SPECIAL SECTION ON HUMAN-DRIVEN EDGE COMPUTING (HEC)

Received February 13, 2020, accepted February 27, 2020, date of publication March 2, 2020, date of current version March 11, 2020.
Digital Object Identifier 10.1109/ACCESS.2020.2977643

Sensing Cloud Computing in Internet of Things: A

Novel Data Scheduling Optimization Algorithm
ZEYU SUN 1,2 , ZHIGUO LV1 , HUIHUI WANG 3 , (Senior Member, IEEE),
ZHIXIAN LI1 , FUQIAN JIA2 , AND CHUNXIAO LAI2
1 Schoolof Computer Science and Information Engineering, Luoyang Institute of Science and Technology, Luoyang 471023, China
2 Schoolof Information Engineering, Henan Institute of Science and Technology, Xinxiang 453003, China
3 Department of Engineering, Jacksonville University, Jacksonville, FL 32211, USA

Corresponding author: Zhiguo Lv ([email protected])

This work was supported in part by the National Natural Science Foundation of China under Grant U1604149, in part by the Henan
Province Education Department Cultivation Young Key Teachers in University under Grant 2016GGJS-158, in part by the Luoyang
Institute of Science and Technology High-Level Research Start Foundation under Grant 2017BZ07, in part by the Henan Province
Education Department Natural Science Foundation under Grant 19A520006, in part by the Key Science and Technology Program of Henan
Province under Grant 192102210249 and Grant 192102210116, and in part by the Scientific Research Fund of Hunan Provincial Education
Department under Grant 19A275.

ABSTRACT In order to addressing the issues of data matching deviation and load imbalance during the data
scheduling process of the Internet of Things, Sensing Cloud Computing in IoT: A Novel Data Scheduling
Optimization (SCC-DSO) Algorithm is proposed in this paper. First, according to the processing capacity
of the IoT working node, a data placement algorithm is designed to reasonably place the input data of the
job. Second, the data scheduling queue is optimized based on the data block storage location information to
reduce non-local execution of data scheduling. Furthermore, a data prefetching method is designed to pull
the data required for non-local data scheduling in advance, and shorten the waiting time of the task for input
data. Finally, simulation experiment evaluated by the task localization execution rate and response time is
performed. The effectiveness and stability of the algorithm is verified compared with other algorithms.

INDEX TERMS Sensing-cloud computing, Internet of Things, data scheduling, optimization algorithm.

I. INTRODUCTION located, that is, local execution of the task, thereby avoiding
The data-intensive computing model brings great challenges the transmission of task input data across the network and
to traditional grid computing [1]–[4]. In the IoT computing improving System network resource utilization [12]– [15].
environment, computing nodes and storage nodes are inde- The Internet of Things, as a typical example of big data
pendent of each other and interconnected through a high- processing systems, is mainly used to perform data-intensive
speed network [5]–[8]. This mode has many advantages, tasks. In the face of huge amounts of data, large-scale data
for example the storage and computing are independent to transmission in the network will seriously degrade system
facilitate the upgrade of the two parts of the resource, and performance. Therefore, data locality research on the IoT
the storage system can be seamlessly connected to different cluster system is of great significance for improving task
computing systems. However, in the era of big data, facing execution efficiency and cluster resource utilization.
the processing of massive data, the grid architecture with In order to improve the reliability of data, most IoT clusters
separate storage and computing may make network trans- adopt a redundant storage strategy. When data is stored in the
mission a bottleneck for system performance [9]–[11]. In an system, multiple copies are created and stored on different
IoT cluster, working nodes serve two roles: compute nodes working nodes. This strategy improves the disaster prepared-
and storage nodes. This architecture brings great flexibility ness of the file system, and also balances the load of each
to data parallel processing. Tasks submitted by users can working node in the cluster.
be scheduled for execution at the node where the data is However, this data placement strategy easily leads to non-
local execution of data scheduling [16]–[19]. In addition,
The associate editor coordinating the review of this manuscript and the input data for data scheduling in the Internet of Things
approving it for publication was Lu Liu . system usually has the same size as a Distributed File System

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see http://creativecommons.org/licenses/by/4.0/
VOLUME 8, 2020 42141
Z. Sun et al.: SCC-DSO Algorithm

