Journal of Industrial Information Integration 33 (2023) 100464

Contents lists available at ScienceDirect

Journal of Industrial Information Integration

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

Full length article

Assessing the traffic scheduling method for time-sensitive networking (TSN)

by practical implementation
Junhui Jiang a ,1 , Yuting Li b ,1 , Xiongfeng Zhang c , Mengmeng Yu d , Chang Dae Lee e ,
Seung Ho Hong c,d ,∗
a
Modeling and Emulation in E-Government National Engineering Laboratory, College of Computer Science and Technology, Harbin Engineering
University, Harbin 150001, China
b School of Computer Science, Hubei University of Technology, Wuhan 430068, China
c Department of Electronic Engineering, Hanyang University, Ansan 15588, South Korea
d
Research and Development Center, Nestfield Company Ltd., Ansan 15588, South Korea
e
Research Center of Electrical and Information Technology, Seoul National University of Science and Technology, Seoul 01811, South Korea

ARTICLE INFO ABSTRACT

Keywords: Time-sensitive networking (TSN), as standardized and maintained by the IEEE 802.1 Task Group, enhances the
TSN real-time and deterministic capabilities of the Ethernet. However, traffic scheduling is not standardized, and
Real-time communication is being extensively researched. Most studies are theoretical in nature; practical validation studies are rare.
Traffic scheduling
To fill this gap, we first constructed a real-world TSN-based process automation system and presented fine-
Industrial applications
grained guidelines regarding how a traffic scheduling method (TSM) can be deployed in industrial facilities.
Practical implementation
We experimentally investigated the feasibility of our method, and compared its performance to that of a
commercial TSN scheduler. The results show that our method precisely schedules traffic to satisfy firm real-
time requirements and achieves ultra-low latency by eliminating the queuing delays. The scheduling calculation
times for 500 random cases show that our TSM can appropriately schedule flows in milliseconds, more than
571-fold faster than the commercial scheduler.

1. Introduction Things. TSN is not confined to industrial automation; it also plays piv-
otal roles in many other areas that demand real-time communication,
The Time-sensitive networking (TSN) is emerging as the principal such as in-vehicle networking [10], aerospace [11], etc.
technology for the next era of industrial communication [1]. TSN Earlier reports explored TSN in terms of time synchronization [12–
exploits the excellent bandwidth and low cost of the Ethernet [2] to 14], simulation [15,16], resource management [17–19], and traffic
compensate for real-time and deterministic capabilities using a series scheduling [20–34]. In [12], a TSN-based time synchronization method
of enhanced standards. The IEEE 802.1 TSN Task Group [3] devel- was presented that guarantees highly accurate clock synchronization
ops and maintains these standards; examples include high-precision across EtherCAT nodes. To extend TSN to wireless networks, a novel,
clock synchronization [4], bounded latency [5], ultra-reliability [6] high-precision, wireless time synchronization strategy was developed
and flexible resource management [7]. TSN provides better standard [13]. In [14], the clock synchronization accuracy of IEEE 802.1AS [4]
compliance and vendor independence, and compatibility, than other was tested by deploying it in a practical industrial system. The major
Ethernet-based real-time communication technologies (e.g., EtherCAT, factors (PHY jitter and clock granularity) that affect clock accuracy
PROFINET, and Sercos III, etc.). Due to the lower cost and ease of were analyzed. Jiang et al. [15] modeled the IEEE 802.1 Qbv [5] using
use, TSN networks can be readily scaled and integrated with other an OMNET++ simulator. They also further validated the simulation in
high-layer communication technologies (e.g., OPC UA [8]). Therefore, a real-world TSN testbed [16]. An OPC UA TSN configuration architec-
TSN plays key roles in the implementation of various industrial ap- ture was proposed in [17] and validated in a laboratory-level manu-
plications [9], and also contributes to the integration of information facturing system. In [18], an SDN-based self-configuration framework
technology and operational technology in the industrial Internet of was presented to automatically configure TSN networks. Simulations

∗ Corresponding author.
E-mail address: [email protected] (S.H. Hong).
1
Co-first author.

https://doi.org/10.1016/j.jii.2023.100464
Received 20 December 2021; Received in revised form 22 January 2023; Accepted 9 April 2023
Available online 14 April 2023
2452-414X/© 2023 Elsevier Inc. All rights reserved.
J. Jiang et al. Journal of Industrial Information Integration 33 (2023) 100464

indicated that TSN networks with different topologies were appropri- increased, runtime increased dramatically, complicating practical im-
ately configured. To improve the stability of industrial cyber–physical plementation in industrial systems. Different from optimization theory,
systems, Lu et al. [19] proposed a multi-hop TSN-integrated control and Hong et al. [35] invented a traffic scheduling method (TSM) using the
transmission architecture and a delay-based joint design strategy. bandwidth allocation concept from [36]. Most previous studies focused
Various TSN scheduling mechanisms have been released, such as on the theory of TSN scheduling via simulation; no scheduling algo-
gating-based Time-aware Shaper (TAS), and Cyclic Queuing and For- rithm has yet been deployed in a real-world manufacturing scenario.
warding (CQF). The TSN task group recently proposed CQF model, Very few studies [37–39] validated the theoretical approach by practi-
and several studies were conducted. Zhang et al. [26] designed a cal implementation. Farzaneh et al. [37] performed a proof-of-concept
CQF-oriented scheduling algorithm with flow judgment conditions and experiment when evaluating the performance of IEEE 802.1Qbv [5].
position diversity (PD)-based variable search bound (FLJ-VB). FLJ-VB In [38], the off-the-shelf TSN devices available from different ven-
is based on divisibility theory to reduce scheduling computation time dors were summarized. Craciunas et al. [39] used a Z3 SMT solver
and enhance load balancing. Guo et al. [30] proposed a mapping score- as the core component of Slate XNS software for scheduling. Thus,
based scheduling algorithm (MSS) by integrating flow sorting and offset these studies performed practical TSN research. However, [37,38] were
searching. MSS enhances scheduling performance and time efficiency. concerned only with proof of the TSN standards and a description of the
However, zero-jitter cannot be guaranteed by the CQF model, be- hardware solutions. Although [39] validated algorithms using an ex-
cause the Ping-Pong queue-based model introduces jitters every two cy- perimental testbed, neither practical industrial applications nor device
cles. Ultra-low jitter is a mandatory requirement for industrial automa- diversity were considered. Specifically, only end stations transmitting
tion systems [31]. Even though TAS can achieve a jitter-free target, or receiving TSN traffic were evaluated. Several key components of
the state-of-art scheduling algorithms are time-consuming, which leads any practical control system (controllers, sensors, and actuators) were
infeasible for practical industrial scenarios. Therefore, TAS-oriented lacking. Thus, the practical feasibility of TSN scheduling remains to
traffic scheduling remains challenging. Optimization theories such as be proven. Also, the three cited works lacked technical details; neither
Satisfiability Modulo Theories (SMT) [20–22,28,29,32,33] and integer algorithm implementation nor TSN testbed development was described.
linear programming (ILP) [23–25,27,34] have been widely applied. To fill these gaps, we developed a TSM [35] for TSN, and deployed
Oliver et al. [20] used an SMT model and an array theory to formalize it in the real world. The various experimental results (i.e., end-to-
scheduling constraints. Jin et al. [21] presented a no-wait scheduling end latencies, queuing delays, and calculation times) were explored to
algorithm using an SMT-based optimal modulus theory. Yu et al. [22] analyze the practical feasibility of our TSM. The main contributions of
applied TSN to virtual machine (VM)-based control networks, and this study are as follows,
presented a lightweight heuristic algorithm for traffic scheduling. Xu
(1) We constructed a realistic TSN system; the robotic process au-
et al. [28,29] focused on designing the TSN scheduling algorithms
tomation scenario is used to verify the effectiveness of the TSM
using SMT theory. In [28], an SMT/OMT-based joint routing and
scheduler.
scheduling algorithm with a co-design constraint set was proposed,
(2) The real-time performance of our TSM was compared to that
which improves network utilization and schedulability. Then, they
of a commercial scheduler; our TSM guarantees the strict real-
continued to research an iterative SMT scheduling algorithm [29].
time requirements of an industrial robotic application, eliminated
They proposed a learning-based approach to partition flows; thus, the
queuing delay, and afforded ultra-low latency.
large-scale scheduling problems are simplified. The simulation results
(3) Our TSM scheduled 500 flows in milliseconds, more than 571-
demonstrated that schedulability and time efficiency were enhanced.
fold faster than the commercial scheduler. In other words, our
An SMT-based TSN scheduling algorithm was proposed in [32] to
TSM has the potential to quickly handle the reconfigurations re-
guarantee deterministic transmission for industrial scenarios. The au-
quired with changes in network infrastructure or custom-oriented
thors tried to split hyper-period into multiple base periods to minimize
applications in the context of Industry 4.0.
the number of GCL entries, thus simplifying the TSN configuration.
Houtan et al. [33] modeled optimization constraints and presented a The rest of this study is organized as follows. Section 2 presents the
new objective function by minimizing the offset. The proposed con- network model. The key principle of TSM is summarized in Section 3.
straints and functions were solved using the SMT solver. The simulation Section 4 describes implementation, i.e., practical traffic scheduling
results showed that the QoS of BE traffic can be improved while and system configuration for a real-world industrial facility. Perfor-
guaranteeing the real-time performance of time-critical (TC) traffic. mance is evaluated in Section 5. Section 6 provides the conclusion.
ILP-based approaches have also been explored. In [23], an ILP-based
algorithm was used to schedule new TSN traffic entering a system. A 2. Network model
commercial ILP solver (CPLEX) was used to solve the proposed formu-
lations. Atallah et al. [24] presented an ILP-based degree of conflict A TSN topology is modeled as a directed weighted graph 𝐺 ≡ (𝑉 , 𝐸),
(DoC)-aware iterative routing and scheduling algorithm for large-scale where 𝑉 is the set of network nodes and 𝐸 = {(𝑣𝑖 , 𝑣𝑗 ) ∣ 𝑣𝑖 , 𝑣𝑗 ∈ 𝑉 } is a
time-sensitive networks. Falk et al. [25] developed ILP formulations set of all directional links of source 𝑣𝑖 and destination 𝑣𝑗 . 𝑉 = (𝑆𝑊 ∪
that modeled the joint routing and scheduling problems, and explored 𝐸𝑆), where 𝑆𝑊 and 𝐸𝑆 denote the TSN switches and end stations
runtime limitations using a state-of-the-art ILP solver and various in- respectively. An example TSN network topology is shown in Fig. 1, in
{ } { }
puts. An enhanced ILP-based scheduling method with offline and online which 𝑆𝑊 = 𝑆𝑊𝑖 ∣ 𝑖 = 1...5 and 𝐸𝑆 = 𝐸𝑆𝑗 ∣ 𝑗 = 1...10 . The 𝑆𝑊 s
phases was proposed in [27] to compute the conflict-free schedules for and 𝐸𝑆s are synchronized via 802.1AS [4] and equipped with a time-
time-sensitive software-defined networks. The authors demonstrated aware shaper (TAS) for traffic scheduling, as defined in 802.1Qbv [5].
that extra ILP constraints do not induce conflicts, while the evaluation Specifically, the TAS has eight prioritized queues with time-aware
results indicated frame loss and configuration overhead were avoided gates. Using a gate control list (GCL), the TAS schedules traffic by
during schedule updates. Chaine et al. [34] presented an ILP-based opening or closing gates at specific times. The GCL is calculated by the
solution, the so-called Egress TT, which contains two schemes: exclusive central network configurator (CNC) and configured to the 𝑆𝑊 s before
queue allocation and size-based isolation. TC traffic can be scheduled the system runs. The real-time applications are executed by the 𝐸𝑆s,
appropriately by Egress TT while eliminating queuing delays. However, which generate a set of time-critical (TC) flows 𝐹 . Each flow 𝑓𝑖 ∈ 𝐹 is
𝑣
significantly higher latency was imposed on TC traffic. characterized by a tuple (𝑠𝑟𝑐𝑖 , 𝑑𝑠𝑡𝑖 , 𝑝𝑖 , 𝑙𝑖 , 𝜑𝑖 , ℎ𝑜𝑝𝑖 𝑗 ), where 𝑠𝑟𝑐𝑖 ∈ 𝐸𝑆 is
All of the studies described above found that optimization theory the source, 𝑑𝑠𝑡𝑖 ∈ 𝐸𝑆 is the destination, 𝑝𝑖 is the transmission period,
solved TSN scheduling problems. However, as the number of inputs 𝑙𝑖 is the traffic size, 𝜑𝑖 is the delay constraint (i.e., maximum allowable

2
J. Jiang et al. Journal of Industrial Information Integration 33 (2023) 100464

Fig. 2. Bandwidth allocation scheme for egress of an arbitrary 𝑆𝑊𝑘 .

