Future Generation Computer Systems 115 (2021) 671–679

Contents lists available at ScienceDirect

Future Generation Computer Systems

journal homepage: www.elsevier.com/locate/fgcs

Information flow perception modeling and optimization of Internet of

Things for cloud services
∗
Jian Liu a,b , Yuanmin Duan a,b , Yuedong Wu a,b , Rui Chen c , , Liang Chen a,b , Geng Chen a,b
a
Key Laboratory of Ministry of Education for Geomechanics and Embankment Engineering, Hohai University, Nanjing 210098, China
b
Geotechnical Engineering Research Center of Jiangsu Province, Nanjing 210098, China
c
Department of Civil and Environmental Engineering, Harbin Institute of Technology, Shenzhen 518055, China

article info a b s t r a c t

Article history: IoT (Internet of Things) information flow perception is the foundation of the IoT architecture and plays
Received 26 November 2019 an important role. Providing cloud services on demand is one of the key issues and core features of
Received in revised form 28 August 2020 cloud services in the IoT information flow perception. Firstly, we study the mixed logic dynamics (MLD)
Accepted 12 October 2020
modeling method of IoT information flow awareness, and use IoT sensor nodes, controlled nodes, and
Available online 14 October 2020
coordination nodes to describe system application scenarios, and use automata for internal information
Keywords: transmission. Secondly, an open queuing method based on large-scale information flow perception
Mixed Logic Dynamic(MLD) modeling and network delay analysis method is proposed. By analyzing the end-to-end delay of the
Information flow perception node path and the average delay analysis of the whole queuing network, the open queuing is obtained.
Open queuing Aiming at the characteristics of distributed data of information flow perception in the IoT, a priority-
Modeling
based queuing network is proposed to model and analyze the aggregation nodes based on embedded
multi-core SoC(System on Chip), which greatly improves the performance of embedded multi-core
SoC.
© 2020 Elsevier B.V. All rights reserved.

1. Introduction to study the cache allocation method that can obtain the best per-
formance, design efficient IoT convergence bamboo points, and
The cloud-oriented IoT is responsible for hosting server appli- effectively reduce the data set congestion problem; to improve
cations in many IoT products, so it needs to be able to withstand and optimize traditional node location methods and to provide
and process a large number of concurrent network messages, more efficient network service performance.
which place high demands on the system architecture and soft- The IoT cloud system provides a cloud computing service of
ware performance of the cloud platform. The cloud-oriented IoT the type of platform as a service (PaaS, Platform as a Service).
provides a platform-as-a-service cloud computing service for IoT The cloud platform system deploys a sufficient number of servers
applications. At the hardware level, the cloud platform provides in the hardware to carry various services in the system, de-
the necessary server resources and network resources for the ploys all kinds of necessary services of the IoT application server
IoT cloud computing service. At the software level, the cloud in software, and implements a complete cloud system of the
platform is based on the popular IoT cloud service solution, which IoT application [1–4]. IoT application product planners produce
terminal products and applications according to the logic and
implements basic functions including application layer message
protocol standards stipulated by the cloud platform, and access
processing, data storage and management, and load balancing,
the cloud platform in various ways, ultimately enabling the IoT
and is packaged and provided to access application developers.
products to be practically applied. In recent years, some research
In recent years, people have done a lot of work in improving the
has been made on the modeling of large-scale IoT information
service quality of large-scale network in the IoT and enhancing
flow perception nodes. At present, typical modeling methods
the system’s emergency response capability. These studies are
include Petri net modeling method and queuing theory modeling
mainly reflected in the layout and path selection of a pair of
method. Petri net modeling method has been successfully applied
large-scale information flow perception modeling. The matter is
in current wireless sensor networks. For example, a non-Markov
to study effective modeling and simulation evaluation methods; stochastic Petri net model [5–8] is proposed which can evaluate
to study the different types of sensory node modeling problems; star, tree and grid. A method for predicting node energy using
the Petri net model [9–12], thereby determining the node that the
∗ Corresponding author. next hop should pass, which can be used to save the energy of the
E-mail address: [email protected] (R. Chen). node. The use of Petri nets to model wireless sensor networks to

https://doi.org/10.1016/j.future.2020.10.012
0167-739X/© 2020 Elsevier B.V. All rights reserved.
J. Liu, Y. Duan, Y. Wu et al. Future Generation Computer Systems 115 (2021) 671–679

