Power-Delay Trade-off for Heterogenous Cloud Enabled Multi-UAV Systems
Power-Delay Trade-off for Heterogenous Cloud
Enabled Multi-UAV Systems
Ruiyang Duan∗, Jingjing Wang∗, Jun Du∗, Chunxiao Jiang†, Tong Bai‡ and Yong Ren∗
∗Department of Electronic Engineering, Tsinghua University, Beijing, 100084, China
†Tsinghua Space Center, Tsinghua University, Beijing, 100084, China
‡University of Southampton, Southampton, SO17 3RY, United Kingdom
Email: .edu.cn
Abstract—Unmanned aerial vehicles (UAVs) have been widely
used in a range of compelling applications. However, some of
them are incompetent in tackling with computation-intensive
tasks due to limited processing capability and battery life. In
this paper, we combine the mobile edge computing and traditional
cloud computing techniques for offloading the tasks from multi-
UAV systems. Specifically, we jointly optimize the task scheduling
and resource allocation in the heterogeneous cloud architecture,
where we strike a power-delay trade-off of the system relying
on the queue theory and Lyapunov optimization, followed by
its optimal strategy analysis in each time slot. Moreover, we
conceive an iterative algorithm with a closed-form solution at each
iteration round in order to reduce the computational complexity.
Finally, numerical results demonstrate both the feasibility and
effectiveness of our proposed scheme. This paper validates that
the heterogeneous cloud structure can be the beneficial for
improving quality-of-service performance of multi-UAV systems.
Index Terms—Heterogenous cloud, task scheduling, resource
allocation, power-delay trade-off.
I. INTRODUCTION
Thanks to their low cost, high flexibility and easy de-
ployment, unmanned aerial vehicles (UAVs) have been wide-
ly used in both military and civilian applications such as
military surveillance, environment monitoring and emergency
rescue [1]. They are usually equipped with a variety of sensors
to fulfill different tasks. However, some of these tasks are
computation-intensive, which poses a great challenge to UAVs
for their limited processing capacity and battery life. Hence,
new solutions should be conceived to handle the explosive
computation demand as well as the ever-increasing require-
ment of computational quality. Mobile edge computing (MEC)
has become a promising technology for releasing the tension
between the computation-intensive tasks and resource-limited
devices [2], where devices can offload their tasks to cloud
servers which are deployed locally, and the cloud servers return
the results to the devices. Relying on MEC, quality-of-service
(QoS), such as power consumption and execution delay, can
be greatly improved.
However, in comparison to traditional cloud computing, the
resource in the edge cloud is restricted by its scale. Therefore,
the resource allocation problem becomes a critical issue in
MEC, which has attracted much attention [3]–[7]. Specially,
in [3], Sardellitti et al. proposed an iterative algorithm based
on successive convex approximation to jointly allocate the
radio resource and computational resource to multiple users
in a MIMO MEC system. Moreover, a power-delay trade-
off problem in a multi-user MEC system was formulated
in [4], where the local processing ability of devices was
considered and an optimal resource allocation scheme was
designed with the aid of Lyapunov optimization. In [5], Liu
et al. studied the delay-optimal task scheduling and resource
allocation problem under power constraints in MEC systems
modeled by a Markov decision process, which were capable
of achieving shorter average execution delay in comparison to
other schemes.
Since the traditional cloud, i.e. the remote cloud, has
abundant computational resource, the heterogenous cloud ar-
chitecture including both the edge cloud and the remote
cloud has been constructed for jointly improving the QoS
performance [8]. To elaborate a little further, in [9], Gelenbe et
al. studied the optimal load sharing problem between a local
and remote cloud, where an optimal scheme was proposed
based on the analysis of the power consumption and services
time in the context of diverse tasks with different types and
requirements. The fair resource allocation problem in the
heterogenous cloud system was investigated in [10], where
a multi-resource allocation mechanism, namely DRFH, was
designed both to guarantee the fairness of members and to
keep the service isolation among users. The delay-bounded
task offloading problem in the heterogenous cloud system was
highlighted in [11] considering both the wireless transmission
delay as well as the execution delay. Zhao et al. modeled
the arrival-service process as an M/M/1 queue, relying on
which the success probability of the delay-bounded task was
derived in the context of a single-user case and a multi-user
case. Additionally, a total power minimization assisted task
scheduling problem was studied in [12] by Gai et al.
