ARTSN: Exact and Adaptive Self-triggered Traffic Scheduling for ARTS Networks Ruide CaoSTU , Shuangping ZhanP , Jiashuo LinH , Yan LiuSP , Chenxi LingSP , Yi WangSPDB , Guoming TangTB
arXiv:2606.13899v1 [cs.NI] 11 Jun 2026
S
Southern University of Science and Technology; T The Hong Kong University of Science and Technology (Guangzhou) U University of California, Riverside; P Pengcheng Laboratory; H Hanshan Normal University; D Heyuan DET
Abstract—Autonomous real-time systems (ARTS), such as selfdriving vehicles and robotic assembly lines, are increasingly deployed to improve efficiency, accuracy, and responsiveness with reduced human intervention. In ARTS networks, self-triggered (ST) traffic—initiated by internal decision-making rather than fixed schedules or external events—is becoming prevalent and plays a critical role in enabling timely autonomous actions. However, existing network schedulers do not adequately support ST traffic due to two inherent challenges: volatility, where bounded processing jitter leads to uncertain arrival times, and absence, where reserved network resources remain underutilized when ST traffic does not materialize. To address these challenges, we propose ARTSN, an ST-tailored scheduling paradigm built upon time-sensitive networking (TSN). ARTSN introduces two key techniques: (1) an exact offline scheduling method that leverages the inferable arrival information of ST traffic for precise time-slot reservation, and (2) an adaptive online slot-release mechanism that dynamically reclaims unused reservations when ST traffic is absent. Extensive experiments on both a TSN simulator and a real-world testbed show that ARTSN significantly improves schedulability, scalability, and efficiency over state-of-the-art methods while maintaining reliable transmission guarantees. Index Terms—Autonomous real-time systems, industrial IoT, deterministic networks, time-sensitive networking
I. I NTRODUCTION Fueled by recent advances in embedded computing power and Deep Neural Networks (DNNs), end-to-end (E2E) control is evolving rapidly. A fundamental shift is where decisions are made. Modern systems increasingly make decisions internally on-device, rather than relying on human operators or cloud offloading. An example is Tesla’s Full Self-Driving (FSD) stack [1], where perception and driving decisions are computed onboard by a single DNN. Similar trends are emerging on smaller platforms as well: on an NVIDIA Jetson Nano, TensorRToptimized DNNs can respond in tens of milliseconds (92.8 ms for a 3D-CNN and 118.4 ms for AlexNet) [2]. This intelligent onboard decision-making capability combines autonomy with real-time responsiveness, giving rise to Autonomous RealTime Systems (ARTS). ARTS are substantially enhancing the efficiency, accuracy, and responsiveness of E2E control while simultaneously minimizing the need for human intervention. As a consequence, ARTS have attracted significant research interest [3] and have been widely applied [4]–[6]. B Corresponding authors: [email protected], [email protected]. This work is supported by the Guangdong Basic and Applied Basic Research Foundation (2024A0101010001), the Guangdong High-Level Talents Special Support Program (2021TX05X205), Pengcheng Laboratory The Major Key Project of PCL (PCL2025A01). This study was conducted by Ruide Cao during his internship at HKUST (GZ).
Fig. 1. Real-world application examples for TT, ET, and ST control.
Control tasks in ARTS are typically divided into three categories based on their triggering causes: time-triggered (TT), event-triggered (ET), and self-triggered (ST) [7]. TT tasks are scheduled at predetermined intervals for predictable and periodic operations. ET tasks are activated in response to specific external events or conditions, allowing for on-demand execution. In contrast, ST tasks are initiated proactively in response to internal state changes or criteria inherent to the system. Some real-world examples are listed in Fig. 1. While TT tasks continue to predominate, ST tasks are increasingly pervasive [8]. The transition from pre-defined and external triggers to internal decision-making highlights the importance of ST tasks. In particular, the ST mechanism has distinct advantages in safety-critical scenarios, where waiting for a clock countdown or external decision-making signals before reacting is unacceptably slow [9]. From a network perspective, these control tasks correspond to distinct traffic types, each demanding strict latency and reliability guarantees. To meet these stringent requirements, some ARTS networks in safety-critical domains rely on hardwiring [10], where devices are connected through dedicated point-to-point physical links. Such a hardwiring approach is inherently inefficient and non-scalable, as adding new devices requires additional dedicated connections. In pursuit of reliable, efficient, and scalable traffic transmission, Time-Sensitive Networking (TSN) has emerged and attracted significant industrial interest due to its high flexibility and wide applicability [11], [12]. As an Ethernet-based networking technology, TSN supports plug-and-play modular connectors. When applied in ARTS networks, TSN can naturally support TT traffic scheduling through its traffic shaping capabilities [13]–[15]. With recent
research efforts, TSN can also be adopted for the ET traffic scheduling [16]. Refer to § II-B for details. When accommodating ST traffic with TSN, however, we are faced with the following threefold problem. 1) Failure to meet real-time scheduling requirements. A straightforward approach is to leave ST traffic unprotected, i.e., transmit along with other best-effort (BE) traffic but with higher priority as in [13]. ST streams may still miss their deadlines due to the blocking of low-priority packets. 2) Inefficient use of network bandwidth. An alternative practical approach is to treat the ST traffic as ET and adopt the ET-enabled traffic scheduling mechanism. Nevertheless, treating ST traffic with inferable arrival times as completely random and unpredictable ET traffic may lead to overly conservative bandwidth reservations, thereby wasting substantial network capacity. 3) Scalability limitation imposed by switch computing power. Another approach is to leave the scheduling of ST traffic to an online scheduler [17], [18]. Due to the limited computing power, however, this is only available for sporadic ST traffic. Deterministic transmission of bursty ST traffic is still not guaranteed (such as the GOOSE burst in substation communication networks [19]). To further bridge the gap between the flexibility required for autonomy and the determinism required for real-time, we are motivated to design an ST-tailored traffic scheduling paradigm. Specifically, by referring to the essential design principles of TSN, we present ARTSN in this work. In support of TT, ET, and ST tasks in ARTS, ARTSN aims to provide reliable, deterministic traffic scheduling, enabling critical traffic from these tasks to be protected with extremely low latency and hard real-time guarantees. To the best of our knowledge, existing scheduling algorithms can only provide feasible but inexact (refer to Def. 3 and 4 in §IV-B) scheduling schemes for ST traffic. This is the first work to include exact ST traffic scheduling on TSN. Our major contributions can be summarized as follows: We extend the flexibility boundaries of the deterministic network system model by introducing the self-triggering mechanism from control systems to the network traffic scheduling context. • We propose ARTSN, an ST-tailored TSN traffic scheduling paradigm that supports all 3 traffic types. An exact traffic scheduling technique that effectively leverages inferable arrival information from ST traffic is presented, along with an adaptive slot release mechanism. • We evaluate ARTSN in both simulations and a real-world testbed. Comprehensive experiments show that ARTSN significantly improves network schedulability, scalability, and efficiency while guaranteeing reliability. •
II. R ELATED W ORK For decades, multiple technologies have been designed to guarantee the transmission of critical traffic over Ethernet’s high speed and good scalability. This section provides a brief literature review on Ethernet-based deterministic networks.