(HDFS) data block. During the submission process of the job, of IoT cluster configuration brings difficulties to users. Then,
the schedulers creates a Task queue, and establish task-to-data a smart computing-based IoT performance forecasting model
block mapping for each data schedule in the queue [20]–[24]. is proposed. During the operation of the job, different num-
For a multi-copy storage strategy, the same data schedule bers of data tasks are set, and the node utilization of the job
may appear in multiple task queues. During the job executing in each configuration is tested, and a polynomial regression
process, when a working node has disengaged resources, prediction models is established based on this. The limita-
it sends a task request to the scheduler. After the scheduler tion of existing performance prediction methods is that most
accepts the request, it searches for a local task for execution models only consider the impact from the cluster configura-
according to the task list of the working node. If the search tion parameters to job execution performance. However, the
fails, the scheduler looks for the unscheduled data schedule impact from the current resource utilization of the cluster
from the task linked list of other nodes followed the same rack to job performance is rarely considered. In addition, these
first and then different racks, resulting in non-local execution existing IoT job performance prediction models are mostly
of the data schedule. oriented to homogeneous IoT clusters, and the prediction
accuracy in heterogeneous environments is not high.
II. RELATED WORK In addition, due to the numerous parameters of the Internet
Paper [25] aimed at solving the problem of how to optimize of Things, the time load required to select the configuration
the allocation of system resources during the operation of parameters suitable for the prediction model is heavy. The
the Internet of Things to reduce the operating cost of the paper [30] analyzed the relationship between user service
job. Based on the load of the job input and the available level agreements and service nodes. Constrained energy con-
system resources, a cost model for the job was established. sumption model based on the reliability of the system is
Based on this model, system performance is optimized. In this given. Under the premise of meeting this agreement, the load
model, the execution of data scheduling is divided into three power of the service node is reduced and the operating cost is
stages: reading data, executing functions, and writing data. saved. Paper [31] adopted (Dynamic voltage and frequency
The three-phase task execution cost is described by a cost scaling, DVFS) technology to dynamically adjust the power
function. The cost of the final data scheduling is a linear of the working nodes to reduce energy consumption. This
combination of the above three cost functions, and the cost method first establishes the system node usage rate model
function of the data task is established in the same way. Based under load balancing conditions, and then uses feedback
on the cost model, the author optimizes the allocation of theory to control the data usage rate to reduce the energy loss
resources under a certain budget. of the system while ensuring the cluster node load balance.
Paper [26] established a prediction model of job execution The paper [32] established a power distribution optimization
time named Reservation FirstFit and Feedback Distribution model based on the queuing theory to address the issue of
(RF-FD) based on configuration parameters of the IoT clus- how to reasonably allocate the power of each node in the
ter. The model includes three parts: job performance mea- cluster on the premise of ensuring node performance. Exper-
surement, model establishment and prediction. First, the job imental results show that different application scenarios have
execution time is counted under different system configu- different power distribution strategies. Under the constraints
ration parameters. Then, a multiple linear regression model of cost, the performance of the entire cluster may not be
is established between the configuration parameters and the optimal when the working node is running at the maximum
job execution time. Finally, the execution time of the job power consumption mode. The power distribution strategy is
is predicted based on the established performance model. affected by factors such as specific application characteristics
This model only considers the maximum data scheduling and working node processor performance characteristics.
concurrency and data task concurrency of each node. In addition, many researchers have developed distributed
A method is proposed in Paper [27] (architecture for data systems based on energy awareness and virtualization tech-
synchronization based on fog computing, RSYNC) which nology to optimize energy consumption in data center cloud
studies the effect of cluster configuration parameters on the computing environments [33]– [36]. In order to improve the
performance of IoT operations, and builds a performance utilization of resources in the cluster system, part of the
prediction model based on it. An automatic parameter tuning research work uses virtual machine scheduling and migration
method is adopted in paper [28] for IoT cluster systems technology to aggregate virtual machines on the premise
to adjust the configuration parameters from three layers: of not reducing user service quality, thereby saving system
the application layer, the Internet of Things layer, and the resources and improving system energy utilization. Whether
resource layer according to the MAPE-K cycle. In this it is a task scheduling strategy or a virtualization-based
method, the author establishes a prediction model about energy consumption optimization strategy, the key issue is to
networked job performance. This model only considers the understand the characteristics of energy consumption when
impact of three factors, such as the number of job input different loads are running in a cluster environment. To this
files, the size of the input data, and the number of nodes, end, the paper [37] established a linear regression model of
on job performance. It is pointed out in paper [29] that the system energy consumption based on data collected by hard-
optimization choices and forecasting the demand of resources ware counters inside the system. The paper [38] established

42142 VOLUME 8, 2020

Z. Sun et al.: SCC-DSO Algorithm

a virtual machine energy consumption prediction model in

order to optimize the energy utilization rate during virtual
machine aggregation. The limitations of the existing methods
are that the factors considered by most energy consumption
prediction models are not comprehensive and the accuracy of
the prediction is not high. In addition, the choice of factors
affecting energy consumption in the model remains at the
empirical level.
This paper focuses on the problem of local optimization of
data during the execution of data scheduling in IoT clusters.
This method considers three factors that affect the efficiency
of localization of data scheduling: i) where data scheduling
input data is placed, ii) data scheduling allocation, and iii) the
dynamic characteristics of the system when the data schedul-
ing is running. The method is mainly divided into three parts:
Firstly, the processing capacity of the working nodes is
evaluated according to the performance prediction model
proposed in section II, and the corresponding number of data
blocks is allocated according to the size of the processing FIGURE 1. Data storage process in the perception cloud architecture.
capacity;
Secondly, according to the metadata information of the
data block stored by each working node, the mapping rela-
tionship between the data schedule and the data block is in the same rack reduces read and write overhead between
generated and the allocation queue of the data schedule is racks; second, storing copies in different racks avoids failure
optimized. caused by rack failures. In addition, during the node selection
Finally, during the data scheduling execution process, process, the storage capacity of the working nodes is also con-
in order to avoid the non-local execution of data scheduling sidered, and the data blocks are stored as evenly as possible
caused by the prediction error of the performance prediction in each storage node to ensure the load balance of the system
model, a data pre-fetching algorithm is designed to pre-fetch during the running process.
the input data required for data scheduling to the working According to the above analysis, it can be seen that the
node where the data scheduling is located by using the char- rack-aware copy strategy is a data storage strategy for a
acteristics of parallel execution of computing and network homogeneous IoT cluster. It takes into account the storage
transmission to reduce the waiting time of tasks for input data. capacity and stores the data evenly to ensure the load bal-
ance of the system [44], [45]. In heterogeneous IoT clus-
ters, the processing capabilities between working nodes are
III. DATA BLOCK PLACEMENT METHOD different, and there may be huge differences in performance
A. RACK-AWARE BASED DATE PLACEMENT STRATEGIES between different working nodes. If a storage strategy that
The file system in the IoT cluster uses a distributed storage only considers storage capacity is adopted, then the work-
method, where each file is cut into data blocks of equal size ing nodes with weak processing capacity become the per-
and stored on different working nodes [38]–[41]. In order formance bottleneck of the operation. It causes non-local
to improve the reliability of the file system and the ability execution of data scheduling and imbalance in system load.
of concurrent access, each data block has multiple copies, As a result, the system performance is reduced.
and these copies are stored on multiple nodes to prevent data
access failure caused by a single node failure. Different file B. HETEROGENEOUS SENSING DATA PLACEMENT
system copy placement strategies are different. For example, ALGORITHM
HDFS file systems in IoT clusters use rack-aware data copy In IoT clusters, the execution time of data scheduling of
placement strategies [42], [43]. When the client uploads a file, the same input data size on different nodes is different. The
the file system first places the data block on the node where performance of the working nodes in processing IoT tasks is
the client is located, and then uses Pipline to store the second affected by several factors, such as the node data frequency,
copy on other nodes in the same rack as the node where the memory size, and disk read/write speeds [46], [47]. In addi-
client is located. Finally, the third copy is stored on those tion, the execution time of the task is also related to the char-
nodes with different racks compared with the node where the acteristics of the task itself, such as the size of the input data,
client is located. whether it meets locality, and so on. When the data scheduling
Figure 1 illustrates the rack-aware data block placement does not satisfy the data locality, the input data needs to
process. The rack-aware data block placement method has the be transmitted from other nodes in the cluster through the
following advantages: first, the strategy of storing two copies network. In this case, the execution time is also affected by