Although many efforts have been done on the research of
UAV networks, of MEC and on heterogenous cloud, few works
focus on the combination of them. Against this background, in
this paper, we study the energy efficient network association
problem with a focus on task scheduling and resource alloca-
tion in heterogenous cloud aided multi-UAV systems, where
UAVs can offload their tasks to a heterogenous cloud system
which is composed of an edge cloud and a remote cloud. Our
main concern is to determine the percentage of the tasks to be
assigned to the edge cloud as well as the computation resource
978-1-5386-8088-9/19/$31.00 ©2019 IEEE
888ddd
高亮
888ddd
高亮
Fig. 1. The structure of the heterogeneous cloud aided multi-UAV system.
to be allocated to different tasks from UAVs. Moreover, the
queue theory is used for modeling the execution delay. Thus,
we formulate the energy efficient network association problem
as a power consumption minimization problem under system
stability constraints. Additionally, a power-delay trade-off is
struck based on the Lyapunov optimization and an iterative
algorithm is designed in order to find the optimal strategy.
The rest of the paper is organized as follows. The system
model is introduced in Section II. In Section III, the power-
delay trade-off is struck based on Lyapnuov optimization con-
sidering the power consumption and execution delay involved.
Moreover, an iterative algorithm is invoked to find the optimal
strategy in each time slot. Simulation results are provided
in Section IV to evaluate the performance of our proposed
scheme, followed by our conclusions in Section V.
II. SYSTEM MODEL
We consider a heterogeneous cloud aided multi-UAV sys-
tem, which consists of multiple UAVs, the edge cloud and the
remote cloud, as shown in Fig. 1. These UAVs carry sensed
data related to different tasks. We assume that the tasks are
computation-intensive, and hence they need to offload the tasks
to two kinds of clouds considered. The UAVs are able to
communicate with the clouds through a radio access network
(RAN). In our model, the edge cloud can be viewed as a small-
scale processing center with limited computational capability
installed at the wireless access point (AP), while the remote
cloud is a large-scale processing center with powerful process-
ing ability with the aid of the easy access to the Internet. The
system is operated under a high dynamic circumstance because
of the high mobility of UAVs. For the simplicity of analysis,
the dynamic of the system is characterized by discrete time
slots, i.e. t ∈ T = {1, 2, 3, …}. In each time slot t, UAVs first
transmit corresponding sensed data to an AP over the air, and
then the data is processed by the virtualized machines (VMs)
either in the edge cloud or in the remote cloud.
A. The UAV Model
We assume that there are total N UAVs to execute N
different tasks in the multi-UAV system, which can be denoted
as N = [1, 2, …, N ]. The tasks are assigned by the ground
control station (GSC). Once being assigned a task, the related
UAV will fly to certain area according to the pre-defined
path, collect related data and then navigate to the nearby AP
to offload the data. These UAVs are equipped with sensor
units, control units, management units and communication
units to fulfill the tasks. Specifically, the communication units
are composed of multiple modules configured by various
protocols [1], so that the UAVs can communicate with the AP
via Wi-Fi, long-term evolution (LTE), etc. as the circumstance
may require. Relying on the high flexibility, these UAVs can
fly close to and hover around an AP to ensure the reliable
wireless transmission. Under this assumption, we neglect the
occurrence of the transmission outage resulted from the poor
channel condition between the AP and UAVs. Let Ai(t) denote
the number of packets arriving at the AP from UAV i in time
slot t, while A(t) = [A1(t), A2(t), …, AN (t)] represent the
packet arrival vector in time slot t, where Ai(t) in different
time slots is independent and identically distributed (i.i.d.). The
number of the arrival packets is bounded between Ai,min and
Ai,max, i.e. Ai,min ≤ Ai(t) ≤ Ai,max. Let λi = E [Ai(t)] be
the average packet arrival rate from UAV i in time slot t.
B. Cloud Execution Model
In our paper, the packets from different UAVs are offloaded
and processed at the clouds. In time slot t, after receiving the
packets transmitted from related UAVs, part of these packets
from UAV i, denoted as Ae,i(t), will be locally processed at
the edge cloud, while Ar,i(t) packets will be tackled at the
remote cloud. Hence, we have Ai(t) = Ae,i(t) +Ar,i(t).
1) Edge Cloud: In the edge cloud, there are total N
VMs exclusively used by N UAVs. When packets reach
the edge cloud, the corresponding VM allocates appropriate
computational resources to support the data processing task.
