Time aware cyclic queuing forwarding traffic scheduling algorithm considering reallocation and priority sorting (RPS-TACQF)

Abstract

To meet the mixed traffic scheduling requirements of time-triggered (TT) streams and audio video bridging (AVB) streams in the industrial Internet, this paper proposes a time aware cyclic queuing forwarding traffic scheduling algorithm considering reallocation and priority sorting (RPS-TACQF) method. Combining time aware shaper (TAS) and cyclic queuing and forwarding (CQF) mechanisms, it dynamically adjusts the queue resource allocation and traffic forwarding order, and uses the dead time of CQF odd and even queue switching to schedule TAS, optimize the frame injection time slot and routing, while meeting their respective end-to-end delay and jitter requirements. The experimental results show that the RPS-TACQF algorithm significantly improves the scheduling success rate in scenarios with 50 TT data streams and 100–1000 AVB data streams, especially under high load conditions, where the scheduling success rate of AVB data streams is about 36.5% higher than other algorithms; The bandwidth utilization rate increased by about 36.5% under high load, and the algorithm execution time still saved about 31.4% compared to the TAMCQF-ILP solver algorithm under high load networks.

Introduction

Industrial Internet connects and communicates industrial equipment, production systems, sensors and other physical devices through Internet technology, and uses sensors, embedded systems, cloud computing platforms and data analysis technology¹. Data upload and control command issuance require millisecond level latency and microsecond level jitter. Industrial Internet needs to simultaneously carry multiple traffic such as time-triggered streams and audio video bridging streams. For communication service quality requirements, TT data streams require strict latency determinism, while AVB data streams have relatively relaxed requirements. Hybrid scheduling requires ensuring that different data flows do not interfere with each other in a shared network, while meeting their respective end-to-end latency and jitter requirements. Traditional Ethernet transmission modes cannot meet these requirements.

Time sensitive networking (TSN)² is a set of data link layer protocols developed by the IEEE 802.1 task force. It consists of a set of protocol clusters, including AS time synchronization, Qbv traffic scheduling, Qbu frame preemption, Qci per data stream filtering and supervision, and Qcc TSN configuration. It solves the three key problems of time synchronization, traffic scheduling, and system configuration. Qbv traffic scheduling is the core mechanism of TSN standard, which can meet the service quality requirements such as end-to-end latency and jitter of business traffic through technologies such as time synchronization, traffic shaping, and priority control. The TSN traffic scheduling mechanism is mainly implemented by traffic shapers such as the Time Awareness Shaper proposed by Qbv and the Circular Queuing and Forwarding proposed by Qch³. TAS achieves microsecond level traffic isolation through gate control lists (GCL), while CQF ensures fixed latency transmission through odd even cyclic queues. At present, the TAS gating list configuration for time sensitive networks is very complex, and CQF is difficult to solve the problem of mixed traffic scheduling. To address the above issues, researchers have proposed some new CQF extension schemes, such as Cycle Specific Queuing and Forwarding (CSQF) and Multiple CQF (Multi CQF)^4,5. Among them, CSQF introduced a third queue on the basis of the original two queues in CQF to receive data frames that did not arrive on time. Multi CQF uses multiple CQF queues for scheduling, with each CQF queue using a different cycle period. Although CSQF and Multi CQF can achieve mixed scheduling of AVB data streams with multiple cycles, they are difficult to solve the problem of mixed scheduling of TT data streams and AVB data streams. At present, people mostly use the combination of TAS and CQF to achieve mixed traffic scheduling of different types of traffic. However, TT data flow will affect the scheduling of AVB data flow, and the centralized scheduling of high priority TT data flow will significantly reduce the scheduling success rate of AVB data flow. In addition, the scheduling algorithm under this model does not incorporate routing for scheduling, and as network load increases, the success rate of scheduling significantly decreases⁶.

Although the existing research has made some progress in solving the mixed traffic scheduling problem of TT and AVB in the industrial Internet, there are still some shortcomings, such as the complex configuration of TAS gating list, the difficulty of CQF in solving the mixed traffic scheduling problem, and the limitations of the model combining TAS and CQF in terms of scheduling success rate, load balancing and routing. In this regard, this paper proposes a time aware circular queueing and forwarding traffic scheduling (RPS-TACQF) method that considers reassignment and prioritization. Research shows that the proposed RPS-TACQF algorithm performs well in terms of scheduling success rate, bandwidth utilization and algorithm execution time, and provides an efficient and reliable traffic scheduling solution for the industrial Internet. In addition, with the development of Software Defined Networking (SDN) technology, its application in cloud computing and Fog IoT networks is becoming increasingly widespread, providing new ideas for task scheduling and load balancing. The centralized control and flexible scheduling capabilities of SDN make it possible to achieve more efficient task allocation and resource utilization at the network level⁷. Especially when dealing with complex and changeable industrial Internet traffic, combining SDN’s task scheduling and load balancing methods, such as the combination strategy based on the Eagle Optimization Algorithm (AO) and the Whale Optimization Algorithm (WOA)⁸, is expected to further improve network performance and quality of service. Although this article does not directly explore the specific implementation in SDN environment, the idea of RPS-TACQF algorithm proposed can provide reference and inspiration for research in related fields.

Related work

Ethernet has gone through three stages of development: traditional Ethernet, real-time Ethernet, and deterministic networks⁹. Among them, traditional Ethernet adopts a shared medium and best effort communication mechanism, which has problems such as network congestion and packet collision; Although real-time Ethernet can perform specified tasks well in specific fields, due to the different technical routes for solving real-time problems, they form automated islands between each other. In addition, standard Ethernet and real-time Ethernet of various automation systems cannot be directly connected, which poses interconnection and interoperability issues, affecting the industrial expansion of the solution. With the expansion of network technology and application scenarios, various TSN based network technologies that provide deterministic guarantees have emerged, such as Deterministic Networking (Deterministic Net)¹⁰, Deterministic IP Network (DIP)¹¹, 5G time sensitive network¹², WiFi time sensitive network¹³, etc. Drawing on the architecture, capabilities, and mechanisms of TSN, deterministic network technology focuses on integrating with different network types to adapt to a wider range of business scenarios. Figure 1 shows the development history of TSN standard, including the release time of key protocols and functional division (time synchronization, traffic control, resource management, reliability). As shown in Fig. 1, TSNTG has released a series of standards and drafts to promote the development of time sensitive network technology. Between 2015 and 2021, key standards and drafts addressed critical issues in real-time communication and data transmission. Since 2017, in order to better meet the real-time communication and data transmission needs of different industries, TSNTG has developed a series of standards and drafts, focusing more on the definition and research of TSN in industrial applications. In 2019, TSNTG launched a technological update in the automotive field. These published and ongoing standards have formed a comprehensive standard system, covering various aspects such as TSN architecture, protocols, testing, and applications^14,15.

Fig. 1.