VOLUME 8, 2020 42143

Z. Sun et al.: SCC-DSO Algorithm

the network bandwidth [48]– [50]. The purpose of measuring

the processing capacity of working nodes is to solve the
placement scheme of data blocks. Therefore, the processing
capacity of nodes discussed in this section mainly refers to the
performance of nodes when data scheduling meets locality.
Definition 1: For an application of the Internet of Things
system, if the input file size is Input (APPi ), the number of
data blocks in the input file can be obtained by the following
formula:
Input (APPi )
B= (1)
b FIGURE 2. Optimal data block placement.

where b is the storage space occupied by a single data block

in the IoT file system. equivalent to solving the following optimization problems:
  
Definition 2: For the node Nodei in the IoT cluster, suppose X f (j)

that any data scheduling mapk in an application Appj is min max t Nodej , Appi , mapk 
scheduled to be executed by the scheduler. The input data size 
k=1

of mapk is m and the data locality is met, then the calculation n
X
rate for data scheduling in the node Nodei coresponding to the s.t. 1 ≤ j ≤ n; f (j) = B (4)
application Appj can be obtained by the following formula: j=1

mk In formula (4), f (j) represents the number of data blocks

P Nodei , Appj , mapk = (2) stored in the node Nodej ; B represents the total number of
tk
input data blocks of the application Appi .
where tk is the execution time of data schedules mapk on 
Nodei . T Nodei , Appp , mapq × f (i)


Definition 3: For the node Nodei in the IoT cluster, if any 
= T Nodej , Appp , mapq × f (j)

data scheduling mapk in an application Appj is scheduled to n
(5)
f (i) = B
 P
be executed by the scheduler, the running time of the task on 1 ≤ i, j ≤ n i 6 = j,


the node Nodei can be obtained by the following formula: i=1
The precondition of the formula (4) is that the degree of
T Nodei , Appj , mapk

copy of the data block is 1. Assume that there are three
S (3) working nodes in the cluster with different processing capa-
σs K (v (mapk ) , v (maps )) + a
P
=
s=1
bilities. Ideally, the data block placement and data scheduling
execution time of each node corresponding to the optimal
In formula (3), v(mapk ) is the feature vector of data solution of the optimization problem represented by formula
scheduling, which includes the hardware characteristics and (4) is shown in Figure 2. From Figure 2 we can see that,
load characteristics of the cluster nodes; K (v(mapk ),v(maps )) under the condition of a single copy, the best data block
is the Gaussian radial basis function; a is the deviation. strategy is to be able to satisfy each node to complete its
From the analysis in Section III, we know that the running assigned data scheduling at the same time. Based on the above
time of the Internet of Things job is composed of two parts: analysis, the optimization problem shown in formula (4) can
node running time and data running time. When the job be transformed into the solution of the following problems:
is in the node execution phase, data scheduling run on the
working nodes in the IoT cluster in parallel method. The node C. DATA PLACEMENT ALGORITHM DATA SCHEDULING
running time depends on the completion time of the work- QUEUE OPTIMIZATION
ing node that performs the data scheduling with the longest In the perceptual cloud computing system, the input data
time. Therefore, reducing the load of the working node can of the job is stored on the working nodes in a multi-copy
reduce overall running time of the network. In the ideal block storage method. During the job initialization phase,
case, when the data schedule assigned by each working node the scheduler builds a data scheduling queue that satisfies the
satisfies data locality, and each working node can complete locality of data according to the distribution of data blocks
the assigned data schedule at the same time, the job can obtain stored in the cluster [8], [51], [52]. The multi-copy storage
the minimum completion time during the network operation mechanism causes the scheduling of the same input data
phase. Suppose that the time taken by the application IoT in the data scheduling queues on different nodes. The main
application APPi to run data scheduling mapk on node Nodej task of optimizing the data scheduling queue is to adjust the
ist(Nodej ,APPi , mapk ), then for an IoT cluster consisting position of each data schedule in the queue based on the data
ofn nodes, the optimal data block placement solution can be block allocation result generated by the SCC-DSO algorithm,

42144 VOLUME 8, 2020

Z. Sun et al.: SCC-DSO Algorithm

X
cost (T (s, M )) = cost (e)
e∈PT (s,u)
X
+ cost (n) (7)
n∈PT (s,u)
bandwidth (PT (s, u)) = min {bandwidth (e) , n ∈ PT (s, u)}
(8)
X
delay-jit (PT (s, u)) = delay-jit (e)
e∈Pt (s,u)
X
+ delay-jit (n) (9)
n∈PT (s,u)
Y
loss-packet (Pt (s, u)) = 1 − (1 − loss-packet (n))
n∈PT (s,u)
(10)

D. DATA PREFETCHING
Local execution of data scheduling can avoid reading data
FIGURE 3. Data scheduling queue optimization. remotely, shortening the task’s waiting time, and improv-
ing job execution efficiency. Adjust the scheduling order of
and to generate an optimized queue for data scheduling, data scheduling according to the principle of maximizing
so that the assignment of tasks among the working nodes locality to optimize the data scheduling queue. However,
does not affect each other and improve data scheduled local the dynamic changes in the work load of the working nodes
execution efficiency. In order to describe the process of data may cause errors between the actual execution schedule and
scheduling queue pre-processing and pre-allocation, assume the predicted schedule of some data scheduling. As a result,
that the IoT cluster is composed of 5 nodes, and the input local data scheduling in the task queue maintained by some
data of a job is composed of 26 data blocks. The degree of nodes is completed earlier or delayed in the later stage of job
data block replication in the file system is 2, and the data execution. This causes tasks migrate from nodes with slow
block distribution obtained by the SCC-DSO algorithm is: progress to nodes with fast progress, then increases the non-
Node1 ={m0 , m1 , m2 , m3 , m4 , m5 }, Node2 ={m6 , m7 , m8 , localized execution rate of data scheduling.
m9 , m10 , m11 }, Node3 ={m12 , m13 , m14 , m15 , m16 , m17 }, The purpose of the node task queue monitoring is to track
Node4 ={m18 , m19 , m20 , m21 }, Node5 ={m22 , m23 , m24 , the execution progress of each task queue, and select the
m25 }. It is described in Figure 3 that the queues before and source and destination nodes for the migration data schedul-
after the initial data scheduling of each working node. ing during the data prefetch process. Node selection results
As shown in Figure 3, after the initial task queue is opti- are affected by several factors, such as the timing of node
mized, the task allocation of each node does not affect each selection. In this section we describe the selection method
other. Because the SCC-DSO algorithm allocates data blocks of source node and destination node through formal methods.
according to the processing capacity of the working nodes, Assume that the number of working nodes in the IoT cluster is
once the working nodes have free computing resources, they n, and the set of nodes is Node={Node1 , Node2 , . . . , Noden }.
send task requests to the scheduler, and the scheduler allo- Definition 4: After the task queue of the working nodes
cates data scheduling according to the optimized task queue, is optimized, the Nodei ’s optimization task set of the data
thereby improving the locality of the data scheduling. scheduling queue can be expressed as:
In order to further optimize the scheduling queue, this QNodei = {mapi1 , mapi2 , . . . , mapis } (11)
paper introduces the ant colony algorithm of artificial intelli-
gence. For the data scheduling model, we have introduced Definition 5: When data migration is implemented,
4 degree functions, which are delay functions: Delay (e): the time for the system to select the source node and the
E → R+ ; delay jitter function: Delay-jit (e): E → R+ ; cost destination node. Assuming that the total number of data
function: cost (e): E → R+ ; and Loss-packet (e): V → R+ . scheduling for IoT jobs is m, the node selection time threshold
Loss or damage may occur during data scheduling. If packet λ can be expressed as:
loss or damage is high, the data will not be complete. n lmm
λ=1−θ × 1≤θ < (12)
X m n
delay (PT (s, u)) = delay (e) Here θ represents the experience value, which can be
e∈Pt (s,u) adjusted according to the type of IoT job. If the average
X
+ delay (n) (6) execution time of data scheduling in the job is short, this value
n∈PT (s,u) can be set to a larger value, and vice versa.