A. Reliable Transmission of Time-Triggered Traffic 1) TSN, supported by Broadcom (BCM89571 [20]), Xilinx, Intel, Cisco, NXP, etc. [21]; 2) Time-Triggered Ethernet (TTEthernet), selected by ESA and NASA for their Orion multi-purpose vehicle [22]; 3) Avionics Full-Duplex Switched Ethernet (AFDX), used in Airbus A350 and A380 airplanes [23]; 4) PROFINET [24]; 5) EtherCAT [25]. These technologies use time-division multiplexing to segregate critical traffic from non-critical traffic. Pre-defined TT critical traffic can get low-latency, high-reliability transmission [26]. Frame loss can be solved through replication rather than retransmission [27]. Among these, TSN has attracted much industrial interest due to its wide applicability [28]–[30]. Recently, Yang et al. [31] proposed an industrial control system architecture and a tasktraffic joint scheduling algorithm in the TSN context, ensuring traffic determinism by assigning task computation to switches. B. Reliable Traffic Transmission in ARTS networks Pre-defined TT traffic lacks the flexibility required for environment-adaptive tasks and is insufficient to support ARTS [32]. In ARTS networks, traffic characteristics are not as fixed as in traditional real-time systems but rather more volatile and episodic [33]. Both online and offline scheduling solutions have been proposed to handle such non-fixed traffic. For online solutions, D-TSN [34] supports event-driven real-time traffic using earliest-deadline-first scheduling, enabling dynamic response to aperiodic arrivals. InNetScheduler [17] utilizes computing resources on switches to schedule latency-sensitive ET traffic, thereby increasing the amount of schedulable traffic without centralized coordination. Hebo [18] loosens arrival time constraints of periodic streams, making TSN applicable to volatile traffic with bounded jitter. For offline solutions, E-TSN [16] enables ET traffic scheduling by allowing ET streams to preempt shareable TT streams. A ”prudent slot reservation” technique is presented to compensate for all possible preemptions. Notably, E-TSN’s extra reservation does not iterate until convergence: when the extended window admits more ET arrivals, further extension is needed, and the lack of this recursive calculation may lead to deadline violations under certain conditions (analyzed in §IV-B). DeepScheduler [35] applies deep reinforcement learning to enhance schedulability and efficiency for largescale networks. GP-TSN [36] leverages causal inference to predict future ET streams by modeling event interactions using a spatio-temporal graph attention mechanism. However, none of the above works address ST traffic natively. Unlike ET streams triggered by unpredictable external events, ST streams are initiated by internal system states and have deterministic triggering relationships with TT streams, enabling more precise scheduling. Yet ST traffic also poses unique challenges: bounded volatility from processing jitter and potential absence when not triggered. Existing ET-oriented approaches rely on probabilistic prediction for stochastic events, which cannot exploit the deterministic structure inherent in ST streams. This gap is the focus of our work.
Fig. 2. Transmission with (above) and without (below) the ST support. Joint considerations allow for the utilization of ST traffic arrival time information.
III. S YSTEM M ODEL A. Network Model We model the network topology with a directed graph G = (N , L). All devices and switches are nodes in N , and all links connecting two nodes are in L ⊆ N 2 . Each pair of connected nodes na , nb is assumed to be connected by two edges ⟨na , nb ⟩ and ⟨nb , na ⟩, following the full-duplex context. Two attributes of the edges are considered: ⟨na , nb ⟩.pd denotes the propagation delay of ⟨na , nb ⟩, and ⟨na , nb ⟩.tu is its smallest time unit. We use {.} to denote a set where the order is irrelevant and ⟨.⟩ to denote a sequence of items. We consider all the network traffic as unicast streams, w.l.o.g., following [37], [38]. The set of physically generated concrete traffic is referred to as actual streams, denoted R = {r1 , r2 , . . . , rn }. For the ith actual stream ri , we represent its path by Pi = ⟨⟨na , nb ⟩, ⟨nb , nc ⟩, . . . , ⟨ny , nz ⟩⟩, and its maximum allowed E2E latency by DiR . Based on how ri is triggered and whether it shares its time slot with ET streams, ri has a type ti ∈ {strict, share, event, self}. While ti = event or self directly indicates ri is ET or ST, strict and share are two mutually exclusive types of TT streams. The priority order is: strict > event = self > share. Strict TT streams have the highest priority and require strict transmission time. ET and ST streams share the same priority level, as both require timely delivery upon arrival. Share TT streams have the lowest priority among critical traffic and can be preempted by ET or ST streams when they are present. We use Ti to denote the period of TT streams and the minimum times between successive triggers for each ET and ST stream. There is a hyperperiod H of all actual streams R, where H = lcm(T1 , T2 , . . . , Tn ), which is the least common multiple of all Ti . Note that ST traffic is inherently bounded per hyperperiod, as ST streams are generated in response to TT streams, instead of being generated spontaneously by the controller. We then discuss every network scenario within its first hyperperiod, since later hyperperiods can be considered repetitions. A network scenario is defined as follows. Definition 1. A network scenario γ = (G, R) consists of a network G = (N , L) and the set of actual streams R = {r1 , r2 , . . . , rn } in G. In a network scenario, TT streams can be scheduled deterministically by setting proper offsets on the time-synchronized end nodes. For ET streams, however, the arrival times are
Fig. 3. Standard Ethernet switch versus TAS switch.
independent of the system time. We denote the jth actual arrival time of ET stream ri as AR i,j . Furthermore, we assume each ST stream rk has a final trigger stream trk , which is the last TT stream involved in triggering rk . Thus, all possible controller processing times of trk can be represented as an interval [Ckmin , Ckmax ] and the processing jitter range is ψk = Ckmax − Ckmin . Based on these formulations, a trigger scenario is defined as follows. Definition 2. A trigger scenario γ ∗ = (G, R, AR , C) is a specific instantiation of a network scenario γ = (G, R). R R Arrival times AR = ⟨AR a,1 , Aa,2 , . . . , Am,n ⟩ and process times C = ⟨Cb,1 , Cb,2 , . . . , Co,p ⟩ such that each arrival of each ET stream ri has an arrival time AR i,j ∈ [0, H) and each trigger of each ST stream rk has a process time Ck,l ∈ [Ckmin , Ckmax ]. A traditional deterministic network scenario with full TT streams corresponds to a single trigger scenario because all streams are fixed, whereas a more flexible network scenario with ET or ST streams may typically involve multiple trigger scenarios. The number of trigger scenarios equals the number of combinations of all possible arrival and processing times. As time is discretely considered, the scenarios are finite. Although jitter in the controller’s processing time introduces some uncertainty, it remains feasible to infer the arrival of ST traffic. Fig. 2 compares the traditional ET-enabled paradigm with our proposed ST-enabled paradigm. While the traditional paradigm independently reserves resources for each stream, we implicitly infer the arrival of each ST stream through the controller processing time and the triggering relationships between streams. Targeted reservations can be made according to those inferences. B. Switch Model The 802.1Q standard [39] requires switches to support eight queues with different priorities. The 802.1AS standard [40] enables global clock synchronization. Building on these standards, the 802.1Qbv [14] ensures real-time performance through gate control. While queues spatially differentiate traffic, gating provides temporal differentiation, allowing for precise traffic scheduling. Fig. 3 illustrates the advancement of the Time-Aware Shaper (TAS) mechanism, which can achieve accurate traffic shaping via time-synchronized gating control. As defined in the TSN framework [41], a Centralized Network Configuration (CNC) is deployed to calculate and distribute the Gate Control Lists (GCL). The TSN switches
then open and close the egress port gates in synchronization with the local time and the GCL to avoid blocking caused by low-priority traffic. The gate opening time of each gate control entry in GCL is called a time slot. C. Problem Definition Based on the system models discussed above, we define the Deterministic Traffic Scheduling Problem. The objective is to determine valid scheduling parameters and resource allocation for every actual stream in the network, ensuring deterministic guarantees for the arrival times of all traffic types. Given: • A network topology G(N , L) with link capabilities. • A set of actual streams R = {r1 , r2 , . . . , rn }, where each stream ri is characterized by its tuple ⟨Ti , DiR , Pi , τi , ti ⟩. For ST streams, the trigger source trk and processing jitter ψk are also given. Find: • Time Slot Assignments Φ: A set of gate opening times for all frame instances that need to be scheduled. The detailed formulation of Φ will be presented in §IV-B, where all actual streams are mapped to schedule streams. • Queue Assignments Q: The priority queue allocation for each stream, Q = {q1 , q2 , . . . , qn }, where qi denotes the queue index for stream ri . Subject to: 1) Hard Real-time Constraints: All streams must complete transmission within their maximum allowed E2E latency DiR and respect their period Ti . 2) Network Constraints: All frames must satisfy the physical constraints of links, including frame sequencing within each stream, propagation delays between adjacent links, and temporal isolation to prevent frame collisions at shared egress ports. 3) Deterministic Constraints: The derived schedule must be valid for all possible trigger scenarios γ ∗ . For any actual arrival time within the processing jitter range [Ckmin , Ckmax ], the reserved time slots are sufficient to guarantee that no deadlines will be missed. Objective: Find a tuple ⟨Φ, Q⟩ such that all constraints are satisfied: ^ ∃⟨Φ, Q⟩ : c(Φ, Q, G, R) = True, (1) c∈C
where C represents the set of all aforementioned constraints. If such a solution exists, the network scenario is deemed schedulable. IV. D ESIGN OF ARTSN While §III-C defines the general deterministic scheduling problem, existing solutions typically achieve only feasible, but not exact, scheduling schemes. In this section, we first identify the design challenges and give an overview of ARTSN in § IV-A. We then introduce the concepts of feasible, exact, and present our design of exact traffic scheduling in § IV-B. We finally propose an adaptive slot release mechanism in § IV-C.
