational tasks, denoted by Ci,O (t) to the
ULPDs are i.i.d. block fading channels. We also assume that
MEC server, while Ci,M (t) task bits will be locally processed.
at the beginning of each time slot, the AP can accurately
Denoting by Qi (t) the queue length of the task buffer at ith
estimate the current channel state information (CSI) of the
device at the beginning of tth time slot, the queue length
downlink channel by receiving the pilot signals from the
dynamics is given by
ULPDs. Similarly, the CSIs of the uplink channels are assumed
to be perfectly estimated by the MEC server at the beginning Qi (t + 1) = max {Qi (t) − Ci (t), 0} + Ai (t) , (4)

Authorized licensed use limited to: Khon Kaen University provided by UniNet. Downloaded on March 02,2021 at 07:40:05 UTC from IEEE Xplore. Restrictions apply.
def
where Ci (t) = Ci,O (t) + Ci,M (t) is the total number of task 4) Queue Buffers at the MEC Server: The MEC server
bits departing from the task buffer of ith device during time maintains N task buffers to store the computational tasks
slot t. offloaded by devices but not yet processed by the MEC
1) Local Computing Mode: Let fi (t) = 0 denote the CPU- server. Let Di (t) denote the queue length of the task buffer
cycle frequency of the ith ULPD at tth time slot, if it has corresponding to the ith device at the beginning of time slot t.
decided to perform local computation Assuming that the computational resources of the MEC server
is shared dynamically among different services, in this paper,
fi (t) ∈ [fimin , fimax ] ∪ {0} , (5) we model the computational resources allocated to each device
where fimin and fimax are the minimum and the maximum fre- as a random variable. More specifically, similar to [7] and [11],
quencies of the processor of ith device, respectively. Denoting we assume that at each time slot, a random amount of tasks
by κi the number of computation cycles per processed bit in from the ith device, denoted by Zi (t), is executed by the server
the ith ULPD, the number of tasks locally computed at time with the maximum Zimax . Therefore, each queue Di (t) evolves
slot t can be expressed as according to the following dynamics

Ci,M (t) = T − μ(t) fi (t)/κi . (6) Di (t + 1) = max{Di (t) − Zi (t), 0} + C̃i,O (t) , (13)

Accordingly, the energy consumption of the processor can be where C̃i,O (t) = min{Ci,O (t), [Qi (t) − Ci,M (t)]+ } is the
written as number of available tasks at the ith device for offloading.

Ei,M (t) = T − μ(t) ρi fi (t)3 , (7) Similar to [5, Lemma.4], we can show that the RHS of eq. (13)
can be replaced by Ci,O (t) without loss of optimality. In the
where ρi is the effective switched capacitance of the proces- case when Ci,O (t) > Qi (t) − Ci,M (t), dummy task inputs can
sor’s chip at the ith device.
def  2 be offloaded using the excessive rate, while these dummy tasks
2) Offloading Mode: Define ḡi (t) = gi (t) /σi2 , where are simply going to be ignored by the MEC server.
gi (t) is the channel gain from the ith device to the MEC
server and σi2 is the decoding noise at the MEC server. We III. P ROBLEM F ORMULATION
assume that ḡi (t) is perfectly estimated by the MEC server and Similar to [12], we define the system utility as
shared with the ith device at the beginning of each time slot

t through a perfect feedback. In addition, we assume that the Usum (t) = Ui (Ai (t)) , (14)
i∈N
channel gain is upper-bounded, and thus we have ḡi (t) ≤ ḡimax .
In the offloading phase, ULPDs can offload some of their where Ui (Ai (t)) is the utility of admitting Ai (t) bits of
computation bits to the MEC server using TDMA protocol. computation tasks at the ith device at time slot t. In this
The number of computation bits offloaded to the MEC server paper, we consider the logarithmic function as the utility,
def
at time slot t by the ith ULPD is given by i.e., Ui (x) = log(1 + x), which corresponds to the pro-
  portional fairness task admission and balances the compu-
ḡi (t)Ei,O (t)
Ci,O (t) = μi,O (t)W log 1 + , (8) tational throughput and the fairness among devices [13].
μi,O (t)
Define Ψ(t) = [μi,H (t), μi,ID (t), μi,O (t), Ei,O (t), fi (t)]N
i=1 as
where Ei,O (t) is the transmission
energy with Ei,O (t) ∈ the control actions of ULPDs at the tth time slot. The problem
[0, Ei,O
max
] and μi,O (t) ∈ M μ(t) is the time duration allocated of time-average utility maximization can be formulated as
to the ith device to offload at the tth time slot, where P1
 P1 : Ūsum = max Ūsum