extend the life cycle of nodes [13–15] has effectively improved

the performance of wireless sensor networks. These methods
have solved the problem of optimizing the network structure
modeling in specific small-scale network environments. It is only
for small-scale node modeling and analysis. For the perception
nodes in the large IoT, data analysis and process reasoning be-
come very complicated. People have made some progress in
applying queuing networks to the IoT technology and information
flow perception modeling. A queuing model for analyzing wire-
less sensor network resource replication strategies to maximize
network transmission rate or ensure that query failures do not
exceed a predetermined threshold is proposed [16–19]. However,
these documents do not discuss the packet buffer capacity of the
human-scale perception nodes in the IoT and how to configure
them to optimize node performance. The M/G/1 queuing model
Fig. 1. Three hierarchical relationship diagrams of Cloud IoT information flow
conducts an analysis to determine the processing of various types behavior modeling.
of queries, and designs the APTEEN (Adaptive Sensitive Periodic
Energy Efficient Sensor Network Protocol) protocol, effectively
alleviating the network generated delay [20–26]. According to the
2.1. IoT perception layer application scenario abstraction mecha-
queuing mechanism, the query processing is effectively scheduled nism based on hierarchical structure
to avoid the problem of excessive local load. Due to the com-
plexity of the actual IoT perception layer and the large amount of The general model modeling analysis based on MLD IoT per-
mathematical planning in the MLD modeling method itself, it is ception behavior includes three aspects: the application scene
difficult to obtain results quickly in most cases, which brings great abstraction mechanism, information flow dynamic interaction be-
difficulties for quantitative analysis. How to simplify the solution havior MLD model description method and formal instance ver-
process of the MLD modeling method, optimize the solution algo- ification method. The object abstraction layer application scene
rithm, and reduce the solution time is another important problem abstraction mechanism is the whole modeling. The analysis of
that needs to be solved when applying the MLD modeling method the application scenario abstraction mechanism is the basis of the
to the perception layer of the Internet of Things. entire modeling analysis. The dynamic interaction behavior de-
Although the research on IoT technology for cloud services has scription is the further deepening result of the scene abstraction.
made some progress, there are still many limitations in large- Formal instance verification is the further deepening of the ap-
scale information flow perception modeling and performance plication scenario abstraction and dynamic behavior description.
Fig. 1 is a three-level relationship diagram of IoT information flow
optimization. The paper adopts queuing network modeling and
behavior modeling.
performance analysis methods, and will conduct new research
As shown in Fig. 1, the basic requirements for implementing
and method exploration from the following aspects. This pa-
the behavioral modeling of the IoT are: (1) model can visual-
per uses the automaton to realize the internal information flow ize and standardize the structure of the perception layer and
operation mechanism of IoT information flow perception, and ab- dynamic behavior; (2) the model can clearly express the over-
stractly realizes the clear expression of the overall structure of IoT all structure; (3) model can reflect the object with the multi-
perception in multi-class terminals, multi-task and multi-mode, terminal, multi-tasking and multi-mode features of the network
and the visualization of perceptual structure and dynamic behav- perception layer. IoT information flow perception has multiple
ior. Based on the open queuing method of large-scale information types of terminals, multiple tasks, and multiple modes. In differ-
flow perception modeling and network delay analysis method, ent application scenarios, the composition, communication mode,
the multi-path selection mechanism is determined. Finally, the and scheduling decisions of the IoT perception measurement and
prioritized queuing network is used to model and analyze the control system are different. How to abstract the application
aggregate nodes based on embedded multi-core SoC. scenario into a unified system model is the primary problem for
MLD modeling.

2. MLD behavior modeling method for IoT information flow 2.2. IoT perception MLD behavior model
perception
Assuming that the perception node, the controlled node, and
This chapter will adopt the hybrid logic dynamic MLD model- the coordinator sensed by the IoT information flow have rela-
ing method, start from the basic structure of the IoT information tively independent information decision-making capabilities, the
flow perception and control system, analyze the behavior and information flow of the perceptual measurement and control sys-
tem can be represented as the information flow diagram shown
control behavior of the IoT, through the application scene ab-
in Fig. 2.
straction mechanism, the information flow dynamic interaction
As shown in Fig. 2, an information collection task on the
behavior MLD model description method, and formal instance
perception node is performed to acquire information of a certain
verification. A scenario is a system function that describes how measurement and control object firstly. After the node decision,
the various components of the system interact to perform the the necessary sampling information is wired (shown by a solid
operations of interest to the user [27–31]. It can also be described line in the figure) or wireless (shown by a broken line in the
as a series of events that may occur in the system [32–36]. The figure). The information can be transmitted to the coordinator
method is based on three aspects and is devoted to constructing a node, when the specific information of multiple perception nodes
unified model framework and testing the feasibility of the model is needed in the specific system; the multi-perception nodes are
through examples. aggregated for local decision making (cluster decision), and then
672
J. Liu, Y. Duan, Y. Wu et al. Future Generation Computer Systems 115 (2021) 671–679

Fig. 2. IoT information flow perception modeling flow chart.