The computational resource is decided by the CPU-cycle
frequency [4]. In this paper, we assume that the available CPU-
cycle frequency allocated to UAV i at time slot t is represented
by fi(t), where fi(t) ≤ fi,max. We define the number of the
packets that can be served by the VM i in the edge cloud as:
μi(t) = Δtfi(t)L
−1
i , (1)
where Δt denotes the duration of the time slot, while Li is
the minimum CPU cycles needed to process one packet from
UAV i. Moreover, the power consumption during the whole
execution process in the edge cloud can be approximately
formulated by:
Pe,i(t) = κf
3
i (t), (2)
where κ represents the capacitance coefficient.
The newly arrived packets will be queued in the buffer of the
edge cloud when the corresponding VM is busy with dealing
with early arrived packets. Let Qi(t) denote the queue length
in the ith VM related to UAV i and we have Qi(0) = 0 at
the beginning. Then the dynamic of the queue length can be
expressed as:
Qi(t+ 1) = max{Qi(t)− μi(t), 0}+Ae,i(t). (3)
Because of the local deployment of the edge cloud, the
execution delay in the edge cloud mainly stems from the
queue. Relying on the Little’s Law [13], the average queueing
delay can be derived from the average queue length, which is
calculated by Eq. (3).
2) Remote Cloud: By contrast, the power consumption in
the remote cloud is closely related to the variation of the
workload, which is formulated by:
Pr,i(t) = Ce
Ar,i(t), (4)
where C is a constant coefficient. As for the execution delay
in the remote cloud, given that it is equipped with a multi-
core high-speed CPU, we assume that the arrived packets
can be processed without the queueing time. Hence, the time
delay largely depends on the Internet transmission process.
Moreover, referring to [14], the Internet transmission delay
follows an empirical distribution. For simplicity, we assume
that Internet transmission process has a deterministic delay D.
III. TASK SCHEDULING AND RESOURCE ALLOCATION
SCHEME
As mentioned before, task scheduling is an important issue
in heterogeneous cloud aided systems. Specifically, if all the
packets are locally processed at the edge cloud, it is difficult
to keep the system stable because both the computational
capability and the buffer capacity in the edge cloud are limited.
On the other hand, if all the packets are offloaded and tackled
at the remote cloud, the power consumption is too high. In this
section, our goal is to find an optimal scheme to determine the
percentage of packets processed at the edge cloud as well as
the amount of computational resources allocated for different
tasks considering both the average power consumption and the
cloud execution delay.
A. Problem Formulation
The power consumption for handling the packets from UAV
i in time slot t is composed of two parts, i.e. the power
consumption in the edge cloud and the power consumption
in the remote cloud. Let Pi(t) = Pe,i(t) + Pr,i(t) denote the
power consumption for fulfilling the task from UAV i in time
slot t. Considering a total number of N UAVs, the system’s
power consumption can be calculated as P (t) =
∑N
i=1 Pi(t).
Then, the average power consumption is given by:
P̄ = lim
t→∞
1
t
E
[
t−1∑
τ=0
P (t)
]
. (5)
Since the Internet transmission delay in the remote cloud
is a constant, the average execution delay largely hinges on
the sojourn time in each edge cloud’s queue. Relying on the
Little’s Law [13], we use the average queue length in the task
buffer as a measurement of the average execution delay, which
is expressed as:
Q̄i = lim
t→∞
1
t
E
[
t−1∑
τ=0
Qi(t)
]
. (6)
Definition 1. The queue system in the edge cloud related to
UAV i is strong stable, if:
lim
t→∞
1
t
E
[
t−1∑
τ=0
Qi(t)
]
< ∞. (7)
Note that if all the queues are stable, the system’s average
service rate is equal to the packet arrival rate, i.e.
lim
t→∞
1
t
t−1∑
τ=0
ui(t) = lim
t→∞
1
t
t−1∑
τ=0
Ae,i(t), (8)
which guarantees that the arrived packets can be processed in
terms of finite time delay.