{Ψ(t)}
M(t) = {μi,O }i∈N : μi,O ≥ 0, μi,O ≤ T −μ(t) . (9)
s.t. (1), (5), (9) − (12)
i∈N
0 ≤ Ei,O (t) ≤ Ei,O
max
, ∀i ∈ N , t ∈ T (15)
def
3) Battery Dynamics: Let us denote by Ei,C (t) = Q̄i < ∞ , ∀i ∈ N (16)
Ei,M (t) + Ei,O (t) the total consumed energy and by Bi (t) the
energy queue length of the ith device at time t, respectively. D̄i < ∞ , ∀i ∈ N , (17)
The energy queue of device i evolves according to where the long-term time-average of any function X(t) is
def T −1
Bi (t + 1) = Bi (t) + Ei,H (t) − Ei,C (t) , (10) defined as X̄ = limT →∞ T1 τ =0 E[X(τ )]. Constraints (16)
and (17) imply that the task queues of all devices and the
with Bi (0) = 0. At any time slot, the following energy- MEC server to be stable [14], which guarantee that all the
causality constraint must be satisfied admitted tasks can be completed in a finite delay.
Ei,C (t) ≤ Bi (t) + Ei,H (t) . (11) The problem P1 is obviously a challenging stochastic op-
timization problem with spatially- and temporally-coupled
Moreover, the amount of harvested energy is limited by the constraints, i.e., eqs. (9) and (10). Instead of solving P1
battery capacity, denoted by Bimax . Therefore, at each time slot directly, we propose an online joint radio and computational
t, we have resource management based on Lyapunov optimization, which
Ei,H (t) + Bi (t) ≤ Bimax . (12) as shown in the following section, enables us to decompose

Authorized licensed use limited to: Khon Kaen University provided by UniNet. Downloaded on March 02,2021 at 07:40:05 UTC from IEEE Xplore. Restrictions apply.
P1 into a series of per-time-slot deterministic optimization Lemma 1: For all feasible control actions and all possible
problems, each can be solved with low complexity. values of Θ(t), we have
Remark 1: We assume that all control actions of the AP 

and the MEC server, collectively denoted by SAP/M (t) = ΔV (Θ(t)) ≤ B − V E Ui (Ai (t))|Θ(t) (21)
{wi (t), μ(t), Z(t)}, are chosen at the beginning of each time

i∈N
slot by the AP and the MEC server according to their limited + Qi (t)E[Ai (t) − Ci (t)|Θ(t)]
radio and computational resources as well as the temporal i∈N

dynamics of all services provided by these two entities. The + Di (t)E[Ci (t) − Zi (t)|Θ(t)]
extension of P1 to consider optimization over μ(t) and over i∈N
Z(t) is straightforward. The former can be optimized using the

bijective optimization method, similar to [3], and the later can + (Bi (t) − Bimax )E[Ei,H (t) − Ei,C (t)|Θ(t)] ,
i∈N
be optimized by imposing bounds on the computational power
consumption at the MEC server, similar to [5]. The more where B is a constant
challenging part, i.e., the problem of optimal beamforming def 1


2 max 2 max 2 max 2
B= i ) + (Ei,H ) + (Ei,C ) + (Zi ) +
(Amax
design for MEC systems, has been recently addressed in the 2
i∈N
literature ( [15, and references therein]). The problem of jointly 
 2  
optimizing SAP (t) and Ψ(t) in an online scenario is quite fimax κ−1
i (T − μmin ) + ξi fimax (T − μmin ) + 2W /ln 2 ,
challenging and left for our future work. For the sake of
simplicity, in this paper, we model SAP/M (t) as a set of i.i.d (22)
random variables with some minimum and maximum values, def def
max
in which Ei,C = Ei,O
max
+ T − μmin ρi (fimax )3 , Ei,H =
without any further assumptions about their statistics. def max 2 def
μmax Msat , Amax
i = μmax W log(1 + h̄max i PAP /δi ) and ξi =
2W Ei,O
max
(T −μmin )ḡimax
A. Lyapunov Optimization Framework ln 2 .
def def
Proof of Lemma 1: The proof is similar to the proof of
Define Xi (t) = Bi (t) − Bimax and let Θ(t) = [5, Lemma 1], and hence omitted.
{Q(t), D(t), B(t)} be the concatenated vector of all task and By the principle of opportunistically minimizing an expec-
energy queues. We define the perturbed Lyapunov function as tation [14], the minimization of the RHS of eq. (21) can
1
be transformed to the following deterministic optimization
[Qi (t)2 + Di (t)2 + (Bi (t) − Bimax )2 ] .
def
L(Θ(t)) = problem to be solved at each time slot
2