VOLUME 8, 2020 42145

Z. Sun et al.: SCC-DSO Algorithm

Definition 6: When the system selects the source node and TABLE 1. The pseudocode of SCC-DSO algorithm.
the destination node according to the node selection time
threshold, the estimated remaining time of the optimized data
scheduling in the task queue.
Let the optimized task queue of the node Nodei ’s be
QNodei . When selecting the source node and the destination
node, the total number of data scheduling in QNodei is m,
the progress of the ongoing data scheduling mk is ρ, and the
input data block of task mk is Bk , the number of remaining
unscheduled data scheduling is r, then the remaining com-
pletion time R(QNodei ) of the queue QNodei is:

R (QNodei ) = T × r + Bk × (1 − ρ) × Vi  
m−r−1
P Bj (13)
Vi = tj (m−r−1) 

j=1

where T represents the predicted execution time of data

scheduling; Vi represents the average execution rate of com-
pleted data scheduling in QNodei .
After selecting the source and destination nodes for the
task migration, it is needed to determine whether the task
is suitable for migration. The completion time of the Map
phase of an IoT job depends on the time taken by the node being executed in the SQueue queue is ms , and its execution
that completed the data scheduling with the most time. That progress is ρs ; the data prefetch delay is ϕ. According to
is the optimization time of the task queue that has the longest formula (8), the remaining completion times of the optimized
remaining completion time in the task queue set. If part of queues SQueue and TQueue can be obtained respectively,
the data in the queue migrates to other nodes (destination which are represented as R(SQueue) and R(TQueue). The
nodes) can reduce the execution time of the entire job Map migration task migration should also meet the following con-
phase. In the best case, the task queues on the source and ditions:
destination nodes have the same remaining completion time R(TQueue) > φ

(14)
after migration. However, in order to ensure the data locality R(SQueue) − T (SNode) > φ
of data scheduling, the nodes must ensure that these tasks
meet the data locality before the nodes perform the migration Here T(SNode) represents the predicted execution time
tasks, that is, the input data required by the tasks has been of data scheduling on the source node. The first condition
successfully prefetched. Therefore, when performing migrat- guarantees the locality of the data scheduling data of the desti-
ing tasks, it is needed to consider the impact of the time taken nation node. The second condition avoids the waste of system
by data prefetch on task migration. If the optimization task resources caused by the source node and the destination node
queue of the destination node is empty within the time of performing the same data scheduling at the same time.
data prefetching, it may cause the destination node to steal In the perceptual cloud computing system, assume the
the data scheduling of other nodes, which may cause non- destination node of the task migration is TNode, the current
local execution of the data scheduling. number of network connections is Tt , the migration task is
Definition 7: The time required for data to travel from the mt, and the internal data prefetch delay of the same node is
remote node to the target node over the network. This value φ1 , The data prefetch delay between two nodes located in the
is an empirical value. Different types of IoT job correspond same rack is φ2 , the data prefetch delay between two nodes
to different prefetch delays. In addition, the value reflects located in different racks is φ3 , and the set of nodes storing c
the relationship between the size of the data block and the copies of prefetch data block B is CNode={CNode1 , CNode2 ,
network transmission rate. When the data block is large, . . . , CNodec }. Assume that a copy storage node of data blockB
the transmission takes up more network resources. Under this is CNodei , and the current network connection number of this
condition, the time should be set to a larger value. Otherwise, node is Ti . CNodei and TNode are located in different racks
the time should be set to smaller value. respectively. The prefetch load factor at node CNodei can be
Let the destination node of the migration task be TNode, expressed as:
the optimized task queue of this node is TQueue; the source PLFactor (TNode, CNodei )
node of the migration task is SNode, and the optimized task
q (15)
= (φi − φt )2 + (Ti − Tt )2
queue of this node is SQueue. Let the selection time thresh-
old of node isλ. At timeλ, the data being executed in the The Source Worker Node Choosing (SWNC) algorithm
TQueue queue is mt and its execution progress is ρt ; the data uses the prefetch load factor to select the source data node:

42146 VOLUME 8, 2020

Z. Sun et al.: SCC-DSO Algorithm

FIGURE 4. Comparison of execution time of different jobs under single copy conditions.

first, query the node location and network load information rate of data scheduling, the first copy placement strategy of
of the copy of the migration task input data block; then, the data block is solved by formula (5). Then, the remaining
calculate the prefetch load factor of each replica node of copies are stored according to the heterogeneous-aware data
the prefetched data block according to formula (15); finally, placement algorithm.
select the appropriate prefetch source data node according to Step 3: According to the metadata information of the data
the prefetch load factor of each replica node. block stored on each working node and the data block alloca-
tion result generated by the SCC-DSO algorithm, the SCC-
E. SCC-DSO ALGORITHM DESCRIPTION DSO algorithm is executed to adjust the position of each
The implementation process based on the data localization data scheduling in the task queue of the node so that the
optimization method is given below: assignment of tasks of each working node does not affect
Step 1: Extract performance characteristics from the com- each other and improves the local execution rate of data
pleted data scheduling log information of each node, con- scheduling.
struct a sample data set of the performance prediction model Step 4: Monitor the task queue of the node, track the
proposed in section II, and train the prediction model to obtain execution progress of the task queue of each node, and select
the values of various parameters in the model. the source node and destination node for the migration data
Step 2: Based on the processing capacity and performance scheduling during the data prefetch process. Then, based on
prediction model of the working nodes in the IoT cluster, the execution status of the task queue in the source node
according to the principle of maximizing the local execution and the destination node, it is determined whether to perform