Table 1
Status and information event table in the perception node model.
State name Identification Information event Identification
Perception node n1 Environmental χi1 Fig. 3. IoT information flow continuous and discrete event description and MLD
detection information modeling.
Pretreatment n2 Perception node χi2
observation data
Node decision n3 Preprocessed data χi3
Data is sent to n4 Data transmission χi4 between the linear function and the logical variable into a mixed
the coordinator after decision integer inequality form. Finally, the continuous and discrete parts
of the system are integrated into a unified framework to describe
the whole system as a mixed logic dynamic model.
If the state event of the perception node energy continuously
sent to the coordinator for total decision, when only a single sense
decreasing in time is regarded as a continuous event, the per-
is needed. When the node information is measured, it is directly
sent to the coordinator. Secondly, the coordinator performs in- ceptual overall measurement, the control behavior logic state
formation processing and scheduling decision tasks, analyzes the transition event, the perception node internal judgment decision
acquired perception information, and formulates corresponding event, and the controlled component execution decision event
control commands according to the system control strategy, and can be regarded as discrete events. Then, the information flow
transmits instructions to the network controller node. Finally, the diagram of the IoT information flow perception and control sys-
controlled node interprets the control command and implements tem is set such that the continuous or discrete state set at a
a complete measurement control process by implementing the certain time t is x(t), and the continuous or discrete input set
received control command after the node decides. that causes the change of the perceived state is u(t), and the
The following describes the internal information flow mecha- perception and control system can be transformed into the IoT
nism of the perception node using the automaton method [37]. information flow continuous and discrete event description. Its
The ith detection time of the perception node is ti ; the internal MLD modeling method diagram is shown in Fig. 3.
operation period of the perception node is Ta ; the maximum As shown in Fig. 3, the information flow perception acquires
detection time is max(Ta ), and the related information event of the measurement and control object information through the
the ith operation is represented as a binary group (aix , ti ), and perception node; the coordinator node formulates the schedul-
the data detection state g1 is experienced. The data preprocessing ing decision strategy according to the measurement and control
state is g2 ; the local decision state is g3 and the data transmission information, and the controlled node performs the correspond-
state isg4 . Then the node internal information flow operation ing decision according to the coordinator. The physical meaning
mechanism can be represented as the automatic node model expressed by the continuous and discrete event description is
diagram, and the sense of Table 1 is the meaning and identity expressed in (1) the overall behavior presents information acqui-
of the corresponding state and information events in the node sition, scheduling decision, decision-making three-state cyclical
model. change, that is, jump-discrete state transition; (2) on the other
It should be noted that the above perception node description hand, the overall behavior at the microscopic level shows con-
is for a node with intelligent information processing and decision tinuous evolution intermittently, alternating between the two
function, and does not have an intelligent function node for evolution, and the formation of intricate dynamic processes.
practical applications, and only needs to simplify the model. (1) Establishment of a continuous dynamic state space mech-
MLD modeling is a unified model to describe the continuous anism model for perception and control behavior. The process of
part and discrete parts of the system to the system evolution, and perceptual measurement and control involves various continuous
is the basis of system scheduling optimization research. Firstly, events, discrete events and their mixed events. It is difficult to
the model analyzes the continuous events and discrete events of express with general differential mathematical equations. It is as-
the IoT perception layer, and then establishes the different op- sumed that the overall behavior of the perceptual layer presents
erational modes of the system corresponding to the system state a nonlinear transition state of information acquisition, scheduling
space mathematical model, and sets and define different auxiliary decision, and execution decision three-state cycle. The state at
logic variables of the system, and builds the propositional logic a certain moment is y(t), and the state at the next moment is
related to the constraints and convert them into integer inequal- y(t+1), resulting in the change of the state of the perceived state
ities. The auxiliary variable transforms the product relationship is as v(t); the state of information acquisition is f1 (y(t), v (t)); the
673
J. Liu, Y. Duan, Y. Wu et al. Future Generation Computer Systems 115 (2021) 671–679

state of the scheduling decision is f2 (y(t), v (t)), and the execution 3. Information flow perception modeling and evaluation
decision is f3 (y(t), v (t)). In some state, the constraint set is Yi ; the method based on queuing network
three-state constraint set is Y = (Y1 , Y2 , Y3 ), and the whole mea-
surement and control behavior are based on different constraints Under the limited hardware consumption, maximizing the
and are transformed between the three-state models. Then the performance of the perceived network in the IoT has become
mathematical model of the perception layer measurement and a research hotspot for improving the perceived network QoS
control behavior represented by the piecewise function is: in recent years. In large-scale wireless sensor networks, node
f [y(t), v (t)] packet buffer size is the main factor affecting node performance
⎧
⎨1
⎪
and link blocking. This paper proposes an approach to model
y(t + 1) = f2 [y(t), v (t)] (1) the IoT information flow perception by using an open queuing
⎪
f3 [y(t), v (t)] network, so that it is in the best working condition and analyzes
⎩
its data packet buffer. According to the topology and operational
This is a macro description of the flow-perception measure- characteristics of the IoT information flow, the arrival, transfer
ment and control behavior. In fact, each state also contains many
and departure relationships of the transport nodes, boundary
sub-process behaviors. Let Xi be a subset of X, it internally contains
nodes and sink nodes are analyzed, and the data flow balance
different sub-process behavior constraints in this state. Let the
equation with congestion is obtained, and the interception of the
information flow sense that each subset Xi (i = 1, 2, 3) in the
IoT information flow is established to queue the network model.
constraint set Xi in each state satisfies:
In order to evaluate the congestion situation in the network, the
Yi ∩ Yj = ∅, ∀i ̸ = j
{
actual effective arrival rate and the transfer rate in the model are
(2) obtained, and the maintenance node is added in the queuing net-
∪Y j i = Y
i=1 work model to obtain an extended equivalent queuing network
Let each subset Xi be represented as a x, u-constrained disjoint model.
convex polyhedron, the system state x(t} and the input u (t} space
convex polyhedron set constrained set T, and the correlation coef- 3.1. Packet queuing relationship
ficient of the x(t) constraint in the polyhedron is S; the correlation
coefficient of the u(t) constraint in the polyhedron is R, namely: The IoT information flow perception applied to environmental
{[ ]
y
} monitoring and control is not always processed due to factors
Yi = : Si (y(t)) + R, v (t) ≤ Ti (3) such as energy consumption. Under the premise of meeting the
v real-time requirements of the monitored objects, the IoT infor-
mation flow monitoring and sleep interval are carried out. The
For the state Xi , the logical variable δi ∈ 0, 1, i = 1, 2, 3 is
longest time elapsed before entering another sleep state in all
introduced, and the system is in the state Xi , δi = 1. Since each
state corresponding to each logical variable does not intersect, nodes sensed by the IoT traffic is called the longest monitoring
there are: period of the wireless sensor network node.

