4322 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 71, NO.

4, APRIL 2022

GreenEdge: Joint Green Energy Scheduling and

Dynamic Task Offloading in Multi-Tier
Edge Computing Systems
Huirong Ma , Graduate Student Member, IEEE, Peng Huang , Graduate Student Member, IEEE,
Zhi Zhou , Member, IEEE, Xiaoxi Zhang , Member, IEEE, and Xu Chen , Senior Member, IEEE

Abstract—As mobile edge computing (MEC) emerges as a are required. However, for devices, they usually have limited
paradigm to meet the ever-increasing computation demands from computing capacity and scant energy supply. Besides, offloading
real-time Internet of Things (IoT) applications in 5 G era, the tasks to the cloud to execute may incur long end-to-end latency
development trends of which are mainly divided into two, with one
being MEC with advanced computing architectures, and the other by the wide-area network (WAN) which is far beyond the
being MEC with high efficiency for sustainable operations. We are stringent latency required by emerging IoT applications such
committed to taking advantage of these two trends to explore a as augmented reality [2]–[4] and interactive gaming [5]. As a
novel multi-tier edge computing scenario with hierarchical task remedy to the above limitations, the recently proposed tech-
offloading and green energy provisioning via leveraging the energy nology, mobile edge computing (MEC) [6]–[9] can extend the
harvesting (EH) technique. Specifically, we focus on the key prob- cloud to the network edge thus to satisfy the stringent timeliness
lem of joint task offloading and energy scheduling in such green
multi-tier edge computing systems. We aim to minimize the task requirement and intensive computation demands of emerging
execution cost by jointly considering the system cost that covers IoT applications [10].
latency, energy consumption, and cloud rental fees. By formulat- A typical MEC system consists of a set of geographically
ing the problem as a stochastic optimization problem, we invoke distributed edge servers which are deployed on the periphery of a
the Lyapunov technique to decompose the long-term optimization network, providing elastic resources such as storage, computing,
problem into a series of one-slot optimization problems which only and network bandwidth [11]. Considering the distance between
use the current system information. To solve the one-slot opti-
mization problem which is a mixed-integer linear problem (MILP) the edge servers and IoT devices, edge servers are often orga-
proved to be NP-hard, we first relax the integer variables into nized in a hierarchical fashion. Therefore, the resource-limited
real ones to obtain the optimal fractional solutions. Considering IoT devices which are heavily loaded can offload the tasks to
the capacity of the physical resources of each edge server, we nearby edge servers, thus to reduce the energy consumption and
propose a resource-constrained randomized dependent rounding processing latency; meanwhile, in such a way, edge servers at
algorithm to properly round up or down the fractional variables the lower tier can flexibly share the stringent workload with
to get a feasible yet near-optimal solution. We conduct rigorous
theoretical analysis and extensive simulations to verify the superior the edge servers with greater capacities in the upper tier. With
performance of the proposed schemes. all the effectiveness of edge servers on reducing the energy
consumption, resource requirement and perceived latency of the
Index Terms—Multi-tier edge computing, task offloading, green IoT devices, however, the low energy efficiency at the edge side
energy scheduling, randomized dependent rounding.
may become the bottleneck towards sustainable edge computing.
Besides, a major limitation of edge servers is that the computing
I. INTRODUCTION capacity of edge servers is still significantly limited.
As a last consideration, 5 G technology involves the deploy-
N THE face of the proliferation of real-time IoT applications
I in 5 G era [1], vast computing capacity and energy supply
ment of a large number of base stations (BSs), to increase the
network coverage and provide higher throughput to the users.
This however results in higher energy consumption, which is
Manuscript received October 23, 2021; revised December 24, 2021; accepted expected to considerably contribute to carbon emissions [12].
January 23, 2022. Date of publication January 31, 2022; date of current version Taking China as an example, the annual network power con-
May 2, 2022. This work was supported in part by the National Science Foun-
dation of China under Grants U20A20159, 61972432 and 62102460, in part by
sumption is expected to exceed 100 billion kWh. As thermal
the Program for Guangdong Introducing Innovative and Entrepreneurial Teams power prevails at present, this power consumption soar will
under Grant 2017ZT07X355, and in part by Pearl River Talent Recruitment result in 27.2 billion kilograms of carbon emissions. Therefore,
Program under Grant 2017GC010465. The review of this article was coordinated there is an urgent need to increase the energy efficiency of
by Dr. Zehui Xiong. (Corresponding author: Xu Chen.)
The authors are with the School of Computer Science and En- edge servers and reduce the carbon emissions in edge tiers [13].
gineering, Sun Yat-sen University, Guangzhou, Guangdong 510006, Thanks to the recent advancements of energy harvesting tech-
China (e-mail: [email protected]; [email protected]; niques [14], [15], off-grid green energy harvested from the
[email protected]; [email protected]; chenxu35@
mail.sysu.edu.cn). environment (such as solar radiation, wind power generation,
Digital Object Identifier 10.1109/TVT.2022.3147027 and kinetic human motion) is embraced as an important type of

0018-9545 © 2022 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: CINVESTAV. Downloaded on October 10,2022 at 19:37:15 UTC from IEEE Xplore. Restrictions apply.
MA et al.: GREENEDGE: JOINT GREEN ENERGY SCHEDULING AND DYNAMIC TASK OFFLOADING IN MULTI-TIER EDGE COMPUTING SYSTEMS 4323

Fig. 1. An illustration of the multi-tier edge computing system with green energy and grid energy.

energy supply for the edge servers in edge tiers in the field. In problem to increase the energy efficiency of edge servers
addition, the emergence of energy harvesting technique helps and reduce the operational cost. We introduce harvested en-
to achieve green computing [16]. As shown in Fig. 1, in our ergy into the edge servers when managing a multi-tier edge
paper, the edge servers in front edge tier can not only harvest computing system and consider minimizing the system cost
the green energy but also use the grid power to supply energy that covers latency, energy consumption, and cloud rental
such as the smart street light system [17] [18] in the smart city fees. In addition, our objective function is more general
which employs green energy powered street lighting as the IoT than those proposed by many related works in that we
network communication node device. Despite the advantages, include various types of energy consumption of two tiers of
the randomness and suddenness of the arrival of green energy edge servers (including green energy and fossil energy), the
introduce new challenges for fully reaping the benefits of edge cloud rental fees, and the latency within the edge servers.
servers in edge tiers to achieve green computing. r Online Algorithm Design: We propose a novel online
Therefore, to embrace the profound benefits enabled by MEC- algorithm framework that nicely integrates the Lyapunov
EH in multi-tier edge computing scenario, a task provider needs optimization technique [19] and an efficient optimization
to address the following major issues jointly: 1) task offload- rounding algorithm. More specifically, through a non-
ing, which determines where to offload the tasks subject to trivial transformation, we decouple the problem into a
the computing capacity of edge servers in different edge tiers; series of one-slot online decision problems which do not
and 2) green energy scheduling, how much green energy is require the future system states as a prior. However, the
used to compute the tasks to minimize the system cost that decoupled one-slot problem is a mixed-integer linear pro-
covers latency, energy consumption, and cloud rental fees. The gramming (MILP) problem which is proved to be NP-hard.
timely decision making regarding these two issues is critical but To efficiently solve it, we further relax the problem, and
challenging, due to temporal variations of system dynamics in then use a novel resource-constrained randomized depen-
a wireless environment, uncertainty in the resulting offloading dent rounding (RCRDR) algorithm to convert the fractional
costs, and the randomness of the arrival of green energy. solution into a feasible and near-optimal solution of the
In this paper, we address the challenge of incorporating green original problem.
energy into edge servers and focus on the joint task offloading r Theoretical Analysis and Experimental Verification: We
and green energy scheduling problem for multi-tiered edge conduct theoretical analysis and trace-driven simulations to
systems. We propose an efficient approximated optimization evaluate the performance of our proposed algorithms. The
framework, which can make online decisions for our joint results show that our algorithms achieve a tunable trade-
optimization problem without the need to access future infor- off between system minimizing long-term costs and the
mation and has a good theoretical performance guarantee. We queue length while effectively reducing the costs. Besides,
summarize the key results and main contributions of our paper as through extensive performance evaluations, we show that
follows: our sub-routine algorithm (RCRDR) significantly outper-
r Problem Formulation: We formulate the problem of joint forms existing optimization rounding algorithms, and the
task offloading and green energy scheduling in the multi- overall scheme can not only achieve superior performance
tier edge computing system as a stochastic optimization gain over existing benchmarks but is also very close to