VOLUME 8, 2020 42147

Z. Sun et al.: SCC-DSO Algorithm

FIGURE 5. Scheduled execution rate of different job data under single copy conditions.

a task migration operation. If the judgment result is true, The copy degree of HDFS file system is an important factor
performes tep 5, otherwise finish the optimization process. affecting the performance of network operations. In addition,
Step 5: Determine whether the migrated data schedule the network load in the cluster nodes will also affect the
stores the input data of the data schedule at the destination prefetch of data. For this reason, the paper first tests the
node. If it contains, the source data node selection operation method proposed in the paper compared with RF-FD [26]
is not required, and go to step 4; otherwise, execute the SCC- and RSYNC algorithm [27]. The results are compared and
DSO algorithm to select a suitable prefetch sourc data nodes. analyzed at single-copy scenarios with different data block
Step 6: Prefetch the input data blocks of the migration task sizes and network loads. Since RSYNC can only be used for
to the destination node through the network. single copy scenarios, the performance of the RF-FD method
Step 7: Establish a mapping relationship between the is tested for multi-copy scenarios. The results are compared
migration task and the pre-fetched data block, and add and analyzed with different data block sizes and network
the task to the data scheduling queue of the destination loads, too.
node.Then, go to step 4.
A. PERFORMANCE OF SCC-DSO ALGORITHM UNDER
IV. EXPERIMENTAL RESULTS AND ANALYSIS SINGLE COPY CONDITION
In this paper, multiple experiments are designed to verify This subsection tests the performance of the SCC-DSO
the effectiveness and stability of the SCC-DSO algorithm. method in a cheap and heavy network overhead environment

42148 VOLUME 8, 2020

Z. Sun et al.: SCC-DSO Algorithm

FIGURE 6. Comparision of execution time of jobs under multiple copies.

when the copy degree of HDFS file system is 1. Since the IoT has a higher localized execution rate than the other two
cluster system is interconnected through the Gigabit network, methods. When the data block size is 64M, the localized
the default network environment is regarded as a cheap net- execution rate of the SCC-DSO method can be as high as
work load environment. A cheap network load means that the 91%, while the localized execution rates of the RF-FD and
nodes have a wider network bandwidth. Each case is tested RSYNC methods are only 62.18% and 85%. The perfor-
for 50 times. The experimental results are average values mance of SCC-DSO and the localization execution rate of
which are calculated for 50 times. data scheduling have been greatly improved compared to
Figure 4 shows the execution time of each test case in the RF-FD method. This is because the data blocks in the
the network operation phase under low network conditions. RF-FD method are randomly placed on the working node
It can be seen from Figure 4 that the SCC-DSO has better according to the utilization rate of the node’s hard disk.
performance than the other two optimization methods. In the The strategy does not take into account the differences in
best case, the performance of the SCC-DSO method obtain computing capabilities of the working nodes in the heteroge-
average improvement of 13% compared with the RF-FD neous IoT, causing nodes with strong computing capabilities
method and improvement of 7% compared with the RSYNC to complete local tasks first, and then ‘‘stealing’’ local data
method. The localized execution rate of data scheduling dur- scheduling of working nodes with weaker computing capabil-
ing the execution of each test case is shown in Figure 5. ities, resulting in large amounts of data scheduling Non-local
It can be seen from Figure 5 that the SCC-DSO method execution.

VOLUME 8, 2020 42149

Z. Sun et al.: SCC-DSO Algorithm

FIGURE 7. Scheduling execution rates of different job data under multiple copies.

The non-local execution of data scheduling results in cross- method still causes some working nodes to perform non-
node transmission of data scheduling, thereby increasing local data scheduling. The SCC-DSO method proposed in
the time of the job during the network operation phase. this paper uses the performance prediction model to evalu-
The performance of the SCC-DSO method and the local- ate the processing capacity of the working node, which is
ization execution rate of data scheduling of the SCC-DSO more accurate than RSYNC’s prediction. In addition, the
method have been improved to a certain extent compared SCC-DSO method can dynamically monitor the execution
with RSYNC method. This is because that the SCC-DSO progress of the local data scheduling list of each working
method selects the input data placement node based on the node, and pre-fetch input data to the node where the non-local
ratio of the computing power of each working node and store data scheduling is located. Thus, the SCC-DSO method can
appropriate number of data blocks in equal proportion. The reduce the waiting overhead of non-local data scheduling for
nodes with strong computing power store a larger number the transmission of input data across the node network.
of data to schedule input data blocks, reducing the number
of data block migrations. However, the task execution time B. PERFORMANCE OF SCC-DSO ALGORITHM UNDER
and the input data size are not a simple linear relationship. MULTIPLE COPIES CONDITION
It also related to the load characteristics and the performance Figures 6 and 7 show the job execution time and localized
characteristics of the working nodes. Therefore, the RSYNC execution rate with adopting various optimization methods