Fig. 1. Example TSN network. separate intervals (i.e., time-critical interval and non-time-critical in-
terval) are assigned to TC and non-TC traffic flows, which eliminates
the transmission interference from non-TC to TC traffic flows. A time-
𝑣 critical interval is further subdivided into multiple individual time
delay), and ℎ𝑜𝑝𝑖 𝑗 is the hop-count from 𝑠𝑟𝑐𝑖 to the node 𝑣𝑗 (𝑣𝑗 ∈ 𝑉 ).
slots, which are allocated to individual TC traffic flows. This allocation
To guarantee real-time performance, an 𝑓𝑖 must be transmitted from
guarantees the independent transmission of different TC flows without
𝑠𝑟𝑐𝑖 to 𝑑𝑠𝑡𝑖 within 𝜑𝑖 . The flow specifications of the example topology
{ } any interference with each other. Therefore, the maximum allowable
are illustrated at the bottom of Fig. 1, where 𝐹 = 𝑓1 , 𝑓2 , 𝑓3 , 𝑓4 , 𝑓5 , 𝑓6 , delay requirement is satisfied and no flow experiences a queuing delay.
𝑓1 = (𝐸𝑆1 , 𝐸𝑆3 , 100 μs, 1500𝐵, 100 μs, 3), and so on. Note that the last The inputs to traffic scheduling are the specifications of the flows
𝑑𝑠𝑡
element of flow specifications shown in Fig. 1 is the hop-count ℎ𝑜𝑝𝑖 𝑖 transmitted within each subnetwork, and the outputs are the TSN
𝑓𝑖
from 𝑠𝑟𝑐𝑖 to 𝑑𝑠𝑡𝑖 , i.e., 𝑣𝑗 = 𝑑𝑠𝑡𝑖 . The routing path 𝑃(𝑣 ,𝑣 ) is an ordered schedules. With Network Partition, each 𝜑𝑖 must be adjusted to 𝜑𝐿
𝑖 𝑗 𝑖 using
sequence of all interconnected nodes along the forwarding path of 𝑓𝑖 Eq. (1). In the 𝐿th subnetwork 𝐺𝐿 ′ , 𝜱 is the vector of 𝜑𝐿 (∀𝑖 = 1, … , 𝑁),
𝑖
from source 𝑣𝑖 to destination 𝑣𝑗 . For example, as shown by the green the elements of which are sorted in ascending order as follows,
𝑓1
arrows in Fig. 1, the path of 𝑓1 from 𝐸𝑆1 to 𝐸𝑆3 is 𝑃(𝐸𝑆 = ( 𝐿 )
{[ ]} 1 ,𝐸𝑆3 ) 𝜱 = [𝜑𝐿 1
, … , 𝜑𝐿 𝐿 𝐿
𝑖 , … , 𝜑𝑁 ], 𝜑𝑖 ≤ 𝜑𝑖+1 , ∀𝑖 = 1, … , 𝑁 − 1. (2)
𝐸𝑆1 , 𝑆𝑊1 , 𝑆𝑊2 , 𝐸𝑆3 .
Next, we summarize the fundamental principles of Traffic Scheduling.
3. Key principles of a TSM There are four steps:
Step 1: Determine the time-division interval (𝑇 𝐷𝐼 𝑘 ) for 𝑆𝑊𝑘 egress.
This section summarizes the key principles of the TSM in [35]. The The TC flows, 𝑓𝑖 , and non-TC flows compete for bandwidth resources
inputs are the network topology 𝐺, TC flows 𝐹 , and link specifications at the time of 𝑆𝑊𝑘 egress. As shown in Fig. 2, to guarantee the real-
(length 𝐿(𝑣𝑖 ,𝑣𝑗 ) and speed 𝑆(𝑣𝑖 ,𝑣𝑗 ) ). The outputs are the GCLs for the TSN time performance of 𝑓𝑖 , the shared bandwidth is divided into multiple
switches and the startup instants of traffic transmission for all TSN time-division intervals. Here, we regard instant 0 as the start of the
TSN schedule. The length of all the 𝑇 𝐷𝐼 𝑘 values sums to the so-called
talkers. Essentially, our TSM contains two routines: Network Partition 𝑘
‘‘hyper-period’’ (𝑇𝑐𝑦𝑐𝑙𝑒 ) of [40]. Scheduling is repeated at intervals
to enhance scalability, and Traffic Scheduling to guarantee real-time
𝑘
of 𝑇𝑐𝑦𝑐𝑙𝑒 . Generally, each 𝑇 𝐷𝐼 𝑘 is further split into the guard band
performance.
interval (𝐺𝐵𝐼 𝑘 ) defined in [5], a TC interval (𝑇 𝐶𝐼 𝑘 ), and two non-TC
3.1. Network partition intervals [the first non-TC interval (𝑁𝑇 𝐶𝐼1𝑘 ) and the second non-TC
interval (𝑁𝑇 𝐶𝐼2𝑘 )]. The 𝑇 𝐶𝐼 𝑘 has multiple time slots (𝑇 𝑆 𝑘 ) dedicated
to transmission of 𝑓𝑖 . The 𝐺𝐵𝐼 𝑘 is always reserved immediately before
The Network Partition routine applies the scheduling (described in
𝑇 𝐶𝐼 𝑘 to prevent transmission interference from non-TC flows. 𝑁𝑇 𝐶𝐼1𝑘
Section 3.2) to the subnetworks of a large-scale TSN. Any arbitrary TSN
is allocated before 𝑇 𝐶𝐼 𝑘 and 𝑁𝑇 𝐶𝐼2𝑘 is allocated after 𝑇 𝐶𝐼 𝑘 . Then, the
network 𝐺 can be partitioned into several subnetworks 𝐺′ which does
lengths of 𝑇 𝐷𝐼 𝑘 , 𝐺𝐵𝐼 𝑘 , and 𝑁𝑇 𝐶𝐼1𝑘 are determined.
not have any branches. For example, in Fig. 1, the network is split into
The 𝑇 𝐷𝐼 𝑘 is the basic scheduling unit for 𝑆𝑊𝑘 , the length of which
four subnetworks (see blue dashed circles), 𝐺′ = (𝐺1′ , 𝐺2′ , 𝐺3′ , 𝐺4′ ). For
is determined by (3), where 𝜑𝑚𝑖𝑛 is the smallest maximum allowable
better understanding, we take the first subnetwork 𝐺1′ as an example, delay. Eq. (3) guarantees real-time performance by restricting the
𝐺1′ ≡ {(𝑉1′ , 𝐸1′ ) ∣ 𝑉1′ ∈ 𝑉 , 𝐸1′ ∈ 𝐸}, where 𝑉1′ and 𝐸1′ are network nodes length of 𝑇 𝐷𝐼 𝑘 to 𝜑𝑚𝑖𝑛 :
and directional links of the subnetwork 𝐺1′ , as defined in Section 2.
( )
Specifically, 𝑉1′ is 𝐸𝑆1 , 𝐸𝑆2 , 𝑆𝑊1 , 𝑆𝑊2 , and 𝐸1′ is [(𝐸𝑆1 , 𝑆𝑊1 ), (𝐸𝑆2 , 𝑇 𝐷𝐼 𝑘 = 𝜑𝑚𝑖𝑛 . (3)
𝑆𝑊1 ), (𝑆𝑊1 , 𝑆𝑊2 )]. Each subnetwork 𝐺′ contains 𝑁 inputs 𝑓𝑖 , 𝑁 out-
As defined in 802.1Qbv [5], 𝐺𝐵𝐼 𝑘 is the duration of maximum
puts 𝑓𝑖 , and 𝐾 TSN 𝑆𝑊 s in series. Accordingly, the maximum allowable
′
transmission unit (MTU)-size frame transmission, and is thus deter-
delay 𝜑𝐿 𝑖 of 𝑓𝑖 in the 𝐿th subnetwork 𝐺𝐿 is: mined by (4), where 𝐿𝑚𝑡𝑢 is the MTU size, and 𝑆(𝑣𝑖 ,𝑣𝑗 ) is the link speed:
𝐾 −1
𝑖 = ⌈𝜑𝑖 ×
𝜑𝐿
𝑀 −1
⌉ (1)
𝐿𝑚𝑡𝑢
′ , and
𝐺𝐵𝐼 𝑘 = . (4)
where 𝐿 is the subnetwork index, 𝐾 is the number of 𝑆𝑊 s in 𝐺𝐿 𝑆(𝑣𝑖 ,𝑣𝑗 )
𝑓𝑖
𝑀 is the total number of 𝑆𝑊 s along 𝑃(𝑠𝑟𝑐 ,𝑑𝑠𝑡 ) . If the real end-to-end
𝑖 𝑖 Next, we determine the length of 𝑁𝑇 𝐶𝐼1𝑘 . The sum of 𝑁𝑇 𝐶𝐼1𝑘 and
latency of the 𝑓𝑖 of each subnetwork does not exceed 𝜑𝐿 𝑖 , the total
𝑠𝑟𝑐 𝑇 𝑜𝑆𝑊𝑘
𝐺𝐵𝐼 𝑘 is equal to the network delay (𝐷𝑝ℎ𝑦𝑖 ) experienced by 𝑓𝑖 .
maximum allowable delay requirement 𝜑𝑖 is satisfied. Thus, as soon as 𝑓𝑖 arrives at 𝑆𝑊𝑘 , it will be immediately transmitted
𝑠𝑟𝑐 𝑇 𝑜𝑆𝑊𝑘
during 𝑇 𝐶𝐼 𝑘 . To determine 𝑁𝑇 𝐶𝐼1𝑘 , the 𝐷𝑝ℎ𝑦𝑖 should be firstly
3.2. Traffic scheduling calculated using Eq. (5),
𝑀𝑖 +1 𝑀𝑖 +1 𝑀𝑖
The basic concept of Traffic Scheduling is to divide the available 𝑠𝑟𝑐 𝑇 𝑜𝑆𝑊𝑘 ∑ ∑ ∑
𝐷𝑝ℎ𝑦𝑖 = 𝑑𝑡𝑟𝑎𝑛𝑠 + 𝑑𝑝𝑟𝑜𝑝 + 𝑑𝑝𝑟𝑜𝑐 (5)
bandwidth into time slots via time-division multiplexing. Next, two 𝑗=1 𝑗=1 𝑗=1

3
J. Jiang et al. Journal of Industrial Information Integration 33 (2023) 100464

where 𝑀𝑖 is the number of 𝑆𝑊 s between 𝑠𝑟𝑐𝑖 and 𝑆𝑊𝑘 . The trans-

mission delay (𝑑𝑡𝑟𝑎𝑛𝑠 ) and propagation delay (𝑑𝑝𝑟𝑜𝑝 ) can be simply
calculated as in [41]. We assume that the processing delay (𝑑𝑝𝑟𝑜𝑐 ) of
a switch is an input. Then, the length of 𝑁𝑇 𝐶𝐼1𝑘 can be given by
𝑠𝑟𝑐 𝑇 𝑜𝑆𝑊𝑘
(6), where 𝐺𝐵𝐼 𝑘 and 𝐷𝑝ℎ𝑦𝑖 have been determined by (4) and (5),
respectively:
𝑠𝑟𝑐 𝑇 𝑜𝑆𝑊𝑘
𝑁𝑇 𝐶𝐼1𝑘 = 𝐷𝑝ℎ𝑦𝑖 − 𝐺𝐵𝐼 𝑘 (6)

Step 2: Determine the time-slot allocation interval (𝑇 𝑆𝐴𝐼𝑖 ) and

number of time slots (𝑇 𝑆𝑇 𝐶𝐷𝑘 ) allocated to 𝑇 𝐶𝐼 𝑘 . To eliminate
queuing delay, we first harmonize the maximum allowable delay re-
quirement for input 𝑓𝑖 , and then allocate the appropriate time slots.
Specifically, let TSAI be the vector, the elements of which are sorted in
ascending order of 𝑇 𝑆𝐴𝐼𝑖 as follows:
𝐓𝐒𝐀𝐈 = [𝑇 𝑆𝐴𝐼1 , … , 𝑇 𝑆𝐴𝐼𝑖 , … , 𝑇 𝑆𝐴𝐼𝑁 ],
(7)
(𝑇 𝑆𝐴𝐼𝑖 ≤ 𝑇 𝑆𝐴𝐼𝑖+1 )
The index 𝑖 denotes the location of each element of TSAI, ranging Fig. 3. The system architecture.
from 0 to 𝑁, where 𝑁 presents the number of the input TSN flows. In
other words, the value of elements in TSAI increases as 𝑖 increases.
𝑇 𝑆𝐴𝐼1 is the smallest 𝑇 𝑆𝐴𝐼𝑖 , and should equal the length of the Then, according to (12) and (13), the stability condition becomes
𝑇 𝐷𝐼 𝑘 defined by (8): (14), which guarantees that the 𝑓𝑖 are scheduled appropriately:

𝑇 𝑆𝐴𝐼1 = 𝑇 𝐷𝐼 𝑘 = 𝜑1 . (8) 𝑇 𝐷𝐼 𝑘 ≥ 𝑁𝑇 𝐶𝐼1𝑘 + 𝐺𝐵𝐼 𝑘 + 𝑇 𝐶𝐼 𝑘 (14)