In this paper, we aim to find a task scheduling and re-
source allocation scheme for the sake of minimizing the
average power consumption under the constraint that all the
tasks can be executed with finite time delay. Let η(t) =
[η1(t), η2(t), ..., ηM (t)], where ηi(t) =
Ae,i(t)
Ai(t)
, and f(t) =
[f1(t), f2(t), ..., fM (t)]. Hence, the problem can be formulated
as:
P1 : min
η(t),f(t)
P̄
s.t. 0 ≤ ηi(t) ≤ 1, i = 1, 2, ..., N, t ∈ T , (9a)
0 ≤ fi(t) ≤ fi,max, i = 1, 2, ..., N, t ∈ T , (9b)
lim
t→∞
1
t
E
[
t−1∑
τ=0
Qi(t)
]
< ∞, i = 1, 2, ..., N. (9c)
To elaborate, (9a) requires that the packets from UAV i should
be processed by either the edge cloud or the remote cloud,
while (9b) denotes that the CPU-cycle frequency meets the
given threshold. Moreover, (9c) allows that the packets tackled
in the edge cloud can be processed within finite time delay.
B. Lyapunov Optimization Based Greedy Solution
P1 is a stochastic optimization problem since η(t) and f(t)
is time-varying. In this section, we design an online algorithm
to determine the optimal task scheduling and resource alloca-
tion strategy in each time slot based on Lyapunov optimization.
Lyapunov optimization is designed to greedily minimize the
queue backlogs by solving a deterministic problem. Detailed
descriptions concerning to Lyapunov optimization can be
referred to [15].
First of all, we define the function L (Q(t)) as the sum of
squares of backlogs in the queue, which is expressed as:
L (Q(t)) =
1
2
N∑
i=0
Q2i (t). (10)
Then, the conditional Lyapunov drift for time slot t can be
given by:
Δ(Q(t)) = E [L (Q(t+ 1))− L (Q(t)) |Q(t)] . (11)
If we minimize Δ(Q(t)) in each time slot, it may stabilize
the whole system. However, the average power consumption is
substantially large. Alternatively, we define a drift-plus-penalty
function ΔV (Q(t)) to strike a trade-off between the power
consumption and queue backlogs, i.e.
ΔV (Q(t)) = Δ (Q(t)) + V · E [P (t)|Q(t)] , (12)
where V is the weighting parameter. A large value of V is
beneficial of optimizing the average power consumption yet
with the expense of large average delay. Thus, the drift-plus-
penalty function provably strikes a balance between the power
consumption and cloud execution delay.
Lemma 1. For any η(t) and f(t) such that ηi(t) ∈
[0, 1] , fi(t) ∈ [0, fi,max] , i = 1, 2, ..., N , ΔV (Q(t)) is upper
bounded by
ΔV (Q(t))
≤ 1
2
E
[
N∑
i=1
(
A2e,i(t) + μ
2
i (t)− 2Ae,i(t)μi(t)
)
|Q(t)
]
− E
[
N∑
i=1
Qi(t) (μi(t)−Ae,i(t))|Q(t)
]
+ V · E [P (t)|Q(t)]
(13)
Proof: Relying on Eq. (3), the difference of the Lyapunov
function in adjacent time slots can be given by:
L (Q(t+ 1))− L (Q(t))
=
1
2
N∑
i=1
(
Q2i (t+ 1)−Q2i (t)
)
=
1
2
N∑
i=1
[(
max {Qi(t)− μi(t), 0}+A2e,i(t)
)2 −Q2i (t)
]
≤ 1
2
N∑
i=1
[
A2e,i(t) + μ
2
i (t)− 2Ae,i(t)μi(t)
]
−
N∑
i=1
[Qi(t) (μi(t)−Ae,i(t))],
(14)
where the last deduction of expression (14) relies on the
conclusion:
(max {Q− μ, 0}+A)2
≤ Q2 + μ2 +A2 + 2Aμ− 2Q(μ−A),
(15)
for any Q ≥ 0, μ ≥ 0, A ≥ 0. Then, calculate the drift-plus-
penalty function defined in Eq.(12), and we can obtain the
upper bound in Lemma 1.
For the sake of achieving a superior task scheduling and
resource allocation scheme, in this paper we conceive the
Algorithm 1 aiming to greedily minimize the upper bound of
the ΔV (t) in each time slot, where
hc (η(t),f(t)) =
1
2
N∑
i=1
[
A2e,i(t) + μ
2
i (t)− 2Ae,i(t)μi(t)
]
−
N∑
i=1
[Qi(t) (μi(t)−Ae,i(t))] + V · P (t).