42150 VOLUME 8, 2020

Z. Sun et al.: SCC-DSO Algorithm

in a narrow network bandwidth environment. The simula- data. Experimental results show that the algorithm in this
tion results show that the SCC-DSO algorithm proposed in paper effectively reduces the execution time of non-local data
this paper is superior in job execution time and localized scheduling. Compared with the RF-FD and RSYNC methods,
execution rate of data scheduling. Under average conditions, the localized execution rate and job performance of the SCC-
compared with the RF-FD method, the SCC-DSO method DSO algorithm have been significantly improved.
improves the performance of the job during the execution
phase by 19.8%. Compared with the RSYNC method, the per- REFERENCES
formance of the execution phase of the job is improved by [1] Z. Sun, G. Zhao, M. Li, and Z. Lv, ‘‘Job performance optimization method
7.6% on average. It can be seen from Figure 6, with the low based on data balance in the wireless sensor networks,’’ Int. J. Online Eng.,
network load, the data execution time of the three methods vol. 13, no. 12, pp. 4–17, Dec. 2017.
[2] T. Wang, H. Luo, X. Zheng, and M. Xie, ‘‘Crowdsourcing mechanism for
has been extended to a certain extent, and the localized exe- trust evaluation in CPCS based on intelligent mobile edge computing,’’
cution rate of data scheduling has been improved to a certain ACM Trans. Intell. Syst. Technol., vol. 10, no. 6, pp. 1–19, Oct. 2019.
extent. The main reason is that in a low network bandwidth [3] Z. Sun, Y. Zhang, Y. Nie, W. Wei, J. Lloret, and H. Song, ‘‘CASMOC:
A novel complex alliance strategy with multi-objective optimization of
environment, the network becomes a bottleneck of system coverage in wireless sensor networks,’’ Wireless Netw., vol. 23, no. 4,
performance. Due to non-local data scheduling, the time pp. 1201–1222, 2017.
overhead of waiting for input data on nodes with powerful [4] J. Zhu, Y. Song, D. Jiang, and H. Song, ‘‘A new deep-Q-learning-based
transmission scheduling mechanism for the cognitive Internet of Things,’’
computing power increases. So its execution efficiency is
IEEE Internet Things J., vol. 5, no. 4, pp. 2375–2385, Aug. 2018.
reduced. In this case, for the RF-FD method, a working node [5] T. Wang, Y. Mei, W. Jia, X. Zheng, G. Wang, and M. Xie, ‘‘Edge-based
with weak computing power needs to perform more local data differential privacy computing for sensor-cloud systems,’’ J. Parallel Dis-
scheduling, which increases the execution time of the job. As trib. Comput., vol. 136, pp. 75–85, Feb. 2020.
[6] J. Lu, Y. Xin, Z. Zhang, X. Liu, and K. Li, ‘‘Game-theoretic design of
shown in Figure 7, for the SCC-DSO and RSYNC methods, optimal two-sided rating protocols for service exchange dilemma in
the input data for data scheduling is stored based on the crowdsourcing,’’ IEEE Trans. Inf. Forensics Secur., vol. 13, no. 11,
node’s computing capacity. Therefore, compared to the wide pp. 2801–2815, Nov. 2018.
[7] H. Song and M. Brandt-Pearce, ‘‘A 2-D discrete-time model of physical
network bandwidth situation, the narrow network bandwidth impairments in wavelength-division multiplexing systems,’’ J. Lightw.
has weaker effect on the local network execution rate for its Technol., vol. 30, no. 5, pp. 713–726, Mar. 1, 2012.
data scheduling. However, the performance of the SCC-DSO [8] W. Tian, K. Haoxiong, Z. Xi, W. Kun, S. A. Kumar, and L. Anfeng, ‘‘Big
method is still better than RSYNC. The main reason is that the data cleaning based on mobile edge computing in industrial sensor-cloud,’’
IEEE Trans. Ind. Informat., vol. 16, no. 2, pp. 1321–1329, Feb. 2020.
data prefetch mechanism is used in the SCC-DSO method, [9] W. Tian, L. Hao, W. Jia, A. Liu, and M. Xie, ‘‘MTES: An intelligent trust
which reduces the waiting time of the working node for non- evaluation scheme in sensor-cloud enabled industrial Internet of Things,’’
local data scheduling input data network transmission. IEEE Trans. Ind. Informat., vol. 16, no. 3, pp. 2054–2062, Mar. 2020.
[10] S. Zeyu, W. Weiguo, W. Huanzhao, C. Heng, and X. Xiaofei, ‘‘A novel
coverage algorithm based on event-probability-driven mechanism in wire-
V. CONCLUSION less sensor network,’’ EURASIP J. Wireless Commun. Netw., vol. 2014,
This paper focuses on the performance degradation of jobs pp. 1–17, Dec. 2014.
[11] T. Liu, B. Wu, H. Wu, and J. Peng, ‘‘Low-cost collaborative mobile charg-
in the IoT cluster during the execution stage. A percep- ing for large-scale wireless sensor networks,’’ IEEE Trans. Mobile Com-
tual cloud based algorithm for data scheduling optimiza- put., vol. 16, no. 8, pp. 2213–2227, Aug. 2017.
tion is proposed. This method places different numbers of [12] X. Liu, ‘‘Node deployment based on extra path creation for wireless
data blocks at working nodes according to their processing sensor networks on mountain roads,’’ IEEE Commun. Lett., vol. 21, no. 11,
pp. 2376–2379, Nov. 2017.
capabilities. The dynamic migration of data scheduling is [13] T. Wang, D. Zhao, S. Cai, W. Jia, and A. Liu, ‘‘Bidirectional pre-
introduced to improve localized execution rate. Compared diction based underwater data collection protocol for end-edge-cloud
with other algorithms, the SCC-DSO algorithm can sense the orchestrated system,’’ IEEE Trans. Ind. Informat., to be published, doi:
10.1109/TII.2019.2940745.
heterogeneous processing capabilities of working nodes and [14] X. Liu, ‘‘A novel transmission range adjustment strategy for energy hole
allocate different sizes of data to working nodes according avoiding in wireless sensor networks,’’ J. Netw. Comput. Appl., vol. 67,
to their processing capabilities by means of performance pp. 43–52, May 2016.
[15] Z. Sun, L. Wei, C. Xu, T. Wang, Y. Nie, X. Xing, and J. Lu, ‘‘An energy-
prediction methods. The reliability of data stored by work- efficient cross-layer-sensing clustering method based on intelligent fog
ing nodes is also taken into account. By obtaining the ini- computing in WSNs,’’ IEEE Access, vol. 7, pp. 144165–144177, 2019.
tial data scheduling list information of each working node, [16] Y. Nie, H. Wang, Y. Qin, and Z. Sun, ‘‘Distributed and morphological
a data scheduling queue optimization algorithm based on operation-based data collection algorithm,’’ Int. J. Distrib. Sensor Netw.,
vol. 13, no. 7, pp. 1–16, Jul. 2017.
data block distribution is proposed. A data scheduling pre- [17] Z. Sun, G. Zhao, and X. Pan, ‘‘PM-LPDR: A prediction model for lost
allocation optimization queue is constructed to ensure that packets based on data reconstruction on lossy links in sensor networks,’’
different working nodes perform local data scheduling in par- Int. J. Comput. Sci. Eng., vol. 19, no. 2, pp. 177–188, 2019.
[18] S. Bitam, S. Zeadally, and A. Mellouk, ‘‘Fog computing job scheduling
allel without mutual interference. By monitoring the runtime optimization based on bees swarm,’’ Enterprise Inf. Syst., vol. 12, no. 4,
information of the task queues of each working node, a data pp. 373–397, Apr. 2017.
prefetch algorithm based on the minimum completion time of [19] X. Liu, ‘‘Routing protocols based on ant colony optimization in wireless
the task queue is proposed. The parallelism of computing and sensor networks: A survey,’’ IEEE Access, vol. 5, pp. 26303–26317, 2017.
[20] M. Mededjel, G. Belalem, and A. Neki, ‘‘Towards a traceability system
network transmission operations is used to reduce the waiting based on cloud and fog computing,’’ Multiagent Grid Syst., vol. 13, no. 1,
time of working nodes for non-local data scheduling input pp. 47–68, Apr. 2017.