The definition of (8) guarantees that the most urgent TC traffic does Step 3: Allocate a specific time slot to 𝑓𝑖 , and determine the starting
not exceed its maximum allowable delay requirement. Therefore, all 𝑠𝑟𝑐
instant of the first message sent from a talker (𝐹 𝑀𝑇 𝐼𝑖 𝑖 ) and a switch
delay requirements are satisfied. We next adopt the window scheduling 𝑆𝑊𝑘
(𝐹 𝑀𝑇 𝐼𝑖 ). Here, the sequence of time-slot allocation is determined
algorithm used in [36] to calculate the 𝑇 𝑆𝐴𝐼𝑖 s (∀𝑖 = 2, … , 𝑁):
from the 𝑓𝑖 with the smallest delay requirement to that with the
𝑇 𝑆𝐴𝐼𝑖 = 𝑘𝑖 × 𝑇 𝑆𝐴𝐼1 , 𝑘𝑖 = 2⌊log2 (𝜑𝑖 ∕𝑇 𝑆𝐴𝐼1 )⌋ (9) largest requirement. Specifically, the 𝑓1 with 𝜑1 is allocated to the
first available time slot in the first 𝑇 𝐷𝐼 𝑘 (marked with red asterisk
The first term in (9) indicates that each 𝑇 𝑆𝐴𝐼𝑖 is derived by in Fig. 2). Then, the remaining 𝑓𝑖 with 𝜑𝑖 values (∀𝑖 = 2, … , 𝑁) are
multiplying an integer 𝑘𝑖 by 𝑇 𝑆𝐴𝐼1 . 𝑘𝑖 is defined as a power of subsequently allocated to the next available time slot, and so on. When
two (2𝑚 ) where 𝑚 is the floor function 𝑚(𝑥) = ⌊𝑥⌋. As denoted in 𝑆𝑊
𝑓𝑖 is scheduled at 𝑆𝑊𝑘 , 𝐹 𝑀𝑇 𝐼𝑖 𝑘 becomes:
(9), the maximum time-slot allocation interval 𝑇 𝑆𝐴𝐼𝑚𝑎𝑥 is the least
𝑆𝑊𝑘
common multiple of the 𝑇 𝑆𝐴𝐼𝑖 , and is thus equal to the hyper-period 𝐹 𝑀𝑇 𝐼𝑖 = (𝑚 − 1) × 𝑇 𝐷𝐼 𝑘 + (𝑁𝑇 𝐶𝐼1𝑘 + 𝐺𝐵𝐼 𝑘 )
(15)
𝑇𝑐𝑦𝑐𝑙𝑒 mentioned above. In other words, TSN scheduling is repeated at + (𝑛 − 1) × 𝑇 𝑆 𝑘
intervals of 𝑇 𝑆𝐴𝐼𝑚𝑎𝑥 .
The time slots (𝑇 𝑆𝑇 𝐶𝐷𝑘 ) allocated to 𝑇 𝐶𝐼 𝑘 are: where 𝑚 and 𝑛 are the indexes of 𝑇 𝐷𝐼 𝑘 and 𝑇 𝑆 𝑘 , respectively. Eq. (15)
⌈𝑁 ⌉ gives the instant at which 𝑓𝑖 is to be transmitted out from egress
∑( )
𝑇 𝑆𝑇 𝐶𝐷 =𝑘 𝑘
𝑇 𝐷𝐼 ∕𝑇 𝑆𝐴𝐼𝑖 (10) of 𝑆𝑊𝑘 . Before 𝑓𝑖 arrives at the egress, it experiences a delay along
𝑓𝑖
𝑖=1 its forwarding path 𝑃(𝑠𝑟𝑐 ,𝑆𝑊 )
. Therefore, after backtracking by that
𝑖 𝑘
𝑠𝑟𝑐𝑖
To ensure that all 𝑇 𝑆𝑇 𝐶𝐷𝑘 values are integers, Eq. (10) rounds up amount of time, we determine the 𝐹 𝑀𝑇 𝐼𝑖 :
∑ ( )
the formula 𝑁 𝑘
𝑖=1 𝑇 𝐷𝐼 ∕𝑇 𝑆𝐴𝐼𝑖 , which expresses the average number 𝑠𝑟𝑐𝑖 𝑆𝑊𝑘 𝑠𝑟𝑐 𝑇 𝑜𝑆𝑊𝑘
𝐹 𝑀𝑇 𝐼𝑖 = 𝐹 𝑀𝑇 𝐼𝑖 − 𝐷𝑝ℎ𝑦𝑖 (16)
of time slots allocated to 𝑓𝑖 during a 𝑇 𝐷𝐼 𝑘 . Then, we determine the
length of 𝑇 𝐶𝐼 𝑘 as follows: 𝑠𝑟𝑐 𝑇 𝑜𝑆𝑊𝑘
where 𝐷𝑝ℎ𝑦𝑖 is the network delay from 𝑠𝑟𝑐𝑖 to 𝑆𝑊𝑘 , as determined
𝑘 𝑘
𝐿𝑓 𝑟𝑎𝑚𝑒 by Eq. (5).
𝑇 𝐶𝐼 = 𝑇 𝑆𝑇 𝐶𝐷 × (11)
𝑆(𝑣𝑖 ,𝑣𝑗 ) Step 4: Generate the TSN schedule (i.e., the GCL). The GCL instructs
𝐿𝑓 𝑟𝑎𝑚𝑒
the TAS to open or close a gate at a specific instant and retain that gate
where 𝑆(𝑣𝑖 ,𝑣𝑗 )
is the time taken to transmit 𝑓𝑖 , 𝐿𝑓 𝑟𝑎𝑚𝑒 is the frame size, status for a certain time. Three key parameters are required:
and 𝑆(𝑣𝑖 ,𝑣𝑗 ) is the link speed. At this point, for a 𝑇 𝐷𝐼 𝑘 , the length
of the 𝐺𝐵𝐼 𝑘 , the 𝑁𝑇 𝐶𝐼1𝑘 , and the 𝑇 𝐶𝐼 𝑘 have been determined. The 1. The instant of gate opening to allow transmission of 𝑓𝑖 . The
𝑠𝑟𝑐 𝑆𝑊
remaining bandwidth is allocated to 𝑁𝑇 𝐶𝐼2𝑘 , the length of which is: 𝐹 𝑀𝑇 𝐼𝑖 𝑖 and 𝐹 𝑀𝑇 𝐼𝑖 𝑘 , as determined in Step 3, inform the
talkers or switches when to start transmission.
𝑁𝑇 𝐶𝐼2𝑘 = 𝑇 𝐷𝐼 𝑘 − (𝑁𝑇 𝐶𝐼𝑖𝑘 + 𝐺𝐵𝐼 𝑘 + 𝑇 𝐶𝐼 𝑘 ) (12) 2. The duration of gate opening, which must allow complete traffic
transmission. As shown in Steps 1 and 2, 𝑇 𝐶𝐼 𝑘 , 𝑁𝑇 𝐶𝐼1𝑘 , and
𝑁𝑇 𝐶𝐼2𝑘 decreases as the input TSN flow increases. However,
𝑁𝑇 𝐶𝐼2𝑘 are the intervals available for transmission of 𝑓𝑖 and
𝑁𝑇 𝐶𝐼2𝑘 cannot become negative, because the number of input flows
non-TC flows, respectively. The 𝐺𝐵𝐼 𝑘 is constructed to avoid
cannot increase infinitely (the bandwidth is finite). If too many 𝑓𝑖 are
input, the bandwidth will be overloaded. To avoid this, the number of transmission interference from non-TC flows. The gate is opened
input TSN flows must not exceed the maximum bandwidth. In other at the beginning of each interval and closed at the end.
words, the length of 𝑁𝑇 𝐶𝐼2𝑘 must be restricted as follows: 3. The cycle time of a GCL. The GCL is repeated at intervals of the
𝑇 𝐴𝑆𝐼𝑚𝑎𝑥 determined in Step 2.
𝑁𝑇 𝐶𝐼2𝑘 ≥ 0 (13)

4
J. Jiang et al. Journal of Industrial Information Integration 33 (2023) 100464

Procedure 1 Network Partition

𝑓
Inputs: 𝐺, 𝐹 , and 𝑃(𝑠𝑟𝑐
𝑖
𝑖 ,𝑑𝑠𝑡𝑖 )
Output: 𝐺′ and 𝐹 ′
1: 𝑏𝑟𝑎𝑛𝑐ℎ𝑁𝑜𝑑𝑒𝑠 ← Obtain branch nodes(𝐺)
𝑓𝑖
2: 𝐺′ ← Partition network(𝑏𝑟𝑎𝑛𝑐ℎ𝑁𝑜𝑑𝑒𝑠, 𝑃(𝑠𝑟𝑐 )
𝑖 ,𝑑𝑠𝑡𝑖 )
3: ′ ′
𝐹 ← Adjust inputs(𝐹 , 𝐺 )
4: return 𝐺′ and 𝐹 ′

4.1.1. Network partition

Fig. 4. A procedure flowchart for the CNC and CUC. As shown in Procedure 1, Network partition generally contains three
subfunctions: Obtain branch nodes, Partition network, and Adjust
inputs. Specifically, line 1 indicates that the routine Obtain branch
nodes utilizes the input topology 𝐺 to find branch nodes, termed
4. Implementation methodology {( ) }
𝑏𝑟𝑎𝑛𝑐ℎ𝑁𝑜𝑑𝑒𝑠 = 𝑆𝑊𝑖 ∣ 𝑆𝑊𝑖 ⊆ 𝑆𝑊 , which are switch nodes.
𝑏𝑟𝑎𝑛𝑐ℎ𝑁𝑜𝑑𝑒𝑠 are the breakpoints at which network 𝐺 is divided
We now present fine-grained guidelines for the implementation into various 𝐺′ without any branches. To achieve this, each element of
of Scheduling Procedure and Configuration agent, which are two cru- 𝑏𝑟𝑎𝑛𝑐ℎ𝑁𝑜𝑑𝑒𝑠 should be interconnected via two switches, and to either
cial components for a TSN-based process automation system. As il- end stations or switches. To find all elements of 𝑏𝑟𝑎𝑛𝑐ℎ𝑁𝑜𝑑𝑒𝑠, Obtain
lustrated in Fig. 3, we constructed a system architecture based on branch nodes implements a filter, as follows:
a fully centralized model [7]. Specifically, this is the TSN network
{( ) }
𝐺 = 𝐸𝑆𝑖 , 𝑆𝑊𝑗 ∣ 𝑖 = 1, … , 8, 𝑗 = 1, 2 where 𝑆𝑊1 and 𝑆𝑊2 are inter- (1) Filter a switch node connected with 𝑁 switches (𝑁 ≥ 3) and 𝑀
connected; 𝐸𝑆1 , 𝐸𝑆2 , 𝐸𝑆3 , and 𝐸𝑆4 are TSN talkers, and 𝐸𝑆5 , 𝐸𝑆6 , end stations (𝑀 ≥ 0).
𝐸𝑆7 , and 𝐸𝑆8 are listeners. The hardware deployment and application (2) Filter a switch node connected with 𝑀 TSN end stations (𝑀 ≥ 1)
scenario will be detailed in Section 4.3. We developed two primary and 𝑁 switches (𝑁 = 2).
system components: a centralized network configurator (CNC) and a For example, we input the topology of Fig. 1 and execute Obtain
centralized user configurator (CUC). The CNC calculates forwarding branch nodes. The 𝑏𝑟𝑎𝑛𝑐ℎ𝑁𝑜𝑑𝑒𝑠 are {𝑆𝑊2 , 𝑆𝑊3 } because 𝑆𝑊3 is
paths and schedules for the TSN 𝐸𝑆s and 𝑆𝑊 s. The CUC is responsible connected to three switches (filter rule 1), and 𝑆𝑊2 to two end stations
for central maintenance of traffic information and the configuration (filter rule 2). This guarantees that each subnetwork does not has any
parameters of all TSN 𝐸𝑆s. To clarify how the CNC and CUC are branch (i.e., the number of input and output flows are the same) if the
developed, we present a flowchart in Fig. 4. The CNC inputs are network is split at the points of the 𝑏𝑟𝑎𝑛𝑐ℎ𝑁𝑜𝑑𝑒𝑠.
highlighted in green; these are the network topology 𝐺, sets of TC After defining the 𝑏𝑟𝑎𝑛𝑐ℎ𝑁𝑜𝑑𝑒𝑠, Partition network splits the net-
𝑓𝑖
flows 𝐹 , link length 𝐿(𝑣𝑖 ,𝑣𝑗 ) , and speed 𝑆(𝑣𝑖 ,𝑣𝑗 ) . The orange blocks are work (line 2) by checking whether the forwarding path (𝑃(𝑠𝑟𝑐 )
𝑖 ,𝑑𝑠𝑡𝑖 )
the three internal CNC routines, i.e., the Routing procedure, Schedul- contains the 𝑏𝑟𝑎𝑛𝑐ℎ𝑁𝑜𝑑𝑒𝑠. If an element of 𝑏𝑟𝑎𝑛𝑐ℎ𝑁𝑜𝑑𝑒𝑠 is detected,
ing procedure, and Configuration agent. We use a well-known shortest the network is split into two subnetworks at the point corresponding
path algorithm [42] as the Routing procedure. Then, the forwarding to that element of 𝑏𝑟𝑎𝑛𝑐ℎ𝑁𝑜𝑑𝑒𝑠. For example, the forwarding path of
𝑓2
paths are calculated and input to the Scheduling procedure, wherein the 𝑓2 is 𝑃(𝐸𝑆 ,𝐸𝑆 )
= {[𝐸𝑆2 , 𝑆𝑊1 , 𝑆𝑊2 , 𝑆𝑊3 , 𝐸𝑆5 ]} and the 𝑏𝑟𝑎𝑛𝑐ℎ𝑁𝑜𝑑𝑒𝑠
2 5
2 𝑓
TSM summarized in Section 3 is implemented de novo. The outputs are {𝑆𝑊2 , 𝑆𝑊3 } (as determined above). Hence, along 𝑃(𝐸𝑆 , the
𝑠𝑟𝑐 2 ,𝐸𝑆5 )
of the Scheduling procedure are GCLs and 𝐹 𝑀𝑇 𝐼𝑖 𝑖 configured for ′ ′
network is split into two subnetworks, [i.e., (𝐺1 , 𝐺2 )]. By repeating this
𝑠𝑟𝑐
TSN devices. The 𝐹 𝑀𝑇 𝐼𝑖 𝑖 is utilized by the CUC to configure TSN process, the whole network is eventually split into the four subnetworks
𝐸𝑆s, whereas the GCLs are configured to 𝑆𝑊 s by the CNC. This indicated by the blue dashed circles in Fig. 1.
procedure is based on the NETCONF (Network Configuration Protocol) After execution of Partition network, the flow specifications are
and YANG (Yet Another Next Generation) protocols, which centrally adjusted (line 3). The maximum allowable delay requirements (𝜑𝑖
manage the configurations for network devices. Moreover, NETCONF values) are adjusted using Eq. (1), and then input into the scheduling
is proposed in RFC 6241 that adopts a simple Remote Procedure Call function. For example, in Fig. 1, the 𝜑2 of 𝑓2 is 200 μs, the number of
(RPC) mechanism to acquire and modify the configurations of network 𝑆𝑊 s involved in the first subnetwork 𝐺1′ is two, and the number of
devices. YANG is a data modeling language published in RFC 7950 that 𝑆𝑊 s along the forwarding path is three. Thus, the adjusted maximum
defines the configuration data, state data, and RPC for network devices. allowable delay 𝜑12 is 100 μs using Eq. (1). Finally, 𝐺′ and 𝐹 ′ are
obtained and input into the scheduling functions.
YANG instances can be encoded into XML format and extracted by
NETCONF to configure network devices. Scheduling and configuration
4.1.2. Traffic scheduling
implementation are discussed below.
We implement Traffic Scheduling for TSN by Procedure 2, where
the four scheduling steps are highlighted in orange. In detail, line 1
4.1. Scheduling procedure utilizes the function 𝑙𝑒𝑛() to obtain the length of 𝐹 ′ , which is shown
as the number (𝑛𝑢𝑚𝑓 𝑙𝑜𝑤) of 𝑓𝑖 . Then, 𝐹 ′ is sorted using the ℎ𝑒𝑎𝑝𝑠𝑜𝑟𝑡
method in ascending order of 𝜑′𝑖 , and the ordered 𝜑′𝑖 are contained in
We implemented TSM to be Scheduling Procedure, which uses the 𝜱. The time-division intervals (i.e., Step 1 of Traffic Scheduling ) are
two routines of Network partition and Traffic scheduling. In general, Net- determined by lines 3 and 4. Specifically, the basic scheduling unit 𝑇 𝐷𝐼
work partition obtains and processes inputs by partitioning network 𝐺 is updated by the smallest value (𝜑′1 ) of all 𝜑′𝑖 . Then, the length of 𝐺𝐵𝐼
into several subnetworks 𝐺′ with no branches. Then, the routine adjusts is determined using Eq. (4).
the input parameters (e.g., the maximum allowable delay requirement Step 2 is implemented by lines 5–15. First, the allocation interval
𝜑𝑖 ) for 𝐺′ . Finally, 𝐺′ and the adjusted parameters are input to Traffic 𝑇 𝑆𝐴𝐼𝑖 is determined. The smallest value of 𝑇 𝑆𝐴𝐼𝑖 is the first element
scheduling, which executes a four-step process to calculate the TSN (𝑇 𝑆𝐴𝐼1 ) in TSAI, which is updated by 𝑇 𝐷𝐼 in line 5. Then, the other
schedule. The details are described below. 𝑇 𝑆𝐴𝐼𝑖 values are gradually calculated by Eq. (9). After all 𝑇 𝑆𝐴𝐼𝑖 have