i = max {L1 , L2 ....Li . . . LN }

Lm
j
[[ ] ]
∑ y (8)
φi (t) = 1, [ϕi = 1] → ∈ Yi (4)
i=1
v In the IoT information flow perception, each node has a certain
packet capacity. If the cache capacity is set too much, it will waste
S y(t) + Ri v (t) ≤ Ti1
⎧ 1 1
⎨ i
⎪ resources. On the contrary, if it is too small, the system will be
i v (t) ≤ Ti
[ϕi = 1] → Si21 y(t) + R21 21 blocked. The packet queuing length theorem for the sink node
(5)
⎪ and the transport node is given. For the IoT information flow
i v (t) ≤ Ti
⎩ 31
Si y(t) + R31 31
perception, if the packet length of any aggregation node is Lm i ,
Combined with the MLD propositional logic rule, an equivalent the following relationship is satisfied:
inequality can be derived: m−1
∑ Lm
Si y(t) + Ri v (t) ∗ Ti ≤ Mi∗ (1 − ϕi (t)), i = 1, 2, 3 . . . (6) βie + βk pkj = i
+ (1 − pii )vi (9)
Tim
i=1
In order to make the flow perception of continuous events and
logical events in the behavioral space model, it is represented by In the IoT information flow perception model, the size of
a uniform expression. Based on the previously introduced logic the node packet buffer corresponds to the length of the waiting
variable ϕi (t), the piecewise function model can be expressed as: queue. When the queued queue length reaches the maximum
bit, the packet flow to the node stops, and the packet arriving at
j
∑ this time has not been accepted and processed. At this time, the
y(t + 1) = fi [y(t), v (t)]ϕi (t) (7) queued network is blocked. In other cases, when a node packet
i=1
is transferred from one queue to the next queue that has reached
the maximum length, it will be stuffed in the queue that has just
Through this model expression, the piecewise function of the
completed the service. At this time, the blocked perception node
IoT information flow perception and control process behavior is
cannot process any other data packet until the target node queue
described by including state variables, input variables and aux-
has an idle data packet buffer before releasing the blocking and
iliary variables, and the IoT information flow perceives the con-
tinuous dynamic characteristics, and logic rules and operations continuing the data transmission.
in the process behavior of the measurement and control process. When the holding node is added to the queuing network,
Constraints are integrated into state equations with mixed integer the data packets that have not been served in time will enter
inequality constraints. Under the unified MLD model framework, the holding node for queuing, waiting for the service station to
the system process behavior can be grasped macroscopically, vacate, and the queuing network model with blocking probability.
and factors such as system operation constraints and qualitative Where A represents the set of all the holding nodes, pbij repre-
knowledge are considered, which lays an important foundation sents the blocking probability that the service station ends the
for the coordinated optimization of measurement and control transfer of the service to node j at node i, pbaj represents the total
systems. blocking probability of reaching node j, and pbej represents the
674
J. Liu, Y. Duan, Y. Wu et al. Future Generation Computer Systems 115 (2021) 671–679

Table 2 Table 3
IoT information flow perception modeling performance index solving algorithm Priority-based scheduling algorithm based on multiple information flows
(algorithm 1). (Algorithm 2).
Input: j, Q represents the queuing network; Input: T represents the task input;
lp
Output: βi is the arrival rate; Output: max(N h );
Step 1: j → Q ; Step 1: If T, go to step 2, otherwise go to step 3;
− Step 2: Determine T→ buffer queue;
βje = βje (1 − pbej ) buffer queue = buffer queue+1;
Step 2: pbej → j Step 2.1: While T is not the maximum number, continue, otherwise go to
−
step 3;
β e = βje ÷j = βje (1 − pbej )/j
T = T+1;
Set the loop j = j + 1
End;
Step 3:Compute the probability that queue is in steady state ni
Step 2.2: Computer max(nij );
pni (1 − p)
pi (ni ) = Step 3: T→ cache queue Hj ;
1 − pNc +1 Step 3.1: If Hj is empty, go to step 1, otherwise go to step 3.2.
ni (1, 2..Nc ) represents the number of packet buffers per node;
Step 3.2: If the length of Hj is max(N h )
Step 4: Calculate the blocking probability βij = βi pij (1 − pij );
then T→ buffer queue;
Step 5: Ioop–, if Ioop<0, go to step 7;
the length of Hj = max(N h ) − 1;
Step 6: Compute the difference
else max(N h ) = the length of Hj +1;
pni (1 − p) e − e
= pi (ni ) = -β = βj ÷j = βje (1 − pbej )/j; end if;
1 − pNc +1 Step 4: Output max(N h ).
If difference<a certain value, go to step 8, otherwise go to step 3 and
proceed;
Step 7: If difference<0, then refine λ0j , go to step 3; Table 4
lp
Step 8: Return λi . Calculating the Equivalent Model Algorithm (Algorithm 3).
Input: j, A represents embedded multicore SoC queuing network;
Output: pbm ;
Step 1: j→A;
blocking probability of the external arriving node j. For node j, −
Step 2: Enter βje = βje (1 − pbej );
the total arrival rate is equal to the sum of the external arrival
Set the total blocking probability of the node,
rate λej and the internal node arrival rate blocking. The external pb m = 0.2, where m = 1, 2, . . . , n
effective arrival rate i1 , i2 . . . iq of the node j is: Step 3: Calculate the arrival and departure rates;
−