Fig. 4. Offline (left) and online (right) scheduling phases of ARTSN.
A. Overview Prior knowledge of ST traffic can be used to improve both traffic schedulability and network efficiency. However, leveraging this knowledge in traffic scheduling requires addressing two design challenges: Volatility. Unlike TT streams with fixed arrival times, ST streams arrive within a bounded jitter window. This jitter arises from the controller’s decision-making process, where different code paths under varying conditions lead to variable processing times. The scheduler must reserve slots covering all possible arrivals within this window—neither a single fixed slot (as for TT) nor slots spanning the entire period (as for ET) suffice. • Absence. ST streams are conditionally triggered and may be absent in certain periods. For example, an autonomous vehicle only issues braking commands upon detecting obstacles—when the path is clear, no corresponding ST stream is generated. When absent, reserved slots still block other traffic, unnecessarily consuming bandwidth. This inefficiency is especially severe in topologies where ST streams converge, such as star and tree topologies. •
To address the two challenges above and improve the performance of ARTS networks, we propose a two-phase traffic scheduling paradigm, ARTSN. Fig. 4 shows an overview of ARTSN in a sample network that includes a CNC, two switches, a sensor S1, a controller C1, and an actuator A1. 1) In the offline scheduling phase, we solve the volatility problem of ST traffic by modeling and constraining to make exact time slot reservations where ST streams are actually possible to arrive. These desired reservations are first calculated at the CNC, which then converts them into queue assignments and GCL configuration profiles. The profiles specify when each egress queue of each device can and cannot transmit. Once the calculations and conversions are complete, the CNC distributes them to all involved network devices. 2) In the online scheduling phase, we solve the problem of idle reservation when ST traffic is absent by reporting its absence along its path in advance of the reserved time slots. Upon receiving a TT stream, the controller determines if a corresponding ST stream will be generated. This determination is made before the earliest possible transmission time for the ST stream, leaving a time gap. Thus, this time gap allows the controller to send a small BE report frame to release subsequent unnecessary reservations in non-triggering cases. If a release succeeds, more BE traffic can be transmitted, thereby improving network efficiency.
Fig. 5. Three stages of an ST stream during one hyperperiod. The green time slot is reserved exactly, while the blue time slots are redundant reservations.
B. Exact Traffic Scheduling To address volatility, exact traffic scheduling is designed in the offline scheduling phase. To clarify exactness, we first define a feasible traffic scheduling scheme as follows. Definition 3. A traffic scheduling scheme is a feasible traffic scheduling scheme iff (i) the scheme can be accomplished by deploying GCL at each egress port, and (ii) for all streams, there is no trigger scenario that may result in a deadline miss. A feasible traffic scheduling scheme ensures that all scheduled streams complete transmission before their deadlines. By treating each ST stream’s minimum consecutive trigger interval as its period, all ST streams can be scheduled as TT streams. With such a transformation, several TT scheduling algorithms can schedule ST traffic by solving a Satisfiability Modulo Theories (SMT) or Integer Linear Programming (ILP) problem [31], [38], [42]. These schemes are feasible when the maximum allowable latency of ST streams exceeds their minimum consecutive trigger interval, but their feasibility is not guaranteed when the latency requirement is more stringent. When the maximum allowable latency of an ST stream is less than its minimum consecutive trigger time, exponentially more time slots need to be reserved for the ST stream. In addition to TT traffic, some approaches treat ST traffic as ET traffic to obtain feasible traffic scheduling schemes. As demonstrated in E-TSN [16], the reliable transmission of ET traffic can be achieved by solving GCL constraints using SMT. E-TSN enables ET traffic to preempt some TT time slots, allowing ET traffic to arrive at any time without incurring deadline misses. While feasible traffic scheduling schemes guarantee the delivery of the scheduled critical streams, they do not account for the impact on background traffic. In practice, the feasible traffic scheduling schemes identified by the above two approaches may result in a large number of idle reservation slots in the network. This underutilization occurs because many reserved time slots do not actually contribute to the transmission of ST streams while occupying the egress port, preventing other streams from being transmitted. We define an exact traffic scheduling scheme as follows. Definition 4. A traffic scheduling scheme is an exact traffic scheduling scheme iff (i) the scheme is feasible, and (ii) for every reserved time slot, there exists at least one trigger scenario where its corresponding egress port is utilized during the slot, i.e., no reserved slot remains unused across all trigger
scenarios. Fig. 5 illustrates the three stages within a hyperperiod from the controller’s perspective. These three stages are separated by the earliest and latest possible completion times for the internal processing, respectively. In the first stage, the controller has not yet received or is processing the internal state information that may trigger the ST stream. In the second stage, the ST stream may arrive at any moment. In the third stage, the ST stream has either been generated or will not appear until the end of the hyperperiod. Therefore, reserving is valid only in the second stage. To obtain an exact traffic scheduling scheme, we introduce schedule streams S = {s1,1 , s1,2 , . . . , sn,m } as a collection of hypothetical streams. The frames of a schedule stream si,j are ⟨n ,n ⟩ ⟨n ,n ⟩ ⟨n ,n ⟩ ⟨n ,n ⟩ Fsi,ja b = ⟨fsi,ja,1 b , fsi,ja,2 b , . . . , fsi,ja,k b ⟩, and their start ⟨n ,n ⟩