5
J. Jiang et al. Journal of Industrial Information Integration 33 (2023) 100464

Procedure 2 Traffic Scheduling

Inputs: 𝐺′ , 𝐹 ′ , 𝑆(𝑣𝑖 ,𝑣𝑗 ) , and 𝐿(𝑣𝑖 ,𝑣𝑗 )
𝑠𝑟𝑐𝑖
Output: 𝐺𝐶𝐿𝑠 and 𝐹 𝑀𝑇 𝐼𝑖
′
1: 𝑛𝑢𝑚𝑓 𝑙𝑜𝑤 ← 𝑙𝑒𝑛(𝐹 )
2: 𝜱 ← ℎ𝑒𝑎𝑝𝑠𝑜𝑟𝑡(𝜑′𝑖 , 𝐹 ′ )
// Step 1: Determine the time-division intervals;
3: 𝑇 𝐷𝐼 ← 𝜑′1
4: 𝐺𝐵𝐼 ← Use Eq. (4)
// Step 2: Determine allocation interval, numbers of time slots
5: 𝑇 𝑆𝐴𝐼1 ← 𝑇 𝐷𝐼
6: 𝑇 𝑆𝐴𝐼𝑖 ← Use Eq. (9)
7: 𝑇𝑐𝑦𝑐𝑙𝑒 ← max(𝑇 𝑆𝐴𝐼𝑖 )
8: 𝑇 𝑆𝑇 𝐶𝐷 ← Use Eq. (10)
9: 𝑇 𝐶𝐼 ← Use Eq. (11)
10: 𝑛𝑢𝑚𝑇 𝐷𝐼 ← 𝑇𝑐𝑦𝑐𝑙𝑒 ∕𝑇 𝐷𝐼
11: 𝑛𝑢𝑚𝑇 𝑆 ← 𝑛𝑢𝑚𝑇 𝐷𝐼 × 𝑇 𝑆𝑇 𝐶𝐷
Fig. 5. Sequencing of the configuration procedure. 12: 𝑎𝑣𝑎𝑖𝑙𝑎𝑏𝑙𝑒𝑇 𝑆 ← 𝑛𝑢𝑚𝑇 𝑆 ∗ [−1]
13: if 𝑇 𝐷𝐼 < 𝑁𝑇 𝐶𝐼1 +𝐺𝐵𝐼+𝑇 𝐶𝐼 then
14: return error
15: end if
been determined, the maximum value is updated to 𝑇𝑐𝑦𝑐𝑙𝑒 , at which // Step 3: Allocate specific time slots to 𝑓𝑖
time the TSN schedule is repeated. Next, 𝑇 𝑆𝑇 𝐶𝐷 and the length of 𝑠𝑟𝑐
16: 𝐹 𝑀𝑇 𝐼𝑖 𝑖 ← 𝑛𝑢𝑚𝑓 𝑙𝑜𝑤 ∗ [0]
𝑇 𝐶𝐼 are calculated in lines 8 and 9. The variable 𝑛𝑢𝑚𝑇 𝐷𝐼 denoting 17: for 𝐹 ′ is not empty do
𝑇 𝑣
the number of 𝑇 𝐷𝐼 in one hyper-period (a 𝑇𝑐𝑦𝑐𝑙𝑒 ), is updated by 𝑇𝑐𝑦𝑐𝑙𝑒 𝐷𝐼
. 18: 𝑓𝑖 =(𝑠𝑟𝑐𝑖 , 𝑑𝑠𝑡𝑖 , 𝑝𝑖 , 𝑙𝑖 , 𝜑′𝑖 , ℎ𝑜𝑝𝑖 𝑗 ) ← heappop(𝐹 ′ )
𝑡𝑎𝑙𝑘𝑒𝑟𝑇 𝑜𝑆𝑊 𝑣
𝑛𝑢𝑚𝑇 𝐷𝐼 will always be an integer [see Eqs. (8) and (9)]. The number 19: 𝐷𝑝ℎ𝑦 𝑘
← Use Eq. (5) with input ℎ𝑜𝑝𝑖 𝑗 .
(𝑛𝑢𝑚𝑇 𝑆) of time slots that can be contained in one 𝑇𝑐𝑦𝑐𝑙𝑒 is calculated 20: 𝑘
𝑁𝑇 𝐶𝐼1 ← Use Eq. (6)
in line 11. Line 12 initializes a tuple 𝑎𝑣𝑎𝑖𝑙𝑎𝑏𝑙𝑒𝑇 𝑆 of length 𝑛𝑢𝑚𝑇 𝑆; this 21: for 𝑆𝑊𝑘 in 𝐺′ do
reflects the time-slot occupancy. All elements are initialized to −1 when 22: 𝑖𝑛𝑑𝑒𝑥𝐴𝑣𝑎𝑙𝑖𝑎𝑏𝑙𝑒𝑇 𝑆 ← 𝑎𝑣𝑎𝑖𝑙𝑎𝑏𝑙𝑒𝑇 𝑆.FindIndex(−1)
all time slots are available. During time slot allocation, an arbitrary 23: 𝑚 ← ceil (𝑖𝑛𝑑𝑒𝑥𝐴𝑣𝑎𝑙𝑖𝑎𝑏𝑙𝑒𝑇 𝑆∕𝑇 𝑆𝑇 𝐶𝐷)
element of 𝑎𝑣𝑎𝑖𝑙𝑎𝑏𝑙𝑒𝑇 𝑆 (e.g., 𝑖th element) can be changed to a positive 24: 𝑛 ← 𝑖𝑛𝑑𝑒𝑥𝐴𝑣𝑎𝑙𝑖𝑎𝑏𝑙𝑒𝑇 𝑆 mod 𝑇 𝑆𝑇 𝐶𝐷
𝑆𝑊
25: 𝐹 𝑀𝑇 𝐼𝑖 𝑘 ← Use Eq. (15) with inputs 𝑚, 𝑛.
integer 𝑗, indicating that a flow with ID of 𝑗 occupies the 𝑖th slot. For 𝑆𝑊
26: 𝑓 𝑚𝑡𝑖𝑡𝑎𝑙𝑘𝑒𝑟 ← Use Eq. (16) with input 𝐹 𝑀𝑇 𝐼𝑖 𝑘 .
example, 𝑎𝑣𝑎𝑖𝑙𝑎𝑏𝑙𝑒𝑇 𝑆 = [1, −1, −1, 2] indicates that 𝑛𝑢𝑚𝑇 𝑆 is 4, the 𝑖
𝑠𝑟𝑐𝑖 𝑠𝑟𝑐𝑖 𝑠𝑟𝑐𝑖
27: if 𝑓 𝑚𝑡𝑖𝑖 == 𝑓 𝑚𝑡𝑖𝑗 in 𝐹 𝑀𝑇 𝐼𝑗 then
first and fourth time slots are allocated to 𝑓1 and 𝑓2 , respectively; the
28: 𝑖𝑛𝑑𝑒𝑥𝐴𝑣𝑎𝑙𝑖𝑎𝑏𝑙𝑒𝑇 𝑆 ← 𝑖𝑛𝑑𝑒𝑥𝐴𝑣𝑎𝑙𝑖𝑎𝑏𝑙𝑒𝑇 𝑆+1;
second and third time slots are available for other flows. Lines 13–15 29: Go to Line 22.
maintain the stability of the TSN system by guaranteeing inequality 30: end if
(14). If the stability condition is not satisfied, the scheduling procedure 31:
𝑠𝑟𝑐
𝐹 𝑀𝑇 𝐼𝑖 𝑖 ← 𝑓 𝑚𝑡𝑖𝑖 𝑖
𝑠𝑟𝑐