Authorized licensed use limited to: CINVESTAV. Downloaded on October 10,2022 at 19:37:15 UTC from IEEE Xplore. Restrictions apply.
4324 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 71, NO. 4, APRIL 2022

the offline optimum with at most 4% higher system cost equipped with energy harvesting module, i.e., previous studies
incurred. aimed at the sustainability of IoT devices while our proposed
The rest of the paper is organized as follows. We discuss scheme aims at the stable operation of edge server in energy
the related work in Section II and present the system model in harvesting environment.
Section III. Then we show the online algorithms in Section IV Definitely, the researches of MEC systems with energy har-
and the performance analysis of our algorithms in Section V. vesting module also has attracted great attention from re-
Section VI evaluates the performance of our proposed algo- searches [28], [29]. Authors in [28] address the dynamic work-
rithms via extensive simulations, while Section VII concludes load offloading and the edge server provisioning of the energy
this paper. harvesting MEC system by proposing an efficient reinforcement
learning-based resource management algorithm. This algorithm
uses a decomposition of the (offline) value iteration and (online)
II. RELATED WORK reinforcement learning, thus achieving a significant improve-
ment of learning rate and run-time performance. By integrating
The computing latency minimizing and the energy consump-
renewable energy into the MEC system, authors of [29] propose
tion minimizing for mobile edge computing systems have at-
an offloading scheduling algorithm that considers the MEC
tracted significant attention in recent years.
battery and quality of service experience (QoE). The algorithm
Many researchers studying mobile edge computing (MEC)
in [29] includes the admission control of the offload request
have devoted themselves to solving the problem of computing
and the calculation of the MEC server frequency. However,
efficiency maximizing and the IoT devices’ energy consumption
none of the above algorithms are suitable for our proposed
minimizing [20]–[24] in the past few years. The authors in [20]
scheme that incorporates harvested energy into edge servers and
mainly study the energy efficiency maximization problem of the
focuses on the joint task offloading and green energy scheduling
MEC systems by combing local computing and data offloading
problem for multi-tiered edge systems. Moreover, different from
into a joint computation algorithm by using a new computing
such works that focus on edge computing systems with flat or
efficiency metric. To solve the problem of computation offload-
two-tiered architectures, our solution is applicable to general
ing and resource allocation in a multi-user wireless powered
multi-tiered edge computing systems with time-varying wireless
MEC system, [21] proposes an iterative algorithm by using a
channel states and the randomness and suddenness of the arrival
Lagrangian dual method and transformed the offloaded delay
of green energy.
into an offloaded data rate constraint. A unified mobile-edge
computing (MEC) and wireless power transfer (WPT) design
framework have been proposed in [22], where the proposed
unified design framework MEC-WPT can pave the way to facil- III. SYSTEM MODEL
itate ubiquitous computing for IoT devices in a self-sustainable We consider a multi-tiered edge computing system, as shown
way. To jointly find optimal caching and offloading scheme, in Fig. 1. We assume that time is slotted, and denote the
authors in [23] exploited resource-based utility function and time slot length and the time slot index set by τ0 and T 
device-number-based relative distance, which helps to effec- {0, 1, . . ., T − 1}, respectively. In Fig. 1, inside the front edge
tively reduce the average service latency and keep a low average tier (FET) are a set of front EH-edge servers (FESs-EH) that
leasing cost. To achieve long-term performance for cooperative offer low latency access to a set of wireless IoT devices. Each
computation offloading, authors in [24] propose a multiagent FES-EH is equipped with an EH module to collect green energy
deep reinforcement learning (DRL) framework in which a scatter and store it in its battery. On the other hand, the backend edge
network is adopted to improve its stability and league learning is tier (BET) comprises a set of backend edge servers (BESs).
introduced for agents to explore the environment collaboratively In our paper, we mainly focus on the energy consumption of
for fast convergence and robustness. The above researches [20]– edge servers (in both FET and BET), the cloud rental fees, and
[24] focus on the energy efficiency of the IoT devices while this communication latency within the edge servers (in both FET
paper focuses on that of MEC server. and BET) and cloud server. First, the energy (including both
With the introduction of green computing [16] and the ad- green energy and grid power energy, i.e., fossil energy) con-
vancements of energy harvesting technology [14], [15], the sumption we consider include two parts: processing energy and
researchers study the IoT devices equipped with energy har- transmission energy. The processing energy is induced by the
vested module [25]–[27]. Authors in [25] model the computation task processing on FET and BET. In practice, data transmissions
offloading process as a Markov decision process (MDP) to cope from FESs-EH to BESs and from BESs to the cloud can incur
with the challenge of the unpredictability nature of the generated energy costs. However, the transmission energy from FESs-EH
data and the harvested energy. The work [26] proposes a deep to BESs is much smaller comparing to the processing energy so
reinforcement learning (DRL) based offloading scheme for an we ignore it. Considering that we assume the BESs communicate
IoT device with EH and a DRL based offloading scheme to with the cloud through wireline connections, so we ignore the
further accelerate the learning speed. A green MEC system with transmission energy from BESs to the cloud. Second, the cloud
EH devices and an efficient computation offloading strategy rental fees we consider only include the task’s processing costs
are investigated in [27], in which the execution latency and in the cloud server. Third, the latency we consider include two
task failure are adopted as the performance metric. However, parts: processing latency and transmission latency. We focus on
these algorithms are not suitable for our proposed EH MEC the transmission latency from IoT devices to FESs-EH, to BESs,
environment where the edge servers in the front edge tier are and to cloud server. And we assume that the task processing in

Authorized licensed use limited to: CINVESTAV. Downloaded on October 10,2022 at 19:37:15 UTC from IEEE Xplore. Restrictions apply.
MA et al.: GREENEDGE: JOINT GREEN ENERGY SCHEDULING AND DYNAMIC TASK OFFLOADING IN MULTI-TIER EDGE COMPUTING SYSTEMS 4325

TABLE I stable. However, considering the mobility of users, they

MAIN PARAMETERS AND THEIR DEFINITIONS
may change across different time slots.
r Mobile Task Model: To characterize the task of a device
i ∈ N , for each time slot t ∈ T , we adopt a parameter tuple
< Ii (t), Yi (t), Oi (t) >, in which Ii (t) is the data size of
task, Yi (t) is the amount of computing resource (i.e., the
number of CPU cycles) required when processing the task,
and Oi (t) indicates the output data size of the task.
r FES-EH Resource Model: The edge computing system
consists of F FESs-EH in FET, let F = {1, 2, . . ., F } be
the sets of FESs-EH. Each FES-EH f ∈ F has the dynamic
and restricted amount of physical resources PfF ront (t)
(in CPU cycles) in a time slot. We use Dif F ronttra
(t)
and Dif F rontrec
(t) to denote the corresponding upload and
download data rates from device i to its associated FES-EH
f respectively.
r BES Resource Model: The edge computing system consists
of B BESs in BET. Let B = {1, 2, . . ., B} be the sets of
BESs. Each BES b ∈ B has the dynamic and restricted
amount of physical resources PbBack (t) (in CPU cycles)
in a time slot. Dib Backtra
(t) and Dib
Backrec
(t) are used to
denote the corresponding upload and download data rates
from device i to its associated BES b respectively.
r Cloud Resource Model: We assumed that the operator will
deploy a cloud server at the network to further facilitate
the execution of tasks. Accordingly, we use DiCloudtra (t)
and DiCloudrec (t) to denote the corresponding upload and
download data rates from device i to cloud server respec-
tively.

B. Task Execution Model

The task execution model is introduced as follows. As de-
scribed above, we assume that at each time slot t ∈ T , each
device will generate a task that needs to be offloaded and
executed.
r FES-EH Offloaded Execution: A device i can offload its
own task to FESs-EH via the cellular link. In this case,
we use eF f
ront
(t) (f ∈ F) to represent the energy con-
sumption in FES-EH f generated by receiving one unit
of input data, calculating the data and transmitting the
calculated result data back to device i. Given the amount of
input data that device i offloads to FES-EH f , the energy
consumption is given by Eif F ront
(t) = Ii (t) ∗ eFf
ront
(t).
each time slot can be completed by the end of the same time And the transmission latency is given by Tif F ront
(t) =
slot, thus we ignore the processing latency. Particularly, the task Ii (t)/DifF ronttra
(t) + Oi (t)/DifF rontrec
(t).
can be divided into multiple sub-tasks which can be finished r BES Offloaded Execution: The task of device i on each
during each time slot. Table I summarizes the key notations used FES-EH can be offloaded to and computed by any of its
throughout this paper. accessible BESs. In this case, we use eBack b (t) (b ∈ B) to
represent the energy consumption in BES b generated by
receiving one unit of input data, calculating the data and
A. Basic Resource Model transmitting the calculated result data back to device i.
In the following, we first illustrate the basic settings. Given the amount of input data that device i offloads to
r IoT Device Resource Model: The edge system consists of BES b, the energy consumption is given by Eib Back
(t) =
N IoT devices, and each IoT device i ∈ N can establish Ii (t) ∗ eBack
b (t). And the transmission latency is given by
a cellular link with its associated base station. During TibBack (t) = Ii (t)/Dib Backtra
(t) + Oi (t)/Dib Backrec
(t).
a time slot, the locations of devices are assumed to be r Cloud Offloaded Execution: The task of device i on each
unchanged and the wireless channels are assumed to be BES can be offloaded to the cloud. In this case, we use