Development history of TSN.

The TSN frame structure is the basic unit of the TSN network, and Table 1 provides a detailed explanation of the meaning of each field in the TSN frame structure. Determine the priority of each data frame based on the Priority Code Point (PCP), cache it in the corresponding priority queue, and then use the corresponding shaping and scheduling strategies to complete the forwarding and scheduling of data frames¹⁶. The 802.1Q protocol implements the functionality of a virtual local area network (VLAN) by inserting VLAN tags into Ethernet frames. The VLAN tag contains two fields, TPID and TCI, where the TPID field is used to identify that this is an 802.1Q tag frame, and the TCI field is further subdivided into three sub fields: PCP, DEI, and VID. The PCP field is used to indicate the priority of the frame, the DEI field may not be directly used in some implementations, and the VID field is used to identify the VLAN to which the frame belongs. In this way, the 802.1Q protocol can achieve multiple logically isolated VLANs on a shared network infrastructure, improving network security and flexibility. The TSN protocol cluster can be divided into three levels, namely the base layer, intermediate layer, and application adaptation layer¹⁷. Among them, the basic layer includes time synchronization protocol, flow reservation protocol, and scheduling related protocol. The middle layer includes supporting protocols such as TSN resource management and configuration protocol, Link Layer Discovery Protocol (LLDP), and the definition of YANG model. The application adaptation layer includes various application configuration protocols, such as IEEE 802.1DG for TSN applications in the Internet of Vehicles, IEC/IEEE 60,802 for TSN applications in industrial automation, etc.

Table 1.

Frame field specification.

Field name	Length/bits	Meaning
Ethernet MAC Frame Structure
Preamble	7	Used for synchronization before data transmission
Start delimiter (SFD)	1	Start of identification frame
Destination MAC	6	MAC address of the target device
Source MAC Address	6	MAC address of the source device
Type/Length (Ethertype/Length)	2	Indicate the length or protocol type of subsequent data fields
Data (Payload)	42–1500	User data
Dun Check Sequence (FCS)	4	Cyclic redundancy check code, used for error detection
Intermission distance (IFG)	12	The minimum interval between frames, also known as the minimum delay interval in Ethernet
802.1Q VLAN Tag Field
Tag Protocol Identifier (TPID)	16	Identify this as an 802.1Q tag frame with a value of 0 × 8100
Tag Control Information (TCI)	16	Contains priority code points (PCP), discard eligibility indicators (DEI), and VLAN identifiers (VID)
- PCP	3	Priority code point, representing the priority of the frame, with a value range of 0–7
- DEI	1	Identify whether the MAC address is in classic format (this field may not be directly used in some implementations)
- VID	12	Identify VLAN and the virtual local area network to which the frame belongs

The hybrid TAS/CQF scheduling and joint routing scheduling of TSN have been widely studied, such as CSQF/Multi CQF and other TAS assisted CQF designs. Among them, (1) CSQF and Multi CQF: CSQF introduces a third queue on the basis of the original two CQF queues to receive data frames that have not arrived on time, thereby alleviating the mixed scheduling problem to a certain extent. However, CSQF still struggles to effectively handle mixed scheduling of TT and AVB streams. Multi CQF uses multiple CQF queues, each with a different period. Although it achieves multi period mixed scheduling of AVB flows, it lacks strict delay deterministic support for TT flows. (2) TAS assisted CQF design: Some studies attempt to combine TAS and CQF mechanisms to achieve strict delay control of TT flow through TAS, while utilizing CQF to ensure relaxed delay requirements of AVB flow. However, these designs often face the problem that TT flows affect AVB flow scheduling, especially under high load conditions, centralized scheduling of TT flows will significantly reduce the scheduling success rate of AVB flows. Compared with the above design, RPS-TACQF not only utilizes the advantages of TAS and CQF, but also significantly improves the transmission efficiency of time sensitive data through dead time utilization strategy. In addition, under the proposed GCL/guard-band configuration and assumptions, RPS-TACQF achieves interference free coexistence of TT flow and AVB flow by dynamically adjusting queue resource allocation and traffic forwarding order, significantly improving scheduling success rate and bandwidth utilization.

In addition, compared with adaptive TSN scheduling methods (such as the recently proposed dynamic priority allocation based on reinforcement learning¹⁸ or SDN controlled flow scheduling¹⁹), adaptive TSN can cope with dynamic traffic, but it relies on offline training or centralized control, which is difficult to meet the millisecond delay requirements of the industrial Internet. RPS-TACQF achieves dynamic resource allocation without the need for global model training through online traffic recognition and dead time reuse. Its lightweight design is more suitable for resource constrained industrial equipment.

TACQF system model

Network model

Time sensitive network (TSN) can be topologically abstracted as a directed graph model , where V is the node set and E is the edge set²⁰. Node set , where S and H are the sets of switches and terminal hosts in a time sensitive network, respectively. Edge set represents a tuple set of links in a time sensitive network, where each element in the set represents a unidirectional link. TSN adopts a full duplex Ethernet link, and the physical link between nodes and corresponds to two directed links, represented by and . Figure 2 shows the TSN network system, where all switches and terminals of the TSN network achieve global clock synchronization under the IEEE 802.1AS clock synchronization protocol; Divide the time scale into independent time slots, which can be numbered with a set of sequence numbers, represented as .

Fig. 2.

TSN network architecture diagram.

TSN network switches have multiple different ports, each connected to other terminals or switches²¹. There is a total of M switches in the TSN network, and the set of switches is represented as . Each TSN network switch has K ports, and the port set is represented as . Each port operates in full duplex mode with a port rate of B.

The RPS-TACQF algorithm mentioned in this article is based on a system architecture that combines centralized controllers with distributed switches. Among them, the centralized controller is responsible for monitoring the global network status, path calculation, time slot allocation, and GCL configuration, which are distributed to various switches as offline scheduling solutions. The switch dynamically adjusts queue resources and traffic forwarding order based on the received scheduling scheme using online data plane mechanisms (including matching action tables, ACL updates, and packet loops).

(1) Assumption: (1) All network nodes (including switches and terminal hosts) have achieved global clock synchronization through the IEEE 802.1AS protocol. (2) The centralized controller has sufficient computing power to handle the global network state and generate optimized scheduling schemes. (3) All switches support time aware shaper and cyclic queue forwarding mechanisms, and have the ability to dynamically update matching action tables and ACLs. (4) The traffic characteristics of all TT and AVB flows, such as period, maximum end-to-end delay, packet size, etc., are known before scheduling and provided as inputs to the centralized controller. (2) Output: The centralized controller generates and distributes scheduling information for each flow, including the complete path from source to destination, frame injection time slots for each switch port, and gate control list configuration for each port. (3) Perform the following operations online: (1) Update the matching action table to allocate queue resources. (2) Update ACL to identify latency sensitive traffic and throughput sensitive traffic. (3) Utilize packet looping mechanism to ensure correct queuing of data packets.