is terminated by 𝑒𝑟𝑟𝑜𝑟. In this case, the network engineer must reduce 32: 𝑟𝑒𝑞𝑢𝑖𝑟𝑒𝑑𝑇 𝑆 ← 𝑇𝑐𝑦𝑐𝑙𝑒 ∕𝑇 𝑆𝐴𝐼𝑖
the input 𝑓𝑖 by reference to the maximum bandwidth constraint. 𝑇
33: 𝑠𝑙𝑜𝑡𝐼𝑛𝑡𝑒𝑟𝑣𝑎𝑙 ← (𝑇 𝑆𝑇 𝐶𝐷 × 𝑇𝑐𝑦𝑐𝑙𝑒
𝐷𝐼
)∕𝑟𝑒𝑞𝑢𝑖𝑟𝑒𝑑𝑇 𝑆
Note that all 𝑆𝑊 s involved in the same subnetwork share the 34: 𝑎𝑙𝑙𝑜𝑐𝑎𝑡𝑒𝐶𝑁𝑇 ← 0
parameters calculated in Steps 1 and 2. As no subnetwork contains 35: for 𝑎𝑙𝑙𝑜𝑐𝑎𝑡𝑒𝐶𝑁𝑇 < 𝑟𝑒𝑞𝑢𝑖𝑟𝑒𝑑𝑇 𝑆 do
a branch, the numbers of input and output flows are the same. In 36: 𝑎𝑣𝑎𝑖𝑙𝑎𝑏𝑙𝑒𝑇 𝑆[𝑖𝑛𝑑𝑒𝑥𝐴𝑣𝑎𝑙𝑖𝑎𝑏𝑙𝑒𝑇 𝑆] ← 𝑓𝑖
other words, the calculation (lines 1–15) is only executed once for 37: 𝑎𝑙𝑙𝑜𝑐𝑎𝑡𝑒𝐶𝑁𝑇 ← 𝑎𝑙𝑙𝑜𝑐𝑎𝑡𝑒𝐶𝑁𝑇 + 1
each subnetwork; this dramatically reduces the time needed to schedule 38: 𝑖𝑛𝑑𝑒𝑥𝐴𝑣𝑎𝑙𝑖𝑎𝑏𝑙𝑒𝑇 𝑆 ← 𝑖𝑛𝑑𝑒𝑥𝐴𝑣𝑎𝑙𝑖𝑎𝑏𝑙𝑒𝑇 𝑆 +
𝑠𝑙𝑜𝑡𝐼𝑛𝑡𝑒𝑟𝑣𝑎𝑙
traffic.
39: end for
Step 3 of Traffic Scheduling is implemented by lines 16–41. This
40: end for
allocates time slots for each 𝑓𝑖 on all switches of the subnetwork. Line 41: end for
𝑠𝑟𝑐
16 initializes a tuple 𝐹 𝑀𝑇 𝐼𝑖 𝑖 of size 𝑛𝑢𝑚𝑓 𝑙𝑜𝑤. Each element of the // Step 4: Output the results
tuple presents a startup instant, at which time an 𝑓𝑖 is transmitted 42: Generate GCLs.
from a TSN talker. Lines 17–41 maintain a loop that individually
schedules the 𝑓𝑖 over 𝑛𝑢𝑚𝑓 𝑙𝑜𝑤 iterations. Each iteration contains two
sub-loops. The first is that of lines 21–40, which traverses every switch
in subnetwork 𝐺′ so that an 𝑓𝑖 can be scheduled on all switches. The time slot is the 𝑛th TS in the 𝑚th TDI. In line 23, 𝑚 is determined
second subloop (lines 35–39) guarantees that all the time slots required by rounding up the quotient of 𝑖𝑛𝑑𝑒𝑥𝐴𝑣𝑎𝑙𝑖𝑎𝑏𝑙𝑒𝑇 𝑆 and 𝑇 𝑆𝑇 𝐶𝐷. 𝑛 is
𝑆𝑊
by a 𝑓𝑖 are completely allocated at egress of a specific 𝑆𝑊𝑘 . The the remainder of 𝑖𝑛𝑑𝑒𝑥𝐴𝑣𝑎𝑙𝑖𝑎𝑏𝑙𝑒𝑇 𝑆 and 𝑇 𝑆𝑇 𝐶𝐷. Next, 𝐹 𝑀𝑇 𝐼𝑖 𝑘 is
𝑠𝑟𝑐
procedure will now be explained line by line. As mentioned earlier, the calculated by Eq. (15). The value of 𝐹 𝑀𝑇 𝐼𝑖 𝑖 is temporarily saved by
𝑠𝑟𝑐
time slots are preferentially allocated to 𝑓𝑖 with smaller 𝜑′𝑖 . Therefore, the auxiliary variable 𝑓 𝑚𝑡𝑖𝑖 𝑖 in line 26.
the 𝑓𝑖 with the smallest value of 𝜑′𝑖 will be popped from 𝐹 ′ using the From lines 27–30, any transmission conflict at talker 𝑠𝑟𝑐𝑖 is checked
𝑠𝑟𝑐
heappop method. Then, the network delay experienced by 𝑓𝑖 and the and avoided. Specifically, line 27 checks if any element in 𝐹 𝑀𝑇 𝐼𝑗 𝑖
𝑠𝑟𝑐𝑖
length of 𝑁𝑇 𝐶𝐼1𝑘 can be determined. Next, for a specific 𝑆𝑊𝑘 in 𝐺′ , is equal to 𝑓 𝑚𝑡𝑖𝑖 which is computed in line 26. If so, transmission
we search for a slot for that 𝑓𝑖 . In particular, by using the 𝐹 𝑖𝑛𝑑𝐼𝑛𝑑𝑒𝑥() conflict is present; to resolve this, 𝑖𝑛𝑑𝑒𝑥𝐴𝑣𝑎𝑙𝑖𝑎𝑏𝑙𝑒𝑇 𝑆 is updated to plus
method, the first occurrence of −1 (which indicates that a time slot one, implying that the time slot initially allocated to 𝑓𝑖 is shifted to
is available) in the tuple 𝑎𝑣𝑎𝑖𝑙𝑎𝑏𝑙𝑒𝑇 𝑆 is sought. This is then updated the next available slot. Then, rescheduling is triggered by executing the
to 𝑖𝑛𝑑𝑒𝑥𝐴𝑣𝑎𝑙𝑖𝑎𝑏𝑙𝑒𝑇 𝑆, i.e., to the index of the first available time slot. procedure commencing at line 22. If there is no transmission conflict,
𝑠𝑟𝑐 𝑠𝑟𝑐
Using the variable 𝑖𝑛𝑑𝑒𝑥𝐴𝑣𝑎𝑙𝑖𝑎𝑏𝑙𝑒𝑇 𝑆, we determine the specific po- 𝑓 𝑚𝑡𝑖𝑖 𝑖 is updated to 𝐹 𝑀𝑇 𝐼𝑖 𝑖 in line 31. Then, using the loop from
sition of the first available time slot in current 𝑇𝑐𝑦𝑐𝑙𝑒 . This position lines 35–39, a time slot is allocated to 𝑓𝑖 every 𝑠𝑙𝑜𝑡𝐼𝑛𝑡𝑒𝑟𝑣𝑎𝑙 slots.
can be represented by the indexes (i.e., 𝑚 and 𝑛) of the 𝑚th 𝑇 𝐷𝐼 𝑘 Allocation will be terminated when all time slots required by an 𝑓𝑖 are
𝑘
and 𝑛th 𝑇 𝑆 𝑘 in 𝑇𝑐𝑦𝑐𝑙𝑒 . In other words, in a 𝑇𝑐𝑦𝑐𝑙𝑒 , the first available allocated. To implement this, we must determine the input parameters

6
J. Jiang et al. Journal of Industrial Information Integration 33 (2023) 100464

(lines 32–34). First, we use the variable 𝑟𝑒𝑞𝑢𝑖𝑟𝑒𝑑𝑇 𝑆 to represent the Table 1
𝑇𝑐𝑦𝑐𝑙𝑒 TSN Traffic specifications.
number of time slots required by an 𝑓𝑖 , as determined by 𝑇 𝑆𝐴𝐼 . As
𝑖 𝐹 𝑠𝑟𝑐𝑖 𝑑𝑠𝑡𝑖 𝑙𝑖 𝑝𝑖 𝜑𝑖
mentioned earlier, a time slot is allocated to 𝑓𝑖 every 𝑠𝑙𝑜𝑡𝐼𝑛𝑡𝑒𝑟𝑣𝑎𝑙 slots.
𝑇𝑐𝑦𝑐𝑙𝑒
𝑇 𝑆𝑇 𝐶𝐷× 𝑇 𝐷𝐼 𝑇 𝑓1 Talker Epson C4 Listener Epson C4 1,500 B 1 ms 1 ms
We calculate 𝑠𝑙𝑜𝑡𝐼𝑛𝑡𝑒𝑟𝑣𝑎𝑙 as 𝑟𝑒𝑞𝑢𝑖𝑟𝑒𝑑𝑇 𝑆 , where (𝑇 𝑆𝑇 𝐶𝐷 × 𝑇𝑐𝑦𝑐𝑙𝑒 𝐷𝐼
) is 𝑓2 Talker Epson SCARA Listener Epson SCARA 1,500 B 1 ms 1 ms
the total number of time slots dedicated to transmission of the TC flows 𝑓3 Talker KUKA Listener KUKA 1,500 B 1 ms 1 ms
of a 𝑇𝑐𝑦𝑐𝑙𝑒 . Finally, these parameters are fed into the loop of lines 35– 𝑓4 Talker Turntable Listener Turntable 1,500 B 1 ms 1 ms

39, which allocates all of the time slots required by 𝑓𝑖 during 𝑇𝑐𝑦𝑐𝑙𝑒 . In
detail, the flow ID of 𝑓𝑖 is updated to the 𝑖𝑛𝑑𝑒𝑥𝐴𝑣𝑎𝑙𝑖𝑎𝑏𝑙𝑒𝑇 𝑆-th element Table 2
of the tuple 𝑎𝑣𝑎𝑖𝑙𝑎𝑏𝑙𝑒𝑇 𝑆. This means that the 𝑖𝑛𝑑𝑒𝑥𝐴𝑣𝑎𝑙𝑖𝑎𝑏𝑙𝑒𝑇 𝑆-th time GCLs obtained from the Cisco CNC.
slot is occupied by 𝑓𝑖 . Then, 𝑎𝑙𝑙𝑜𝑐𝑎𝑡𝑒𝐶𝑁𝑇 is updated to plus 1 indicat- Schedule (μs) 𝑆𝑟𝑐𝑖
𝑇𝑂𝑝𝑒𝑛 𝑆𝑟𝑐𝑖
𝑇𝐶𝑙𝑜𝑠𝑒
𝑆𝑊
𝑇𝑂𝑝𝑒𝑛1
𝑆𝑊
𝑇𝐶𝑙𝑜𝑠𝑒1
𝑆𝑊
𝑇𝑂𝑝𝑒𝑛2
𝑆𝑊
𝑇𝐶𝑙𝑜𝑠𝑒2
ing that time-slot allocation has been performed once. Following this, 𝑓1 0 14 35 49 70 84
the next time slot allocated to 𝑓𝑖 is determined by adding 𝑠𝑙𝑜𝑡𝐼𝑛𝑡𝑒𝑟𝑣𝑎𝑙 to 𝑓2 0 14 935 949 984 998
𝑖𝑛𝑑𝑒𝑥𝐴𝑣𝑎𝑙𝑖𝑎𝑏𝑙𝑒𝑇 𝑆. The loop will not be terminated until 𝑎𝑙𝑙𝑜𝑐𝑎𝑡𝑒𝐶𝑁𝑇 𝑓3 4 18 950 963 984 998
is equal to the 𝑟𝑒𝑞𝑢𝑖𝑟𝑒𝑑𝑇 𝑆, which means that time slots required by an 𝑓4 4 18 49 63 84 98
𝑓𝑖 are completely allocated. The TSN schedules (GCLs) are generated
in line 42 using the parameters calculated above. Table 3
Each interval allocated in one 𝑇 𝐷𝐼 corresponds to an entry of GCL. GCLs obtained from our traffic scheduling method.
For example, a 𝑇 𝐷𝐼 𝑘 usually contains four intervals (see Fig. 2), namely Schedule (μs) 𝑆𝑟𝑐𝑖
𝑇𝑂𝑝𝑒𝑛 𝑆𝑟𝑐𝑖
𝑇𝐶𝑙𝑜𝑠𝑒
𝑆𝑊
𝑇𝑂𝑝𝑒𝑛1
𝑆𝑊
𝑇𝐶𝑙𝑜𝑠𝑒1
𝑆𝑊
𝑇𝑂𝑝𝑒𝑛2
𝑆𝑊
𝑇𝐶𝑙𝑜𝑠𝑒2
𝑁𝑇 𝐶𝐼1𝑘 , 𝐺𝐵𝐼 𝑘 , 𝑇 𝐶𝐼 𝑘 , and 𝑁𝑇 𝐶𝐼2𝑘 ; hence, a 𝑇 𝐷𝐼 𝑘 corresponds to a 4- 𝑓1 0 12.336 33.74 46.076 67.48 79.816
entry GCL with different parameters of gate states and time intervals. 𝑓2 12.336 24.672 46.076 58.412 79.816 92.152
As discussed, TSN scheduling is repeated at intervals of 𝑇𝑐𝑦𝑐𝑙𝑒 𝑘 (hyper- 𝑓3 24.672 37.008 58.412 70.748 92.152 104.488
𝑘 𝑘 𝑓4 37.008 49.344 70.748 83.048 104.488 116.824
period), and one 𝑇𝑐𝑦𝑐𝑙𝑒 contains multiple same 𝑇 𝐷𝐼 . Therefore, the
length of GCL is equal to the number of GCL entries contained in one
𝑘
𝑇𝑐𝑦𝑐𝑙𝑒 , which can be determined as follows,
𝑘 4.3. Experimental deployment
𝑇𝑐𝑦𝑐𝑙𝑒
𝑙𝑒𝑛𝐺𝐶𝐿 = 𝑛𝑢𝑚𝐼𝑛𝑡𝑒𝑟𝑣𝑎𝑙 × (17)
𝑇 𝐷𝐼 𝑘 Based on the above, we constructed a real-world TSN-based process
where 𝑛𝑢𝑚𝐼𝑛𝑡𝑒𝑟𝑣𝑎𝑙 is the number of intervals allocated in one 𝑇 𝐷𝐼, automation system with the architecture shown in Fig. 3. The left side
𝑇𝑘
and 𝑇 𝑐𝑦𝑐𝑙𝑒
𝐷𝐼 𝑘
denotes the number of 𝑇 𝐷𝐼 𝑘 contained in one hyper-period. of Fig. 6 shows the entire system, and the right side shows the industrial
𝑘
𝑇 𝐷𝐼 𝑘 and 𝑇𝑐𝑦𝑐𝑙𝑒 can be determined by (3) and (9), respectively. TSN facilities. The system features three type-independent industrial
Next, we specify the queue allocation at each interval (NTCIs, robots (the Epson C4, Epson SCARA, and KUKA R540 robots) and one
GBI, and TCI). As discussed, TAS contains eight prioritized queues to turntable. The turntable conveys an iron plate, and two of the three
transmit TC and non-TC traffic. Wherein queues 6 and 7 are dedicated robots are selected in turns to execute successive pick-and-place tasks.
to managing TC traffic, the others (i.e., queues 0 to 5) are utilized to Meanwhile, the third robot executes the waiting task, which requires
manage non-TC traffic [9]. Therefore, we choose queue 7 to transmit the robot to wait for the trigger signal to switch the current task to the
TC traffic, and queues 0–5 to transmit non-TC traffic. During NTCIs, pick-and-place tasks. Notably, one of the main goals of this research
queues 6 and 7 are set to be closed, while the other queues are open is to validate the proposed TSM using this real industrial scenario.
for non-TC traffic. All queues are closed to prevent transmission conflict More information on the detailed implementation of the robotic process
during GBI. TC traffic will be transmitted during TCI; thus, queue 7 is automation applications can be found in [44]. These TC applications
opened, and others are closed. are executed by four TSN talkers that generate the TC flows controlling
the three robots and turntable. The traffic specifications are listed in
4.2. Configuration procedure Table 1. Specifically, four TC flows 𝑓𝑖 of equal length (𝑙𝑖 = 1,500 B)
are periodically generated by talkers (𝑝𝑖 = 1 ms). The delay constraint
Configuration is performed jointly by the CNC and CUC; the CNC
(𝜑𝑖 ) of 𝑓𝑖 is 1 ms, which indicates that 𝑓𝑖 must be transmitted to the
configures TSN 𝑆𝑊 s and the CUC configures 𝐸𝑆s. Fig. 5 shows the
corresponding listener within 1 ms. All physical links have the same
sequence. Six components are involved (colored boxes at the top). The
maximum bandwidth (1 Gbps), in other words, the queues carrying TC
blue arrows indicate information interaction between two objects. Self-
traffic at the egress of all TSN devices are of the same size. To verify our
messages are shown in blue; black and green indicate that information
TSM, we developed a traffic generator (TG) to establish a high traffic
is transferred from left to right and right to left, respectively. The
load by adjusting the transmission interval. The interference traffic
Configuration agent (CA) routine of CNC is triggered when the scheduled
𝑠𝑟𝑐 generated by the TG stresses the link between 𝑆𝑊1 and 𝑆𝑊2 , which
results (GCLs and 𝐹 𝑀𝑇 𝐼𝑖 𝑖 ) are obtained. CA executes two internal
𝑠𝑟𝑐 is shared by all traffic. Without effective scheduling, TC flows will
sub-routines: RESTful server to transfer 𝐹 𝑀𝑇 𝐼𝑖 𝑖 to the CUC, and net-
conf client to configure the GCLs as TSN 𝑆𝑊 s. Completion of scheduling be disturbed by massive interference traffic, compromising real-time
initializes the CUC and CNC responsible for TSN configurations. Next, service.
the RESTful server and netconf server are discovered by the RESTful client As shown on the right of Fig. 6, the device cabinet contains all
of the CUC and netconf client of the CNC, respectively. When the con- TSN equipment, which is split into three layers. On the top layer,
nections are confirmed, the CNC immediately configures the 𝑆𝑊 s via LS1028ardb boards [45] serve as the TSN 𝐸𝑆s that transmit or receive
two steps. First, pyang [43], a Python-based toolbox for the YANG data TC flows. In the middle, two Cisco switches (IE 4K) [46] serve as TSN
model, is used to transform the GCLs to XML-based configuration data. 𝑆𝑊 s switching both scheduled and unscheduled traffic according to
Second, the netconf client remotely configures the 𝑆𝑊 s by transmitting the TSN schedules. NUC mini-PCs are contained in the bottom layer,
configuration data to the netconf server at the target 𝑆𝑊 s. In parallel, wherein the CUC and CNC are executed. The CUC has a web-based user
𝑠𝑟𝑐
the RESTful server sends the scheduled 𝐹 𝑀𝑇 𝐼𝑖 𝑖 data to the CUC interface, as shown in the bottom right of Fig. 6; the network topology
via the POST method. When the CUC obtains the configuration data, is that of Fig. 3. Open-source software [47] was used as the TG and
the netconf client configures 𝐸𝑆s. The CUC configuration procedure is executed on an NUC PC to generate large amount of interference traffic.
identical to that of the CNC; CUC configuration will not be terminated Hence, we established a high traffic load scenario to verify scheduling
until all 𝐸𝑆s are configured. performance.