β¯je = βje (1 − pbej ) (10) pb m = 0.2*βje = βje (1 − pbej );

−
Step 4: The blocking probability = pbm ∗ pbm−1 .... ∗ βje = βje (1 − pbej );
The actual effective data flow arrival rate from node i to node Loop = loop–;
j is: Step 4.1: Obtain the node blocking probability;
Node blocking probability = pbm + βje ;
βij = βi pij (1 − pij ) (11) Step 4.2: Calculate the effective utilization rate of the node;
Effective utilization rate = pbm ∗ βje ;
It is convenient to describe the arrival of nodes in the equiv-
Step 4.3: Recalculate the total blocking probability of the node.
alence queuing network model by adding the infinite capacity of pbm = pbm + pbm−1 . . . + pb1 ;
the retention node in the queuing network model to release the Loop–;
congestion of the service desk, and to correspond to the blocked If loop<0 is established, go to step 5.
time when the processing of the service station is maintained in Step 5: output pbm = pbm + pbm−1 . . . + pb1 .
the node. An approximate iterative algorithm is applied to the
obtained queuing network model to solve the total arrival rate of
each node. Next, the digital result of the algorithm is solved for between nodes, adding a limited number of holding nodes to the
the IoT information flow perception modeling performance index. queuing network.
The algorithm is briefly described as follows (see Table 2).
The amount of computation of the algorithm is mainly re- j → M/M 1/N (12)
flected in the loop from step 3 to step 6. In addition to the
In embedded multi-core SoC design, scheduling strategy is
program cycle screwing, and determining that the difference be-
an important indicator that affects its performance. Designing
tween the average position and the internal arrival time interval
scheduling policies based on task priorities can effectively im-
of the two calculation nodes is less than a certain value, the loop
prove system operation efficiency. Each execution core has an
control variable loop is also used for control. This is mainly to
consider that if the initial value of the blocking probability of independent public priority buffer. The priority-based scheduling
each node is unreasonable, it may cause the iterative process to algorithm based on multiple information flows is (see Table 3):
diverge, which can make the program exit normally.
3.3. Queuing evaluation algorithm
3.2. IoT information flow perception modeling and scheduling strat-
egy based on prioritized queuing network Next, this chapter establishes a queuing network model for
embedded multi-core SoC aggregation nodes, and analyzes its
In an embedded multi-core SoC, a task with priority can be external arrival rate, service rate and transition probability. Ac-
processed by multiple execution cores, and the priority of this cording to the design model evaluation algorithm, the mainte-
task may be different in different execution cores. The embed- nance node is added, and the queued network model with priority
ded multi-core SoC task flow arrival time interval approximates is extended to obtain an equivalent queuing network. Then, the
the Poisson distribution, and the processing time of each ex- equivalent model is calculated to obtain the node utilization and
ecution core obeys the exponential distribution, and each ex- blocking probability of each priority (see Table 4).
ecution core acts as a node. Therefore, the embedded multi- It can be seen from Algorithm 3 that the calculation amount
core SoC can be built into a queuing network model, and the in a single cycle mainly focuses on the calculation of the different
situation of each node is described by M/M/1/N type queues. priority arrival rates of each node and the calculation of the
Reference [30,31] proposes a solution to the task flow blocking blocking probability. In addition, it has a great relationship with
675
J. Liu, Y. Duan, Y. Wu et al. Future Generation Computer Systems 115 (2021) 671–679

Table 5
Model input parameters.
i CAai ui CSi NCi
1 12 42 15 1-31
2 13 41 17 1-31
3 13 42 17 1-31
4 11 41 12 1-31
5 6 31 11 1-31
6 9 31 11 1-31
7 15 92 31 1-31

Fig. 5. Boundary node utilization and packet buffer curve.

Fig. 4. Perceptual node utilization and packet buffer curve.

the number of iterations of the algorithm. Different initial value

settings will have a great impact on the convergence speed of the
algorithm. In the program loop control, the loop control variable
loop is also used for control in addition to the judgment is less
than a certain value. The main consideration here is that if the
initial value of the total blocking probability of the node is set
unreasonably, it may cause the iteration process to diverge, and Fig. 6. Convergence in the perception of utilization and packet cache
the program can be normally exited. relationship curve.