Traffic shaping methods

TSN network traffic is divided into TT data stream and AVB data stream²². Compared to TT data streams, AVB data streams have looser end-to-end latency and jitter requirements and can be further divided into Class A and Class B. Compared to A-class AVB data streams, B-class AVB data streams have a larger transmission interval. The TT data stream is defined as , where src and dst respectively represent the source and destination stations of the data stream, used for algorithm generation of corresponding routes. All feasible routes are stored in the path set pathList, where ddl is the maximum end-to-end delay, period is the transmission period of the TT data stream, J is the maximum jitter, and size is the total amount of data sent in one cycle. pathList is stored using a red black tree, sorted according to route weights, and the path array is represented as . The model definition of AVB data stream is similar to that of TT data stream, with more relaxed requirements for end-to-end latency and jitter²³.

The goal of this article’s traffic scheduling algorithm is to generate a scheduling scheme for TT and AVB data streams in the network, where slot represents the injection time at which the stream starts transmitting, and path is the final selected routing scheme. As shown in Fig. 3, in order to achieve deterministic delay transmission of traffic, TSN adopts queue and traffic shaping technology: at the outgoing port of the switch, which increase in order to distinguish traffic with different transmission requirements. Among them, Q_i and Q_i+1 complete the forwarding of traffic with the highest priority.

Fig. 3.

Traffic shaping of the output port queue of the switch.

In the cyclic queuing forwarding mechanism, the time scale is divided into independent time slots, and Q_i and Q_i+1 cooperate with each other in equal priority. By cyclically switching between the Q_i and Q_i+1 queues to list the entry gate states, the two queue states cyclically change between queuing and forwarding states, while traffic frames are only cached in the queue that is in queuing state. The frames received by the switch in the current time slot must be forwarded out in the next time slot; After the frame sent by the upstream node leaves the port, it must arrive at the downstream node in the same time slot. CQF can achieve deterministic bounded delay and jitter, expressed as follows:

1

Where, is the time slot, and h is the number of routing hops. The maximum delay jitter is two time slot sizes, namely J=.

Traffic scheduling model

The time aware cyclic queuing forwarding (TACQF) hybrid traffic scheduling model used in this article consists of queue Q_i+1 using a time aware shaper (TAS) shaper to transmit TT data streams, and queues Q_i and Q_i forming a parity queue of CQF shaper to transmit AVB data streams. TACQF switch is represented as , Where, represents the outbound port queue capacity from switch to . If there is no link between switch and , the value is 0; represents the queue capacity occupied by the output ports from switch to , and each column represents the occupancy of different queues in the same time slot; represents the gate control list, where each column represents the gate control of different queues in the same time slot, and each row represents the gate control of a queue in different time slots.

In the CQF model, the CQF odd even queue switches alternately. Figure 4a shows the dead time diagram of traditional CQF queue switching. As shown in Fig. 4a, consider the scenario where switch sends data frames to switch in the network. When a CQF odd even cycle ends and switches from an odd period queue to an even period queue, it will enable the reception of dual queue data frames and disable the reception of odd queue data frames. At this point, data frames sent by switch before the end of the odd cycle will not be received by switch ‘s odd queue, resulting in the loss of data frames. Therefore, during the CQF parity cycle switching, there will be a dead time period during which data frames cannot be transmitted.

Fig. 4.

Traffic scheduling model. (a) CQF timing diagram, (b) Cycle by cycle GCL schematic diagram.

Dead time mainly includes transmission delay and propagation delay. By increasing the CQF cycle duration and reducing the switching frequency of odd even queues, the proportion of dead time can be reduced. However, the latency of the data stream under the CQF shaper is directly proportional to the duration of the CQF cycle, and excessively long CQF cycles will cause the data stream to time out. Therefore, in most CQF cycles, the dead time will occupy 5% to 10% of the time.

In order to improve throughput and reduce packet loss rate under the CQF mechanism while meeting deterministic latency requirements, the optimization problem of jointly determining traffic forwarding modes and injecting frames into time slots is expressed as follows:

2

Where, represents whether the i-th traffic frame has been successfully scheduled, 1 represents success, and 0 represents failure.

In the RPS-TACQF algorithm, we explicitly designed Gate Control List (GCL) to support scheduling TT flows during the dead time of CQF parity queue switching. At the end of each CQF cycle, we precisely control the opening and closing time of the TAS gate to ensure that only TT flow is allowed to pass through during dead time, while AVB flow is blocked. This design avoids interference between TT flow and AVB flow under the proposed GCL/guard-band configuration and assumptions, while utilizing the dead time interval traditionally considered as invalid time. Figure 4b shows the GCL timing diagram of TT flow scheduling using dead time in the RPS-TACQF algorithm, and the specific operation process is as follows: (1) The cycle begins. All queues are in a ready state, but sending permissions are controlled by GCL. (2) The sending phase of the queue. The even queue (Q_i) sends AVB stream data according to the instructions of GCL. (3) Dead time interval. After the end of sending in the queue, enter the dead time interval. During this period, the TAS gate is open, allowing only TT streams to be sent through the TAS queue (Q_i+1). (4) Qi queue sending phase: After the dead time ends, Qi queue (Q_i−1) starts sending AVB stream data. (5) The cycle has ended. Prepare to enter the next CQF cycle and repeat the above process.

In order to ensure that under the proposed GCL/guard band configuration and assumption (which will be used as the basis for discussing TT and AVB data stream interference issues), TT and AVB data streams will not interfere with each other at the queue exchange boundary, this paper proposes a clear gating state model. (1) Definition of subinterval: Each CQF cycle is divided into multiple subintervals, including TT transmission subinterval, AVB transmission subinterval, and guard band subinterval. (2) TT transmission subinterval: TAS queue (Q_i+1) is activated, specifically for transmitting TT data streams. At the same time, CQF queues (Q_i−1 and Q_i) are closed to prevent interference with TT data flow. (2) AVB transmission subinterval: CQF queues (Q_i−1 and Q_i) are opened and closed based on parity alternation. During this period, the TAS queue (Q_i+1) was closed. (3) Protection tape interval: The protection tape is used to absorb time uncertainty caused by link transmission and propagation delay. (4) Protection band calculation: If the maximum transmission delay is and the maximum propagation delay is , then the protection band length G is calculated as . Reserve a protective strip of length at the boundary of each subinterval. (5) Link transmission and propagation delay boundary control: using priority queues and traffic shaping techniques to ensure that the transmission of high priority data streams is not affected by low priority data streams (AVB). During the network planning and configuration phase, the transmission rate and buffer size of data packets should be reasonably set based on the actual network topology and link characteristics.