⟨n ,n ⟩
⟨n ,n ⟩
transmission times are Φ = ⟨ϕsi,ja ,1 b , ϕsi,ja ,2 b , . . . , ϕsi,ja ,k b ⟩. ⟨n ,n ⟩
⟨n ,n ⟩
The transmission of fsi,ja,k b takes τsi,ja,k b amount of time, and the maximum allowed E2E latency of si,j is Di,j . Since the transmission time is constant (given fixed bandwidth and maximum frame size), each slot can be represented by its start ⟨n ,n ⟩ time only: the interval is simply [ϕsi,j ,k , ϕsi,j ,k + τsi,ja,k b ]. All the schedule streams are mapped from actual streams. Each TT actual stream is mapped to a single schedule stream, regardless of whether it shares time slots. Each ET stream maps to N schedule streams, while each ST stream is mapped to M schedule streams. Such one-to-many mappings are intended to characterize the volatility of their arrivals for scheduling. By discretizing all arrival possibilities of an ET actual stream ri into N cases, the arrival time of its jth mapped schedule stream si,j can be written as Ai,j = Ti × Nj . For ST streams, Ptr [last]
Ptr [last]
Ai,j = ϕstr i,1 ,last + τstr i,1 ,last + Cimin + ψi × ( i
i
j ), M
(2)
where ψi = Cimax − Cimin indicates the processing time interval of ST stream si , and tri is the TT stream that may ultimately trigger si . With the above notations, an exact traffic scheduling scheme can be obtained by Alg. 1. Generally, we first convert all actual streams into schedule streams respectively, then formulate the schedule streams scheduling problem with a set of constraints, and finally solve the problem using an SMT solver. The inputs of Alg. 1 are actual streams R and the network G. An empty schedule stream set S is first initialized. During the traversal of each actual stream, the stream’s type determines the rules according to which it will be converted. 1) A strict stream retains the original values of all its attributes to ensure that it gets strictly scheduled. 2) In contrast, the transmission time for each share stream is extended based on its per-hop ET demand. The extension must be computed recursively: when a share stream’s ⟨n ,n ⟩ transmission time τi,1a b increases, the number of pos⟨n ,n ⟩ sible ET preemptions ⌈τi,1a b /Tj ⌉ may also increase, requiring further extension. The algorithm iterates until a fixed point is reached, at which point no transmission time
streams can only start after arriving at the first node.
Algorithm 1: Exact traffic scheduling of ARTSN Input: Network G, actual streams R Output: Start times Φ, Queue assignment Q 1 S ← {}; 2 forall ri ∈ R do 3 if ti ∈ {strict, share} then 4 Initialize si,1 according to ri ; 5 if ti = share then 6 repeat 7 forall rj ∈ R where (tj = event ∧ ⟨na , nb ⟩ ∈ Pi ∩Pj ) do ⟨n ,nb ⟩
8
τet ← τj,1a
9
τi,1a
10
until no τ changes
⟨n ,nb ⟩
P [1]
i ∀si,j ∈ S, (ti ∈ {strict, share}) ∨ (ϕsi,j ,1 ≥ Ai,j ).
Frame Sequence Constraint. Even when multiple streams mix at the same path, the sending sequence of their frames must maintain the original sequence within each stream. ⟨n ,n ⟩
a ,nb ⟩ ∀si,j ∈ S, ∀⟨na , nb ⟩ ∈ Pi , ∀fsi,ja,k b ∈ Fs⟨n : i,j
⟨n ,n ⟩
τi,1a b Tj
⟨n ,nb ⟩,orig
← τi,1a
⟨n ,n ⟩
b if fsi,ja,k b is not fsi,ja,last :
⟨n ,n ⟩
⟨n ,n ⟩
(5)
⟨n ,n ⟩
b ϕsi,ja ,k b + τsi,ja,k b ≤ ϕsi,ja ,k+1 .
⟨n ,n ⟩
×
End-to-End Latency Constraint. The difference between a frame’s arrival time and its complete time must not exceed the maximum allowed E2E latency.
;
+ τet ;
∀si,j ∈ S, P [last]
11 12
14
15
Add all generated si,j to S;
16 17
i (ti ∈ {event, self} ∧ ϕsi,j ,last − Ai,j ≤ Di,j ) ∨
else if ti = event then Initialize ⟨si,1 , si,2 , . . . , si,N ⟩ with Di,j = DiR − TNi , Ai,j = Ti × Nj ; else if ti = self then Initialize ⟨si,1 , si,2 , . . . , si,M ⟩ with i Di,j = DiR − ψ M , Ai,j set by Eq. (2);
13
(4)
(6)
Pi [last] Pi [f irst] (ti ∈ {strict, share} ∧ ϕsi,j ≤ Di,j ). ,last − ϕsi,j ,1
Frame Overlap Constraint. When two frames are going to be sent from the same egress port of the same switch, they can overlap in only two cases: 1) These two frames are derived from the same ET or ST stream; 2) One of them is ET, and the other is shareable TT. Other than the two cases, the sending time windows of the two frames must not overlap.
Schedule constrained S using SMT solver for Φ, Q; return Φ, Q
a ,nb ⟩ a ,nb ⟩ ∀⟨na , nb ⟩ ∈ L, ∀Fs⟨n , Fs⟨n , i ̸= k, i,j k,l
{ti } ∪ {tk } ̸= {event, share} : ⟨n ,n ⟩
changes. This recursive extension, overlooked in prior work such as E-TSN [16], ensures that share streams have sufficient reservation to accommodate all possible ET preemptions without missing their deadlines. 3) Random arrivals of every event stream are uniformly divided into N parts for schedule consideration. The j-th schedule stream assumes arrival at time Ai,j = Ti × Nj , with deadline Di,j = DiR − TNi . 4) Similarly, the randomness of the occurrence time of each self stream is considered by means of M uniform possible intervals. The deadline for each possibility also subtracts the interval length. Assigning larger values for N and M will lead to more frequent redundant reservations, enhancing real-time performance but eroding overall network utilization. We formalize the criteria for exact scheduling schemes (there can be more than one) as the following constraints: Offset Range Constraint. Each frame needs to be transmitted within its period. Therefore, a frame cannot start earlier than the period itself nor later than the latest possible start time to complete transmission within the period. ⟨n ,n ⟩
a ,nb ⟩ ∀si,j ∈ S, ∀⟨na , nb ⟩ ∈ Pi , ∀fsi,ja,k b ∈ Fs⟨n : i,j
⟨n ,n ⟩
⟨n ,n ⟩
0 ≤ ϕsi,ja ,k b ≤ Ti − τsi,ja,k b .