Authorized licensed use limited to: CINVESTAV. Downloaded on October 10,2022 at 19:37:15 UTC from IEEE Xplore. Restrictions apply.
4326 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 71, NO. 4, APRIL 2022

ccloud to represent the cost of receiving one unit of input

data, calculating the data and transmitting the calculated re-
sult data back to device i. Considering the bandwidth usage
cost and the performance cost of the WAN is high, the cost
in cloud server is typically more expensive than EFSs-EH
and BESs. Given the amount of input data that device i
offloads to the cloud server, the resource rental cost in time
slot t of the cloud server can be expressed as: CiCloud (t) =
Ii (t) ∗ ccloud , and the transmission latency is given by
TiCloud (t) = Ii (t)/DiCloudtra (t) + Oi (t)/DiCloudrec (t).
Fig. 2. Battery state dynamics. With unknown value of EfF ront (t) and
C. Task Offloading Constraints Efgreen (t), therefore the system needs to solve the joint task offloading and
energy scheduling problem.
Next, we will formulate the constraints for the task of-
floading problem. Specifically, for each time slot t ∈ T ,
let πi (t) (∀i ∈ N ) be a binary indicator that is 1 if de-
t. Besides, rf (t) is denoted as the amount of harvested green
vice i has a task to be offloaded, and 0 otherwise (i.e.,
energy at time slot t. Note that rf (t) is a time-varying and
the task queue of device i is empty). In addition, we de-
random variable depending on the ambient environment. And
note a binary vector xi (t) = [xi1 (t), . . ., xii (t), . . ., xiF (t),
the green energy rf (t) harvested by the EH module of FES-EH
xiF +1 (t), xiF +2 (t), . . ., xiF +B (t), xiF +B+1 (t)] for each de-
f is subjected to the availability constraint
vice i to denote the task offloading decision making at time
slot t. Specifically, when j ∈ {1, 2, . . ., F } and xij (t) = 1, it
rf (t) ∈ [0, rfmax ], (5)
represents the task is executed on the FET; when j ∈ {F +
1, F + 2, . . ., F + B} and xij (t) = 1, it represents the task is where rfmax denotes the upper-bound of the possible amount of
executed on the BET; otherwise, j = F + B + 1 and xij (t) = green energy that can be harvested by FES-EH f at each time
1, it represents the task is executed on the cloud. Note that at slot t.
each time slot, the task of device i is assigned to one and only For realistic considerations, the collected green energy is not
one execution node. We then have the task offloading constraints the sole source of energy for each FES-EH’s battery which also
as follow: can use the grid electricity (fossil energy) when the harvested
xij (t) ∈ {0, 1}, ∀i ∈ N , ∀j ∈ {1, 2, . . ., F + B + 1}. (1) energy cannot fulfill the task executions’ needed energy. And we
use the variable Efgreen (t) to denote the harvested green energy

N
used to execute the tasks in FES-EH f which is also an energy
xij (t)Yi (t) ≤ PjF ront (t), ∀j ∈ {1, 2, . . ., F }. (2) scheduling decision. Therefore, we have
i=1


N Efgreen (t) ≥ 0, ∀f ∈ {1, 2, . . ., F }. (6)
xij (t)Yi (t) ≤ PjBack (t), ∀j ∈ {F+1, F + 2, . . ., F+B}.
i=1 Besides, in FES-EH f , the needed green energy for execution
(3) does not exceed the total energy consumption (i.e., EfF ront (t),

 where we have EfF ront (t) = N i=1 xif (t)Eif (t)):
F ront
xij (t) = πi (t), ∀i ∈ N . (4)
j∈{1,2,...,F +B+1}
Efgreen (t) ≤ EfF ront (t), ∀f ∈ {1, 2, . . ., F }. (7)
The constraint (2) represents that the physical resources in the
FES-EH j (j ∈ {1, 2, . . ., F }) is restricted by PjF ront (t). The Considering the energy level of the EH module at time slot t,
constraint (3) represents that the physical resources in the BES we further have:
j (j ∈ {F + 1, F + 2, . . ., F + B}) is restricted by PjBack (t).
The constraint (4) represents that during a task offloading round Efgreen (t) ≤ Qf (t), ∀f ∈ {1, 2, . . ., F }. (8)
a device will offload at most one task.
Thus, the dynamics of the renewable energy level of each EH
module in FES-EH f follows:
D. FESs-EH Battery Model and Green Energy Constraints
For the EH battery model, we assume that each FESs-EH Qf (t + 1) = Qf (t) − Efgreen (t) + rf (t), ∀f ∈ {1, 2, . . ., F },
in FET is equipped with an EH module (which can be a solar (9)
panel, a wind turbine, a kinetic tile, or a combination of them) as where rf (t) can be viewed as the “arrivals” of the energy queue
shown in Fig. 1. Before powering FES-EH, the harvested energy Qf (t), and Efgreen (t) can be viewed as the “service rate” of
is buffered at battery to ensure sustained energy supply, since such a queue. Note that, if the harvested green energy cannot
the ambient renewable energy sources appear to be random and be fully utilized, the virtual energy queue may accumulate and
bursty in nature. even grow indefinitely over time. To avoid this, we must keep the
For each FES-EH f ∈ {1, 2, . . ., F }, we use Qf (t) to denote queue stable, i.e., limT →∞ E{Qf (T )}/T = 0 (Qf (T ) < ∞).
the energy level of the EH module at the beginning of time slot The FESs-EH battery state dynamics is illustrated in Fig. 2.

Authorized licensed use limited to: CINVESTAV. Downloaded on October 10,2022 at 19:37:15 UTC from IEEE Xplore. Restrictions apply.
MA et al.: GREENEDGE: JOINT GREEN ENERGY SCHEDULING AND DYNAMIC TASK OFFLOADING IN MULTI-TIER EDGE COMPUTING SYSTEMS 4327

E. System Cost Minimization Problem

In this paper, we assume that for each FES-EH f ∈ F, the cost
coefficient of using green energy is α and the cost coefficient
of using grid power energy (fossil energy) is β. Note that, α
is smaller than β. Besides, we use λei , λti ∈ {0, 1} to denote
the weighting parameters of the costs of energy consumption
and the cloud rental fees, and the execution latency for devices’
decisions making, respectively.
Therefore, for each time slot t ∈ T , the FESs-EH’ costs
caused by executed tasks of all IoT devices including
transmission latency costs and energy costs in FES-EH f can
 F
be denoted as C F ront (t) = N f =1 xif (t)λi Tif
t F ront
(t) +
F green
i=1
green
f =1 λi pf (t)[αEf (t) + β(Ef (t) − Ef (t))];
e F ront F ront
Fig. 3. The summary of the proposed optimization process.
the BES’ costs caused by executed all IoT devices’ tasks includ-
ing transmission latency costs and energy costs in BES b can be
 F +B feasible yet near-optimal solution. The proposed optimization
denoted as C Back (t) = N i=1 b=F +1 xib (t)[λi Tib
t Back
(t) + algorithm workflow is summarized in Fig. 3.
λ i pb
e Back
(t)Eib (t)]; the cloud costs caused by executed
Back

all IoT devices’ tasks include transmission latency costs and A. Problem Transformation Via Lyapunov Optimization
cloud resource rental costs, and can be denoted as C Cloud (t) =
N To proceed, we define the following key concepts in Lyapunov
i=1,c=F +B+1 xic (t)[λi Ti (t) + λei CiCloud (t)].
t Cloud
optimization tailored to the formulated problem P1.
Note that with the capability of energy harvesting, our objec-
1) Queue Stability: In order to stabilize the virtual queue
tive is to make joint task offloading and green energy scheduling
(i.e., Qf (T ) < ∞), the quadratic Lyapunov drift function is
decisions to minimize the system total costs, which can be
defined as:
formulated as:
1
F
L(Θ(t))  Qf (t)2 , (10)
1 2
min lim f =1
T →∞ T


T −1 here Θ(t) < Qf (t) > |f ∈F , the vector of battery queue back-
E[C F ront (t) + C Back (t) + C Cloud (t)] (P1) logs on time slot t. We further define the conditional Lyapunov
t=0 drift, Δ(Θ(t)) as the expected change in the Lyapunov function
from one time slot to the next:
s.t. (1) − (9).
Δ(Θ(t))  E[L(Θ(t + 1)) − L(Θ(t))|Θ(t)]. (11)
By the Lyapunov drift theorem [19], if a policy minimizes
IV. ALGORITHM DESIGN
the Lyapunov drift in each time slot, then all queue backlogs
To solve the problem P1, a direct way is to gradually op- are consistently pushed towards a small size, which potentially
timize each individual time slot to achieve the optimization maintains the stabilities of all queues.
goal of minimizing the total system costs. However, consider- 2) Joint Queue Stability and System Costs Minimization:
ing the dynamic coupling between the EH module buffer of We next consider the joint queue stability and the objective
each FES-EH in different time slots as well as the dynamic optimization (i.e., minimizing the time-average costs for task
system states (such as the arrival of available green energy execution of all IoT devices). Following the Lyapunov optimiza-
and the connections between devices and each FES-EH), it is tion framework [19], this is equivalent to minimize an upper
not realistic to solve problem P1 offline. Therefore, we devote bound of the following drift-plus-penalty function:
to propose an online approach that can efficiently make task
offloading and green energy scheduling decisions on-the-fly Δ(Θ(t)) + V E[C F ront (t) + C Back (t) + C Cloud (t)|Θ(t)],
without foreseeing the future. We first apply the Lyapunov op- where V is a non-negative control parameter that achieves a
timization technique [19] to decouple problem P1 into one-slot desirable tradeoff between the optimality and queue backlogs.
deterministic MILP subproblem P2. Unfortunately, the resulted The following Lemma provides an upper bound of the drift-plus-
decomposed subproblem P2 in each time slot is NP-hard. As a penalty function.
result, the standard Lyapunov optimization approach relying on Lemma 1: For all possible values of Δ(Θ(t)) and under any
getting optimal solution from the decomposed problems fails task offloading policy x(t) and green energy scheduling decision
to work. Hence, we will develop an approximated Lyapunov E(t) for all time slot t, the drift-plus-penalty function satisfies:
optimization scheme to overcome this key issue. Specifically,
we further relax the integer constraint to get problem P3, and Δ(Θ(t))
by using the simplex method [30], we get the optimal fractional
+ V E[C F ront (t) + C Back (t) + C Cloud (t)|Θ(t)]
solution and finally propose an efficient resource-constrained
randomized dependent rounding RCRDR algorithm to get a ≤ B + y(t)+