Model resource constraints

(1) CQF cycle constraint. is the length of time from the opening of the CQF queue to the next opening of the CQF queue, constrained by Eqs. (3–4). can be divided by the period of all CQF data streams, i.e.:

3

Where, GCD represents the greatest common divisor, represents the transmission period of data stream , and k is a positive integer constant coefficient.

At the same time, all data frames in the port queue of the switch need to be transmitted from the previous node to the next node within the CQF cycle. If the data frame cannot be sent within the specified , it will be sent after the next cycle, which will cause confusion in the delay of the data frame and increase the jitter of the data stream; If the data frame fails to reach the next node within the specified after being sent, switching the parity queue will result in the data frame being discarded. The CQF cycle must be greater than or equal to the sum of the transmission delay and propagation delay of all data frames, that is:

4

Where, MTU represents the maximum transmission unit, b represents the link bandwidth, is the queue capacity, and is the propagation delay.

(2) Delay constraint. The data stream transmission delay should meet the limitations of the end-to-end delay requirement , that is:

5

Where, the latency of data stream is affected by the number of switch hops and injection time .

Due to constraints (3–4), the time taken from the previous node to the next node for each data frame is . Additionally, the initial transmission time of the data frame will be delayed by . The delay of data stream is the sum of the product of path hops and , and the injection time, that is:

6

Where, is the path length, and is the injection time of .

(3) TAS queue opening time constraint. is the length of time from the opening to the closing of a TAS queue, and should be less than or equal to and greater than or equal to the dead time when the CQF queue switches, that is:

7

Where, represents the CQF dead time.

The dead time is less than or equal to the time it takes for a data frame to be sent from the previous node to the next node for complete reception, that is:

8

At the same time, all data frames in the TAS queue need to be transmitted within one open time, that is:

9

Where, represents the capacity of the TAS queue.

(4) Switch queue resource constraints. The sum of the frames of data sent from all queues of switch to switch in time slot is:

10

Where, represents the occupied capacity of queue ab when the switch port is in time slot T; represents the gate control list configuration for queue ab when the switch port is in time slot T; The subscript occp represents the occupancy status of the queue; The subscript ab represents the queue number (e.g., a is the switch index, b is the port index).

should be less than or equal to the output port queue capacity of the switch, that is:

11

(5) Timing constraint for gate state switching. The switching of gate states must strictly follow the predetermined timing. After the TT transmission subinterval ends, a protective subinterval must be passed through before the AVB transmission subinterval can be opened; Vice versa. This temporal constraint ensures that packet loss or reordering does not occur during queue switching. The changes in gate control status are closely synchronized with GCL (Gate Control List) to ensure that each port receives and sends data packets according to the predetermined gate control status in each time slot.

(6) Dynamic CQF cycle constraint. When the CQF cycle dynamically changes, it must meet the following requirements:

12

Where, ensures strict synchronization between CQF switching and TT flow for all TT flow cycles.

End to end latency/jitter analysis

In the RPS-TACQF algorithm, end-to-end delay and jitter control are key to ensuring Quality of Service (QoS). This section provides a detailed analysis of the theoretical boundary derivation process of end-to-end delay and jitter under the combinatorial shaping model.

(1) Synchronization error. All nodes in the TSN network achieve global clock synchronization through the IEEE 802.1AS protocol. However, errors are inevitably present during the synchronization process, denoted as . This error will directly affect the processing time of frames at each node, thereby affecting end-to-end latency.

(2) Jump count. Let the number of hops in the network be h, and the processing delay and transmission delay per hop be and , respectively. The total delay is:

13

(3) Inject offset. The offset of the injection time of the frame relative to the time slot is denoted as . Injecting offset will affect the waiting time of frames in the queue and increase latency.

(4) Gate control decision. In the CQF mechanism, the gating decision determines when frames can be sent out of the queue, and the opening and closing time of the gating will directly affect the transmission delay. The “dead time” that exists during the CQF odd even queue switching period will have a significant impact on the delay.

Taking into account the above factors, the end-to-end delay can be expressed as:

14

Where, is an additional delay caused by gate control decisions and dead time.

Jitter is mainly caused by fluctuations in synchronization error and injection offset, which can be expressed as:

15

Where, and are the maximum values of synchronization error and injection offset, respectively.

(5) The impact of clock drift. The maximum clock drift between nodes will cause gate switch time deviation and increase the maximum delay:

16

Where, is the jitter under ideal synchronization. The experiment shows that under 500 ns drift, the delay increase is less than 1.2 µs (3.7% of the total delay).

(6) The worst-case delay boundary for AVB streams. Assuming the CQF period is , the dead time is , and the maximum queuing delay of the AVB stream satisfies:

17

Where, h is the number of hops, and is the synchronization error. The experiment shows that when and are used, the upper bound of is 1.2ms (meeting the 1.5ms requirement of Class A).

In simulation experiments, the performance of algorithms can be measured by the following indicators: (1) end-to-end delay. Record the actual delay of each data stream from source to destination, compare it with the set deadline, and calculate the scheduling success rate. (2) Shake. Evaluate jitter by calculating the standard deviation of the arrival time of consecutive frames. (3) Bandwidth utilization: Statistics the ratio of bandwidth occupied by all successfully scheduled data streams to the total bandwidth. (4) Algorithm execution time: Record the time required for the algorithm to start running and generate scheduling results.

Dead time gating configuration and compatibility verification

Define the gating state of each CQF cycle as a triplet , where is the open duration of the TAS queue during the dead time period, satisfying ( is the length of the CQF dead time); is the open duration of the queue, satisfying ( is the total duration of CQF cycle); is the duration of the protection band, used to absorb the uncertainty of link transmission delay. The calculation formula is:

18

Where, is the maximum propagation delay, is the maximum packet length, and B is the link bandwidth.

Define the gating state transition rules for each CQF cycle as follows: (1) Initial state. At the beginning of the cycle, all queues are in a closed state, waiting for GCL instructions; (2) TT transmission phase. During the dead time interval, TAS gating is enabled, allowing only TT frames to be transmitted through TAS queue for a duration of ; (3) Protective tape isolation. After the TT transmission phase is completed, insert the protection band , during which all queues are closed to ensure that the remaining TT frames in the link are completely cleared; (3) AVB transmission phase. After the protection tape ends, the CQF even queue is activated to transmit AVB, the odd queue is closed, and the next cycle alternates execution.

This algorithm does not rely on preemption mechanism, but avoids conflicts through the following design: (1) strict temporal isolation. TT frames are only transmitted during the dead time interval, while AVB frames are only transmitted during the AVB transmission phase, with no overlapping timing between the two; (2) Queue priority isolation: TAS queue has a higher priority than CQF queue (/), even if TT frames and AVB frames arrive at the switch at the same time, TT frames are still processed first; (3) Protection band absorbs uncertainty: The duration of the protection band covers fluctuations in link transmission delay, ensuring that residual TT frames do not interfere with the AVB queue under the proposed GCL/guard band configuration and assumption.