7
J. Jiang et al. Journal of Industrial Information Integration 33 (2023) 100464

Fig. 6. A real-world TSN-based process automation system.

Fig. 7. Illustration of the time-slot allocations of Tables 2 and 3.

𝑖𝑣 𝑖 𝑣
5. Performance evaluation and analysis of opening (𝑇𝑂𝑝𝑒𝑛 ) and closing (𝑇𝐶𝑙𝑜𝑠𝑒 ), respectively, of the timed gates
𝑠𝑟𝑐
that transmit 𝑓𝑖 . The orange arrow indicates the instant (𝐹 𝑀𝑇 𝐼𝑖 𝑖 )
We now examine the numerical results of the real-world TSN-based when 𝑓𝑖 is generated; the transmission window is then immediately
process automation system (Fig. 6). The system is configured by GCLs opened to send 𝑓𝑖 out.
𝑠𝑟𝑐
derived using our TSM and Cisco CNC software [48] (for comparison). As shown in Fig. 7(a), our TSM calculates 𝐹 𝑀𝑇 𝐼𝑖 𝑖 for each
Experiments were conducted on a PC with an Intel Core i7-8565U CPU 𝑓𝑖 , starting at 0 time-instant and increasing by 12.336 μs until the
and 16 GB of RAM. The TSM and Cisco CNC inputs were identical, so calculation is complete. Four dedicated time slots are continuously
the network topology 𝐺 = {(𝐸𝑆𝑖 , 𝑆𝑊𝑗 ) ∣ 𝑖 = 1, … , 8, 𝑗 = 1, 2} (see Fig. 3) allocated to 𝑓𝑖 to eliminate queuing delay and avoid transmission
𝑠𝑟𝑐
and traffic specifications 𝐹 are the same as those in Table 1. The dotted conflict. However, Cisco CNC randomly assigns an 𝐹 𝑀𝑇 𝐼𝑖 𝑖 [Fig. 7(b)]
arrows with different colors in Fig. 3 denote the forwarding paths for to 𝑓𝑖 at 𝑆𝑊1 , which causes significant queuing delay, although the
4 TSN flows (e.g., the purple dotted arrows indicate the forwarding time-slot allocations do not overlap. We discuss the queuing delay later.
path of 𝑓1 ), which are respectively input into TSM and Cisco CNC. The TSN listeners are connected to the individual ports of 𝑆𝑊2 , which
The outputs of the Cisco CNC and TSM are the GCLs summarized in means that the TSN flows 𝑓𝑖 do not compete for bandwidth resources.
Tables 2 and 3, respectively, which specify the schedules for 𝑓1 –𝑓4 . Hence, our TSM and Cisco CNC use the same policy to allocate time
𝑣𝑖
The tables contain the opening time-instants (𝑇𝑂𝑝𝑒𝑛 ) and closing time- slots at 𝑆𝑊2 ; an appropriate slot is allocated when 𝑓𝑖 is ready for
𝑣𝑖
instants (𝑇𝐶𝑙𝑜𝑠𝑒 ) for the prioritized queue. Specifically, at the egress of transmission.
𝑣𝑖 𝑣𝑖
node 𝑣𝑖 , transmission of 𝑓𝑖 will commence at 𝑇𝑂𝑝𝑒𝑛 and end at 𝑇𝐶𝑙𝑜𝑠𝑒 . We used the bandwidth allocation scheme in Section 3 to plot fine-
The remaining time is fully utilized to transmit non-TC flows. grained time slot allocation diagrams of 𝑆𝑊1 and 𝑆𝑊2 . As shown in
To facilitate understanding of the GCLs (Tables 2 and 3), the corre- Figs. 7(c) and 7(d), the slots appropriately transmitted both TC and
sponding time slot allocation diagrams are presented in Fig. 7. Specif- non-TC flows within a hyper-period (1,000 μs). Allocation was repeated
ically, as shown in Figs. 7(a) and 7(b), four timelines represent the at intervals of 1,000 μs. As shown in Fig. 7(c), the TSM accurately cal-
schedules for 𝑓𝑖 . Each timeline contains three boxes that indicate the culates the 𝑁𝑇 𝐶𝐼1𝑘 (blue boxes) and 𝐺𝐵𝐼 𝑘 (gray boxes) using Eqs. (4)
time slot allocations at the TSN talkers (𝑠𝑟𝑐𝑖 ) and switches (𝑆𝑊1 and and (6). When 𝐺𝐵𝐼 𝑘 finishes, 𝑇 𝐶𝐼 𝑘 commences immediately, and
𝑆𝑊2 ). The rising and falling edges of each box are the time-instants successively allocates time slots (orange boxes) for transmission of 𝑓𝑖 .

8
J. Jiang et al. Journal of Industrial Information Integration 33 (2023) 100464

Fig. 8. End-to-end latency (𝑇𝑒2𝑒 ) comparison: TSM and Cisco CNC results.

This guarantees that TSN flow is timely; there is no queuing when 𝑆𝑊 s

egress. For example, 𝑓2 is generated at 12.336 μs and arrives at 𝑆𝑊1 ,
and a time slot is promptly allocated when 𝑓2 is ready for transmission
at 46.076 μs. The remaining slots 𝑁𝑇 𝐶𝐼21 (blue box) are fully utilized
by non-TC flows. However, as Fig. 7(d) shows, the Cisco CNC randomly
assigns non-overlapping time slots for 𝑓𝑖 . This avoids transmission
conflicts but introduces substantial queuing delays. As an example, 𝑓2
is generated at 0, and arrives at the egress of 𝑆𝑊1 after approximately
35 μs. However, 𝑓2 is then queued until a slot is allocated at 935 μs.
Please note that both TSM and Cisco CNC allocate all remaining slots
except 𝑇 𝐶𝐼 𝑘 and 𝐺𝐵𝐼 𝑘 to 𝑁𝑇 𝐶𝐼 𝑘 so that the bandwidth can be fully
utilized by non-TC traffic. Moreover, to fully eliminate the transmission
interference, both TSM and Cisco allocate specific slots (see orange
boxes marked with a red asterisk) to TC traffic flows at egresses of 𝑆𝑊2 ,
even TSN 𝐸𝑆s are attached to the separate egresses of 𝑆𝑊2 . According
to Eq. (17), the length of GCL configured to 𝑆𝑊1 is four (by TSM) and
seven (by Cisco CNC). As shown in Figs. 7(c) and 7(d), the 𝑛𝑢𝑚𝐼𝑛𝑡𝑒𝑟𝑣𝑎𝑙
of TSM is four [i.e., 𝑁𝑇 𝐶𝐼11 (0 μs to 21.404 μs), 𝐺𝐵𝐼 1 (21.404 μs to
Fig. 9. Comparison of scheduling performance with different delay constraints.
33.74 μs), 𝑇 𝐶𝐼11 (33.74 μs to 83.084 μs), and 𝑁𝑇 𝐶𝐼21 (83.084 μs to
1000 μs)], and Cisco CNC generate seven 𝑛𝑢𝑚𝐼𝑛𝑡𝑒𝑟𝑣𝑎𝑙, namely 𝑁𝑇 𝐶𝐼11
(0 μs to 22 μs), the first 𝐺𝐵𝐼 1 (22 μs to 35 μs), 𝑇 𝐶𝐼11 (35 μs to 63 μs),
and 𝑁𝑇 𝐶𝐼21 (63 μs to 922 μs), the second 𝐺𝐵𝐼 1 (922 μs to 935 μs), 𝑇 𝐶𝐼12 where green boxes denote a valid TSN schedule can be obtained,
(935 μs to 948 μs), and 𝑁𝑇 𝐶𝐼31 (984 μs to 1000 μs). 𝑇 𝐷𝐼 1 and 𝑇𝑐𝑦𝑐𝑙𝑒
1 are whereas red ones indicate a feasible schedule cannot be calculated.
equal to 1000 μs. Both TSM and Cisco CNC generate the GCLs with the Specifically, our TSM can properly schedule TSN traffic when 𝜑𝑖 varies
same length of four for 𝑆𝑊2 .
from 1 ms to 100 μs. On the other hand, Cisco CNC can only schedule
Then, we further evaluated the performance of TSM and Cisco CNC
TSN traffic when 𝜑𝑖 is higher than 200 μs. Once the 𝜑𝑖 becomes tight,
by using the same inputs but gradually reducing the delay constraint
i.e., less than 200 μs, Cisco CNC cannot obtain a valid schedule due
(𝜑𝑖 ) from 1 ms to 100 μs in steps of 10 μs. It is worth noting that
we assume that 𝜑𝑖 is equal to the transmission period (𝑝𝑖 ), so the to starckOverFlowError is thrown, which is caused by the limitation of
values of 𝑝𝑖 will also change as 𝜑𝑖 changes. Therefore, 100 sets of stack size. Afterwards we compared the performance of our TSM and
comparison experiments were conducted, and for each set, we executed the Cisco CNC based on three key indicators: the end-to-end latency
𝑆𝑊
TSM and Cisco CNC to calculate the schedules. Fig. 9 plots the results, 𝑇𝑒2𝑒 , queuing delay 𝑇𝑞𝑢𝑒𝑓𝑘 , and calculation time 𝑇𝑐𝑎𝑙 .
𝑖

9
J. Jiang et al. Journal of Industrial Information Integration 33 (2023) 100464

Table 4
Queuing delay results (𝜇s)
Scheduler 𝑆𝑊1
𝑇𝑞𝑢𝑒 𝑆𝑊1
𝑇𝑞𝑢𝑒 𝑆𝑊1
𝑇𝑞𝑢𝑒 𝑆𝑊1
𝑇𝑞𝑢𝑒 𝑆𝑊2
𝑇𝑞𝑢𝑒 𝑆𝑊2
𝑇𝑞𝑢𝑒 𝑆𝑊2
𝑇𝑞𝑢𝑒 𝑆𝑊2
𝑇𝑞𝑢𝑒
𝑓1 𝑓2 𝑓3 𝑓4 𝑓1 𝑓2 𝑓3 𝑓4

TSM 0 0 0 0 0 0 0 0
Cisco CNC 1.26 901.26 912.26 11.26 1.26 15.26 0.26 1.26

5.1. End-to-end latency

The end-to-end latency (𝑇𝑒2𝑒 ) is the transmission time of 𝑓𝑖 from a

talker (𝑠𝑟𝑐𝑖 ) to a listener (𝑑𝑠𝑡𝑖 ). 𝑇𝑒2𝑒 is critical for evaluating the real-
time performance of industrial systems. We used ProfiShark 1G+ delay
measurement devices to measure 𝑇𝑒2𝑒 . Two devices were respectively
deployed at the egress of the TSN talker and the ingress of the TSN
listener. 𝑓𝑖 was captured and timestamped 𝑇𝑠𝑟𝑐𝑖 when it was transmitted
by the talker. Then, 𝑓𝑖 was timestamped again when it reached the
ingress of the TSN listener (𝑇𝑑𝑠𝑡𝑖 ). Hence, 𝑇𝑒2𝑒 is expressed as:

𝑇𝑒2𝑒 = 𝑇𝑑𝑠𝑡𝑖 − 𝑇𝑠𝑟𝑐𝑖 (18)

Fig. 10. Calculation time (𝑇𝑐𝑎𝑙 ) of TSM and Cisco CNC for real-world testbed (shown
We measured the 𝑇𝑒2𝑒 of the TSN testbed in Section 4.3. A total of in Fig. 6).