i∈N
(18) P2 : min Qi (t)Ai (t) − V Ui (Ai (t)) + Xi (t)Ei,H (t)
{Ψ(t)}
Additionally, the conditional Lyapunov is given by i∈N

def   + Yi (t)Ci,O (t) − Xi (t)Ei,O (t)


Δ Θ(t) = E L(Θ(t + 1)) − L(Θ(t))|Θ(t), S(t) , (19) i∈N

where theexpectation is taken over the system state vector + −Qi (t)Ci,M (t) − Xi (t)Ei,M (t)
i∈N
S(t) = h(t), μ(t), Z(t), g(t), {wi (t) } and the control
actions Ψ(t) s.t. : (1), (5), (9), (12), (15),
given queue lengths Θ(t). Intuitively, by mini-
mizing Δ Θ(t) , the task queues Qi (t) and Di (t) are pushed def
where Yi (t) = Di (t) − Qi (t). Since at any time slot t with
towards zero to stabilize the task queues and to guarantee the
given μ(t), the timing parameters of SWIPT phase, namely
stability
constraints
(eqs. (16) and (17)). Moreover, minimiz-
{μH (t), μID (t)}, can be decoupled from the timing parameters
ing Δ Θ(t) pushes the energy queues Bi (t) towards their
of computing phase, μO (t), the minimization problem P2 can
capacities Bimax . As will be shown later in Lemma 2 and
be separated into three sub-problems as follows.
Theorems 2 and 3, by a proper choice of Bimax , the energy
causality constraints (eq. (11)) can be satisfied. Thus, from B. Optimal Harvesting Energy and Task Allocation
now on, we remove these constraints. Now, we define the drift- The optimal harvesting energy and task allocation can be
minus-utility function as obtained via the following optimization problem


def PDL : min Qi (t)μi,ID (t)Ri,DL (t)
ΔV (Θ(t)) = Δ(Θ(t)) − V E Ui (Ai (t)) , (20) {μH (t),μID (t)}
i∈N + Xi (t)μi,H (t)Pi,DL (t) − V Ui (μi,ID (t)Ri,DL (t))
in which V > 0 is a control parameter which tunes the s.t. : μi,H (t) + μi,ID (t) ≤ μ(t)
trade-off between the sum-utility and the lengths of task 0 ≤ μi,H (t) ≤ min{−Xi (t)/Pi,DL (t), μ(t)} (23)
buffers which characterizes the execution delay according to
0 ≤ μi,ID (t) ≤ min{Amax
i /Ri,DL (t), μ(t)} .
the Little’s theorem [16]. Following Lyapunov optimization
framework, instead of minimizing ΔV (Θ(t)), we minimize First, we note that the problem PDL is a convex optimiza-
the following upper bound on ΔV (Θ(t)). tion problem since the objective function is convex and its

Authorized licensed use limited to: Khon Kaen University provided by UniNet. Downloaded on March 02,2021 at 07:40:05 UTC from IEEE Xplore. Restrictions apply.
where Ẽi,O (t) = μḡiOi (t)
(t) W Yi (t)ḡi (t)
def
constraints are linear, and therefore can be solved by any [ Xi (t) ln 2 − 1].
standard convex optimization algorithm [17]. The following Proof of Lemma 3: The proof is similar to the proof of
lemma reveals useful properties about the optimal solution of [18, Theorem 1], and hence omitted.
PDL . From eq. (28), we can conclude that if
Lemma 2:
W Yi (t)ḡi (t)
• The ith device may harvest energy at time slot t only if Xi (t) ≤ , (29)
Xi (t) < 0, i.e., set μi,H (t) = 0 if Xi (t) ≥ 0. ln 2
• Assuming Bi (0) = 0, we always have Xi (t) ≤ 0, and the optimal transmission energy, and therefore the optimal