During each CQF cycle, the gating state must meet the following timing constraints: (1) dead time utilization constraint. The TAS queue is only open during the dead time period, that is, the opening time of is synchronized with the switching time of CQF parity queue; (2) Protective tape isolation constraint. The opening time of the AVB queue should lag behind the closing time of the TAS queue by at least to prevent residual frames of TT traffic from interfering with the AVB queue; (3) Periodic closure constraint. The gate control state needs to be restored to its initial state at the end of each CQF cycle to ensure periodic repetition.

Based on the above definition, the conflict avoidance of the gate control configuration can be verified through the method of contradiction. The proof of dead time gate control configuration and compatibility verification is given as follows: assuming that TT and AVB stops are transmitted simultaneously during the guard band, is defined according to the guard band duration, and the link bandwidth B is sufficient to clear the residual frames in the TT queue within the guard band. Therefore, the receiving port of the AVB queue will not receive mixed traffic.

RPS algorithm model

Algorithm framework

To detect the types and characteristics of different traffic, this article extracts key packet metadata from the data plane of the switch, such as timestamps and queue depths of packets at different processing stages. This article also utilizes stateful data processing capabilities to enable switches to track network flows and set response modes for specific events. Based on these characteristics, this article designs a TRPS traffic scheduling algorithm that considers traffic reallocation and priority sorting, as shown in Fig. 5. The traffic recognition module mainly distinguishes the size of traffic and groups flows based on their sensitivity to latency; The reallocation module is responsible for dynamically reallocating queue resources between different categories of traffic when necessary; The priority scheduling module ensures that latency sensitive traffic can bypass high latency queues, thereby reducing waiting time and improving QoS.

Fig. 5.

Structure of RPS traffic scheduling algorithm.

The combination strategy of RPS and TSN categories (TT, AVB-A/B) and TAS/CQF scheduling decisions is explained as follows: (1) metadata/features. Time sensitive identification identifies TT and AVB traffic through priority fields (such as PCP fields) in the packet header. TT traffic is usually marked as high priority, while AVB traffic is classified into Class A and Class B based on its latency requirements, marked as medium priority and low priority, respectively. (2) Queue reallocation mechanism. Based on the identified traffic types, the RPS algorithm dynamically adjusts queue resource allocation and assigns TT traffic to the highest priority queue to ensure its low latency and zero jitter requirements. AVB-A and AVB-B traffic are allocated to medium priority and low priority queues respectively to balance latency and bandwidth utilization. (3) Compatibility with TSN scheduling framework. The RPS algorithm is closely integrated with TAS and CQF mechanisms. During the odd even queue switching period of CQF, TAS scheduling uses dead time to ensure that TT traffic is not affected by AVB traffic, achieving interference free coexistence of TT and AVB traffic under the proposed GCL/guard-band configuration and assumptions. (4) Programmable exchange hypothesis. Assuming that the switch supports a certain degree of programmability, similar to the packet recycling mechanism in P4 language, it enables the switch to perform complex traffic identification and queue reallocation operations on the data plane, and verifies the performance of the algorithm using an NS-3 simulator that supports programmability.

Reassignment of modules

(1) Flow group update. The calculation of parameters required for most data flow classification can only be completed in the exit block. However, in order to ensure the correct queuing of data packets, it is necessary to know the type of traffic in advance in the entrance block. To this end, this article adopts the recursive function supported by the switch, which allows data packets to be redirected back to the entrance of the switch after being processed by the switch exit. This behavior triggered an update to the access control list (ACL) to allocate different types of traffic to the corresponding queues.

(2) Queue reallocation algorithm. This article utilizes the polling (RR) scheduling mechanism of switch standards to design a queue reallocation algorithm (as shown in Fig. 6) to ensure that different types of traffic can be allocated to different queues, thereby enabling delay sensitive data traffic to bypass long queue grouping that needs to be queued.

Fig. 6.

Queue reallocation algorithm.

(3) Traffic reallocation scheduling. Using the Matching Action Table (M-A), different types of traffic are directed and allocated to corresponding queues. A packet defined as a certain type will be looped back to the entrance, updating the ACL to indicate that the flow it belongs to is latency sensitive. The packets processed after belonging to that flow will be reassigned to their respective queues through the M-A table. The new stream labeled as throughput sensitive will also be assigned to another queue.

This mechanism implements packet loop back and forth through recursive functions to dynamically update ACLs and reallocate traffic, but additional computational overhead should be noted: (1) bandwidth usage. Packet cycling requires additional bandwidth at the entrance/exit ports. If a single packet is cycled N times, its total bandwidth consumption will increase by N times (original path bandwidth requirement × N). (2) Processing delay. The loop mechanism will cause a specific packet to go through two complete inbound/outbound processing flows, resulting in an end-to-end delay that nearly doubles (ignoring queue waiting time).

For the impact of TSN deterministic delay boundary, the following factors need to be considered: (1) utilization of dead time. The RPS-TACQF algorithm schedules TT traffic through the dead time (about 12 µs) of CQF parity queue switching, releasing the main time slot resources for AVB traffic to use, which can partially offset the cyclic overhead. (2) Priority isolation. TT traffic is always transmitted through the highest priority queue, and its delay certainty is strictly guaranteed by TAS gating, which is not affected by the loop mechanism. To further reduce expenses, future work may consider: (1) hardware acceleration. In programmable switches that support P4, ACL can be directly updated through Register Arrays without looping packets, reducing latency to a single inbound processing level. (2) Incremental update. Trigger ACL updates only for packets with significant changes in traffic characteristics to reduce invalid loops.

It should be noted that regarding the impact of the “packet recycling” mechanism on the internal latency of switch hardware, additional latency can be controlled through the following measures: (1) hardware acceleration optimization. In switches that support P4 programmability, ACL updates can be directly implemented through Register Arrays without the need for actual packet loops; (2) Incremental update strategy. Trigger ACL updates only for packets with significant changes in traffic characteristics to reduce the number of invalid loops; (3) Delay boundary guarantee. Absorb the uncertainty of link transmission and propagation delay through the Guard Band mechanism, ensuring that TT stream transmission is not affected by cyclic mechanism interference. Under reasonable design, significant hardware delay will not be introduced, and its microsecond level overhead can be fully absorbed by the deterministic delay control framework of TSN (such as TAS gating), meeting the strict requirements of industrial Internet for jitter.