(16)
In Algorithm 1, solving the stochastic optimization problem
P1 is reduced to deal with a deterministic optimization
problem P2 in each time slot. By changing the value of V , we
can achieve the power-delay trade-off. Moreover, the proposed
algorithm is capable of minimizing the power consumption and
of driving the queue backlog towards a low congestion state.
Algorithm 1 A Lyapunov Optimization Based Task Schedul-
ing and Resource Allocation Algorithm
1: At the begin of the time slot t, obtain {Qi(t)} and {Ai(t)}.
2: Calculate
P2 : {η(t),f(t)} =argmin hc (η(t),f(t))
s.t. (9a), (9b).
3: Update {Qi(t)} according to Eq. (3) and set t = t+ 1.
C. A Low-complexity Iterative Algorithm
In this section, we provide a low-complexity iterative al-
gorithm to solve P2 based on its structure feature. Since
P2 is a combinational optimization problem in terms of
two variables, the minimization optimization can be achieved
by searching the solution in accordance with the decreasing
gradient direction of either of them. Hence, by fixing one of
the two variables, we can obtain the optimal value of the other
one.
Plugging Eq. (1), Eq. (2) and Eq. (4) into Eq. (14), and we
have:
hc (η(t),f(t))
=
1
2
N∑
i=1
[η2i (t)A
2
i (t) + Δt
2f2i (t)L
−2
i
−2ηi(t)Ai(t)Δtfi(t)L−1i ]
−
N∑
i=1
[
Qi(t)
(
Δtfi(t)L
−1
i − ηi(t)Ai(t)
)]
+ V
N∑
i=1
[
κf3i (t) + Ce
(1−ηi(t))Ai(t)
]
.
(17)
Optimal CPU-cycle Frequency: By fixing η(t), the objec-
tive function of P2 reduces to:
hc (η(t),f(t)|η(t))
=
1
2
N∑
i=1
[
Δt2f2i (t)L
−2
i − 2ηi(t)Ai(t)Δtfi(t)L
−1
i
]
−
N∑
i=1
Qi(t)Δtfi(t)L
−1
i + V
N∑
i=1
κf3i (t).
(18)
Hence, the single-variable optimization problem can be given
by:
P3 : min
η(t)
hc (η(t),f(t)|η(t))
s.t. 0 ≤ fi(t) ≤ fi,max, i = 1, 2, ..., N. (19a)
Since fi(t)
′s are i.i.d., the optimal solution can be achieved
by solving
∂hc(η(t),f(t)|η(t))
∂fi(t)
= 0, which is expressed as:
f∗i (t) =
−Δt2L−2i +
√
Δt4L−4i + 12V κΔtL
−1
i (ηi(t)Ai(t) +Qi(t))
6V κ
.
(20)
We can verify that
∂2hc(η(t),f(t)|η(t))
∂fi
2(t)
∣∣∣
fi(t)=f
∗
i
(t)
≥ 0. Hence,
f∗i (t) is the global minimum point. Because fi(t) is bounded
by fi,max, thus the optimal solution can be given by:
f̂i(t) = min {f∗i (t), fi,max} , i = 1, 2, ..., N. (21)
Optimal Data Diffluence: By fixing f(t), the objective of
of P2 reduces to:
hc (η(t),f(t)|f(t))
=
1
2
N∑
i=1
[
η2i (t)A
2
i (t)− 2ηi(t)Ai(t)Δtfi(t)L−1i
]
+
N∑
i=1
[Qi(t)ηi(t)Ai(t)] + CV
N∑
i=1
e(1−ηi(t))Ai(t).
(22)
Then, the optimization problem can be rewritten as:
P4 : min
η(t)
hc (η(t),f(t)|f(t))
s.t. 0 ≤ ηi(t) ≤ 1, i = 1, 2, ..., N. (23a)
Similarly, P4 can be solved by solving:
∂hc (η(t),f(t)|f(t))
∂ηi(t)
= Ai(t)CV e
(1−ηi(t))Ai(t) − ηi(t)A2i (t) +Ai(t)Δtfi(t)L−1i
−Qi(t)Ai(t) = 0
(24)
Eq. (24) is a transcendental equation, which can be easily
solved by numerical methods such as Newton’s method or
secant method. Let η∗i (t) be the solution of Eq. (24), and
we can also verify that η∗i (t) is the global minimum point.