4. Experiments and results

node link is serious, resulting in lower node utilization. Next, as
4.1. Experimental parameter setting the node cache capacity increases, the curve is basically parallel
to the horizontal axis. This point can be used as the setting value
In the IoT information flow-perception network of cloud ser- of the cache size.
vices, the setting of the packet buffer capacity of the node is an As shown in Fig. 6, as the node packet buffer capacity in-
important factor affecting the efficiency of the whole network. creases, the curve of the starting node rises faster. It can be seen
If the packet buffer capacity is too small, some links in the that when the buffer capacity is small, the utilization rate of the
system will be seriously blocked, and the data processing and sink node is relatively low, and then the curve is basically parallel
transmission efficiency will be low. to the horizontal axis, as the node cache capacity increases. Here,
In the experiment, according to the IoT information flow per- the curve has an inflection point. This point can be used as a
ception data statistics actually used for temperature monitoring, reference value for the aggregated sense cache size to optimize
the input parameters of the queuing network model are shown the utilization of the node.
in Table 5, where CAai represents the external arrival variance; ui A queuing network model is established for the embedded
represents the service desk processing rate, and CSi represents the multi-core SoC, and the utilization ratio of each execution core
service desk processes the variance, and NCi indicates the task under different priorities is calculated by the algorithm 2, and
cache size, which ranges from 1 to 30. the curve between the node utilization rate and the task cache
As shown in Fig. 4, in the IoT information flow perception capacity is obtained as follows:
node, there are node 1, node 2, node 3 and node 4 bit trans- As shown in Fig. 7, each curve undergoes a rapid or gradual
mission nodes, and their role in the IoT is mainly to collect data change process, eventually tending to balance, and node uti-
and transmit it to other nodes, along with the cache capacity the lization is gradually increased, indicating the IoT traffic-aware
increase of node 1 is faster, and the other three stages are slower. internal blocking condition also changes as the task cache ca-
As shown in Fig. 5, as the node cache capacity increases, the pacity changes, as the number of task buffers allocated to each
two curves tend to be steeper in the initial stage, the node 5 rises execution core increases.
faster than node 6, according to this situation. When the packet Fig. 8 Shows the delay of the service packet scheduling pro-
buffer capacity setting of the node 5 is small, the blocking on the cessing for different scheduling algorithms. The horizontal axis
676
J. Liu, Y. Duan, Y. Wu et al. Future Generation Computer Systems 115 (2021) 671–679

Fig. 9. Comparison of average delays of packet scheduling processing under

Fig. 7. Relationship between utilization and task cache. different algorithms.

Fig. 10. Comparison of data transmission acquisition under different scheduling

of sensor nodes.

As shown in Fig. 9, the average delay of packet schedul-

ing processing is about 0.02s–0.5s. In the IoT information flow
perception model, the dynamic adjustment of the probability
partition of the service queue evaluation index is realized, which
makes the interaction between queues effectively reduce. The
queue service can select high priority queues with high proba-
bility, and avoid the problem that low priority queues with short
deadline information cannot get timely service.
In Fig. 10, the node data value range is evenly distributed
between 0 and 2 s, and the node sends once every 0.5 s. If this
method is adopted, the data transmission amount is about 40, and
the mean value of the information loss coefficient (that is, the
data at a certain time under the scheduling policy) is 0.125. The
results show that the IoT information flow perception model is
feasible and effective. When the loss coefficient of the scheduling
information is small, it can ensure that the node transmits data
Fig. 8. Packet scheduling processing delay under different scheduling algorithms. within the deadline and saves energy.
As shown in Fig. 11, the execution rate distribution of each
execution core before and after the adaptive adjustment algo-
represents the simulation time and the vertical axis represents rithm is actually tested. It can be seen that after adding adaptive
the delay time (unit: hour t). adjustment to Core1–Core6 (as shown in the blue histogram in
As shown in Fig. 8, under the condition that the deadline the figure), the task size of the priority assigned in Core2–6 is
of each node satisfies the uniform distribution of 0.1t–1.3t, the significantly higher than that without adaptive adjustment (as
simulation time is 1.4 h. Within 0.2 h of interception, the param- shown in the figure).
eters of each node successfully processed under the algorithm are As shown in Fig. 12, in the experiment, the data packets sent
1.1t/1.2t = 91.7%, 1.3t/1.4t = 92.8%. and received by the node are counted to obtain the true arrival
677
J. Liu, Y. Duan, Y. Wu et al. Future Generation Computer Systems 115 (2021) 671–679

analyzing different types of nodes. You can configure the appro-

priate task cache size separately, so that nodes can get the best
performance with minimal hardware overhead. The embedded
multi-core SoC is used to optimize the cache size of multiple
execution cores according to the characteristics of IoT distributed
data, and a load balancing algorithm is designed to effectively
improve the utilization and load balancing of multiple execution
cores. The paper proposes a method for modeling large-scale
information flow perception nodes in the IoT by using queuing
networks. It can model the general distribution of standard func-
tions such as Poisson arrival and exponential services, but it does
not conform to the distribution of standard functions. The system
should fit the function of arrival and service law, and the follow-
up work of the paper will be studied in the function of special
distribution.

CRediT authorship contribution statement

Fig. 11. Performance evaluation of an embedded multi-core SoC queuing net-
work model. (For interpretation of the references to color in this figure legend,
Jian Liu: Methodology and Writing. Yuanmin Duan: Data
the reader is referred to the web version of this article.)
curation. Yuedong Wu: Formal analysis. Rui Chen: Supervision.
Liang Chen: Validation. Geng Chen: Investigation.

Declaration of competing interest

The authors declare that they have no known competing finan-

cial interests or personal relationships that could have appeared
to influence the work reported in this paper.

Acknowledgments

This research was partly supported by the National Natural

Science Foundation of China (No. 51608152); the Fundamental
Research Funds for the Central Universities (No. 2018B00714);
the Fundamental Research Funds for the Central Universities
(2015B25914); Shenzhen Science and Technology Innovation
Commission (Grant No. JCYJ20170811160740635); National Key
Research and Development Program of China (No.
Fig. 12. Comparison of total node arrival rate.
2017YFC0404803).