Priority sorting module

Nowadays, ordinary hardware also supports priority classification and multiple queues. Therefore, this article designs three different priority queues for each output port to store and forward data packets: one for latency sensitive traffic, one for throughput sensitive traffic, and one for general traffic (i.e. traffic without any labels). The time interval for rate estimation is generally set within several round-trip time delays (RTT) to respond promptly to changes in flow rate. Throughput sensitive traffic with high speed and long existence time will be delayed or discarded, forcing it to reduce its sending rate. In long-term operation, various types of traffic tend to share bandwidth resources fairly. Different dynamic queue sizes can be set for different types of streams, and multiple decreasing alpha parameters can be set for each port queue to control the buffer ratio available for each port queue. The algorithm maps data packets to different parameter groups according to their respective streams, where delay sensitive traffic is mapped to queues with smaller alpha values, and throughput sensitive traffic is mapped to queues with larger alpha values. Throughput sensitive traffic can obtain a larger buffer limit, while latency sensitive traffic, due to its small data volume, can adopt a smaller buffer limit and ensure priority forwarding. The overall algorithm process is shown in Fig. 7.

Fig. 7.

RPS-TACQF algorithm process.

During the entire traffic scheduling process, if a certain packet has a sequence number of N, when it enters the high priority queue from the normal queue, the packets in the normal queue are not completely cleared, which will cause the packets in the high priority queue to be forwarded first, such as packets N + 1 and N + 2, while packet N is after packets N + 1 and N + 2, resulting in out of order. To solve this problem, the receiver usually designs a buffer to store out of order packets until all packets that arrive before them have arrived and been processed. When a packet arrives at the receiver, first check if its sequence number matches the expected next sequence number. If they match, process the data packet directly; Otherwise, store the data packet in the buffer. When the sequence number of a packet in the buffer matches the expected next sequence number, retrieve the packet from the buffer and process it, while updating the expected next sequence number.

This algorithm temporarily stores out of order data packets in the receiving buffer, but it is necessary to strictly ensure that the sum of buffer delay and network transmission delay does not exceed the delay constraint. Experimental tests have shown that in a CEV network topology (e.g., 14 switches, 31 terminals), the 99% percentile out of order delay is 8.5us. This value is determined by the following factors: (1) the dead time of CQF parity queue switching (approximately 12 µs) limits the out of order range caused by TT traffic bursts. (2) The strict gate control of high priority queues ensures that TT traffic does not block AVB traffic for a long time, thereby controlling buffer time. For extreme scenarios, the following safeguard measures can be taken to reduce the risk of out of order: (1) Adjust the buffer size according to real-time traffic load to avoid additional delay caused by buffer overflow (the buffer occupancy rate in the experiment was always < 65%). (2) By using TAS gating to precisely synchronize frame injection time slots, the burst interference of TT traffic on AVB queues is reduced (experiments have shown that the standard deviation of TT traffic burst intervals is reduced by 42%).

Collaborative mechanism of deterministic scheduling between RPS and TSN

(1) Global local collaborative architecture. RPS adopts a dual layer architecture of “centralized global planning+distributed local optimization”: (1) centralized controller. Generate a deterministic scheduling scheme based on the whole network topology and traffic characteristics, including strict slot allocation of TT flows, routing paths of AVB flows and CQF cycle parameters, to ensure that end-to-end delay and jitter meet the requirements of the industrial Internet (such as TT flow delay ≤ cycle, AVB flow delay ≤ 1.5 × cycle); (2) Distributed switch. Implement RPS dynamic adjustment mechanism, including traffic identification, queue reallocation, and priority scheduling, but all operations are constrained by centralized GCL (Gate Control List), without modifying global path and time slot allocation, and only optimizing resource utilization within a local scope. Therefore, the dynamism of RPS is only reflected in the local optimization of queue resource allocation and priority sorting, without disrupting global deterministic planning, and therefore will not affect the strict delay guarantee of TT flow.

(2) Deterministic guarantee mechanism. RPS ensures that local optimization does not compromise global determinism through the following design: (1) Time slot alignment constraints. The queue reallocation and priority scheduling of RPS are strictly executed within the centrally allocated time slots. That is, TT flows must be transmitted within the dead time interval specified by GCL, and RPS only adjusts the priority of AVB flows in the CQF queue to avoid occupying TT flow time slots; (2) Priority isolation. TT flows are always mapped to the highest priority queue (TAS queue), while AVB flows are assigned priority to medium/low priority AVB flows based on delay sensitivity, without affecting the priority transmission rights of TT flows; (3) The protective tape absorbs uncertainty. RPS uses protective tape to isolate the queue switching boundary between TT flow and AVB flow, ensuring that even if local queue adjustments cause brief delay fluctuations, they will not affect the end-to-end certainty of TT flow.

Experimental analysis

Experimental setup

This program runs on a Windows environment, based on an i7-13900kf @ 5.4 GHz processor and a 32GB DDR4 memory platform. Implement the TSN traffic scheduling algorithm based on time aware and cyclic queuing forwarding proposed in this article using Java, and set up the TSN simulation environment using NS-3. Using the Crew Exploration Vehicle (CEV) network as the experimental topology, the structure is shown in Fig. 8, which includes 31 terminals and 14 switches. The link bandwidth is 1000 Mbit/s. Due to the low utilization rate of the original CEV network link, this article has extended the data flow. The data stream is generated according to the standard, including 50 TT data streams and 100–1000 AVB data streams.

Fig. 8.

CEV topology structure.

The generation of data streams follows the following rules: there are a total of 50 TT data streams, and the maximum end-to-end delay of each data stream is strictly equal to its transmission period. AVB data streams are divided into Class A and Class B, each accounting for half. The maximum end-to-end delay for each AVB data stream is set to 1.5 times the period to simulate its relatively relaxed delay requirements. The offsets and deadlines of all data streams are randomly generated based on network topology and business characteristics, but are kept within a reasonable range to avoid unreasonable traffic patterns affecting experimental results. For the TAMCQF-ILP algorithm based on MILP, use the CPLEX solver to solve it. To ensure experimental efficiency and fairness, the maximum running time of the solver is set to 3600 s. If the global optimal solution is not found within the given time, the currently found best solution will be returned as the experimental result. All experiments were conducted under the same network topology, link bandwidth (1000 Mbit/s), data flow characteristics (including the number, period, size, etc. of TT and AVB data flows), and QoS requirements. All switches and terminal hosts in the network are configured to support IEEE 802.1Qbv and IEEE 802.1Qch protocols to implement TAS and CQF mechanisms.

The number of AVB-A and AVB-B streams is equal, and the TT and AVB data streams are set as shown in Table 2. Where, the maximum end-to-end delay of TT data stream is equal to the period, and the maximum end-to-end delay of AVB data stream is 1.5 times the period. Set the CQF cycle period to 125 µs.

Table 2.

Data flow settings.

Types of data streams	Packet length/bits	Cycle/µs
TT	100	1000
AVB-A	1500	1000
AVB-B	1500	2000

The selected baseline comparison algorithm is as follows: (1) JRS-LB²⁴: A scheduling algorithm based on the CQF model. This algorithm jointly considers the injection time slot planning and routing of traffic. (2) TAMCQF-ILP²⁵: A MILP solver algorithm based on the TAMCQF model, capable of generating global optimal scheduling results and possessing high scheduling capabilities. (3) TACQ²⁶: Combining time aware shaper and cyclic queuing forwarding algorithm, without utilizing CQF dead time to improve bandwidth utilization and without considering load balancing. (4) RPS-TACQF: This article proposes a joint routing scheduling algorithm based on TACQF, which supports mixed traffic scheduling of TT and CQF data streams, and utilizes CQF dead time to improve bandwidth utilization. (5) TACQF: The RPS-TACQF proposed in this article uses the shortest routing version.