Considering ηi(t) ∈ [0, 1], the optimal solution can be given
by:
η̂i(t) = max {0,min {η∗i (t), 1}} , i = 1, 2, ..., N. (25)
Relying on the optimal solution obtained by Eq. (21) and
Eq. (25), we provide a low-complexity iterative algorithm to
find the optimal solution for P2. The algorithm is detailed in
Algorithm 2.
IV. SIMULATION RESULTS
In this section, numerical results are provided to evaluate
the performance of our proposed scheme. The length of one
time slot is set to be 30 ms. We assume that Ai,min = 10,
Ai,max = 100 and Ai(t) is uniformly distributed within the
range of [10, 100]. Further set κ = 10−27, L = 3 × 106,
fi,max = 3 GHz and C = 10
−20. The simulation results are
averaged over 100 time slots.
Algorithm 2 A Low-complexity Iterative Algorithm for Solv-
ing P2
1: Initialize η(t), set k = 0 and ε as the threshold.
2: Calculate {f̂ki (t)} according to Eq. (21).
3: Calculate {η̂ki (t)} by solving Eq. (25).
4: Calculate hkc according to Eq. (17).
5: repeat
6: Update {f̂k+1i (t)} according to Eq. (21).
7: Update {η̂k+1i (t)} by solving Eq. (25).
8: Update hk+1c according to Eq. (17).
9: Update k = k + 1.
10: until
∣∣hkc − hk−1c ∣∣ ≤ ε
11: Obtain f̂
opt
i (t) = f
k
i (t), η
opt
i (t) = η̂
k
i (t)
1 5 10 15 20
Number of Iterations
4500
5000
5500
6000
6500
7000
7500
8000
h c
N = 5
N = 10
N = 20
Fig. 2. The convergence performance of Algorithm 2.
Fig. 2 presents the convergence performance of Algorithm
2, where we set V = 4000. It can be inferred that given a
set of randomly generated η0(t), the proposed algorithm is
capable of achieving the convergence in 2 steps. Moreover,
the increment of UAVs almost has no influence on the rate
of convergence. Therefore, we validate that our algorithm
can achieve fast convergence with lower computational cost,
comparing with solving P2 with interior-point methods, .
Fig. 3 shows the relationship between the average power
consumption of task execution/average queue length per user
and the control parameter, where we set N = 10. It can be
observed that the average power consumption decreases with
the increment of V, and converges to the minimum value when
V is sufficiently large. In contrast, the average queue length per
user increases with V and also converges when V is sufficiently
large. It can be explained that when V is sufficiently large,
the optimal data diffluence tends to be a constant, so do the
average power consumption and average queue length. These
results verify the tradeoff between the power consumption and
execution delay caused by the control parameter.
Fig. 4 depicts the comparison of the average pow-
er consumption with respect to queue length perfor-
mance of systems between the scenarios with and with-
0 1 2 3 4
Control parameter V
×104
0
10
20
30
40
50
60
70
A
ve
ra
ge
p
ow
er
c
on
su
m
pt
io
n
0 1 2 3 4
Control parameter V
×104
0
100
200
300
400
500
600
700
800
900
A
ve
ra
ge
q
ue
ue
le
ng
th
p
er
u
se
r
Fig. 3. Average power consumption and queue length versus the control
parameter.
0 500 1000 1500 2000 2500
Average queue length per user
0
50
100
150
200
250
300
A
ve
ra
ge
p
ow
er
c
on
su
m
pt
io
n
With Remote Cloud
Without Remote Cloud
V=50
V=4000
Fig. 4. The comparison of average power consumption and queue length
performance of systems with and without the remote cloud.
out the remote cloud. Here we set N = 10 and V ∈
{50, 100, 200, 300, 400, 500, 1000, 2000, 4000}. It is shown
that in general, the average power consumption decreases
with the increment of average queue length for the both
systems. Therefore, it is necessary to carefully select a suitable
V to balance the power consumption and execution delay.
Moreover, the falling range of the power consumption of the
system without the remote cloud is much larger than the
system with the remote cloud, which indicates that the control
parameter has more influence on the system without the remote
cloud. In addition, more power is consumed for the system
without the remote cloud while leading to longer delay. This
is because the edge cloud cannot stabilize its task buffers, due
to its limited computational capacity if not equipped with the
remote cloud. In this case, Fig. 4 verifies that the superiority
of heterogenous cloud system on improving the quality of
computation experience.