Authorized licensed use limited to: CINVESTAV. Downloaded on October 10,2022 at 19:37:15 UTC from IEEE Xplore. Restrictions apply.
4328 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 71, NO. 4, APRIL 2022


N 
F  +B
N F
Algorithm 1: On-line JTOES algorithm for solving P1.
xif (t)wif (t) + xib (t)wib (t)
i=1 f =1 i=1 b=F +1 Input: Qf (0), V , rfmax , πi (t), PfF ront (t), PbBack (t);
Output: xij , Efgreen (t);

N 
F 
N
+ xic (t)wic (t) + xif (t)wif (t) 1: for each time slot t ∈ {0, 1, 2, . . ., T − 1} do
i=1,c=F +B+1 f =1 i=1
2: Obtain xif (t), xib (t), xic and Efgreen (t) by solving
P2.

F
3: Qf (t + 1) = Qf (t) − Efgreen (t) + rf (t).
+ Efgreen (t)wf (t), 4: end for
f =1

with wif (t) = V λti Tif

F ront
(t),
C. Approximate Algorithm for P2
wib (t) = V [λti TibBack (t) + λei pBack
b (t)Eib
Back
(t)],
Due to the constraint (1)-(4) and (6), P2 is a mixed integer
wic (t) = V [λti TiCloud (t) + λei CiCloud (t)], linear programming (MILP) problem which can be proved to be
NP − hard in Theorem 1.
wif (t) = V βλei pF
f
ront
(t)Eif
F ront
(t), Theorem 1: The optimization problem P2 is NP-hard.
wf (t) = V λei pF ront
(t)[α − β] − Qf (t). (12) Proof: We first need to introduce the uncapacitated facility
f
location (UFL) problem [32]:

Here B = 1/2 F f =1 ((Ef ) + (rfmax )2 ) is a constant
max 2  
value across all time slots, Efmax
= maxt Efgreen (t) and y(t) = min cij xij + fi yi
F j∈F i∈N i∈N
f =1 Qf (t)rf (t) involves no offloading indicators. Therefore, 
we do not consider them in the joint task offloading and green s.t. xij = 1 for each j ∈ F,
energy scheduling algorithm design. Due to the page limit, their i∈N
specific expressions and the detailed proof is given in our online
xij ≤ yi ,
technical report [31].
xij ∈ [0, 1], yi ∈ {0, 1} for each i ∈ N , j ∈ F.
B. Online Algorithm for P1 (14)

According to the transformed problem and Lemma 1, the As we know, in a typical uncapacitated facility-location prob-
key technique of approximated Lyapunov optimization is con- lem, we are given a set of facilities i ∈ N with facility-opening
sistently minimizing the drift-plus-penalty upper bound (i.e., costs fi and a set of clients with demands j ∈ F, and we want
the last five terms on the right-hand side) of (12), subject to to open facilities (e.g., the green energy scheduling decisions
constraints (1)-(9), which can be approved to achieve a good Ejgreen (t) in our problem) and assign clients to open facilities
performance for the problem in P1. Now, we rearrange the (e.g., the task offloading decisions xij (t) in our problem) so as
formulations in (12) for an unfold form P2, and then propose the to minimize the sum
 of facility-opening costs (e.g., the green en-
online Joint Task Offloading and Energy Scheduling (JTOES) ergy used costs f ∈F Efgreen (t)wf (t)) and client-assignment
algorithm for solving P1.  
costs (e.g., the task offloading costs i∈N f ∈F xif (t)wif (t)
in our problem).
P2 :
The UFL problem is known to be NP-complete (see [32]). To

N 
F  +B
N F prove the NP − hardness of our problem P2, we first prove
min xif (t)wif (t) + xib (t)wib (t) that the UFL problem can be reduced to the next problem P̃:
x(t),E(t)
i=1 f =1 i=1 b=F +1
 
P̃ : min xif (t)wif (t) + Efgreen (t)wf (t)

N 
F 
N
x(t),E(t)
+ xic (t)wic (t) + xif (t)wif (t) i∈N f ∈F f ∈F

i=1,c=F +B+1 f =1 i=1 s.t. constraints (1) − (9), (15)


F
+ Efgreen (t)wf (t) where wif (t) = V λti Tif
F ront
(t) and wf (t) =
f =1 V λ i pf
e F ront
(t)[α − β] − Qf (t).
s.t. constraints (1) − (9). (13) Therefore, the problem P̃ is NP − hard . We can see when
the set of F , B, and cloud degrades to F , some constraints are
The detailed performance analysis of Algorithm 1 would be relaxed to unlimited in problem P̃. So it is obviously to see that
shown in section V. Unfortunately, even if we get the one slot the problem P̃ is a special case of our problem P2. Thus our
minimization problem P2, it still is NP-hard and intractable problem P2 is much harder than the traditional UFL problem
to find the optimal solutions. We will prove the NP-hardness because of the constraints (2), (3) and (9), clearly our problem
and then apply a relaxation and randomized dependent rounding P2 is NP − hard .
technique to obtain a near-optimal solutions for P2. Therefore we prove the NP − hardness of problem P2. 

Authorized licensed use limited to: CINVESTAV. Downloaded on October 10,2022 at 19:37:15 UTC from IEEE Xplore. Restrictions apply.
MA et al.: GREENEDGE: JOINT GREEN ENERGY SCHEDULING AND DYNAMIC TASK OFFLOADING IN MULTI-TIER EDGE COMPUTING SYSTEMS 4329

Considering the NP-hardness of P2, we can easily know Adopting the above independent rounding solution may al-
that it is computationally infeasible to find the optimal solu- ways generate an infeasible policy. That is, with a certain
tion when the system scale is getting large. Fortunately, for probability, all variables are rounded up which may violate
a MILP problem, there is plenty of heuristic algorithms that the capacity constraint of FESs-EH in FET and BESs in BET,
can be used to get sub-optimal solutions. And the core ideas of or all variables are rounded down and all the tasks needed to
these solutions are first to get the optimal fractional solutions be sent to the cloud to execute which would incur high cloud
by relaxing the integral constraint of the MILP problem, then rental fees. To address the above deficiency, we resort to ex-
propose heuristic algorithms to round the fractional solutions ploiting the inherent dependence of variables ẋij (t) to improve
into integral ones based on the special structure of the problems. the optimality and further to propose a resource-constrained
Unfortunately, most of these algorithms ([32], [33] and etc) can randomized dependent rounding (RCRDR) solution. Obviously,
only get parameterized or constant rounding competitive ratios the term “dependent” underscores the fact various random
which also are hard to generalize to common problems. decisions we make are highly dependent. Our key idea is
In [34], the pipage rounding technique is proposed to round that a rounded-down variable will be compensated by another
the optimal policies, which incurs a constant competitive ra- rounded-up variable, ensuring that the physical resource con-
tio and an additional constant rounding gap with the optimal straint in both FET and BET can be satisfied even after the
policies. And inspired by this, we propose an extended efficient rounding.
dependent rounding algorithm tailored to our problem by taking We give a simple expression of RCRDR scheme for P2. For
into account the edge resource constraints. Solving the integral each task i, we introduce two sets: the rounded set S̃i+ (t) =
variables in a MILP problem makes the problem much harder, {k|ẋik (t) ∈ {0, 1}} and the floating set S̃i− (t) = {k|ẋik (t) ∈
and in reality, once the integral decisions are made, the remaining [0, 1]}. Intuitively, S̃i+ (t) denotes the set of tasks with binary
task offloading and green energy used problem may be optimally x̃ij (t) while S̃i− (t) denotes the set of tasks with fractional ẋij (t)
solved by classical linear programming methods. Therefore, and thus should be rounded. For each fractional ẋik (t), we intro-
getting the integral variables matters a lot. duce a probability coefficient pik and a weight coefficient wik
We first relax the integral constraint (1), obtaining the frac- associated with it. We initialize pik = ẋij (t), and wik = Ci (t).
tional optimization problem P3 as follows: Our proposed RCRDR scheme runs a series of rounding
P3 : iterations by rounding the elements in S̃i− (t) with fractional
xij (t). Specially, at each iteration, we randomly select two