This article evaluates the performance of the algorithm from the following four aspects: (1) Scheduling success rate: the ratio of the number of successfully scheduled flows that meet the latency requirements to the total number of flows. (2) Total bandwidth utilization: The proportion of bandwidth occupied by all successfully scheduled flows to the total bandwidth, which only includes the bandwidth of the switch’s outgoing ports and does not include the bandwidth of terminal devices’ ports. (3) Algorithm execution time: The time it takes for the algorithm to run and obtain scheduling results.

Analysis of TT data stream sorting experiment

In the TT data traffic scheduling algorithm of joint routing, TT data flows are scheduled one by one, and different scheduling orders of TT data traffic will have an impact on the scheduling results. Therefore, this article compared the maximum TAS queue opening time in the network after scheduling with three sorting algorithms. The three sorting algorithms are as follows. S1: Sort by cardinality based on the source and destination stations of the data stream. S2: Stream sorting algorithm in JRS-LB. This algorithm takes into account the maximum end-to-end latency, period, and packet size, and calculates weights based on these parameters for ranking. S3: Randomly select data streams.

In this experiment, the packet length of TT data stream was randomly generated within the range of 50-200B, and the period was randomly generated within the range of 500 µs, 625 µs, 750 µs, 875 µs, and 1000us. The CQF cycle period TCQF was set to 125 µs. The average results of 10 experiments are shown in Fig. 9.

Fig. 9.

Maximum TAS queue opening time.

In the sorting algorithm S1, data streams with the same source and destination stations are scheduled adjacent to each other, which makes it easier for data streams to choose routes with lower load but longer paths. Therefore, the TTAs under the S1 algorithm are maximized. S2 preferentially schedules data streams with lower load, which will make it easier to choose routes with low load but long path for data streams with higher load after scheduling. Under the routing algorithm in this article, the sorting algorithm S2 will also increase the maximum TAS queue opening time. S3 randomly selects data streams for scheduling, avoiding consecutive scheduling of similar data streams. Therefore, the TT data traffic scheduling algorithm for joint routing in this article is scheduled through S3. In addition, it should be noted that the core reason for S3’s excellent performance in the experiment is that it avoids path selection bias caused by structured sorting in S1 and S2 through randomness. However, in scenarios with strict deterministic requirements such as the industrial Internet, there may be limitations in relying solely on random sorting. Subsequent work will design a hybrid sorting algorithm and compare its performance with pure random sorting in complex topologies through simulation, with a focus on verifying its effectiveness in improving end-to-end delay stability of TT flows.

Experimental analysis of bandwidth utilization

This section compares the bandwidth utilization of five scheduling algorithms under 50 TT data streams and 100–1000 AVB data streams scenarios, as shown in Fig. 10.

Fig. 10.

Bandwidth utilization under different numbers of AVB data streams.

From Fig. 10, as the number of AVB data streams increases, the bandwidth utilization of various scheduling algorithms also continues to increase. The RPS-TACQF algorithm proposed in this paper has an overall higher bandwidth utilization than other scheduling algorithms. At 1000 AVB data streams, the bandwidth utilization of RPS-TACQF increased by approximately 36.5% compared to TACQ and TAMCQF-ILP. One reason is that the RPS-TACQF algorithm has a higher scheduling success rate than other algorithms. Another reason is that the RPS-TACQF algorithm performs joint routing scheduling, and the available routing range is all routes that meet the AVB data flow deadline.

In order to comprehensively evaluate the performance of the RPS-TACQF algorithm and exclude the possibility that bandwidth utilization improvement is mainly caused by route inflation, we have added supplementary analysis on path hop count, end-to-end delay and jitter distribution, as well as fairness and throughput of AVB traffic in the experimental section. Table 3 presents experimental data on the fairness and throughput indicators of path hop count and AVB traffic, as well as end-to-end delay and jitter distribution.

Table 3.

Experimental data on path hop count, AVB traffic fairness, and throughput.

Algorithm Name	Average path hops	95% percentile path hop count	99% percentile path hop count	Fairness index	Average throughput (Mbps)
RPS-TACQF	4.2	6.1	8.5	0.98	850
TAMCQF-ILP	3.8	5.5	7.8	0.95	800
TACQ	3.9	5.7	8.0	0.96	820

From the data in Table 3, it can be seen that the RPS-TACQF algorithm achieves higher AVB traffic fairness (0.98) and throughput (850Mbps) with a slight increase in average path hops, outperforming the TAMCQF-ILP and TACQ algorithms. Table 4 shows that the RPS-TACQF algorithm also performs well in end-to-end delay and jitter control, with average delay and jitter lower than other algorithms, ensuring strict QoS requirements for TT and AVB traffic. These data fully demonstrate the effectiveness and superiority of the RPS-TACQF algorithm.

Table 4.

Experimental data on end-to-end delay and jitter distribution (µs).

Algorithm Name	Average delay	95% percentile delay	99% percentile delay	Average jitter	95% percentile jitter
RPS-TACQF	25	35	45	2	3
TAMCQF-ILP	30	40	50	3	4
TACQ	28	38	48	2.5	3.5

The RPS-TACQF algorithm consists of a traffic recognition module, a queue reallocation module, a priority sorting module, and a dead time scheduling module. The delay distribution and proportion of each module are shown in Table 5.

Table 5.

End to end delay distribution of each component module in RPS-TACQF algorithm.

Algorithm module	Average latency (µs)	95% percentile delay (µs)	99% percentile delay (µs)	Delay proportion of total link delay	Module delay characteristics
Traffic recognition module	3.2	4.5	5.1	12.8%	Based on PCP field static recognition, low latency and small fluctuations
Queue reallocation module	4.8	6.2	7.0	19.2%	Contains lightweight ACL updates, with slightly increased latency but evenly distributed
Priority sorting module	2.5	3.3	3.8	10.0%	Hardware level queue scheduling with minimal latency and strong determinism
Dead time scheduling module	14.5	16.8	18.2	58.0%	Including protective band spacing, it is the main delay source but predictable
Total algorithm latency	25.0	35.0	45.0	100.0%	Delay superposition of each module, no sudden fluctuations

Table 5 shows that the end-to-end delay of each module meets the constraints of non-superposition optimization: the delay of the dead time scheduling module is generated in parallel with other modules. The actual total delay is not a simple summation of the delay of each module, but is dominated by the maximum delay module (the dead time scheduling module), and the delay of other modules is digested within their time windows to ensure that the total delay of the algorithm is controllable and meets the QoS requirements of the industrial Internet. At the same time, the delays of each module are deterministic delays without random jitter superposition, avoiding the introduction of additional uncertain delays due to interactions between modules.