1,000 packets were captured during the peak traffic period, i.e., when
the TG generated voluminous jamming traffic. In such a case, the
shared link between 𝑆𝑊1 and 𝑆𝑊2 was completely occupied. The
numerical 𝑇𝑒2𝑒 results are plotted in Fig. 8; the orange and green
lines are the 𝑇𝑒2𝑒 values of 𝑓𝑖 scheduled by our TSM and Cisco CNC,
respectively, and the red lines are the 𝜑𝑖 values. No measured 𝑇𝑒2𝑒
exceeds the 𝜑𝑖 ; both the TSM and Cisco CNC appropriately schedule 𝑓𝑖
to satisfy real-time performance requirements. The 𝑇𝑒2𝑒 results of TSM
and Cisco CNC in Figs. 8(a) and 8(d) are quite similar. Because at the
egress of 𝑆𝑊1 , Cisco CNC allocations for 𝑓1 (from 35 μ𝑠 to 49 μ𝑠) and
𝑓4 (from 49 μ𝑠 to 63 μ𝑠) are coincidentally similar to those of our TSM,
i.e., 𝑓1 (from 33.74 μ𝑠 to 46.076 μ𝑠) and 𝑓4 (from 70.748 μ𝑠 to 83.084
μ𝑠). But those of Figs. 8(b) and 8(c) differ significantly because TSM
eliminates queuing delays whereas Cisco CNC does not (see below).

5.2. Queuing delay

𝑆𝑊
The queuing delay 𝑇𝑞𝑢𝑒𝑓𝑘 is the time over which 𝑓𝑖 is ‘‘stuck’’ in 𝑆𝑊𝑘 ,
𝑖
and is expressed as:
𝑆𝑊 𝑆𝑊 𝑠𝑟𝑐 𝑇 𝑜𝑆𝑊𝑘 𝑠𝑟𝑐𝑖 𝑆𝑊
𝑇𝑞𝑢𝑒𝑓𝑘 = 𝑇𝑂𝑝𝑒𝑛𝑘 − 𝐷𝑝ℎ𝑦𝑖 − 𝐹 𝑀𝑇 𝐼𝑖 − 𝑇𝑞𝑢𝑒𝑓𝑘−1 (19)
𝑖 𝑖

𝑆𝑊
where 𝑇𝑂𝑝𝑒𝑛𝑘 is the instant when the gate opens to transmit 𝑓𝑖 (i.e., at the
𝑠𝑟𝑐 𝑇 𝑜𝑆𝑊
egress of 𝑆𝑊𝑘 ). 𝐷𝑝ℎ𝑦𝑖 𝑘
is the physical delay when 𝑓𝑖 is transmitted
𝑠𝑟𝑐
from 𝑠𝑟𝑐𝑖 to 𝑆𝑊𝑘 . 𝐹 𝑀𝑇 𝐼𝑖 𝑖 is the start time of 𝑓𝑖 transmission at
𝑆𝑊
talker 𝑠𝑟𝑐𝑖 . These three parameters were calculated above. 𝑇𝑞𝑢𝑒𝑓𝑘−1 is the
𝑖
queuing delay experienced by 𝑓𝑖 during its previous hop. We assume
𝑆𝑊0 𝑆𝑊𝑘
that 𝑇𝑞𝑢𝑒𝑓 is 0 when 𝑘 is 1. A larger 𝑇𝑞𝑢𝑒𝑖 results in a greater end-to-end
𝑖
latency; 𝜑𝑖 may be exceeded and real-time performance is not possible.
Using Eq. (19), we calculated the queuing delays and summarize them
in Table 4. TSM eliminates these delays by carefully scheduling 𝑓𝑖 , but
the use of Cisco CNC is associated with significant delays; in particular,
𝑆𝑊 𝑆𝑊
𝑇𝑞𝑢𝑒𝑓1 and 𝑇𝑞𝑢𝑒𝑓1 are both about 900 μs.
2 3

5.3. Calculation time

The calculation time 𝑇𝑐𝑎𝑙 is the time between scheduler startup and Fig. 11. Calculation time (𝑇𝑐𝑎𝑙 ) comparison: TSM and Cisco CNC.
calculation of a feasible TSN schedule. 𝑇𝑐𝑎𝑙 is important for operational
performance in real industrial systems. Changes in product service re-
quirements necessitate adjustments in service-oriented applications and CNC 100 times to schedule 𝑓𝑖 on the TSN testbed. We plot 𝑇𝑐𝑎𝑙 results
network infrastructure. These changes may be frequent and the TSN
in Fig. 10, where x-axis denotes the running number, whereas Y-axis
schedule must be recalculated every time. Therefore, a TSN scheduler
presents the calculation time in seconds. Orange and green lines denote
with a high 𝑇𝑐𝑎𝑙 is unacceptable. To measure 𝑇𝑐𝑎𝑙 , we used time, a
popular Linux tool, which returns the elapsed time between invocation 𝑇𝑐𝑎𝑙 of Cisco CNC and our TSM, respectively. Black dotted lines show
and termination of a program [49]. We used both TSM and Cisco the average values, i.e., 12.3 ms and 15.6 s, which are the averaged 𝑇𝑐𝑎𝑙

10
J. Jiang et al. Journal of Industrial Information Integration 33 (2023) 100464

Table 5 In future, we will extend our TSM to schedule incremental traf-

Calculation time (𝑇𝑐𝑎𝑙 ) comparison: TSM and Cisco CNC. fic during system runs, while guaranteeing configuration consistency
Flow numbers Our TSM (ms) Cisco CNC (s) x-fold (i.e., coexistence of old and new configurations) during configuration
10 30.94 17.69 571-fold updates. This will meet the flexible configuration requirements of vari-
20 32.69 21.86 668-fold ous industry 4.0 applications (e.g., plug-and-produce application [44]).
30 31.68 26.15 825-fold
40 31.91 30.45 954-fold
50 33.35 33.82 1014-fold CRediT authorship contribution statement
60 34.00 36.04 1060-fold
70 35.53 40.20 1131-fold Junhui Jiang: Investigation, Methodology, Software, Validation,
80 37.83 43.81 1158-fold Formal analysis, Writing – original draft. Yuting Li: Methodology,
90 38.57 47.96 1243-fold
Software, Writing – review & editing. Xiongfeng Zhang: Resources,
100 38.78 56.45 1455-fold
110 38.31 82.78 2160-fold Visualization. Mengmeng Yu: Resources, Writing – review & editing.
120 39.72 105.41 2653-fold Chang Dae Lee: Software, Resources. Seung Ho Hong: Conceptualiza-
tion, Supervision, Writing – review & editing, Project administration,
Funding acquisition.

values of TSM and Cisco CNC, respectively. Our TSM was much faster Declaration of competing interest
than Cisco CNC.
We also evaluated the effect of the number of input TSN flows on The authors declare that they have no known competing finan-
𝑇𝑐𝑎𝑙 . The input setups (network configuration and traffic specifications) cial interests or personal relationships that could have appeared to
for TSM and Cisco CNC are identical and specified below. We input 𝑓𝑖 influence the work reported in this paper.
flows, where 𝑖 ranged from 10 to 500 (step size = 10), and generated
50 sets of cases. Each set contained 10 cases of different topologies, Data availability
each of which was of the same network size (20 𝑆𝑊 s and 60 𝐸𝑆s)
but exhibited random connections. Hence, 500 cases were ultimately The authors do not have permission to share data.
generated. We used NetworkX [50] to create the random topologies.
The talkers 𝑠𝑟𝑐𝑖 and listeners 𝑑𝑠𝑡𝑖 were randomly distributed among Acknowledgment
the 𝐸𝑆s. For each case, using the same traffic specification, TSM and
Cisco CNC were deployed to calculate the TSN schedules, respectively. This work was supported by Institute for Information & communi-
The flow size is 1,500 Bytes, flow period 𝑝𝑖 was randomly selected cations Technology Promotion(IITP) grant funded by the Korea govern-
from 4, 8, and 16 ms, and the deadline 𝜑𝑖 was equal to the period. ment(MSIP) (No.2020-0-00299, Development & Proof of Open Manu-
facturing Operation System with ICT Convergence).
Fig. 11 plots the 𝑇𝑐𝑎𝑙 results, where the x-axis denotes the number of
flows and the y-axis denotes 𝑇𝑐𝑎𝑙 , which increases with the number of
References
flows. Specifically, Fig. 11(a) shows that the TSM can schedule 500
flows within 137 ms. Fig. 11(b) shows that Cisco CNC schedules 120 [1] S. Schriegel, J. Jasperneite, A migration strategy for profinet toward ethernet
flows in about 110 s. When the number of flows reaches 130, the Cisco TSN-based field-level communication: An approach to accelerate the adoption
CNC process is terminated because the starckOverFlowError switch is of converged IT/OT communication, IEEE Ind. Electron. Mag. (2021) http:
//dx.doi.org/10.1109/MIE.2020.3048925.
thrown; this is caused by the limitation of the stack size. To intuitively
[2] IEEE Standard for Ethernet, IEEE Standard, 2018, 802.3-2018.
show how fast our TSM is than Cisco CNC, we present Table 5 to show [3] IEEE 802.1. Time-Sensitive Networking (TSN). [Online]. Available: https://1.
specific values of 𝑇𝑐𝑎𝑙 when flow numbers ranging from 10 to 120. We ieee802.org/tsn/.
divide 𝑇𝑐𝑎𝑙 of Cisco CNC by that of our TSM and present this value [4] IEEE standard for local and metropolitan area networks – Timing and synchro-
nization for time-sensitive applications, 2018, IEEE Standard 802.1AS-Rev/D7.2.
(i.e., x-fold) in the third column. For example, to schedule 10 TSN flows [5] IEEE standard for local and metropolitan area networks – Bridges and bridged
(see the first row of Table 5), our TSM needs 30.94 ms, while Cisco CNC networks - Amendment 25: Enhancements for scheduled traffic, 2016, IEEE
17.69𝑠
requires 17.69 s Thus, our TSM is at least 571-fold ( 30.94𝑚𝑠 ) faster than Standard 802.1Qbv-2015, Mar.
Cisco CNC. Moreover, when the network is fixed, our TSM can schedule [6] IEEE standard for local and metropolitan area networks–Frame replication and
elimination for reliability, 2017, IEEE Standard 802.1CB.
more traffic flows. [7] IEEE standard for local and metropolitan area networks–Stream Reservation Pro-
tocol (SRP) enhancements and performance improvements, 2018, IEEE Standard
802.1Qcc.
6. Conclusion
[8] D. Bruckner, M.P. Stănică, R. Blair, et al., An introduction to OPC UA TSN for
industrial communication systems, Proc. IEEE 107 (6) (2019) 1121–1131.
To bridge the gap between theoretical research and practical imple- [9] Time-sensitive networking profile for industrial automation. IEC/IEEE Standard,
2018, 60802/rev6p0.
mentation, we deployed a TSM for a practical TSN system. We present
[10] IEEE Standard. P802.1DG – TSN profile for automotive in-vehicle ethernet
fine-grained guidelines that aid practical implementation of TSM. We communications, 2020, [Online]. Available: https://1.ieee802.org/tsn/802-1dg/.
constructed a real-world TSN-based process automation system and [11] IEEE Standard. P802.1DP – TSN for Aerospace Onboard Ethernet Communica-
specified the technical details. Finally, we experimentally studied TSM tions. [Online]. Available: https://1.ieee802.org/tsn/802-1dp/.
[12] D.B. Mateu, D. Hallmans, M. Ashjaei, A.V. Papadopoulos, J. Proenza, T. Nolte,
performance by measuring the end-to-end latency, queuing delay, and Clock synchronization in integrated TSN-EtherCAT networks, in: 2020 25th IEEE
calculation time, and compared the TSM data to those of the Cisco CNC. International Conference on Emerging Technologies and Factory Automation,
Our conclusions are as follows: ETFA, 2020, pp. 214–221, http://dx.doi.org/10.1109/ETFA46521.2020.9212153.
[13] A.M. Romanov, F. Gringoli, A. Sikora, A precise synchronization method for
future wireless TSN networks, IEEE Trans. Ind. Inf. 17 (5) (2021) 3682–3692,
(1) Our TSM guarantees robust real-time performance for TC traffic
http://dx.doi.org/10.1109/TII.2020.3017016.
flows, even when traffic loads are high; [14] M. Gutierrez, W. Steiner, R. Dobrin, S. Punnekkat, Synchronization quality of
(2) TSM eliminates queuing delays of TC traffic flows because IEEE 802.1as in large-scale industrial automation networks, in: Proc. IEEE RTAS,
scheduling is fine-grained; IEEE, Pittsburgh, PA, USA, 2017.
[15] J. Jiang, Y. Li, S.H. Hong, A. Xu, K. Wang, A Time-Sensitive Networking (TSN)
(3) When presented with 500 random cases, our TSM required only a
simulation model based on OMNET++, in: 2018 IEEE International Conference
very short computation time to generate a feasible schedule. The on Mechatronics and Automation, ICMA, 2018, pp. 643–648, http://dx.doi.org/
TSM was about 571-fold faster than the Cisco CNC. 10.1109/ICMA.2018.8484302.