(3)
Arrival Time Constraint. The transmission of ET and ST
a ,nb ⟩ a ,nb ⟩ a ,nb ⟩ ∀fs⟨n ∈ Fs⟨n , fsk,la,h b ∈ Fs⟨n , i,j ,g i,j k,l
∀x ∈ {1, 2, . . . , lcm(Ti , Tk )/Ti },
(7)
∀y ∈ {1, 2, . . . , lcm(Ti , Tk )/Tk } : ⟨n ,n ⟩
⟨n ,n ⟩
a ,nb ⟩ (ϕ⟨n + x × Ti ≥ ϕsk,la ,hb + y × Tk + τsk,la,h b ) ∨ si,j ,g
⟨n ,n ⟩
a ,nb ⟩ a ,nb ⟩ (ϕsk,la ,hb + y × Tk ≥ ϕ⟨n + x × Ti + τs⟨n ). si,j ,g i,j ,g
Queue Assignment Constraint. The switches achieve spatial isolation of streams by queues. We reserve one dedicated queue for ET streams and one for ST streams. TT streams that do not share time slots with ETs use the queues from N SHL to N SHH , and TT streams that perform time slot sharing with ETs use the queues from SHL to SHH . ∀si,j ∈ S : (ti = strict ∧ N SHL ≤ qi ≤ N SHH ) ∨ (ti = self ∧ qi = ST ) ∨
(8)
(ti = event ∧ qi = ET ) ∨ (ti = share ∧ SHL ≤ qi ≤ SHH ). In addition to the above constraints, the frame isolation constraint (Appendix A) and the adjacent link constraint (Appendix B) are also needed in line with [16], [38]. With all these constraints, an exact traffic scheduling scheme can be obtained with the Z3 solver [43]. According to the obtained scheme, Alg. 1 eventually outputs the times Φ when each frame starts to be transmitted in every hyperperiod and the queue assignment Q = {q1 , q2 , . . . , qn } of all actual streams.
Fig. 6. Three phases of the adaptive slot release mechanism.
network resource overhead (for transmitting RP frames) and hardware support from the switches. For these reasons, the mechanism should be activated selectively on certain switches at certain times. For example, as shown in Fig. 8, the worst waste occurs at the central switch, where enabling the mechanism yields the greatest benefit. We call it low-intrusive because other switches do not require any changes. Given the minimum Ethernet frame size of 64 bytes, we designed the RP frame to be 70 bytes. Since the RP frames are transmitted as BE traffic without real-time guarantees, release failures may occur. Thus, the mechanism should be activated only when ST streams are relatively heavily loaded and volatile, as the payback and success rate of releasing are higher. The mathematical expectation of the release gain for each ST stream applied should be at least 70 bytes. V. E VALUATION
Fig. 7. Mapping modification.
Fig. 8. Concentrated waste.
C. Adaptive Slot Release To address the resource waste caused by the absence of ST streams, we design an adaptive slot release mechanism (Fig. 6) with three phases: a processing phase, a reporting phase, and a releasing phase. The structure of this section is organized accordingly. 1) Processing Phase: The controller extracts the payloads of the TT frames and performs calculations to make decisions. Distinguished from the time when switches and NICs process the data frame headers, this processing is handled by applications running on top of the controller. After processing, whether the corresponding ST stream needs to be triggered can be determined from the next system action indicated by the controller’s processing result. 2) Reporting Phase: As soon as the controller becomes aware that it no longer needs to produce the ST stream, it can attempt to report its absence to downstream switches by transmitting a small report frame. This frame (marked as RP in Fig. 6) requires only a unique identifier, along with the ID of the reported ST stream and the hyperperiod. The controller sends this frame as BE traffic. Since this RP frame is sent from the same source but prior to the ST stream, it is expected to arrive at each hop ahead of the reserved time slots of the ST stream, although this is not guaranteed. 3) Releasing Phase: Once a downstream switch receives a frame reporting the absence of a particular ST stream and specifying the current hyperperiod, the mapping relationship between frame priority and switch queues is modified during that absence. By this means, the switches enable BE traffic to enter the ST queue for transmission during the absence of ST streams, as Fig. 7 shows. While modifying the GCL to open the BE queue gate does not take effect in real-time, this mapping modification operation can achieve the same effect in microseconds with hardware support. This mechanism can adapt to the triggering conditions of ST streams to dynamically regulate network transmission, thereby improving efficiency. However, it requires consideration of
This section starts with environment setup and baselines (§V-A and §V-B), followed by comprehensive performance comparisons in terms of schedulability (§V-C), scalability (§V-D), efficiency (§V-E), and concludes with the reliability validation (§V-F). Six key metrics are included: 1) number of schedulable scenarios, 2) memory consumption, 3) runtime, 4) theoretical utilization, 5) E2E latency, and 6) E2E jitter. Our experiments lead to the following key findings: • ARTSN’s schedulability significantly outperforms other approaches in ST-contained scenarios, enabling effective scheduling even when all other approaches fail. • ARTSN consumes the least memory and the shortest runtime under increasing network loads, demonstrating the best scalability with the slowest overhead growth. • ARTSN is exceptionally efficient in ST-contained scenarios. It exactly reserves all valid time slots and may even achieve zero waste through dynamic release. • Verified by hardware in practice, ARTSN can guarantee critical transmissions under background traffic disturbances. In our 5-hop testbed, it achieved a 47.18 µs worstcase latency and a 1.01 µs maximum jitter over 5 minutes. A. Environment Setup Referring to the case studies on substation communication networks [19] and in-vehicle networks [44], we conduct experiments with two representative topologies, as in Fig. 9. Each local switch and all its connected end-devices form a bay. There are 4 bays in each topology. The sensor, controller, and actuator bays are connected to three of their corresponding end-devices, while the mixed bay is connected to one of each kind. The difference between the two topologies is that, in the TABLE I S TREAM PATTERN OF EVALUATION
Number Type Size (bytes) Period (ms) Probability
5
10 5 share
5
10
5
10 5 strict 500 * U 20 5 10 20 100%
1
10 50%
1 1 self 1500 * U 10 10 70% 80%
Fig. 9. Line topology (left) and star topology (right) with 4 bays.
(a) Line topology
(b) Star topology
Fig. 11. Schedulability on the same set of network scenarios.
(a) Line topology Fig. 10. Testbed with 5 TSN switches and 2 TSN NICs.
line topology, the local switches are directly connected to a single line, whereas in the star topology, they are collectively connected to a central switch. According to the IEC 61850 requirements for GOOSE traffic on real-time substation control [45] and the event tracking performance of embedded SCADA systems [19], the stream workload is set as in Table I. All links have 1 Gbps bandwidth. A total of 43 streams are in our test networks: 40 are TT streams, and 3 are ST streams. To accommodate more practical application scenarios, the 3 ST streams are set to different sizes or trigger chances. We use a utilization factor U to scale the streams from 500 bytes (1 Ethernet frame) to 150,000 bytes (100 Ethernet frames) to achieve different network utilization rates. Each stream has an E2E deadline equal to its period, and the hyperperiod is 20 ms. The ST streams share a 1 ms processing-time jitter, meaning a 1 ms interval between the earliest and latest completion times for the internal ST-triggering processing. We set up a simulated environment on tsnkit [46], which is a high-fidelity scheduling and benchmark toolkit for TSN. Simulation is done with an Intel i7-12700KF CPU, 12 cores with a clock speed of 3.60 GHz, and 32 GB DDR4 memory. We further built a testbed in accordance with IEEE TSN standards as in §III-B for actual network measurements to better assess the reliability of ARTSN scheduling. It contains 5 switches and 2 NICs, as in Fig. 10. The switches and Network Interface Controllers (NICs) are both built on AMD KintexTM UltraScaleTM FPGA platforms [47]. B. Baselines We compare ARTSN with the two most relevant state-ofthe-art deterministic scheduling methods. 1) Period: Fully deterministic schedules for multi-hop 802.1Qbv-compliant networks are formalized as an SMT problem in Period [38]. To compare, ST streams are scheduled as TT streams, with their periods set to 1 ms instead of 10 ms, to achieve lower queueing delay by 10× reservations.
(b) Star topology
Fig. 12. Memory consumption under different network loads.