V. CONCLUSIONS
In this paper, we studied the energy efficient network
association problem in heterogenous cloud aided multi-UAV
systems. We dedicate on the optimization of task scheduling
and resource allocation. We formulated the problem as a power
consumption minimization problem under system stability
constraints. A power-delay trade-off was struck as well as
an optimal solution relying on Lyapunov optimization was
designed. Numerical results showed that our proposed scheme
can reach the optimal solution within a few iterative rounds
and can yield an improved QoS performance.
ACKNOWLEDGEMENT
This work is supported by the China Postdoctoral Sci-
ence Foundation (2018M640130), the National Natural Sci-
ence Foundation of China (91338203), the Pre-research
Fund of Equipments of Ministry of Education of China
(6141A02022615).
REFERENCES
[1] J. Wang, C. Jiang, Z. Han, Y. Ren, R. G. Maunder, and L. Hanzo,
“Taking drones to the next level: Cooperative distributed unmanned-
aerial-vehicular networks for small and mini drones,” Ieee VehIcular
Technology MagazIne, vol. 12, no. 3, pp. 73–82, Jul. 2017.
[2] 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,
no. 11, pp. 1–16, Sep.
[3] S. Sardellitti, G. Scutari, and S. Barbarossa, “Joint optimization of
radio and computational resources for multicell mobile-edge computing,”
IEEE Transactions on Signal and Information Processing over Networks,
vol. 1, no. 2, pp. 89–103, Jun. 2015.
[4] Y. Mao, J. Zhang, S. Song, and K. B. Letaief, “Power-delay tradeoff in
multi-user mobile-edge computing systems,” in IEEE Global Commu-
nications Conference (GLOBECOM), Washington, DC, Feb. 2016, pp.
1–6.
[5] J. Liu, Y. Mao, J. Zhang, and K. B. Letaief, “Delay-optimal computation
task scheduling for mobile-edge computing systems,” in IEEE Interna-
tional Symposium on Information Theory (ISIT), Barcelona, Spain, Jul.
2016, pp. 1451–1455.
[6] C. Jiang, Y. Chen, K. R. Liu, and Y. Ren, “Renewal-theoretical dynamic
spectrum access in cognitive radio network with unknown primary
behavior,” IEEE Journal on Selected Areas in Communications, vol. 31,
no. 3, pp. 406–416, Feb. 2013.
[7] C. Jiang, Y. Chen, Y. Gao, and K. R. Liu, “Joint spectrum sensing and
access evolutionary game in cognitive radio networks,” IEEE transac-
tions on wireless communications, vol. 12, no. 5, pp. 2470–2483, Mar.
2013.
[8] Z. Sanaei, S. Abolfazli, A. Gani, and R. Buyya, “Heterogeneity in mobile
cloud computing: Taxonomy and open challenges,” IEEE Communica-
tions Surveys & Tutorials, vol. 16, no. 1, pp. 369–392, May. 2014.
[9] E. Gelenbe, R. Lent, and M. Douratsos, “Choosing a local or remote
cloud,” in The Second Symposium on Network Cloud Computing and
Applications (NCCA), London, UK, Mar. 2012, pp. 25–30.
[10] W. Wang, B. Liang, and B. Li, “Multi-resource fair allocation in
heterogeneous cloud computing systems,” IEEE Transactions on Parallel
and Distributed Systems, vol. 26, no. 10, pp. 2822–2835, Oct. 2015.
[11] T. Zhao, S. Zhou, X. Guo, and Z. Niu, “Tasks scheduling and resource
allocation in heterogeneous cloud for delay-bounded mobile edge com-
puting,” in IEEE International Conference on Communications (ICC),
Paris, France, Jul. 2017, pp. 1–7.
[12] K. Gai, M. Qiu, and H. Zhao, “Energy-aware task assignment for mobile
cyber-enabled applications in heterogeneous cloud computing,” Journal
of Parallel and Distributed Computing, vol. 111, pp. 126–135, Jan. 2018.
[13] J. D. Little and S. C. Graves, “Little’s law,” in Building Intuition.
Springer, 2008, pp. 81–100.
[14] G. Hooghiemstra and P. Van Mieghem, “Delay distributions on fixed
internet paths,” Delft University of Technology, Report, vol. 20011020,
2001.
[15] M. J. Neely, “Stochastic network optimization with application to
communication and queueing systems,” Synthesis Lectures on Commu-
nication Networks, vol. 3, no. 1, pp. 1–211, 2010.