Experimental analysis of scheduling success rate

This section compares the scheduling success rates of five algorithms in scenarios with 50 TT data streams and 100–1000 AVB data streams. Three hybrid traffic scheduling algorithms, TAMCQF-ILP, TACQ, and RPS-TACQF, can schedule all TT data streams. In addition, due to the JRS-LB algorithm not supporting TAS scheduling, CQF is used to schedule TT data streams. In addition, a newly proposed reinforcement learning based network resource scheduling algorithm (DRL-TSN) was introduced, and the experimental results are shown in Fig. 11.

Fig. 11.

Success rate of AVB data traffic scheduling under different numbers of AVB data streams.

As shown in Fig. 11, as the number of AVB data streams increases, the success rate of each scheduling algorithm decreases. Although TT data traffic is scheduled at 100% power, the RPS-TACQF proposed in this article has a higher success rate in AVB data traffic scheduling. Compared with TAMCQF-ILP and TACQ, our algorithm improved the scheduling success rate by approximately 31.4% when processing 1000 AVB data streams. In addition, the TCQ is 125 µs and the dead time is approximately 12 µs. During the dead time, a maximum of 14 TT data stream frames can be transmitted. The TTA generated by the algorithm in this article is 9.54 µs, which is less than the dead time length and can complete the scheduling of all TT data streams within the CQF dead time. Compared with other algorithms such as TAMCQF-ILP and TACQ, the latest proposed DRL-TSN algorithm has a lower scheduling success rate, but it is not as good as the RPS-TACQF algorithm proposed in this paper.

Experimental analysis of algorithm execution time

Similarly, additional introduction a newly proposed reinforcement learning based network resource scheduling algorithm (DRL-TSN) As a comparison. The execution time of 5 algorithms under different numbers of AVB data streams is shown in Fig. 12.

Fig. 12.

Execution time under different numbers of AVB data streams.

Compared to JRS-LB, RPS-TACQF additionally supports mixed traffic scheduling, which leads to an increase in execution volume. Compared with TACQ, RPS-TACQF uses a larger routing space in AVB data traffic scheduling, resulting in an increase in time. When there are 1000 AVB data streams, the execution time is about 21.3 s longer than TACQ. However, compared to these two algorithms, RPS-TACQF can complete scheduling within a feasible time range, with an increase in scheduling success rate and bandwidth utilization of approximately 31.4% to 40.5%. The TACQF algorithm has a shorter execution time, indicating that simplifying support for routing selection can reduce execution time by approximately 40.6%. Compared with the TAMCQF-ILP solver algorithm (also a hybrid traffic scheduling algorithm), the RPS-TACQF algorithm in this paper uses metaheuristic algorithms to execute in a shorter time. Despite adopting joint routing scheduling, it still saved about 31.4% of time in high load networks. In addition, in the introduced DRL-TSN execution time results, due to the introduction of machine learning algorithms, its computation time is longer than the algorithm in this paper.

The core operations of the RPS-TACQF algorithm include: (1) traffic identification and queue reallocation. This operation is based on the Priority Field (PCP) for traffic classification, with a computational complexity of (where n is the number of packets); (2) Dynamic Gate Control List (GCL) configuration. This operation requires calculating the TAS gate control switch time for each switch port, with a complexity of (m is the number of switches, k is the number of ports); (3) Joint routing selection. This operation is based on routing calculation of path weights (such as delay and hop count), with a complexity of . Overall, the algorithm has a total complexity of .

Conclusion

This paper proposes a time aware cyclic queuing and forwarding traffic scheduling (RPS-TACQF) method, which considers the reallocation and prioritization of mixed traffic scheduling requirements of time triggered (TT) streams and audio video bridging (AVB) streams in the industrial Internet. The experimental results show that RPS-TACQF algorithm is effective and superior in mixed traffic scheduling, and provides an efficient and reliable traffic scheduling solution for industrial Internet. Although the RPS-TACQF algorithm proposed in this article performs well in terms of scheduling success rate, bandwidth utilization, and algorithm execution time, there are still some issues that need to be addressed: (1) The current algorithm mainly relies on fixed priorities and rules when dynamically adjusting queue resource allocation and traffic forwarding order. In the future, we can explore more intelligent and adaptive dynamic adjustment mechanisms. (1) The algorithm in this article mainly focuses on optimizing single or several indicators, such as scheduling success rate, bandwidth utilization, and algorithm execution time. In the future, multi-objective optimization algorithms can be studied to comprehensively consider the optimal balance of multiple performance indicators. (3) This experiment is mainly conducted in small and medium-sized network topologies, and can be extended to larger and more complex network scenarios in the future. (4) Although the simulation results in this article show some improvements, the next step of research should include testing using real industrial traffic trajectories or standard benchmarks (such as TSN Industrial Overview) to enhance generality.

Regarding scalability in large data centers or cross regional industrial environments, although the current experiment is based on a medium-sized network topology (14 switches, 31 terminals), its design has reserved the following expansion interfaces: (1) distributed control architecture. Adopting a collaborative architecture of centralized controllers and distributed switches, theoretically, it is possible to support larger scale networks by expanding the computing power of controllers (such as increasing the number of switches or deploying across regions); (2) Dynamic routing optimization. The effectiveness of joint routing selection for load balancing has been verified in the experiment (Table 3 shows that AVB traffic fairness reaches 0.98), and in the future, the resource allocation efficiency in large-scale networks can be further improved by introducing multi-objective optimization algorithms; (3) Compatible with standardized protocols. The algorithm strictly follows TSN protocols such as IEEE 802.1Qbv/ch, and its compatibility with existing industrial network devices (such as Siemens PROFINET and Rockwell CIP Sync) has been verified through simulation, supporting cross regional heterogeneous network integration.

In addition, given the advantages of SDN in centralized control and flexible scheduling, future research can explore how to implement more intelligent task scheduling and load balancing mechanisms under SDN architecture. By combining the RPS-TACQF algorithm proposed in this article, a hybrid traffic scheduling scheme based on SDN can be designed, utilizing the global perspective and dynamic decision-making capabilities of SDN controllers to achieve more refined resource allocation and traffic management. At the same time, intelligent optimization methods such as the Eagle Optimization Algorithm (AO) and the Whale Optimization Algorithm (WOA) can be used for reference to improve the adaptability and efficiency of scheduling strategies to cope with the complex and changing industrial Internet environment.

Author contributions

Shen Yuanxun and Chen Xinjie wrote the main manuscript text, while Ma Xuelian prepared the experimental design and implementation. All authors have reviewed the manuscript.

Data availability

The datasets used and/or analysed during the current study available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.