2) E-TSN: Based on Period, E-TSN [16] further modeled and addressed ET traffic scheduling. E-TSN extends the time slots of each shareable TT stream for the total possible ET traffic that could occur within those slots, and allows ET streams to preempt shareable TT streams, thus guaranteeing their transmission. To compare, ST streams are considered ET during scheduling, i.e., they can arrive at arbitrary moments. C. Schedulability We randomly generated 100 network scenarios for each topology, with U ranging from 1 to 50 (two scenarios per U ), to test the scheduling capabilities of the methods. There is no timeout limit set. For each method, a scenario is unschedulable if there is no feasible scheme that satisfies all its constraints in the entire solution space. A scenario is schedulable as long as one feasible scheme is found. For the same 100 network scenarios in the line topology, the schedulability of ARTSN, Period, and E-TSN is shown in Fig. 11(a). The sets of successfully scheduled scenarios exhibit a containment relationship. The 16 scenarios successfully scheduled by E-TSN are fully contained within the 42 successfully scheduled by Period, which in turn are contained within the 97 successfully scheduled by ARTSN. There are only three scenarios that ARTSN cannot schedule, and they are also not schedulable with other methods. Fig. 11(b) demonstrates a similar situation in the star topology. ETSN solves two more scenarios, and ARTSN solves one more. ARTSN’s superior schedulability stems from lower reservation overhead: leveraging deterministic TT–ST triggering, it avoids excessive bandwidth reservations that cause Period and E-TSN to fail under high utilization. D. Scalability We compare the scalability of the scheduling methods from both (1) memory consumption and (2) runtime perspectives. As U ranges from 1 to 10, the network is under increasing traffic pressure. The rates at which memory and runtime
TABLE II U TILIZATION AND RESERVATION CORRESPONDING TO U FACTORS U Actual ARTSN* ARTSN Period E-TSN
1 2.00% 2.01% 2.08% 4.60% 11.60%
9 18.00% 18.01% 18.72% 41.40% 104.4%
20 40.00% 40.01% 41.60% 92.00% 232.0%
40 80.00% 80.01% 83.20% 184.0% 464.0%
50 100.00% 100.01% 104.00% 230.00% 580.00%
(a) Line topology
(a) Net throughput gain (b) Star topology Fig. 13. Runtime for SMT solving under different network loads. The missing bars in E-TSN are due to scheduling failures or to timeouts.
overhead increase with traffic pressure are key indicators of scalability. Due to limited computing resources, a 4-hour timeout is set, as in [48]. The memory consumption of the scheduling methods under different network loads in both line and star topologies is shown in Fig. 12. In all scenarios, the memory consumption of the three methods shows the same magnitude relationship: E-TSN consumes the most memory, ARTSN the least, and Period is always between them. In the U = 8 case of line topology, E-TSN consumes 2062.492 MB, Period consumes 647.891 MB, and ARTSN consumes only 359.926 MB. Their memory consumption grows as the network load increases, but at different rates. The consumption of E-TSN shows a nearly linear increase from the U = 4 case in both topologies, whereas ARTSN is almost unaffected by network load. Fig. 13(a) and Fig. 13(b) show the runtime of the scheduling methods for the line and star topologies, respectively. In all cases, ARTSN takes at most 13.67 seconds to find a feasible solution, while Period takes quarters, and E-TSN takes hours. In the line topology, the ARTSN runtime increases slightly as network load increases exponentially, whereas it remains almost constant in the star topology. The runtime of Period and E-TSN shows nearly linear growth on a logarithmic scale, i.e., exponential growth on a linear scale. It is worth noting that, even though the network reservation for E-TSN with U of 9 exceeds 100% (as in Table II), it still schedules successfully in the star topology. This is because E-TSN allows some time slots to overlap each other. We attribute the dramatically better scalability of ARTSN to the fact that its exact constraints on the ST streams, based on the relationship between the TT streams and the triggered ST streams, significantly reduce the solution space. A smaller search space leads to an easier search and a lower overhead.
(b) Max efficiency increase
Fig. 14. Results of enabling the adaptive slot release mechanism.
E. Efficiency We define network utilization numerically as PS ⟨na ,nb ⟩ si,j τi,j , U tilization := max H ⟨na ,nb ⟩
(9)
which represents the proportion of time that the most heavily used link is utilized for transmissions within one hyperperiod. Specifically, it measures how heavily loaded a network is. Table II compares the actual network utilization with the reservation of the three scheduling algorithms. The actual network utilization can be derived from Table I. In the case of U = 1, there are 90 TT frames of 500 bytes and 6 ST frames with varying sizes and trigger probabilities in a 20 ms hyperperiod. The expected load is 50 KB per hyperperiod. Given that a 1 Gbps link can transmit 2.5 MB in 20 ms, the actual utilization is 50 KB / 2.5 MB = 2.00%. With the exact traffic scheduling ability, ARTSN can guarantee this 2.00% actual traffic at a cost of 2.08% reservations. So we say ARTSN has an efficiency of 2.00/2.08 = 96.15%. With U up to 40, the actual utilization reaches 80.00%. ARTSN can complete scheduling at 83.20%, with only 3.20% extra reservations. In contrast, Period and E-TSN require 184% and 464% reservations, respectively, which cause them to fail in scheduling. Only 43.48% and 17.24% of the reserved slots are utilized in Period and E-TSN. The efficiency gap arises because Period treats ST as fully periodic (reserving slots throughout the hyperperiod), and E-TSN treats ST as fully aperiodic (reserving at all possible arrival times). In contrast, ARTSN reserves slots only within the bounded jitter window where ST traffic may arrive, achieving near-optimal efficiency. Enabling the adaptive slot release mechanism may further reduce unnecessary reservations, as shown by ARTSN* in Table II. When U = 1, the expected unused reservation of one hyperperiod is (50%∗500 B+30%∗1500 B+20%∗1500 B)∗ (20 ms/10 ms) = 2000 B, while the overhead of sending the
(a) Line topology, U = 1
(b) Line topology, U = 8
(c) Star topology, U = 8
Fig. 15. Average E2E latency of the nine selected streams. Each stream is from one of the different stream classes listed in Table I. Streams 1, 2, 3 are shareable TT, streams 4, 5, 6 are strict TT, and streams 7, 8, 9 are ST. The error bars show the range from minimum to maximum latency.