N   +B
N F
F
elements k1 and k2 from the floating S̃i− (t), and let the prob-
min xif (t)wif (t) + xib (t)wib (t) ability of one of these two elements round to 0 or 1, then
x(t),E(t)
i=1 f =1 i=1 b=F +1
used the parameter γ1 and γ2 to denote, which can also adjust

N 
F 
N the value of pik1 and pik2 . Note that after adjustment, at least
+ xic (t)wic (t) + xif (t)wif (t) one of pik1 and pik2 ∈ {0, 1}, and then set x̃ik1 (t) = pik1 or
i=1,c=F +B+1 f =1 i=1 x̃ik2 (t) = pik2 . By doing so, at each iteration, the number of
elements in S̃i− (t) would decrease at least by 1. Finally, when

F
+ Efgreen (t)wf (t) S̃i− (t) has only one element in the last iteration, we set the
f =1  pik = 1 if the total physical resource
last element constraint
in FET k∈F x̃ik (t)Ci (t) is smaller than PkF ront (t) or if the
s.t. constraints (2) − (9), total physical resource constraint in BET k∈B x̃ik (t)Ci (t) is
xij (t) ∈ [0, 1], ∀i ∈ N , ∀j ∈ {1, 2, . . ., F + B + 1}. (16) smaller than PkBack (t) when k ∈ {F + 1, F + 2, . . ., F + B},
otherwise set pik = 0. In Algorithm 2 we summarize the above
Since the above problem P3 is a standard linear programming detailed RCRDR algorithm.
problem with linear constraints, it can be optimally solved in We should note that our proposed RCRDR algorithm is cost-
polynomial time, by taking existing linear programming opti- efficient and computation-efficient, in terms of that it would
mization technique such as the simplex method [30]. Therefore, not offload all tasks to the cloud to increase the cloud rental
we invoke the simplex method to solve the linear programming fees, and it only has the complexity of O(|N ||H|) to make the
P3 (relaxed-P2) and obtain the optimal fractional policies rounding process because there are at most |N ||H| fractional
(ẋ(t), Ė green (t)). variables, where H = F + B + 1. Therefore, the complexity of
Solving the relaxation problem P3 for each time slot t, we Algorithm 1 is O(|N ||H||T |). We will illustrate three important
can obtain the optimal fractional solution which we denote to properties of the RCRDR algorithm and their proofs in the
be ẋ. Considering the integrality constraint (9i) of the original following.
problem P2, it is of great importance to choose a properly r Continuous reduction property. After each iteration, there
rounding scheme to round the fractional solutions ẋ to get the will be at least one of the two selected variables ẋik1 (t) and
binary solution x̃. For this purpose, the independent randomized ẋik2 (t) rounded to be an integer.
rounding method is a straightforward solution, the core idea of r Weighted conservation property. That is, after the main
which is to round each fractional ẋij (t) to binary x̃ij (t) by loop of each iteration, the sum of the products of the
treating xij (t) as the probability of rounding ẋij (t) up, i.e., selected two elements’ probabilities and their weights re-
Pr[x̃ij (t) = 1] = ẋij (t). mains unchanged. Moreover, the RCRDR algorithm can

Authorized licensed use limited to: CINVESTAV. Downloaded on October 10,2022 at 19:37:15 UTC from IEEE Xplore. Restrictions apply.
4330 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 71, NO. 4, APRIL 2022

|N | maxi∈N Ci (t), and maxi∈N Ci (t) is the maximum

Algorithm 2: The Proposed RCRDR for Solving P2.
computing resources required for the task i ∈ N . This
Input: Qf (0), V , rfmax , πi (t), PfF ront (t), PbBack (t); property plays the key role in ensuring the capacity con-
Output: the rounded policies ẋ(t) and green energy straint Eq. (2) and Eq. (3) would not be violated when run-
scheduling decisions Ė green (t); ning RCRDR to round the fractional solution. Also, we can
1: Invoke the simplex method to solve the LP problem see that our resource-constrained randomized dependent
P3 and obtain the optimal fractional policies rounding algorithm fully considers resource constraints in
(ẋ(t), Ė green (t)); edge servers.
2: for each task i ∈ N do r Marginal distribution property. In the main while loop,
3: let S̃i+ (t) = {k|ẋik (t) ∈ {0, 1}}, the probability of each element ẋik (t) ∈ S̃i− (t) satisfies
S̃i− (t) = {k|ẋik (t) ∈ [0, 1]}; Pr(x̃ik (t)=1)=ẋik (t), Pr(x̃ik (t)=0)=1-ẋik (t). This prop-
4: for each offloadee k ∈ S̃i+ (t) do erty indicates that the task would not be offloaded to several
5: set x̃ik (t)=ẋik (t); servers and will be assigned only once when rounding the
6: end for fractional policies.
7: for each offloadee k ∈ S̃i− (t) do The detailed proof is given in our online technical report [31].
8: set pik = ẋik (t) and wik = Ci (t); We should note that the inequality (17) values a lot since it es-
9: end for tablishes the connection between rounding computing resources
10: while S̃i− (t) > 1 do sizes and optimal fractional computing resources sizes, which
11: Randomly select two elements k1 , k2 from S̃i− (t); also will be used as a bridge to analyze the gap of our RCRDR
wik2 algorithm in the next section.
12: Define γ1 = min{1 − pik1 , pik },
wik1 2
wik2 V. PERFORMANCE ANALYSIS
γ2 = min{pik1 , (1 − pik2 )};
wik1 In this section, we will illustrate the performance of our
γ2 proposed algorithms. We first give the rounding gap of P2
13: With the probability set, pik1 = pik1 + γ1 , and then presents the performance analysis of P1 using the
γ1 + γ2
wik1 Lyapunov optimization technique [19].
pik2 = pik2 − γ1 ;
wik2
γ1 A. Rounding Gap of RCRDR
14: With the probability set, pik1 = pik1 − γ2 ,
γ1 + γ2 We first give our conclusion of the rounding gap of RCRDR
wik1 in the following theorem:
pik2 = pik2 + γ2 ;
wik2 Theorem 2: After applying the randomized dependent round-
15: If pik1 ∈ {0, 1}, then set x̃ik1 (t) = pik1 ; ing RCRDR algorithm to get the final integral solutions x̃(t)
16: S̃i+ (t) = S̃i+ (t) ∪ k1 , S̃i− (t) = S̃i− (t) \ k1 ; ˜ (t), the maximum cost of P2 is
and to further get the E green
17: If pik2 ∈ {0, 1}, then set x̃ik2 (t) = pik2 ; proportional to the optimal value and has an additional constant
18: S̃i+ (t) = S̃i+ (t) ∪ k2 , S̃i− (t) = S̃i− (t) \ k2 ; deviation, i.e.,
19: end while
C̃(xij (t)) + C̃(xif (t)) + C̃(E green (t))
20: if |S̃i− (t)|=1 then
21: Set x̃ik1 (t) = ẋik1 (t) the only one element k in ≤ φCtopt + Π. (18)
S̃i− (t);
22: S̃i+ (t) = S̃i+ (t) ∪ k, S̃i− (t) = S̃i− (t) \ k; Here C̃(xij (t)) + C̃(xif (t)) + C̃(E green (t)) is the cost of
23: end if problem P2 after applying the rounding decisions x̃(t) and
24: end for Ẽ green (t), and Ctopt is the optimal cost of original problem P2
25: Apply again the simplex method to problem P2 when in time slot t. φ and Π are constants which will be specified in
knowing x̃(t) and obtain Ẽ green (t). our online technical report [31] in which the detailed proof of
Theorem 2 will be given.
The basic idea is to leverage the relationship between the
finally achieve that: optimal fractional solution x̃ik (t) and the rounded solution
Ẽjgreen (t), which has been characterized by (17). Then, we
 
x̃ik (t)Ci (t) further take this connection as a bridge to bound each cost terms
i∈N k∈F+B
of problem P2.
 
≤ ẋik (t)Ci (t) + |N | max Ci (t), (17) B. Performance Analysis of JTOES
i∈N
i∈N k∈F+B
This subsection presents the performance analysis of JTOES
which means that the total physical resources constraints using the Lyapunov optimization technique [19] while using the
in both FET and BET after running RCRDR are bounded RCRDR to solve the subproblem P2. The following theorem
by the total physical resources of fractional policies plus gives our conclusions.