therefore Bi (t) ≤ Bimax . transmission time allocation should be zero, i.e., Ei,O (t) = 0.
• If
D. Optimal Local CPU Frequencies
V ≤ Qi (t) ln 2 , (24)
The optimal CPU-cycle frequencies for each device can be
then optimal task admission is equal zero, i.e., set obtained by solving the following optimization problem
μi,ID (t) = 0.
Proof of Lemma 2: Please refer to Section VI-A. PM : min −Qi (t)Ci,M (t) − Xi (t)Ei,M (t) . (30)
fi (t)∈[fimin ,fimax ]∪0
C. Optimal Transmission Power and Time Duration
Since PM is a convex optimization problem, its solution,
The optimal Ei,O (t) and μi,O (t) can be obtained by solving denoted by fi (t), is achieved either at the stationary point or
the following optimization problem at the boundaries points, i.e, {0, fimin , fimax }. Taking the first-

PO : min Yi (t)Ci,O (t) − Xi (t)Ei,C (t) order derivation of the above objective function, we find the
Ei,o (t),μi,O (t)
i∈N stationary point, denoted by f˜i (t), of the objective function as
s.t. : {μi,O (t)} ∈ M(t) (25) 
−Qi (t)
f˜i (t) =
def
Ei,O (t) ∈ [0, Ei,O
max
]. (26) .
3ρi κi Xi (t)
def
First, we define Np = {i : Di (t) < Qi (t)}. Since Xi (t) ≤ 0 Besides zero, the objective function has another positive root,

according to Lemma 2, for any device i with Yi (t) ≥ 0, i.e., def −Qi (t) ˜
given by fi (t) =
o
ρi κi Xi (t) . Note that fi (t) < fi (t). The
o
i∈ / Np , the corresponding term in objective function of PO
optimal CPU-cycle frequencies are given as follows:
is monotonically increasing w.r.t. the transmission energy, and

• If Xi (t) = 0, then fi (t) = fi
max
therefore the optimal transmission time allocation and energy .
should be equal to zero. Thus, Ei,O 
(t) = 0 if i ∈
/ Np , which • If Xi (t) < 0, then we have:

intuitively means that offloading from ith device at time t is ⎪

⎨0 fimin ≥ fio (t)
not optimal if the queue length at the MEC server, Di (t),
is larger than Qi (t), the queue length at that device. Since

fi (t) = fi min
f˜i (t) < fimin < fio (t)


PO is a convex optimization problem, standard optimization ⎩min{f˜i (t), fimax } f˜i (t) ≥ fimin .
solvers such as the interior point method can be used to (31)
find optimal solution [17]. However, in order to gain insights From eq. (31), we can conclude that if
on the structure of the optimal solution, we propose here
an alternating optimization algorithm for solving PO . Since Qi (t)
Xi (t) ≤ − 2 , (32)
the feasible region of PO is a Cartesian product of those fi,min ρi κ i
of {Ei,O (t)} and {μi,O (t)}, we first derive the optimal time then fi (t) = 0, and thus Ei,M

(t) = 0.
allocations {μi,O (t)} for a given {Ei,O } and then we derive
the set of optimal transmission energy for a given set of time E. Performance Analysis
allocations. In this section, we analyse the performance of our algorithm.
Lemma 3: For problem PO , we have: First, we derive upper bounds on the backlogs of task queues
• for a given set of {Ei,O (t)}i∈Np satisfying Ei,O (t) ∈ at devices and MEC server.
[0, Ei,O
max
], the optimal transmission time allocations for Theorem 1: For any V > 0 and any initialization of task
devices in Np are given by queues satisfying 0 ≤ Qi (0) ≤ Qmax and 0 ≤ Di (0) ≤ Dimax ,
i
ḡi (t)Ei,O (t) ∀i, where
μi,O (t) =  [T − μ(t)] . (27)
j∈Np ḡj (t)Ej,O (t) Qmax
def
= V /ln 2 + Amax
i i
• for a given {μi,O (t)} ∈ M(t), the optimal transmission def
Dimax = V /ln 2 + Amax
i + Ci,O
max
, (33)
energy set is given by
⎧   def W Ei,O ḡi
max max
⎨min max{Ẽ (t), 0}, E max Xi (t) < 0
max
and Ci,O = ln 2 , we have
 i,O i,O
Ei,O (t) =
⎩E max Xi (t) = 0 , 0 ≤ Qi (t) ≤ Qmax
i
i,O
(28) 0 ≤ Di (t) ≤ Dimax , ∀i, t. (34)