References
rate of each node. It can be seen from the figure that there is a
[1] A. Ahmad, M.H. Rehmani, H. Tembine, et al., IEEE access special section
certain difference in the node arrival rate between the simulation editorial: Optimization for emerging wireless networks: IoT, 5G, and smart
data obtained by the queuing software simulation and the real grid communication networks, IEEE Access 5 (2017) 2096–2100.
experimental data obtained from the actual IoT platform. When [2] D. Vu, N. Dao, Y. Jang, et al., Joint energy and latency optimization for
upstream IoT offloading services in fog radio access networks, Trans.
the node inflection point is reached (such as Experiment 3), the
Emerg. Telecommun. Technol. 30 (8) (2019) e3497–e3504.
node arrival rate of the experimental data is significantly higher [3] Y. Zhang, H. Wang, Y. Xie, An intelligent hybrid model for power flow
than the simulation data, which indicates that the IoT queuing optimization in the cloud-IOT electrical distribution network, Cluster
modeling and simulation software designed in this paper has Comput. 2017 (2) (2017) 1–10.
practical application value. [4] C. Mascolo, C. Mascolo, C. Mascolo, et al., Low-resource multi-task audio
perception for mobile and embedded devices via shared deep neural net-
work representations, Proc. Acm Inter. Mob. Wearable Ubiquitous Technol.
5. Conclusion 1 (3) (2017) 50–63.
[5] S. Zhengyang, D. Yucong, W. Shixiang, et al., Processing optimization of
typed resources with synchronized storage and computation adaptation in
This paper studies the IoT aware MLD modeling method in- fog computing, Wirel. Commun. Mob. Comput. 2018 (2) (2018) 1–13.
cluding the IoT information flow application scenario abstraction [6] J. Kim, H. Hwangbo, Sensor-based optimization model for air quality
mechanism and the IoT information flow-aware information dy- improvement in home IoT, Sensors 18 (4) (2018) 959–972.
namic interaction behavior, which realizes the clear expression of [7] D. Amaxilatis, O. Akrivopoulos, G. Mylonas, et al., An IoT-based solution for
monitoring a fleet of educational buildings focusing on energy efficiency,
the Internet of Things-aware overall structure system and system
Sensors 17 (10) (2017) 2296–2312.
process behavior. The whole measurement and control behavior [8] J. Wang, J.M. Li, Z. Ru, et al., Locating hazardous gas leaks in the
of the Internet of Things is transformed into MLD model. Large- atmosphere via modified genetic, MCMC and particle swarm optimization
scale information flow perception modeling and performance algorithms, Atmos. Environ. 157 (2017) 27–37.
optimization method based on queuing network for IoT is pro- [9] X. Xu, M. Chao, Z. Wang, A value-oriented iterative service modeling
process, Lect. Notes Bus. Inf. Process. 122 (2017) 104–116.
posed. The delay type queuing network delay analysis method is
[10] H. El-Sayed, S. Sankar, M. Prasad, et al., Edge of things: The big picture
used to obtain the sensor node. The best path and alternative path on the integration of edge, IoT and the cloud in a distributed computing
for data transfer can be realized by classifying and qualitatively environment, IEEE Access 6 (99) (2018) 1706–1717.