Authorized licensed use limited to: CINVESTAV. Downloaded on October 10,2022 at 19:37:15 UTC from IEEE Xplore. Restrictions apply.
MA et al.: GREENEDGE: JOINT GREEN ENERGY SCHEDULING AND DYNAMIC TASK OFFLOADING IN MULTI-TIER EDGE COMPUTING SYSTEMS 4331

Theorem 3: By applying JTOES to P1 and RCRDR to the TABLE II

SIMULATION SETUP AND SYSTEM PARAMETERS
subproblem P2, the time-average system cost satisfies:
T −1
1 B
E{C(x(t), E green (t))} ≤ + φCtopt + Π, (19)
T t=0 V
and the time-average harvested green energy satisfies:
T −1
1 
E{Qf (t)}
T t=0
f ∈F

B
≤
η
φ[C max (x(t), E green (t)) − C opt (x(t), E green (t))]
+V .
η
(20)
1 T −1
Here C opt (x(t), E green (t)) = limT →∞ t=0 E{C(x
opt
T
(t), (E ) (t))} is the optimal time-average cost,
green opt

C max (x(t), E green (t)) is the largest cost, and η ≥ 0 is a

constant which represents the long-term queue length surplus
achieved by some stationary control policies. And as we have

specified above that B = 1/2 F f =1 ((Ef ) + (rfmax )2 ) and
max 2
green
Efmax = maxt Ef (t). Due to the limitation of pages, the
constants φ, Π and η will be specified in our online technical TABLE III
report [31]. AVERAGE RUNNING TIME AMONG DIFFERENT ROUNDING ALGORITHMS
Besides, the proof of Theorem 3 will be given in our online
technical report [31]. By adjusting the parameter V , we can
balance between the performance gap from the offline optimum
and the expected energy queue length. Performance evaluation
in Section VI corroborates that our proposed approximated
algorithm is highly efficient, with at most 4% higher system
cost than the optimal solution.

VI. PERFORMANCE EVALUATION Dynamic execution with greedy energy scheduling (Dynamic
In this section, we evaluate the performance of our proposed execution (GD)), Dynamic execution under no energy harvesting
JTOES algorithm and RCRDR algorithm via extensive trace- (Dynamic execution (NoEH)), which minimize the execution
driven simulations. The parameters about tasks’ and wireless cost at the current time slot. They work as follows:
connections set in our simulation are based on the commonly 1) Cloud execution: The tasks are offloaded to cloud server
adopted wireless environment settings that have been used to execute.
in [35], [36]. The harvested green energy amount is assumed 2) Dynamic execution (GD): In this situation, the green en-
to follow a uniformly distribution, and the detailed value range ergy is scheduled greedily, which means that the Efgreen (t)
is similar to the setting in [37]. is not a decision variable, the tasks are executed by
We simulate a hierarchal edge computing system with 5 FESs- offloading to the execution modes with the minimum
EH and 2 BESs. For each FES-EH f , its accessible BES set Bf costs. While in this situation, the FES-EH’ costs can
is chosen uniformly randomly from the power set of the BESs be denoted in two circumstances, case 1: the green
set with size |Bf | = 3. In this paper, similar to many existing energy in FES-EH f is capable of providing the en-
studies such as [38] [39], we assume that the backhaul capacities ergy to execute task i, then C̃iF ront (t) = λti Tif F ront
(t) +
between the front-edge tier and the backend edge-tier as well α ∗ λ i pf
e F ront
(t)EifF ront
(t), case 2: otherwise we have:
as the capacities between the backend edge tier and the cloud C̃iF ront (t) = λti Tif
F ront
(t) + β ∗ λei pF
f
ront
(t)Eif
F ront
(t).
tier are fixed parameters. We set the time slot length τ0 = 1 3) Dynamic execution (NoEH): The FESs-EH are not
second. During each time slot, the tasks arrive at the system equipped with the energy harvesting ability and the
in the unit of packets, each with a size of [500,2000] KB. All operators select the minimum costs of offloading mode
results are averaged over 50000 time slots. The values of the for each task. While in this situation, the FESs’ costs, the
other parameters are listed in Table II . BESs’ costs, and the cloud rental fees caused by executed
For performance comparison, we consider three benchmark IoT device i’ tasks at time slot t can be denoted as
policies, namely, Cloud server execution (Cloud execution), C̃iF ront (t) = λti Tif
F ront
(t) + λei pF
f
ront
(t)Eif
F ront
(t),

Authorized licensed use limited to: CINVESTAV. Downloaded on October 10,2022 at 19:37:15 UTC from IEEE Xplore. Restrictions apply.
4332 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 71, NO. 4, APRIL 2022

Fig. 4. Reduction ratio of costs vs control parameter V under different Fig. 5. The dynamic of average energy queue vs time slot t under different
methods. methods.

CiBack (t) = λti TibBack (t) + λei pBack

b (t)Eib
Back
(t),
CiCloud
(t) = λi Ti
t Cloud
(t) + λi Ci
e Cloud
(t). Operators
select the minimum value of C̃iF ront (t), CiBack (t),
CiCloud (t) and offload the tasks to the appropriate mode
to execute.
In our experiment, Cloud execution is used as the benchmark
of system cost saving rate.

A. Cost Trade-Off
We conduct the simulations under different values of V to
Fig. 6. The dynamic of energy queue vs control parameter V under JTOES.
evaluate the cost-queue trade-off performance of JTOES.
As shown in Fig. 4, when V increases from 0.1 to 10, the
time-average system cost reduction ratio increases until it gets
to a minimum which means that the system cost is getting
smaller and finally approaches to a minimum under this system
circumstances. Specifically, when V changes from 2 to 7, the
system cost decreased dramatically. This phenomenon verifies
the (19) in Theorem 3, Section V, where the time-average cost
is proportional to the values of 1/V . Therefore, the larger V
we take, the smaller system costs we get. Fig. 4 also shows
that our JTOES achieves excellent performance in system cost
comparing to the dynamic execution (GD) algorithm and dy-
namic execution (NoEH) algorithm. We can see that the dynamic
execution (NoEH) algorithm achieves the smallest cost reduc- Fig. 7. Average system costs in JTOES vs different β under different methods.
tion ratio when comparing to JTOES and dynamic execution
(GD), this proves that our scheduling green energy method can
rest simulations in which the value of V is not the variable. From
effectively reduce the system costs.
Fig. 5 and Fig. 6 we can see no matter what V is, the change curve
To further prove the effective performance in energy queue
of the energy queue will gradually become stable over time,
scheduling of our JTOES, we depict the instant and time-average
which shows that our JTOES can provide sustainable energy for
energy queue backlog in Fig. 5 in which that our JTOES out-
system task execution when the maximum battery capacity is
performs the dynamic execution (GD) algorithm in the green
fixed, for which the dynamic execution (GD) algorithm cannot
energy Efgreen (t) scheduling.
fulfill.
In Fig. 6, we can see that when V increases, the time-average
energy queue backlog grows and almost proportionally to the
B. Impact of β, Tasks Number N , Edge Servers Number F on
values of V . This phenomenon verifies the (20) in Theorem 3,
Section V, where the queue length is almost proportional to the System Performance
values of V . And we can see as the value of V increases, the Fig. 7 shows that the influence of the cost-efficient of using
energy queue backlog gradually changes little, this is because in grid β in FET on system performance. In our simulation, the
our simulation the tasks arriving ration is not a variable value thus cost-efficient of using green energy α is set to be 0.5, and
a bigger value of V may not influence too much on the changes we set the value of β from 0.5 to 20. A larger β means a
of our system costs. To balance the trade-off between the task higher price to buy the grid power in FET and it will influence
execution cost and the queue backlog, we choose V = 7 for the the system costs. Specifically, setting β = α = 0.5 achieves

Authorized licensed use limited to: CINVESTAV. Downloaded on October 10,2022 at 19:37:15 UTC from IEEE Xplore. Restrictions apply.
MA et al.: GREENEDGE: JOINT GREEN ENERGY SCHEDULING AND DYNAMIC TASK OFFLOADING IN MULTI-TIER EDGE COMPUTING SYSTEMS 4333

Fig. 8. Average System costs vs different tasks number N under different Fig. 10. Time-average value of problem P2 under different tasks number N .
methods.

shows the fluctuation of the computing capacity in FET influence

the system cost significantly.