11
J. Jiang et al. Journal of Industrial Information Integration 33 (2023) 100464

[16] J. Jiang, Y. Li, S.H. Hong, M. Yu, A. Xu, M. Wei, A simulation model for Time- [41] J. Kurose, K. Ross, Computer Networking: A TOP-DOWN APPROACH, seventh
Sensitive Networking (TSN) with experimental validation, in: 2019 24th IEEE ed., Pearson, 2017, pp. 62–68.
International Conference on Emerging Technologies and Factory Automation, [42] Michael L. Fredman, Robert Endre Tarjan, Fibonacci heaps and their uses in
ETFA, 2019, pp. 153–160, http://dx.doi.org/10.1109/ETFA.2019.8869206. improved network optimization algorithms, J. ACM 34 (3) (1987) 596–615.
[17] Y. Li, J. Jiang, C. Lee, S.H. Hong, Practical implementation of an OPC UA TSN [43] pyang. PyPI. [Online]. Available: https://pypi.org/project/pyang/.
communication architecture for a manufacturing system, IEEE Access 8 (2020) [44] X. Ye, J. Jiang, C. Lee, N. Kim, M. Yu, S.H. Hong, Toward the plug-and-
200100–200111, http://dx.doi.org/10.1109/ACCESS.2020.3035548. produce capability for industry 4.0: An asset administration shell approach, IEEE
[18] N. Sertbaş Bülbül, D. Ergenç, M. Fischer, SDN-based self-configuration for Ind. Electron. Mag. 14 (4) (2020) 146–157, http://dx.doi.org/10.1109/MIE.2020.
time-sensitive IoT networks, 2021, arXiv e-prints, arXiv-2103. 3010492.
[19] X. Lu, Q. Xu, J. Zhang, C. Chen, Joint design of control and transmission for [45] Layerscape LS1028A reference design board, 2023, [Online]. Available:
industrial CPS under time sensitive networking, in: IECON 2022–48th Annual https://www.nxp.com/design/qoriq-developer-resources/layerscapels1028a-
Conference of the IEEE Industrial Electronics Society, Brussels, Belgium, 2022, reference-design-board:LS1028ARDB. (Accessed 17 January 2023).
pp. 1–6, http://dx.doi.org/10.1109/IECON49645.2022.9968459. [46] Cisco Industrial Ethernet 4000 Series Switches, 2023, [Online]. Available:
[20] R.S. Oliver, S.S. Craciunas, W. Steiner, IEEE 802.1 Qbv gate control list synthesis https://www.cisco.com/c/en/us/products/switches/industrialethernet-4000-
using array theory encoding, in: 2018 IEEE Real Time and Embedded Technology series-switches/index.html. (Accessed: 17 January 2023).
and Applications Symposium, RTAS, IEEE, 2018, pp. 13–24. [47] S.P. Ostinato Packet Generator, 2021, [Online]. Available: https://ostinato.org/.
[21] X. Jin, C. Xia, N. Guan, P. Zeng, Joint algorithm of message fragmentation and
[48] CISCO, Time-Sensitive Networking: A Technical Introduction White Paper, 2017,
no-wait scheduling for time-sensitive networks, IEEE/CAA J. Autom. Sinica 8 (2)
[Online]. Available: https://www.cisco.com/c/dam/en/us/solutions/collateral/
(2021) 478–490, http://dx.doi.org/10.1109/JAS.2021.1003844.
industry-solutions/white-paper-c11-738950.pdf.
[22] Q. Yu, H. Wan, X. Zhao, Y. Gao, M. Gu, Online scheduling for dynamic VM
[49] time(1), Linux manual page, 2021, [Online]. Available: https://man7.org/linux/
migration in multicast time-sensitive networks, IEEE Trans. Ind. Inf. 16 (6)
man-pages/man1/time.1.html.
(2020) 3778–3788, http://dx.doi.org/10.1109/TII.2019.2925538.
[50] NetworkX — NetworkX documentation. [Online]. Available: https://networkx.
[23] N.G. Nayak, F. Dürr, K. Rothermel, Incremental flow scheduling and routing
org/.
in time-sensitive software-defined networks, IEEE Trans. Ind. Inf. 14 (5) (2018)
2066–2075, http://dx.doi.org/10.1109/TII.2017.2782235.
[24] A.A. Atallah, G.B. Hamad, O.A. Mohamed, Routing and scheduling of time-
triggered traffic in time-sensitive networks, IEEE Trans. Ind. Inf. 16 (7) (2020) Junhui Jiang received the B.S. degree in electronic information engineering from the
4525–4534, http://dx.doi.org/10.1109/TII.2019.2950887. Harbin University of Science and Technology, Harbin, China, in 2017, and the Ph.D.
[25] J. Falk, F. Dürr, K. Rothermel, Exploring practical limitations of joint routing degree in electronic engineering from Hanyang University, Seoul, South Korea, in 2022.
and scheduling for TSN with ILP, in: 2018 IEEE 24th International Conference He is currently an Assistant Professor with the Modeling and Emulation in
on Embedded and Real-Time Computing Systems and Applications, RTCSA, 2018, E-Government National Engineering Laboratory, College of Computer Science and
pp. 136–146, http://dx.doi.org/10.1109/RTCSA.2018.00025. Technology, Harbin Engineering University, Harbin. His research interests include
[26] Y. Zhang, Q. Xu, L. Xu, C. Chen, X. Guan, Efficient flow scheduling for industrial information and communication technologies in smart manufacturing and Industrial
Time-Sensitive Networking: A divisibility theory-based method, IEEE Trans. Ind. Internet of Things, with a special focus on time sensitive networking.
Inf. 18 (12) (2022) 9312–9323, http://dx.doi.org/10.1109/TII.2022.3151810. Dr. Jiang was a recipient of the Young Professionals Award from the IEEE
[27] Z. Pang, X. Huang, Z. Li, S. Zhang, Y. Xu, H. Wan, X. Zhao, Flow scheduling for International Conference on Emerging Technologies and Factory Automation in 2019.
conflict-free network updates in time-sensitive software-defined networks, IEEE
Trans. Ind. Inf. 17 (3) (2021) 1668–1678, http://dx.doi.org/10.1109/TII.2020.
2998224. Yuting Li received the B.S. degree in electronic information engineering from Wuhan
[28] L. Xu, Q. Xu, Y. Zhang, J. Zhang, C. Chen, Co-design approach of scheduling University of Science and Technology, Wuhan, China, in 2015, and the Ph.D. degree
and routing in time sensitive networking, in: IEEE Conference on Industrial from the Department of Electronic Engineering, Hanyang University, Seoul, South
Cyberphysical Systems, Vol. 1, ICPS, 2020. Korea, in 2021.
[29] L. Xu, et al., Learning-based scalable scheduling and routing co-design with He is currently an Assistant Professor with the School of Computer Science, Hubei
stream similarity partitioning for Time-Sensitive Networking, IEEE Internet University of Technology, Wuhan. His research interests include smart manufacturing,
Things J. 9 (15) (2022) 13353–13363, http://dx.doi.org/10.1109/JIOT.2022. Industrial Internet of Things, time-sensitive networking, and artificial intelligence.
3143829.
[30] M. Guo, C. Gu, S. He, Z. Shi, J. Chen, MSS: Exploiting mapping score for CQF
start time planning in Time-Sensitive Networking, IEEE Trans. Ind. Inf. 19 (2) Xiongfeng Zhang received the B.S. degree in automation from the Wuhan University
(2023) 2140–2150, http://dx.doi.org/10.1109/TII.2022.3206815. of Science and Technology, Wuhan, China, in 2016, and the Ph.D. degree in electronic
[31] A. Nasrallah, et al., Ultra-Low Latency (ULL) networks: The IEEE TSN and IETF engineering from Hanyang University, Seoul, South Korea, in 2022.
DetNet standards and related 5G ULL research, IEEE Commun. Surv. Tutor. 21 (1) His research interests include 5G communication, artificial intelligence (deep
(2019) 88–145, http://dx.doi.org/10.1109/COMST.2018.2869350, Firstquarter. reinforcement learning), smart grid (demand response), and smart manufacturing.
[32] Q. Li, D. Li, X. Jin, Q. Wang, P. Zeng, A simple and efficient Time-Sensitive
Networking traffic scheduling method for industrial scenarios, Electronics 9 (12)
(2020). Mengmeng Yu received the B.S. degree in communication engineering from the
[33] B. Houtan, M. Ashjaei, M. Daneshtalab, M. Sj¨odin, S. Mubeen, Synthesising Wuhan University of Technology, Wuhan, China, in 2008, the M.S. degree in detection
schedules to improve QoS of best effort traffic in TSN networks, in: International technique and automation equipment from the Chongqing University of Posts and
Conference on Real-Time Networks and Systems, RTNS, 2021. Telecommunications, Chongqing, China, in 2011, and the Ph.D. degree in electronic
[34] P.-J. Chaine, M. Boyer, C. Pagetti, F. Wartel, Egress-TT configurations for TSN engineering from Hanyang University, Seoul, South Korea, in 2015.
networks, in: International Conference on Real-Time Networks and Systems, She is currently a Principal Research Engineer with Nestfield Company Ltd.,
RTNS, 2022. Ansan, South Korea. From 2015 to 2018, she was a Postdoctoral Researcher under
[35] S.H. Hong, Y. Li, Central network configurator, and Time-Sensitive Networking the BK21 PLUS Program (BK21+) with Hanyang University, and from 2019 to 2021
control system including the same, 2020, PCT/KR2020/009308, PCT Patent she was a Research Assistant Professor with the Research Institute of Engineering
Application. and Technology, Hanyang University. Her research interests include carbon neutrality,
[36] S.H. Hong, Scheduling algorithm of data sampling times in the integrated sustainable manufacturing, smart grid, and game theory.
communication and control system, IEEE Trans. Contr. Syst. Technol. 3 (2) Dr. Yu was a recipient of the Outstanding Researcher Award of BK21+ and the
(1995) 225–230. Best Paper Award from the Workshop on Smart City Infrastructure and Applications
[37] M.H. Farzaneh, A. Knoll, Time-sensitive networking (TSN): An experimental in 2016. She is recognized as the Highly Cited Paper author (Web of Science top 1%
setup, in: 2017 IEEE Vehicular Networking Conference, VNC, 2017, pp. 23–26, most cited worldwide) from 2020 to 2022.
http://dx.doi.org/10.1109/VNC.2017.8275648.
[38] S. Chouksey, H.S. Satheesh, J. Åkerberg, An experimental study of TSN-
NonTSN coexistence, in: 2021 IEEE 11th Annual Computing and Communication Chang Dae Lee received the B.S. degree in electronic engineering from Kangwon
Workshop and Conference, CCWC, IEEE, 2021, pp. 1577–1584. National University, Kangwon, South Korea, in 2017, and the Ph.D. degree in electronic
[39] S.S. Craciunas, R.S. Oliver, W. Steiner, Demo abstract: Slate XNS–An online engineering from Hanyang University, Seoul, South Korea, in 2022.
management tool for deterministic TSN networks, in: 2018 IEEE Real-Time and He is currently a Postdoctoral Researcher with the Research Center of Electrical
Embedded Technology and Applications Symposium, RTAS, IEEE, 2018, pp. and Information Technology, Seoul National University of Science and Technology,
103–104. Seoul. His research interests include information and communication technologies and
[40] I. Ripoll, R. Ballester-Ripoll, Period selection for minimal hyperperiod in periodic Industrial Internet of Things, with special focuses of their applications in smart grid
task systems, IEEE Trans. Comp. 62 (9) (2013) 1813–1822. and smart manufacturing

12
J. Jiang et al. Journal of Industrial Information Integration 33 (2023) 100464

Seung Ho Hong ([email protected]) received the B.S. degree in mechanical of Standards and Technology, Gaithersburg, MD, USA, and the Vienna University of
engineering from Yonsei University, Seoul, South Korea, in 1982, the M.S. degree Technology, Vienna, Austria, and a Guest Professor with the Shenyang Institute of
in mechanical engineering from Texas Tech University, Lubbock, TX, USA, in 1985, Automation, Shenyang, China; the Chinese Academy of Sciences, Beijing, China; the
and the Ph.D. degree in mechanical engineering from Pennsylvania State University, Chongqing University of Posts and Telecommunications, Chongqing, China; and the
University Park, PA, USA, in 1989. Zhejiang University, Hangzhou, China. His research interests include industrial commu-
He is an Emeritus Professor with the Department of Electronic Engineering, nication networks, Industrial Internet of Things, smart manufacturing, cyber–physical
Hanyang University, Ansan, South Korea, and an Executive Advisor/Fellow of Nestfield systems, and artificial intelligence.
Company Ltd., Ansan. He was a Visiting Scholar with the U.S. National Institute

13