RP frames is (50% + 30% + 20%) ∗ 70 B ∗ (20 ms/10 ms) = 140 B. RP frames are transmitted as BE traffic. In an ideal situation where every release is successful, ARTSN could achieve a 1860 B net throughput gain for each hyperperiod, as Fig. 14(a) shows. On the contrary, 140 B is incurred as ineffective overhead if all releases fail. The threshold success rate of a positive throughput gain is 7%. Therefore, when the observed success rate exceeds a threshold, ARTSN adaptively activates the slot release mechanism for higher efficiency. The maximum achievable increase in efficiency depends on the amount of ST traffic and its triggering probabilities. Fig. 14(b) shows how the maximum efficiency increase changes with the U factor varying from 1 to 50. The overhead of sending RP frames is constant, while the releasable time slots scale up. This enables the maximum efficiency increase to grow with the ST traffic load as a concave function, until it B = 2/52 ≈ 3.846%. converges to limU →∞ U ∗2U KB−140 ∗52 KB F. Reliability To verify its reliability, we used ARTSN to schedule 9 cases, with U ranging from 1 to 40. All scheduling schemes scheduled by ARTSN were verified through both the tsnkit simulator and our 5-hop testbed. In the simulator, each scheme was repeatedly verified over 100 hyperperiods with randomized ST triggering patterns. On the testbed, we first used the NIC-plugged host as the CNC to solve and broadcast GCL, then performed 5-minute E2E observations for each scheme. A worst-case latency of 47.18 µs and a maximum jitter of 1.01 µs were observed, well within the 5 ms deadline constraint. Both results confirm that all types of critical traffic were well protected from 1 Gbps background interference traffic during transmission, with zero deadline misses observed. We selected 9 streams for a detailed E2E latency comparison, as shown in Fig. 15. As constrained, no deadlines were missed in any scenario. The latency of strict TT streams was always constant, while the latency of shareable TT streams in E-TSN and ST streams fluctuated within a safe range due to the randomness of preemption and ST triggering. Benefiting from the preemption of shareable TT streams, E-TSN achieved the minimum possible transmission latency for most ST streams (except for stream 9 in Fig. 15(c), where ARTSN performs better) but at the cost of a significantly larger jitter. In comparison, the scheduling of Period is more stable, while the scheduling of ARTSN is the most deterministic.
In addition, ARTSN schemes always have the lowest upper bounds of latency on ST streams. In terms of reliability, the upper bounds (worst cases) are significantly more important than the average values and the lower bounds. Compared to the time synchronization error of < 10 ns per hop (measured with oscilloscope), the processing delay, transmission delay, and propagation delay of ∼ 10 µs per hop, the queuing delay can easily amount to 1000+ µs (inferred from Fig. 15). The dominance of queuing delay and the decisive role of scheduling schemes are evident. Through exact scheduling schemes, ARTSN guarantees the lowest latency upper bound, thereby ensuring transmission reliability. VI. C ONCLUSION As real-time systems race toward full autonomy, ST control becomes the key enabler—yet it introduces unique challenges to the ARTS network that existing schedulers cannot address. In this work, we present ARTSN to fill the void. It provides a scheduling paradigm that addresses, for the first time, the volatility and absence challenges in handling ST traffic. The integration of exact traffic scheduling and the adaptive slot release mechanism optimizes network resource utilization. Through both simulated and real-world experiments, ARTSN demonstrates significant improvements in schedulability, scalability, and efficiency without compromising reliability. This work bridges a critical gap between autonomous control and deterministic networking, enabling safer and wider ARTS deployments in industrial automation, smart grids, embodied intelligence, and other fields. A PPENDIX A. Frame Isolation Constraint ∀⟨nc , nd ⟩ ∈ L, ∀si,j , sk,l ∈ S, i ̸= k, c ,nd ⟩ c ,nd ⟩ c ,nd ⟩ c ,nd ⟩ ∀fs⟨n ∈ Fs⟨n , ∀fs⟨n ∈ Fs⟨n , i,j ,u i,j k,l ,w k,l
∀x ∈ {1, . . . , lcm(Ti , Tk )/Ti }, ∀y ∈ {1, . . . , lcm(Ti , Tk )/Tk } : c ,nd ⟩ (ϕ⟨n si,j ,u × ⟨nc , nd ⟩.tu + x × Tk ≤ a ,nc ⟩ ϕ⟨n si,j ,u × ⟨na , nc ⟩.tu + ⟨na , nc ⟩.pd + y × Ti )∨ c ,nd ⟩ (ϕ⟨n sk,l ,w × ⟨nc , nd ⟩.tu + y × Ti ≤ b ,nc ⟩ ϕ⟨n si,j ,u × ⟨nb , nc ⟩.tu + ⟨nb , nc ⟩.pd + x × Tk )∨
(qi ̸= qk )
(10)
B. Adjacent Link Constraints ∀si,j ∈S, ∀⟨na , nb ⟩, ⟨nb , nc ⟩ ∈ Pi : a ,nb ⟩ b ,nc ⟩ , 0) − Fs⟨n o ← max( Fs⟨n i,j i,j
⟨n ,n ⟩
b ,nc ⟩ ∀fsi,jb,k c ∈ Fs⟨n : i,j
⟨n ,n ⟩
(11)
ϕsi,jb ,kc × ⟨nb , nc ⟩.tu − ⟨na , nb ⟩.pd ≥ ⟨n ,n ⟩
⟨n ,n ⟩
b b (ϕsi,ja ,k+o + τsi,ja,k+o ) × ⟨na , nb ⟩.tu
R EFERENCES [1] Tesla, “Full Self-Driving (Supervised),” https://www.tesla.com/fsd, accessed: 2026-01-18. [2] T. P. Swaminathan, C. Silver, T. Akilan, and J. Kumar, “Benchmarking deep learning models on nvidia jetson nano for real-time systems: An empirical investigation,” Procedia Computer Science, vol. 260, pp. 906– 913, 2025. [3] M. Yang, “Avoiding pitfalls when using nvidia gpus for real-time tasks in autonomous systems,” in Proceedings of the 30th ECRTS, 2018. [4] R. Cao, J. Ye, J. Zhang, Q. You, C. Tang, Y. Liu, and Y. Wang, “An adaptive uav scheduling process to address dynamic mobile network demand efficiently,” in 2024 DATE. IEEE, 2024, pp. 1–6. [5] J. De Winter, J. Beckers, G. Van de Perre, I. El Makrini, and B. Vanderborght, “Single assembly sequence to flexible assembly plan by autonomous constraint generation,” Robotics and Computer-Integrated Manufacturing, vol. 79, p. 102417, 2023. [6] M. Jover, M. Barranco, and J. Proenza, “Opportunities and specific plans for migrating from prp to tsn in substation automation systems,” in 2023 IEEE 28th ETFA. IEEE, 2023, pp. 1–4. [7] W. P. Heemels, K. H. Johansson, and P. Tabuada, “An introduction to event-triggered and self-triggered control,” in 2012 IEEE 51st CDC. IEEE, 2012, pp. 3270–3285. [8] X. Yi, K. Liu, D. V. Dimarogonas, and K. H. Johansson, “Dynamic event-triggered and self-triggered control for multi-agent systems,” IEEE Transactions on Automatic Control, vol. 64, no. 8, pp. 3300–3307, 2018. [9] A. Adimoolam, I. Saha, and T. Dang, “Safe self-triggered control based on precomputed reachability sequences,” in Proceedings of the 26th ACM HSCC, 2023, pp. 1–12. [10] S. Gent, P. G. Peón, T. Frühwirth, and