VOLUME 8, 2020 42151

Z. Sun et al.: SCC-DSO Algorithm

[21] Z. Sun, X. Xing, B. Song, Y. Nie, and H. Shao, ‘‘Mobile intelligent com- [43] Y. Zhang, ‘‘Classified scheduling algorithm of big data under cloud com-
puting in Internet of Things: An optimized data gathering method based on puting,’’ Int. J. Comput. Appl., vol. 41, no. 4, pp. 262–267, 2019.
compressive sensing,’’ IEEE Access, vol. 7, pp. 66110–66122, 2019. [44] W. Xiuran and W. Feng, ‘‘Massive data balance scheduling in cloud com-
[22] X. Liu and P. Zhang, ‘‘Data drainage: A novel load balancing strategy puting environment,’’ Int. J. Mechatronics Appl. Mech., vol. 2019, no. 5,
for wireless sensor networks,’’ IEEE Commun. Lett., vol. 22, no. 1, pp. 100–105, 2019.
pp. 125–128, Jan. 2018. [45] Y. Zhang, W. Huang, T. Zhang, and T. Zhang, ‘‘A novel topology optimiza-
[23] Z. Sun, L. Li, X. Xing, Z. Lv, and N. N. Xiong, ‘‘A novel nodes deploy- tion theory and parallel data analysis model based resource scheduling
ment assignment scheme with data association attributed in wireless sensor algorithm for cloud computing,’’ Recent Adv. Elect. Electron. Eng., vol. 11,
networks,’’ J. Internet Technol., vol. 20, no. 2, pp. 509–520, 2019. no. 4, pp. 449–456, Nov. 2018.
[24] T. Wang, P. Wang, S. Cai, Y. Ma, A. Liu, and M. Xie, ‘‘A unified trust- [46] R. Xie and X. Jia, ‘‘Data transfer scheduling for maximizing throughput of
worthy environment establishment based on edge computing in indus- big-data computing in cloud systems,’’ IEEE Trans. Cloud Comput., vol. 6,
trial IoT,’’ IEEE Trans Ind. Informat., to be published, doi: 10.1109/ no. 1, pp. 87–98, Jan. 2018.
TII.2019.2955152. [47] G. Yang, T. Liang, X. He, and N. Xiong, ‘‘Global and local reliability-
[25] Q. Li, W. Wu, Z. Sun, L. Wang, J. Huang, and X. Zhou, ‘‘A novel hierar- based routing protocol for wireless sensor networks,’’ IEEE Internet Things
chal scheduling strategy for rendering system,’’ in Proc. Int. Conf. Iden- J., vol. 6, no. 2, pp. 3620–3632, Apr. 2019.
tificat., Inf., Knowl. Internet Things (IIKI), Beijing, China, Oct. 2015, [48] L. A. Tawalbeh, W. Bakheder, and H. Song, ‘‘A mobile cloud computing
pp. 206–209. model using the cloudlet scheme for big data applications,’’ in Proc. IEEE
[26] T. Wang, J. Zhou, A. Liu, M. Z. A. Bhuiyan, G. Wang, and W. Jia, ‘‘Fog- 1st Int. Conf. Connected Health, Appl., Syst. Eng. Technol. (CHASE),
based computing and storage offloading for data synchronization in IoT,’’ Wsahington, DC, USA, Jun. 2016, pp. 73–77.
IEEE Internet Things J., vol. 6, no. 3, pp. 4272–4282, Jun. 2019. [49] W. Tian, B. M. Z. Alam, W. Guojun, Q. Lianyong, W. Jie, and H. Thaier,
[27] J. Tang, A. Liu, J. Zhang, N. Xiong, Z. Zeng, and T. Wang, ‘‘A trust-based ‘‘Preserving balance between privacy and data integrity in edge-assisted
secure routing scheme using the traceback approach for energy-harvesting Internet of Things,’’ IEEE Internet Things J., to be published, doi: 10.1109/
wireless sensor networks,’’ Sensors, vol. 18, no. 3, p. 751, Mar. 2018. JIOT.2019.2951687.
[28] T. Wang, L. Qiu, A. K. Sangaiah, G. Xu, and A. Liu, ‘‘Energy-efficient [50] T. Qiu, K. Zheng, H. Song, M. Han, and B. Kantarci, ‘‘A local-
and trustworthy data collection protocol based on mobile fog comput- optimization emergency scheduling scheme with self-recovery for a smart
ing in Internet of Things,’’ IEEE Trans Ind. Informat., vol. 16, no. 5, grid,’’ IEEE Trans Ind. Informat., vol. 13, no. 6, pp. 3195–3205, Dec. 2017.
pp. 3531–3539, May 2020, doi: 10.1109/TII.2019.2920277. [51] W. Li, H. Song, and F. Zeng, ‘‘Policy-based secure and trustworthy sensing
[29] Z. Sun, X. Xing, B. Yan, and Z. Lv, ‘‘CMTN-SP: A novel coverage-control for Internet of Things in smart cities,’’ IEEE Internet Things J., vol. 5, no. 2,
algorithm for moving-target nodes based on sensing probability model in pp. 716–723, Apr. 2018.
sensor networks,’’ Sensors, vol. 19, no. 2, p. 257, Jan. 2019. [52] Z. Sun, R. Tao, L. Li, and X. Xing, ‘‘A new energy-efficient multi-target
[30] H. Song and M. Brandt-Pearce, ‘‘Range of influence and impact of phys- coverage control protocol using event-driven-mechanism in wireless sen-
ical impairments in long-haul DWDM systems,’’ J. Lightw. Technol., sor networks,’’ Int. J. Online Eng. (iJOE), vol. 13, no. 2, pp. 53–67,
vol. 31, no. 6, pp. 846–854, Mar. 15, 2013. Feb. 2017.
[31] Z. Sun, G. Zhao, and X. Xing, ‘‘ENCP: A new energy-efficient nonlinear
coverage control protocol in mobile sensor networks,’’ EURASIP J. Wire-
less Commun. Netw., vol. 2018, no. 1, pp. 1–15, Jan. 2018.
[32] X. Liu, ‘‘Survivability-aware connectivity restoration for partitioned
wireless sensor networks,’’ IEEE Commun. Lett., vol. 21, no. 11,
pp. 2444–2447, Nov. 2017.
[33] Z. Sun, X. Xing, T. Wang, Z. Lv, and B. Yan, ‘‘An optimized clustering
communication protocol based on intelligent computing in information- ZEYU SUN received the B.S. degree in computer
centric Internet of Things,’’ IEEE Access, vol. 7, pp. 28238–28249, 2019. science and technology from the Henan University
[34] H. Tao, M. Z. A. Bhuiyan, M. A. Rahman, T. Wang, J. Wu, S. Q. Salih,
of Science and Technology, in 2003, the M.Sc.
Y. Li, and T. Hayajneh, ‘‘TrustData: Trustworthy and secured data col-
degree from Lanzhou University, in 2010, and
lection for event detection in industrial cyber-physical system,’’ IEEE
Trans Ind. Informat., vol. 16, no. 5, pp. 3311–3321, May 2020, doi:
the Ph.D. degree from Xi’an Jiaotong Univer-
10.1109/TII.2019.2950192. sity, in 2017. He is currently an Assistant Pro-
[35] Z. Sun, Y. Zhang, X. Xing, H. Song, H. Wang, and Y. Cao, ‘‘EBKCCA: fessor with the School of Computer and Infor-
A novel energy balanced k-coverage control algorithm based on probability mation Engineering, Luoyang Institute of Science
model in wireless sensor networks,’’ KSII Trans. Internet Inf. Syst., vol. 10, and Technology, Luoyang, Henan, China. He is
no. 8, pp. 3621–3640, 2016, doi: 10.3837/tiis.2016.08.011. also the External Master Tutor with the School
[36] Y. Wu, H. Huang, N. Wu, Y. Wang, M. Z. Alam Bhuiyan, and T. Wang, of Information Engineering, Henan Institute of Science and Technology,
‘‘An incentive-based protection and recovery strategy for secure big data Xinxiang, Henan. His research interests include wireless sensor networks,
in social networks,’’ Inf. Sci., vol. 508, pp. 79–91, Jan. 2020. mobile computing, and the Internet of Things. He is a member of the China
[37] Z. Sun, Z. Lv, Y. Hou, C. Xu, and B. Yan, ‘‘MR-DFM: A multi-path Computer Federation.
routing algorithm based on data fusion mechanism in sensor networks,’’
Comput. Sci. Inf. Syst., vol. 16, no. 3, pp. 867–890, 2019.
[38] T. Wang, L. Qiu, A. K. Sangaiah, A. Liu, M. Z. A. Bhuiyan, and Y. Ma,
‘‘Edge computing based trustworthy data collection model in the
Internet of Things,’’ IEEE Internet Things J., to be published, doi:
10.1109/JIOT.2020.2966870.
[39] Sun, Li, Wei, Li, Min, and Zhao, ‘‘Intelligent sensor-cloud in fog computer:
A novel hierarchical data job scheduling strategy,’’ Sensors, vol. 19, no. 23, ZHIGUO LV received the B.S. degree in applied
p. 5083, Nov. 2019.
electronic technology from Henan Normal Univer-
[40] H. Teng, Y. Liu, A. Liu, N. N. Xiong, Z. Cai, T. Wang, and X. Liu,
sity, in 2000, the M.S. degree in communication
‘‘A novel code data dissemination scheme for Internet of Things through
mobile vehicle of smart cities,’’ Future Gener. Comput. Syst., vol. 94,
and information system from the Guilin University
pp. 351–367, May 2019. of Electronic Technology, in 2008, and the Ph.D.
[41] L. Yunchang, L. Chunlin, L. Youlong, S. Yanling, and Z. Jing, ‘‘Schedul- degree from Xidian University, in 2019. He is
ing multimedia services in cloud computing envieronment,’’ Enterprise Inf. currently a Lecturer with the School of Computer
Syst., vol. 12, no. 2, pp. 218–235, 2018. and Information Engineering, Luoyang Institute
[42] J. Yue, M. Xiao, and Z. Pang, ‘‘Distributed fog computing based on of Science and Technology. His research interests
batched sparse codes for industrial control,’’ IEEE Trans Ind. Informat., include compressive sensing, wireless sensor net-
vol. 14, no. 10, pp. 4683–4691, Oct. 2018. works, and MIMO systems.