C. Approximation Performance of RCRDR in JTOES

In this subsection, we conduct substantial simulations to
evaluate the approximation performance of RCRDR for JTOES.
We compare RCRDR with the following benchmarks:
r Optimal Solution (Opt): Directly solves P2 by using a
commercial-grade MILP solver (dual-simplex solver) to
compute the optimal solutions;
r Independent Randomized Rounding (IRR): Ignoring the
Fig. 9. Time-average system costs under different values of F . computing capacity in FET and BET, we rounded the
fractional optimal solutions with possibility Pr[x̃ij (t) =
the smallest system cost, and the time-average cost in four 1] = ẋij (t);
methods all increasing as the value of β increases. The JTOES r Top-i: Places the Top-i tasks according to the fractional
algorithm we proposed always achieves better performance than optimal solutions (sort the ẋij (t) in a descending order, and
the dynamic execution (GD) algorithm and dynamic execution then round the biggest fractional element to 1 one by one)
(NoEH) algorithm. As we have verified aforementioned, the while satisfying the computing capacity of the FESs-EH
time-average cost is proportional to the values of 1/V , in Fig. 7, in FET and BESs in BET.
JT OES(V = 10.0) achieves the bigger time-average system As shown in Fig. 10 , we present the time-average objective
cost than JT OES(V = 5.0). And we choose β = 5 for the rest value of the problem P2 while using different rounding policies.
simulations. We can see that the RCRDR algorithm always achieves the
The effect of the number of devices N on the system perfor- best approximation performance to the optimal solutions Opt
mance is shown in Fig. 8. It can be seen that with the increase of under different values of N . The rounding policy IRR always
N , the performance of our proposed algorithm JTOES is better achieves a good performance in minimizing the problem P2,
than the other two algorithms (compared with dynamic execu- only slightly larger than our rounding policy RCRDR, how-
tion (GD) algorithm and dynamic execution (NoEH) algorithm, ever, it cannot guarantee the computing capacity constraint
the system task execution cost is decreased by more than 6% due to its probabilistic nature (when all the xij (t) = 1, j ∈
and 15% respectively). All these results show that our proposed {1, 2, . . ., F, F + 1, . . ., F + B}, the FESs-EH in FET or BESs
algorithm JTOES in this paper has a good scalability. in BET can not have enough computing resources to compute all
Fig. 9 shows that the time-average system cost under different the tasks). In contract, our rounding policy RCRDR can finely
number of FESs-EH in FET (the different computing capacity inherit the probabilistic of IRR and use the dependence of xij
in FET). When the number of FESs-EH in FET continues to to satisfy the computing constraint in FET and BET.
increase, our JTOES and the benchmarks all have decreasing Besides achieving the near-optimal performance, the RCRDR
system costs. This is because the more computing resources the algorithm also does well in the average running time. In Fig. 11,
FET has, they can serve more tasks itself and there is no need to we show the average running time of the benchmarks. We can see
offload the tasks to the BET or the public cloud, avoiding a large that the average running time of the optimal solver Opt increases
offloading latency. Notice that our JTOES always achieves the dramatically when the tasks number grows larger, that is because
smallest system cost among the different computing capacity in of the NP-hardness of P2. However, the running time of our
FESs-EH (at least 10% lower than dynamic execution (GD)). RCRDR only increases proportionally with the tasks number.
Specifically, when the value of F increases from 5 to 8, the Besides, our RCRDR always achieves the shortest running time
system cost of JTOES decreases by 11.5% dramatically, which when comparing with IRR algorithm and Top-i algorithm.This

Authorized licensed use limited to: CINVESTAV. Downloaded on October 10,2022 at 19:37:15 UTC from IEEE Xplore. Restrictions apply.
4334 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 71, NO. 4, APRIL 2022

system and consider minimizing the system cost that covers

latency, energy consumption, and cloud rental fees. We further
develop a novel approximated dynamic optimization framework
to overcome the uncertainty of system information, only using
the current system information to minimize the system costs
meanwhile ensuring the stability of the harvested energy buffer
queues. The rigorous performance analysis reveals the rounding
gap and the extensive performance evaluation corroborate the
excellent performance on our framework. For the future work,
we are going to explore the possibility of integrating the green
energy prediction techniques into the holistic algorithm design.

Fig. 11. Average running time vs different tasks number N under different
methods. REFERENCES
[1] Z. Xiong, Y. Zhang, D. Niyato, R. Deng, P. Wang, and L.-C. Wang,
“Deep reinforcement learning for mobile 5G and beyond: Fundamentals,
applications, and challenges,” IEEE Veh. Technol. Mag., vol. 14, no. 2,
pp. 44–52, Jun. 2019.
[2] R. Azuma, Y. Baillot, R. Behringer, S. Feiner, S. Julier, and B. MacIntyre,
“Recent advances in augmented reality,” IEEE Comput. Graph. Appl.,
vol. 21, no. 6, pp. 34–47, Nov./Dec. 2001.
[3] Z. Xiong et al., “UAV-assisted wireless energy and data transfer with deep
reinforcement learning,” IEEE Trans. Cogn. Commun. Netw., vol. 7, no. 1,
pp. 85–99, Mar. 2021.
[4] H. Yang, A. Alphones, Z. Xiong, D. Niyato, J. Zhao, and K. Wu, “Artificial-
intelligence-enabled intelligent 6G networks,” IEEE Netw., vol. 34, no. 6,
pp. 272–280, Nov./Dec. 2020.
[5] W. Hu et al., “Quantifying the impact of edge computing on mobile
applications,” in Proc. 7th ACM SIGOPS Asia-Pacific Workshop Syst., New
York, NY, USA: Association for Computing Machinery, 2016, pp. 1–8.
[6] X. Chen, Q. Shi, L. Yang, and J. Xu, “ThriftyEdge: Resource-efficient
Fig. 12. The dynamic of time-average system costs vs time slot t under edge computing for intelligent IoT applications,” IEEE Netw., vol. 32,
different rounding methods. no. 1, pp. 61–65, Jan./Feb. 2018.
[7] H. Ma, Z. Zhou, and X. Chen, “Predictive service placement in mobile
edge computing,” in Proc. IEEE/CIC Int. Conf. Commun. China, 2019,
shows that our RCRDR has good scalability and can be used to pp. 792–797.
[8] T. Koketsu Rodrigues, J. Liu, and N. Kato, “Offloading deci-
handle a large system where contains a big number of tasks. sion for mobile multi-access edge computing in a multi-tiered 6G
In Fig. 12, we show the detailed time-average system costs network,” IEEE Trans. Emerg. Topics Comput., to be published,
among different approximation policies vs time slot t. We can see doi: 10.1109/TETC.2021.3090061.
[9] T. K. Rodrigues, J. Liu, and N. Kato, “Application of cybertwin for
as time slot t increases, the time-average system cost achieved offloading in mobile multi-access edge computing for 6G networks,” IEEE
by RCRDR and Opt is very close, and the approximate ratios Internet Things J., vol. 8, no. 22, pp. 16231–16242, Nov. 2021.
between RCRDR and Opt to maintain in a very small value [10] Y. Xiao and M. Krunz, “QoE and power efficiency tradeoff for fog
(only 4% larger than the optimal solutions). This shows that the computing networks with fog node cooperation,” in Proc. INFOCOM
IEEE Conf. Comput. Commun., 2017, pp. 1–9.
RCRDR can always achieve a good performance in system cost [11] W. Shi, J. Cao, Q. Zhang, Y. Li, and L. Xu, “Edge computing: Vision and
minimization. challenges,” IEEE Internet Things J., vol. 3, no. 5, pp. 637–646, Oct. 2016.
Moreover, we further extend the number of F to evaluate the [12] N. Bhushan et al., “Network densification: The dominant theme for wire-
less evolution into 5G,” IEEE Commun. Mag., vol. 52, no. 2, pp. 82–89,
performance of RCRDR. From TABLE III we can see when the Feb. 2014.
number of tasks and FESs-EH in FET continues to increase, the [13] D. Li, K. Ota, Y. Zhong, M. Dong, Y. Tang, and J. Qiu, “Towards
running time of Opt increases dramatically, especially when the high-efficient transaction commitment in a virtualized and sustainable
RDBMS,” IEEE Trans. Sustain. Comput., vol. 6, no. 3, pp. 507–521,
system size is larger than N = 75, F = 6, Opt takes nearly 38 Jul.-Sep. 2021.
times longer than RCRDR. Therefore, the RCRDR can provide [14] S. P. Beeby, M. J. Tudor, and N. M. White, “Energy harvesting vibration
a good trade-off between performance and time complexity, es- sources for microsystems applications,” Meas. Sci. Technol., vol. 17,
pecially when the system size is large. Though the IRR achieves no. 12, pp. R175–R195, Oct. 2006.
[15] X. Lu, P. Wang, D. Niyato, D. I. Kim, and Z. Han, “Wireless networks with
a small running time, it violates the computing resource capacity RF energy harvesting: A contemporary survey,” IEEE Commun. Surv. Tut.,
in both FET and BET. vol. 17, no. 2, pp. 757–789, Apr.-Jun. 2015.
[16] Q. Li and M. Zhou, “The survey and future evolution of green computing,”
in Proc. IEEE/ACM Int. Conf. Green Comput. Commun., 2011, pp. 230–
VII. CONCLUSION 233.
[17] M. C. V. S. Mary, G. P. Devaraj, T. A. Theepak, D. J. Pushparaj, and J. M.
In this paper, we study the joint task offloading and energy Esther, “Intelligent energy efficient street light controlling system based
scheduling problem in an edge computing system with multiple on IoT for smart city,” in Proc. Int. Conf. Smart Syst. Inventive Technol.,
tiers in which the edge servers in FET can harvest the green 2018, pp. 551–554.
[18] Y.-C. Chang and Y.-H. Lai, “Campus edge computing network based on
energy. We introduce harvested energy (green energy) into IoT street lighting nodes,” IEEE Syst. J., vol. 14, no. 1, pp. 164–171,
the edge servers when managing a multi-tier edge computing Mar. 2020.