Authorized licensed use limited to: Khon Kaen University provided by UniNet. Downloaded on March 02,2021 at 07:40:05 UTC from IEEE Xplore. Restrictions apply.
Algorithm 1 Our Proposed Online Resource Management IV. S IMULATION R ESULTS
Algorithm for SWIPT-Based MEC systems
In this section, we evaluate the performance of our proposed
(1) Initialization: set Bimax according to eq. (36) and algorithm. In our simulations, the downlink and uplink channel
Qi (0) = Bi (0) = Di (0) = 0. gains are modeled as the Rician fading model with Rician
(2) Repeat at each time slot t ∈ T : factor of 3.5dB and the path-loss model follows the ITU
• Calculate the backlog of virtual queues: Xi (t) and indoor channel model [3]. Also, we have assumed that the
Yi (t), ∀i ∈ N . MEC server is co-located with the AP, and thus ḡi (t) =
 
• Observe the system state S(t). hi (t)2 /σ 2 . In addition, we assume that devices are located
 i
• For each device i, solve PDL to obtain μi,DL (t) and at the distance between dmin and dmax from the AP/MEC
μi,H (t), ∀i ∈ N . server. For the case of dmin = dmax , devices are are assumed
 
• Solve PO to obtain μi,O (t) and Ei,O (t), ∀i ∈ N . to be equally spaced by (dmax − dmax )/N . For each time slot

• Solve PM to obtain fi (t), ∀i ∈ N . t, we set T = 1s, and μ(t) is uniformly distributed over
• Update Qi (t) and Di (t) according to eqs. (4) and (13) [0.2, 0.8]. At the MEC server, the number of task departure
and Bi (t) according to eq. (10). Zi (t) is uniformly distributed within [0, 0.5Dimax ]. Besides,
the other parameters are: Ei,O max
= 80mJ, W = 0.2MHz,
−28
PAP = 1W, ρi = 10 , κi = 100cycles/bit, Lloc
max
i = 100bit,
Proof of Theorem 1: Please refer to Section VI-B. Li = 8bit, fimin = 1MHz, fimax = 1GHz, σi2 , δi2 = −105dB
In the following theorem, we derive the required battery and ḡimax = E[ḡi (t)] + 30dB for all i ∈ N . Finally, EH
capacity in such a way that any device will not locally parameters are chosen according to [10] as Msat = 20mW,
compute or offload its tasks if the available energy is less a = 6400 and b = 0.003.
than the maximum energy consumption of each device, i.e.,
A. Theoretical Result Verification
Bi (t) ≤ Ei,Cmax
, where Ei,Cmax
is defined in Lemma 1, and
therefore, energy availability constraints in eq. (11) are always As a baseline scheme with which to compare our proposed
satisfied. algorithm, we consider the joint optimization algorithm with
Theorem 2: We define equal offloading time (JOET), in which access to the up-
link channel has been evenly distributed among the devices,
def Qmax
i W Qmax max
i ḡi which is equivalent to our proposed algorithm if we set
Ximax = max{ , }. (35)
(fi )2 ρi κi
min ln 2 μi,O (t) = (T − μ(t))/N . We set NAP = 4, N = 4, dmin = 5
By setting and dmax = 15. Fig. 3 shows that the average number of
admitted computational tasks increases monotonically with the
Bimax = Ei,C
max
+ Ximax , (36) control parameter V . Moreover, the average sum of task queue
lengths increases with V as shown in Fig. 4. Additionally,
the energy consumption of ith device at any time slot t is
we can see that the increment rates of average of admitted
always smaller than the available energy, i,e., Ei,C (t) ≤ Bi (t).
computational tasks and that of average sum of task queue
Proof of Theorem 2: Please refer to Section VI-C
backlogs are decreasing with V which is consistent with