678
J. Liu, Y. Duan, Y. Wu et al. Future Generation Computer Systems 115 (2021) 671–679

[11] N. Jiang, Y. Deng, A. Nallanathan, et al., Reinforcement learning for real- Jian Liu was born in Jiangsu, China PRC in 1986. He
time optimization in NB-IoT networks, IEEE J. Sel. Areas Commun. 37 (6) received his Ph.D. degree in geotechnical engineering
(2019) 1–13. from Hohai University, China in 2015 and worked as
[12] P. Wang, F. Ye, X. Chen, A smart home gateway platform for data collection a PostDoc researcher in the Hong Kong University of
and perceptionness, IEEE Commun. Mag. 56 (9) (2018) 87–93. Science and Technology afterwards. He joined Hohai
University as an associate research in January 2018
[13] L. Liu, C. Zheng, X. Guo, et al., Multiobjective optimization for computation
and associate professor in 2019. He has published 35
offloading in fog computing, IEEE IoT J. 5 (1) (2018) 283–294.
papers and holds 11 patents. His research interests in-
[14] R. Zhang, J. Tao, Data-driven modeling using improved multi-objective clude unsaturated water flow in soil, internet of things
optimization based neural network for coke furnace system, IEEE Trans. in geotechnical engineering and soft soil foundation
Ind. Electron. 64 (4) (2017) 3147–3155. reinforcement technology. Email: [email protected]
[15] C. Yang, D. Puthal, S.P. Mohanty, et al., Big-perception-data curation for the
cloud is coming: A promise of scalable cloud-data-center mitigation for
next-generation IoT and wireless sensor networks, IEEE Consum. Electron.
Mag. 6 (4) (2017) 48–56. Yuanmin Duan was born in Anhui, China PCR in
[16] D. Shin, V. Sharma, J. Kim, et al., Secure and efficient protocol for route 1998. She received her B.E. degree in civil engineering
optimization in PMIPv6-based smart home IoT networks, IEEE Access 5 from Anhui Agricultural University, China in 2019 and
(99) (2017) 11100–11117. Currently studying for master’s degree in geotechnical
[17] L. Jian, L. Yun, Z. Zhang, et al., Towards green IoT networking: Performance engineering in Hohai University. Email: 17356515281@
optimization of network coding based communication and reliable storage, 163.com
IEEE Access 5 (99) (2017) 8780–8791.
[18] S. Guo, S. Shao, W. Yao, et al., Cross stratum resources protection in fog-
computing-based radio over fiber networks for 5G services, Opt. Fiber
Technol., Mater. Devices Syst. 37 (2017) 61–68.
[19] O. Alsaryrah, I. Mashal, T.Y. Chung, Bi-objective optimization for energy
perception IoT service composition, IEEE Access 6 (2018) 26809–26819.
[20] L. Xin, X. Zhang, Rate and energy efficiency improvements for 5G-based Yuedong Wu was born in Fujian, China PRC in 1969. He
IoT with simultaneous transfer, IEEE IoT J. 6 (4) (2019) 5971–5980. received his M.A. degree and Ph.D. degree in geotech-
[21] R. Deng, R. Lu, C. Lai, et al., Optimal workload allocation in fog-cloud nical engineering from Hohai University, China in 1999
computing toward balanced delay and power consumption, IEEE IoT J. 3 and 2008. He joined Hohai University as a full-time
(6) (2017) 1171–1181. lecturer in August 1995, and is currently the deputy di-
[22] L. He, K. Ota, M. Dong, Learning IoT in edge: Deep learning for the IoT rector of Geotechnical Engineering at Hohai University.
with edge computing, IEEE Netw. 32 (1) (2018) 96–101. His research interests include soft foundation treatment
and basic engineering and geotechnical engineering
[23] Y. Lin, J. Yang, Z. Lv, et al., A self-assessment stereo capture model
test technology. E-mail: [email protected]
applicable to the internet of things, Sensors 15 (8) (2015) 20925–20944.
[24] J. Li, G. Feng, W. Wei, et al., PSOTrack: A RFID-based system for ran-
dom moving objects tracking in unconstrained indoor environment, IEEE
Internet Things J. 5 (6) (2018) 4632–4641.
[25] W. Wei, M. Guizani, S.H. Ahmed, et al., Guest editorial: Special section on
Rui Chen was born in Fujian, China PRC in 1978. He
integration of big data and artificial intelligence for internet of things, IEEE
received his Ph.D. degree in civil engineering from
Trans. Ind. Inf. 16 (4) (2020) 2562–2565.
the Hong Kong University of Science and Technology,
[26] W. Wang, F. Xia, H. Nie, et al., Vehicle trajectory clustering based on China in 2007. He received his Bachelor degrees in
dynamic representation learning of internet of vehicles, IEEE Trans. Intell. hydraulic and hydro-Power Engineering from Tsinghua
Transp. Syst. (2020) http://dx.doi.org/10.1109/TITS.2020.2995856. University in 2001. He joined Harbin Institute of Tech-
[27] F.G. Brundu, E. Patti, A. Osello, et al., IoT software infrastructure for energy nology, Shenzhen as an associate professor in 2011 and
management and simulation in smart cities, IEEE Trans. Ind. Inf. 13 (2) professor in 2019. E-mail: [email protected]
(2017) 832–840.
[28] R. Mahmud, K. Ramamohanarao, R. Buyya, Latency-perception application
module management for fog computing environments, Acm Trans. Internet
Technol. 19 (1) (2018) 1–21.
[29] T. Roubicek, G. Tomassetti, Thermomechanics of hydrogen storage in Liang Chen, male, born in Jiangsu Province of PR.
metallic hydrides: modeling and analysis, Discrete Contin. Dyn. Syst. Ser. China in October 1976, Professor of Hohai University of
B 19 (7) (2017) 2313–2333. China. Graduated from Physics Department of Nanjing
[30] A. Alelaiwi, A collaborative resource management for big IoT data University in 1996. Research direction include: seepage
processing in Cloud, Cluster Comput. 20 (5) (2017) 1–9. theory and leakage test; seepage deformation mecha-
[31] G. Liang, J. Qiu, Optimization technology in cloud manufacturing, Int. J. nism; solute transport theory and solid waste landfill
Adv. Manuf. Technol. 97 (2) (2018) 1–13. underground pollution test. Email: [email protected].
[32] C. Tang, X. Wei, S. Xiao, et al., A mobile cloud based scheduling strategy cn
for industrial IoT, IEEE Access 6 (99) (2018) 7262–7275.
[33] L. Deng, D. Li, Z. Cai, et al., Smart IoT information transmission and security
optimization model based on chaotic neural computing, Neural Comput.
Appl. 2019 (2) (2019) 1–14.
[34] M.C. Chen, S.Q. Lu, Q.L. Liu, Uniform regularity for a Keller–Segel–Navier– Geng Chen was born in Shandong, China PRC in
Stokes system, Appl. Math. Lett. 107 (2020) 106476. 1987. He received his Ph.D. degree in geotechnical
[35] M.C. Chen, S.Q. Lu, Q.L. Liu, Global regularity for a 2D model of electro- engineering from Hohai University, China in 2015 and
kinetic fluid in a bounded domain, Acta Math. Appl. Sin. Engl. Ser. 34 (2) joined Hohai University April 2016. He has published
(2018) 398–403. 19 papers and holds 15 patents. His research inter-
ests include ultra-soft soil consolidation, foundation
[36] A. Quilliot, H. Toussaint, Prospective network flow models and algorithms
treatment and environmental geotechnical engineering.
for bin packing problems, Electron. Notes Discrete Math. 36 (2018) 49–56.
Email: [email protected]
[37] D. Shuang, H. Xin, J. Wang, Multiobjective optimization model for service
node selection based on a tradeoff between quality of service and resource
consumption in mobile crowd perception, IEEE IoT J. 4 (1) (2017) 258–268.

679