42152 VOLUME 8, 2020

Z. Sun et al.: SCC-DSO Algorithm

HUIHUI WANG (Senior Member, IEEE) received FUQIAN JIA received the B.S. degree in Inter-
the Ph.D. degree in electrical engineering from net of Things project from the Henan Institute
The University of Virginia, Charlottesville, VA, of Science and Technology, in 2018, where he is
USA, in August 2013. In 2011, she was currently pursuing the M.Sc. degree. His research
an Engineering Intern with Qualcomm Inc. interests include wireless sensor networks, artifi-
In August 2013, she joined the Department cial intelligence, and intelligent control.
of Engineering, Jacksonville University, Jack-
sonville, FL, USA, where she is currently an
Associate Professor and the Founding Chair of
the Department of Engineering. She is the author
of more than 50 articles and holds one U.S. patent. Her research interests
include cyber-physical systems, the Internet of Things, healthcare and
medical engineering based on smart materials, robotics, haptics based on
smart materials/structures, ionic polymer metallic composites, and micro-
electromechanical systems.

ZHIXIAN LI received the M.Sc. degree from CHUNXIAO LAI received the B.S. degree in
the Department of Information and Control Engi- computer science and technology from Anyang
neering, Xi’an University of Architecture and Normal University, in 2013. He is currently pur-
Technology, in 2005. He is currently an Associate suing the M.Sc. degree with the Henan Institute
Professor with the School of Computer and Infor- of Science and Technology. His research interests
mation Engineering, Luoyang Institute of Sci- include wireless sensor networks and information
ence and Technology, Luoyang, Henan, China. technology.
His research interests include wireless sensor net-
works, virtual reality, and the Internet of Things.

VOLUME 8, 2020 42153