the (O(1/V ), O(V )) trade-off between utility and backlog
Let Ūsum denote the achieved objective value of P1 under our established in Theorems 1 and 3. Numerical results for JOET
proposed algorithm. The following theorem provides a lower algorithm, illustrated in Figs. 3 and 4, show that even distri-
bound on the performance of our algorithm. bution of offloading time allocation provides less opportunity
Theorem 3: Assume that the system state is i.i.d over time1 . for devices located further from the AP and consequently
We have: leads to higher task queue backlogs. Comparing to the task
 P1 B
Ūsum ≥ Ūsum − , (37) admission utility, the impact of offloading time allocation
V optimization is more severe on the task queue lengths which
where B is defined in eq. (22). can be explained by the fact that optimization of offloading
Proof of Theorem 3: Please refer to Section VI-D. time allocation directly affects the queue task lengths while
Remark 2: Theorem 3 shows that our algorithm can achieve it impacts indirectly on the number of admitted tasks through
a time-average utility within O(1/V ) of the optimal solution the task stability constraints.
of P1 . Also, in Theorems 1 and 2, we have shown that the
B. Performance Evaluation
solution of our algorithm is feasible to P1 . However, as shown
in Theorems 1 and 2, the task and energy queues under our In this section, we will show the impact of two system
algorithm have upper bounds of O(V ), which means that there parameters on the performance of our proposed algorithm.
is a trade-off between the time-average utility and time-average The impact of different average distances between the AP
task queue occupancy. and devices on the average of admitted computational tasks
is shown in Fig. 5. Clearly, increasing the average distance
1 By using multi-slot drift analysis in [14, Section 4.9], similar result can between the AP and the devices decreases both downlink
be obtained for the case of ergodic non i.i.d. system state. and uplink channel gains, and therefore reduces the amount

Authorized licensed use limited to: Khon Kaen University provided by UniNet. Downloaded on March 02,2021 at 07:40:05 UTC from IEEE Xplore. Restrictions apply.
105
106 18
1.8

Average number of admitted tasks (bps)


16
Average number of admitted tasks (bps)
1.6
14
1.4
12
1.2 10

1 8 dmin=5, dmax=5
Our Algorithm
6 dmin=5, dmax=15
0.8 JOET: Equal Offloading Time Allocation
dmin=15, dmax=15
4
0.6 dmin=15, dmax=25
2
0.4
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
V 1015
V 1015

Fig. 3. Impact of different values of V on average number of admitted tasks.


Fig. 5. Average number of admitted tasks for various average distances
between the AP and the devices for NAP = 4.
108
3.5
Average queue backlog at the ULPD (bits)

10 6
3 2

Average number of admitted tasks (bps)


dmin=5, dmax=5
2.5 JOET: Equal Offloading Time Allocation dmin=5, dmax=15
1.5 dmin=15, dmax=15
Our Algorithm
2
dmin=15, dmax=25

1.5 1

0.5
0.5
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
V 1015
0
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Fig. 4. Impact of different values of V on average sum of task queue lengths. V 10 15

Fig. 6. Average number of admitted tasks for various average distances


of harvested energy and the number of received command between the AP and the devices for NAP = 1.
messages. Additionally, since offloading and local computation
capabilities of devices decrease as the average distance in-
creases, less number of tasks can be executed which decreases tem state and the current task and energy queue backlogs,
the computational throughput. Finally, the impacts of different the optimal time allocations for EH and command message
number of antennas at the AP is studied in Fig. 6. For higher decoding, optimal CPU-cycles frequencies for local compu-
number of antennas at the AP, the AP can use maximum tation, the optimal time and energy allocation for offloading
ratio transmission (MRT) to increase the amount of harvested are obtained. Our theoretical analysis demonstrated that our
energy as well as the number of received command messages, proposed algorithm can provide a bounded gap to the optimal
and therefore the amount of admitted tasks at the devices. solution at the cost of linear increase in task queue lengths and
Furthermore and due to the MRC (maximum ratio combining) required battery capacities. More specifically, we showed that
gain, the uplink achievable sum rate is higher in multiple- the performance of our algorithm obeys the classical trade-off
antenna scenario, and therefore, the devices can offload more (O(1/V ), O(V )) between the system utility and the average
tasks to the MEC server which increases the computational task queue backlogs. Finally, numerical results have also been
throughput. provided to evaluate the performance of our algorithm.
V. C ONCLUSION
In this paper, the problem of radio and computational
ACKNOWLEDGEMENT
resource allocation for multi-user SWIPT-based MEC systems
has studied. By leveraging Lyapunov optimization technique,
an online time-average utility maximization algorithm has This work was supported by F.R.S.-FNRS under the EOS
been developed, where utility of system compromises both program (EOS project 30452698), by INNOVIRIS under the
computational throughput and fairness. Under our proposed COPINE-IOT project and by UCL under the ARC SWIPT
algorithm, at each time slot and based on the current sys- project.