Authorized licensed use limited to: CINVESTAV. Downloaded on October 10,2022 at 19:37:15 UTC from IEEE Xplore. Restrictions apply.
MA et al.: GREENEDGE: JOINT GREEN ENERGY SCHEDULING AND DYNAMIC TASK OFFLOADING IN MULTI-TIER EDGE COMPUTING SYSTEMS 4335

[19] M. J. Neely, “Stochastic network optimization with application to commu- Peng Huang (Graduate Student Member, IEEE) re-
nication and queueing systems,” Synth. Lectures Commun. Netw., vol. 3, ceived the bachelor’s degree in 2019 from Sun Yat-sen
no. 1, pp. 1–211, 2010. University, Guangzhou, China, where he is currently
[20] H. Sun, F. Zhou, and R. Q. Hu, “Joint offloading and computation energy working toward the master’s degree with the School
efficiency maximization in a mobile edge computing system,” IEEE Trans. of Computer Science and Engineering. His research
Veh. Technol., vol. 68, no. 3, pp. 3052–3056, Mar. 2019. interests include mobile edge computing and SLAM
[21] J. Feng, Q. Pei, F. R. Yu, X. Chu, and B. Shang, “Computation offloading (simultaneous localization and mapping).
and resource allocation for wireless powered mobile edge computing
with latency constraint,” IEEE Wireless Commun. Lett., vol. 8, no. 5,
pp. 1320–1323, Oct. 2019.
[22] F. Wang, J. Xu, X. Wang, and S. Cui, “Joint offloading and computing
optimization in wireless powered mobile-edge computing systems,” IEEE Zhi Zhou (Member, IEEE) received the B.S., M.E.,
Trans. Wireless Commun., vol. 17, no. 3, pp. 1784–1797, Mar. 2018. and Ph.D. degrees from the School of Computer Sci-
[23] K. Peng, J. Nie, N. Kumar, C. Cai, and Y. Zhang, “Joint optimization ence and Technology, Huazhong University of Sci-
of service chain caching and task offloading in mobile edge computing,” ence and Technology, Wuhan, China, in 2012, 2014,
Appl. Soft Comput., vol. 103, no. 3, 2021, Art. no. 107142. and 2017, respectively. He is currently a Research Fel-
[24] Z. Li, M. Xu, J. Nie, J. Kang, W. Chen, and S. Xie, “Noma-enabled low with the School of Data and Computer Science,
cooperative computation offloading for blockchain-empowered Internet Sun Yat-sen University, Guangzhou, China. Since
of Things: A learning approach,” IEEE Internet Things J., vol. 8, no. 4, 2016, he has been a Visiting Scholar with the Univer-
pp. 2364–2378, Feb. 2021. sity of Göttingen, Göttingen, Germany. His research
[25] Z. Wei, B. Zhao, J. Su, and X. Lu, “Dynamic edge computation offloading interests include edge computing, cloud computing,
for Internet of Things with energy harvesting: A learning method,” IEEE and distributed systems. He was nominated for the
Internet Things J., vol. 6, no. 3, pp. 4436–4447, Jun. 2019. 2019 CCF Outstanding Doctoral Dissertation Award, the sole recipient of 2018
[26] M. Min, L. Xiao, Y. Chen, P. Cheng, D. Wu, and W. Zhuang, “Learning- ACM Wuhan & Hubei Computer Society Doctoral Dissertation Award, and the
based computation offloading for IoT devices with energy harvesting,” recipient of the Best Paper Award of IEEE UIC 2018.
IEEE Trans. Veh. Technol., vol. 68, no. 2, pp. 1930–1941, Feb. 2019.
[27] Y. Mao, J. Zhang, and K. B. Letaief, “Dynamic computation offloading
for mobile-edge computing with energy harvesting devices,” IEEE J. Sel.
Areas Commun., vol. 34, no. 12, pp. 3590–3605, Dec. 2016. Xiaoxi Zhang (Member, IEEE) received the B.E.
[28] J. Xu, L. Chen, and S. Ren, “Online learning for offloading and autoscaling degree in electronics and information engineering
in energy harvesting mobile edge computing,” IEEE Trans. Cogn. Com- from the Huazhong University of Science and Tech-
mun. Netw., vol. 3, no. 3, pp. 361–373, Sep. 2017. nology, Wuhan, China, in 2013, and the Ph.D. degree
[29] L. Park, C. Lee, W. Na, S. Choi, and S. Cho, “Two-stage computation in computer science from The University of Hong
offloading scheduling algorithm for energy-harvesting mobile edge com- Kong, Hong Kong, in 2017. She is currently an Asso-
puting,” Energies, vol. 12, no. 22, 2019, Art. no. 4367. ciate Professor with the School of Computer Science
[30] J. A. Nelder and R. Mead, “A simplex method for function minimization and Engineering, Sun Yat-sen University (SYSU),
comput,” Comput. J., no. 4, p. 4, 1965. Guangzhou, China. Before joining SYSU, she was a
[31] H. Ma, P. Huang, Z. Zhou, and X. Chen, “Greenedge: Joint green energy Postdoctoral Researcher with the Department of Elec-
scheduling and dynamic task offloading in multi-tiered edge computing trical and Computer Engineering, Carnegie Mellon
systems,” [Online]. Available:https://bit.ly/3oY2CRE. University, Pittsburgh, PA, USA. Her research interests include optimization and
[32] F. A. Chudak and D. B. Shmoys, “Improved approximation algorithms for algorithm design for networked systems, including cloud and edge computing
the uncapacitated facility location problem,” SIAM J. Comput., vol. 33, networks, NFV systems, and distributed machine learning systems.
no. 1, pp. 1–25, 2003.
[33] I. Baev, R. Rajaraman, and C. Swamy, “Approximation algorithms for data
placement problems,” SIAM J. Comput., vol. 38, no. 4, pp. 1411–1429,
2008.
[34] A. A. Ageev and M. I. Sviridenko, “Pipage rounding: A new method of Xu Chen (Senior Member, IEEE) received the Ph.D.
constructing algorithms with proven performance guarantee,” J. Combi- degree in information engineering from the Chinese
natorial Optim., vol. 8, no. 3, pp. 307–328, 2004. University of Hong Kong, Hong Kong, in 2012. He is
[35] H. Ma, Z. Zhou, and X. Chen, “Leveraging the power of prediction: Pre- currently a Full Professor with Sun Yat-sen Univer-
dictive service placement for latency-sensitive mobile edge computing,” sity, Guangzhou, China, and the Vice Director with
IEEE Trans. Wireless Commun., vol. 19, no. 10, pp. 6454–6468, Oct. 2020. National and Local Joint Engineering Laboratory.
[36] X. Chen, B. Proulx, X. Gong, and J. Zhang, “Exploiting social ties for From 2012 to 2014, he was a Postdoctoral Research
cooperative D2D communications: A mobile social networking case,” Associate with Arizona State University, Tempe, AZ,
IEEE/ACM Trans. Netw., vol. 23, no. 5, pp. 1471–1484, Oct. 2015. USA. From 2014 to 2016, he was a Humboldt Scholar
[37] W. Fang, X. Zhao, Y. An, J. Li, Z. An, and Q. Liu, “Optimal scheduling Fellow with the Institute of Computer Science, Uni-
for energy harvesting mobile sensing devices,” Comput. Commun., vol. 75, versity of Goettingen, Goettingen, Germany. He was
no. 1, pp. 62–70, 2016. the recipient of the prestigious Humboldt Research Fellowship awarded by
[38] T. Yang, N. Liu, W. Kang, and S. ShamaiShitz, “Converse results for the Alexander von Humboldt Foundation of Germany, 2014 Hong Kong Young
downlink multicell processing with finite backhaul capacity,” IEEE Trans. Scientist Runner-up Award, 2016 Thousand Talents Plan Award for Young
Inf. Theory, vol. 65, no. 1, pp. 368–379, Jan. 2019. Professionals of China, 2017 IEEE Communication Society Asia-Pacific Out-
[39] J. Kim, H. Lee, S.-E. Hong, and S.-H. Park, “Flexible connectivity standing Young Researcher Award, 2017 IEEE ComSoc Young Professional
level control via rate splitting in fog RAN downlink with finite-capacity Best Paper Award, Honorable Mention Award of 2010 IEEE international
backhaul,” in Proc. Int. Conf. Inf. Commun. Technol. Convergence, 2020, conference on Intelligence and Security Informatics (ISI), Best Paper Runner-up
pp. 984–988. Award of 2014 IEEE International Conference on Computer Communications
(INFOCOM), and Best Paper Award of 2017 IEEE Intranational Conference
Huirong Ma (Graduate Student Member, IEEE) re- on Communications (ICC). He is currently the Area Editor of the IEEE
ceived the B.S. degree from the School of Com- OPEN JOURNAL OF THE COMMUNICATIONS SOCIETY, an Associate Editor for
puter Science and Technology, Soochow University, the IEEE TRANSACTIONS WIRELESS COMMUNICATIONS, IEEE INTERNET OF
Suzhou, China, in 2017. She is currently work- THINGS JOURNAL, IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, and
ing toward the Ph.D. degree with the School of IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS (JSAC) series on
Data and Computer Science, Sun Yat-sen University, Network Softwarization and Enablers.
Guangzhou, China. Her research interests include
edge computing, edge intelligence, and edge relia-
bility.

Authorized licensed use limited to: CINVESTAV. Downloaded on October 10,2022 at 19:37:15 UTC from IEEE Xplore. Restrictions apply.