D. Etz, “Hosting functional safety applications in factory networks through time-sensitive networking,” in 2020 25th IEEE ETFA, vol. 1. IEEE, 2020, pp. 230–237. [11] T. Zhang, G. Wang, C. Xue, J. Wang, M. Nixon, and S. Han, “Timesensitive networking (tsn) for industrial automation: Current advances and future directions,” ACM Computing Surveys, vol. 57, no. 2, pp. 1–38, 2024. [12] C. Xue, T. Zhang, A. Loveless, and S. Han, “Keepon: Supporting deterministic traffic on standard nics,” in 23rd USENIX NSDI, 2026. [13] Forwarding and Queuing Enhancements for Time-Sensitive Streams, IEEE/ISO/IEC Std., 2009. [14] Enhancements for scheduled traffic, IEEE/ISO/IEC Std., 2018. [15] Cyclic queuing and forwarding, IEEE/ISO/IEC Std., 2019. [16] Y. Zhao, Z. Yang, X. He, J. Wu, H. Cao, L. Dong, F. Dang, and Y. Liu, “E-TSN: Enabling Event-triggered Critical Traffic in Time-Sensitive Networking for Industrial Applications,” in 2022 IEEE 42nd ICDCS. IEEE, 2022, pp. 691–701. [17] X. Zhuge, X. Cai, X. He, Z. Wang, F. Dang, W. Xu, and Z. Yang, “InNetScheduler: In-network scheduling for time- and event-triggered critical traffic in TSN,” in IEEE INFOCOM 2024 - IEEE Conference on Computer Communications, May 2024. [18] X. Jiang, Z. Wang, X. Yang, Y. Jiao, T. Yu, X. Wang, W. Fu, Y. Sun, and Z. Sun, “Hebo: Fpga-based transfer time planning for volatile traffic in tsn,” in 2024 IEEE/ACM 32nd IWQoS. IEEE, 2024. [19] T. Docquier, Y.-Q. Song, and V. Chevrier, “On the relevance of tsn for substation communication networks,” in 2022 IEEE ETFA. IEEE, 2022. [20] Broadcom Inc., “Automotive multigigabit ethernet switch with multilayer security bcm89586m,” May 2022, accessed: 2024-0601. [Online]. Available: https://www.broadcom.com/products/ethernetconnectivity/automotive/switches/bcm89586m [21] Allied Market Research, “Time-Sensitive Networking Market Report,” Jun. 2024, accessed: 2024-06-25. [Online]. Available: http://www.alliedmarketresearch.com/time-sensitive-networking-market
[22] A. T. Loveless, “On ttethernet for integrated fault-tolerant spacecraft networks,” in AIAA SPACE 2015 Conference, 2015, p. 4526. [23] A. Soni, J.-L. Scharbarg, and J. Ermont, “A hybrid approach to wctt analysis in a real-time switched ethernet network,” in IEEE 30th RTAS, 2024. [24] PI Organization, “PROFINET,” Jun. 2024, accessed: 2024-06-25. [Online]. Available: https://www.profinet.com/ [25] M. Ghandour, S. Jleilaty, N. Ait Oufroukh, S. Olaru, and S. Alfayad, “Real-time ethercat-based control architecture for electro-hydraulic humanoid,” Mathematics, vol. 12, no. 9, p. 1405, 2024. [26] R. Garreau, M. Ladeira, E. Grolleau, H. Bauer, F. Ridouard, and P. Richard, “Link between real-time scheduling and time-triggered networks,” in 2023 IEEE RTSS. IEEE, 2023, pp. 397–410. [27] Frame replication and elimination for reliability, IEEE/ISO/IEC Std., 2019. [28] X. Wang, H. Yao, T. Mai, S. Guo, and Y. Liu, “Reinforcement learningbased particle swarm optimization for end-to-end traffic scheduling in tsn-5g networks,” IEEE/ACM Transactions on Networking, vol. 31, no. 6, pp. 3254–3268, 2023. [29] Z. Wang, X. He, X. Zhuge, S. Xu, F. Dang, J. Xu, and Z. Yang, “Enabling network diagnostics in time-sensitive networking: Protocol, algorithm, and hardware,” in 2024 IEEE/ACM IWQoS. IEEE, 2024. [30] J. Sasiain, D. Franco, A. Atutxa, J. Astorga, and E. Jacob, “Towards the integration and convergence between 5g and tsn technologies and architectures for industrial communications: A survey,” IEEE Communications Surveys & Tutorials, 2024. [31] Z. Yang, Y. Zhao, F. Dang, X. He, J. Wu, H. Cao, Z. Wang, and Y. Liu, “CaaS: Enabling Control-as-a-Service for Time-Sensitive Networking,” in IEEE INFOCOM 2023. IEEE, 2023, pp. 1–10. [32] R. Cao, Z. Qi, Q. He, C. Ling, Y. Wang, and G. Tang, “Sagkit: A python sag toolkit for response time analysis of hybrid-triggered jobs,” in 2025 IEEE 45th International Conference on Distributed Computing Systems Workshops (ICDCSW). IEEE, 2025, pp. 339–344. [33] H. Li, C. Lu, and C. Gill, “Predicting latency distributions of aperiodic time-critical services,” in 2019 IEEE RTSS. IEEE, 2019, pp. 30–42. [34] G. Patti, L. L. Bello, and L. Leonardi, “Deadline-aware online scheduling of tsn flows for automotive applications,” IEEE Transactions on Industrial Informatics, vol. 19, no. 4, pp. 5774–5784, 2022. [35] X. He, X. Zhuge, F. Dang, W. Xu, and Z. Yang, “Deepscheduler: Enabling flow-aware scheduling in time-sensitive networking,” in IEEE INFOCOM 2023. IEEE, 2023, pp. 1–10. [36] J. Li, K. Xie, J. Wen, G. Xie, and W. Liang, “Event-triggered traffic scheduling in time-sensitive networks based on causal inference,” in IEEE/ACM 33rd IWQoS. IEEE, 2025, pp. 1–10. [37] W. Steiner, “An evaluation of smt-based schedule synthesis for timetriggered multi-hop networks,” in 2010 IEEE RTSS. IEEE, 2010. [38] S. S. Craciunas, R. S. Oliver, M. Chmelı́k, and W. Steiner, “Scheduling real-time communication in ieee 802.1 qbv time sensitive networks,” in Proceedings of the 24th RTNS, 2016, pp. 183–192. [39] Bridges and Bridged Networks, IEEE Std., 2022. [40] Timing and Synchronization for Time-Sensitive Applications, IEEE Std., 2020. [41] Stream Reservation Protocol (SRP) enhancements and performance improvements, IEEE/ISO/IEC Std., 2020. [42] J. Falk, F. Dürr, and K. Rothermel, “Exploring practical limitations of joint routing and scheduling for tsn with ilp,” in 2018 IEEE 24th RTCSA. IEEE, 2018, pp. 136–146. [43] Z3 Theorem Prover, “The z3 theorem prover,” Jun. 2024, accessed: 2024-06-11. [Online]. Available: https://github.com/Z3Prover/z3 [44] Y. Peng, B. Shi, T. Jiang, X. Tu, D. Xu, and K. Hua, “A survey on in-vehicle time-sensitive networking,” IEEE Internet of Things Journal, vol. 10, no. 16, pp. 14 375–14 396, 2023. [45] T. Docquier, Y.-Q. Song, V. Chevrier, L. Pontnau, and A. Ahmed-Nacer, “Iec 61850 over tsn: traffic mapping and delay analysis of goose traffic,” in 2020 25th IEEE ETFA, vol. 1. IEEE, 2020, pp. 246–253. [46] C. Xue, T. Zhang, Y. Zhou, M. Nixon, A. Loveless, and S. Han, “Real-time scheduling for 802.1 qbv time-sensitive networking (tsn): A systematic review and experimental study,” in 2024 IEEE 30th RTAS. IEEE, 2024, pp. 108–121. [47] AMD, “AMD Kintex UltraScale FPGAs,” Jul. 2024, accessed: 202407-29. [Online]. Available: https://www.amd.com/en/products/adaptivesocs-and-fpgas/fpga/kintex-ultrascale.html [48] M. Nasri and B. B. Brandenburg, “An exact and sustainable analysis of non-preemptive scheduling,” in 2017 IEEE RTSS. IEEE, 2017.