Authorized licensed use limited to: Khon Kaen University provided by UniNet. Downloaded on March 02,2021 at 07:40:05 UTC from IEEE Xplore. Restrictions apply.
VI. A PPENDIX R EFERENCES
A. Proof of Lemma 2 [1] C. Li, W. Chen, H. Tang, Y. Xin and Y. Luo, “Stochastic Computation
Resource Allocation for Mobile Edge Computing Powered by Wireless
First part is straightforward as the objective function is Energy Transfer,” in Ad Hoc Networks, vol. 93, May 2019.
monotonically decreasing w.r.t. Ei,H (t) if Xi (t) ≥ 0. The [2] F. Wang, H. Xing and J. Xu, “Optimal Resource Allocation for Wireless
Powered Mobile Edge Computing with Dynamic Task Arrivals,” ICC
second part can be proved by induction. For t = 0, from eq. 2019 - 2019 IEEE International Conference on Communications (ICC),
(23), we have Xi (0) + Ei,H (0) ≤ 0. Assume that Xi (t) ≤ 0. Shanghai, China, pp. 1-7, 2019.
Therefore: [3] N. Janatian, I. Stupia and L. Vandendorpe, “Optimal resource alloca-
tion in ultra-low power fog-computing SWIPT-based networks,” 2018
IEEE Wireless Communications and Networking Conference (WCNC),
Xi (t + 1) = Bi (t) + Ei,H (t) − Ei,C (t) − Bimax Barcelona, pp. 1-6, 2018.
≤ Xi (t) + Ei,H (t) ≤ 0 , (38) [4] S. Huang, B. Lv and Rui Wang, “MDP-Based Scheduling Design for
Mobile-Edge Computing Systems with Random User Arrival,” arXiv
where the last inequality is due to the constraint (23) imposed preprint arXiv:1904.13024, 2019.
[5] Y. Mao, J. Zhang, S. H. Song and K. B. Letaief, “Stochastic Joint Radio
in PDL . For the third part, given μi,H (t), the optimal ID and Computational Resource Management for Multi-User Mobile-Edge
time duration should be the minimizer of convex function Computing Systems,” in IEEE Transactions on Wireless Communica-
def
J(t) = Qi (t)μi,ID (t)Ri,DL (t) − V log 1 + μi,ID (t)Ri,DL (t) tions, vol. 16, no. 9, pp. 5994-6009, Sept. 2017.
[6] G. Zhang, W. Zhang, Y. Cao, D. Li and L. Wang, “Energy-Delay
under the relevant constraints. Taking the derivative of J(t) Tradeoff for Dynamic Offloading in Mobile-Edge Computing System
w.r.t. μi,DL (t), it can be shown that its stationary point falls With Energy Harvesting Devices,” in IEEE Transactions on Industrial
below zero if V ≤ Qi (t) ln 2. Informatics, vol. 14, no. 10, pp. 4642-4655, Oct. 2018.
[7] H. Wu, X. Lyu and H. Tian, “Online Optimization of Wire-
less Powered Mobile-Edge Computing for Heterogeneous Indus-
B. Proof of Theorem 1 trial Internet of Things,” in IEEE Internet of Things Journal. doi:
At t = 0, eq. (34) holds. We prove eq. (34) by induction. 10.1109/JIOT.2019.2932995
[8] F. Wang, J. Xu, X. Wang and S. Cui, “Joint Offloading and Computing
Assume that Qi (t) ≤ V / ln 2 + Amaxi . Then: Optimization in Wireless Powered Mobile-Edge Computing Systems,”
• If Qi (t) ≤ V / ln 2, then Qi (t + 1) ≤ Qi (t) + Ai (t) ≤ in IEEE Transactions on Wireless Communications, vol. 17, no. 3, pp.
1784-1797, March 2018.
V / ln 2 + Amax
i . [9] L. Li, R. Cai, H. Jiang and X. Su, “Rate-Energy Tradeoff for SWIPT
• If V / ln 2 ≤ Qi (t) ≤ V / ln 2 + Ai , then according to
max
Systems with Multi-User Interference Channels under Non-Linear En-
Lemma 2, then Ai (t) = 0, and therefore, Qi (t + 1) ≤ ergy Harvesting Model,” 2019 IEEE 89th Vehicular Technology Con-
ference (VTC2019-Spring), Kuala Lumpur, Malaysia, pp. 1-6, 2019.
Qi (t) ≤ V / ln 2 + Amax
i . [10] E. Boshkovska, D. W. K. Ng, N. Zlatanov and R. Schober, “Practical
Now we turn to Di (t). First, we note that due to the inequality Non-Linear Energy Harvesting Model and Resource Allocation for
SWIPT Systems,” in IEEE Communications Letters, vol. 19, no. 12,
log(1 + x) ≤ x/ log 2, we have Ci,O (t) ≤ Ci,O
max
. Similarly, we pp. 2082-2085, Dec. 2015.
consider two cases: [11] L. Huang and M. J. Neely, “Utility Optimal Scheduling in Energy-
Harvesting Networks,” in IEEE/ACM Transactions on Networking, vol.
• If Di (t) ≤ V / ln 2 + Ai , then Di (t + 1) ≤ Di (t) +
max
21, no. 4, pp. 1117-1130, Aug. 2013.
Ci,O ≤ V / ln 2 + Ai + Ci,O
max max max
. [12] X. Lyu, W. Ni, H. Tian, R. P. Liu, X. Wang, G. B. Giannakis and A.
• If Di (t) ≥ V / ln 2 + Ai , then we have Di (t) ≥ Qi (t),
max Paulraj, “Optimal Schedule of Mobile Edge Computing for Internet of
Things Using Partial Information,” in IEEE Journal on Selected Areas
and therefore i ∈ / Np . From Section III-C, we have in Communications, vol. 35, no. 11, pp. 2606-2615, Nov. 2017. n IEEE
Ci,O (t) = 0, and thus Di (t + 1) ≤ Di (t) ≤ V / ln 2 + Communications Surveys and Tutorials, vol. 16, no. 1, pp. 5-24, First
Amax
i + Ci,O
max
. Quarter 2014.
[13] H. Shi, R. V. Prasad, E. Onur and I. G. M. M. Niemegeers, “Fairness
in Wireless Networks:Issues, Measures and Challenges,” i
C. Proof of Theorem 2 [14] M. J. Neely, Stochastic Network Optimization with Application to
We consider two cases. 1) Assume that for ith device at Communication and Queueing System. Morgan&Claypool, 2010.
[15] J. Liu, K. Xiong, P. Fan, Z. Zhong and K. B. Letaief, “Optimal
time slot t, we have Bi (t) ≤ Ei,C max
. Based on eq. (36), we Design of SWIPT-Aware Fog Computing Networks,” IEEE INFOCOM
have Xi (t) ≤ −Xi . From eqs. (29), (32) and (35), it is clear
max
2019 - IEEE Conference on Computer Communications Workshops
 
that for this device at time t, we have Ei,O (t) = Ei,M (t) = 0. (INFOCOM WKSHPS), Paris, France, pp. 13-19, 2019.
 [16] D. Bertsekas and R. Gallager, Data Networks. New Jersey: PrenticeHall,
Thus, if Bi (t) ≤ Ei,C , then Ei,C (t) = 0. 2) The other case,
max
Inc., 1992.
i.e., Bi (t) ≥ Ei,C
max
, is trivial. [17] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge Uni-
versity Press, New York, NY, USA, 2004.
D. Proof of Theorem 3 [18] Z. Hadzi-Velkov, I. Nikoloska, G. K. Karagiannidis and T. Q. Duong,
“Wireless Networks with Energy Harvesting and Power Transfer: Joint
First, we define another optimization problem, denoted by Power and Time Allocation,” in IEEE Signal Processing Letters, vol.
P2 by removing eqs. (11) and (12) and replacing eq. (10) with 23, no. 1, pp. 50-54, Jan. 2016.
[19] F. Amirnavaei and M. Dong, “Online Power Control Optimization for
Ēi,C = Ēi,H , where Ēi,C and Ēi,H are long-term time-averages Wireless Transmission With Energy Harvesting and Storage,” in IEEE
of Ei,C (t) and Ei,H (t). Similarly to [19, Section 3], we can Transactions on Wireless Communications, vol. 15, no. 7, pp. 4888-
P2 P1
show P2 is indeed a relaxed version of P1 i.e., Ūsum ≥ Ūsum . 4901, July 2016.
Following the same approach as in [19, Theorem 1], we can
P2 P1
show that ΔV (Θ(t)) = B − V Ūsum ≤ B − V Ūsum . The rest
of proof is similar to the proofs of [19, Theorem 1] and [5,
Theorem 1], and hence omitted due to the lack of space.

Authorized licensed use limited to: Khon Kaen University provided by UniNet. Downloaded on March 02,2021 at 07:40:05 UTC from IEEE Xplore. Restrictions apply.

You might also like