PaperlayerView exact revision on arXiv

Source-linked AI summary

Decentralized Latency Models and Information Flow for TSN Shapers - A Tutorial

Alexej Grigorjew, Stefan Geissler, Jürgen Schmitt, Tobias Hossfeld

arXiv:2610.11999v1cs.NI

TL;DR

Decentralized latency analysis remains challenging for heterogeneous networks, where bounds often depend on global information or homogeneous shaping assumptions. This tutorial formalizes shaper-to-shaper delay segments, local information flows, and five latency models; its evaluation finds that model suitability depends on topology and traffic conditions.

  • Problem

    Decentralized heterogeneous deployments need local latency bounds, but current latency models assume homogeneous upstream shaping.

    Source: §VIII · p. 23

  • Method

    The paper defines shaper-to-shaper delay segments, specifies local information flows, and derives five decentralized latency models for SP, ATS, CQF, and CBS.

    Source: §IX · p. 24

  • Results

    SP NC consistently ranked among the top performers; ATS was robust, while CQF and CBSA NC were highly situational.

    Source: §IX · p. 24

  • Takeaways & Limitations

    The authors recommend SP NC over SP SIMPLE when its computational complexity is acceptable, and describe ATS as a robust option with simple delay calculations.

    Source: §IX · p. 24

    Current latency models assume homogeneous upstream shaping, so hybrid deployments require theoretical adjustments.

    Source: §VIII · p. 23

Abstract

from arXiv · show

Recent advancements in the fields of industrial automation and in-vehicle communication have also increased the requirements on their underlying networks. Traditional field bus technologies are being replaced by regular Ethernet devices with features from the IEEE Time-Sensitive Networking working group. Their standards are designed to achieve zero congestion loss and deterministic latency bounds, but the specific guarantees can differ depending on the shapers used and how the bounds are calculated. After several contributions to the standardization efforts, this work presents a number of decentralized latency models that can be applied locally on the switches, without requiring global knowledge from the rest of the network. It includes a formal definition of delay segments that allows to incorporate interoperability between different shapers and latency models in the future. It presents pseudo-code suggestions for the latency bounds computation of Strict Priority, Credit-Based Shaping, Asynchronous Traffic Shaping, and Cyclic Queuing and Forwarding, all of which can be applied locally on the devices. Finally, five models were implemented and evaluated with respect to their efficiency in terms of latency, jitter, and number of accepted streams. The results illustrate how different shapers are suited for different networks and traffic distributions.

I. INTRODUCTION

TSN aims to bring open, converged real-time networking to Ethernet, but decentralized latency models and the information they require remain underdeveloped. This tutorial surveys that gap, establishes notation, and lays out questions on local bounds, signaling, interoperability, and shaper choice.

I. INTRODUCTION

  • Decentralized latency models are rare, and many existing worst-case latency computations require global topology and stream information, limiting their fit for local admission control.The tutorial frames local mathematical foundations as a basis for interoperability across disparate technologies.Source: §I · p. 1
  • The tutorial asks how to compute local bounds, acquire the needed information, support shaper transitions, compare models fairly, and recommend shapers for application scenarios.These five research questions structure the tutorial's contribution.Source: §I · p. 2

II. OVERVIEW ON RELATED WORK

  • A review of 311 publications, after excluding 160 unrelated papers, finds that decentralized delay analysis is scarce and that most latency analyses examine a single shaper.Only four publications explicitly distinguished decentralized latency analysis; two focused primarily on path finding, while the others restricted their scope to CBS and SP.Source: §II · pp. 2–4

Lessons learned

  • The tutorial addresses the gap between formal latency models and deployment by detailing models, local information acquisition, reference implementations, and admission-control comparisons.It states that no other known publication spans these categories.Source: Lessons learned · p. 4

III. NOTION AND NOTATION

  • Latency jitter means the full possible range between maximum and minimum end-to-end latency, and the tutorial uses shaper for four TSAs: SP, ATS, CQF, and CBS.NC denotes Network Calculus; latency threshold denotes a pre-configured upper bound.Source: §III · p. 4

IV. DELAY MODEL AND DELAY SEGMENTS

The paper separates transmission, propagation, queuing, and processing delays, then defines per-hop segments between shapers to support decentralized bounds. It describes the local information needed to calculate those bounds and exchange it during reservation.

A. Types of Delay

  • The delay model distinguishes transmission, propagation, queuing, and processing; the latency algorithms cover transmission and queuing, while the other two require separate bounds.Transmission is counted at the receiver and equals frame size divided by link bandwidth.Source: A. Types of Delay · p. 5
  • Queuing is split into artificial shaper delay and priority queuing, separating delays by cause rather than requiring separate physical buffers.Priority queuing serves the highest traffic class that is currently nonempty at line rate.Source: A. Types of Delay · p. 5

B. Delay Segments

  • A delay segment runs from an upstream shaper's eligibility point to the downstream shaper's eligibility point, with the downstream switch computing its worst-case bound.These boundaries give each segment a well-defined workload and let analysis use the traffic properties enforced by shaping.Source: B. Delay Segments · p. 5
  • Post-shaping segment boundaries let analyses use formal traffic properties directly and simplify shaper transitions in heterogeneous networks.Shapers can enforce token-bucket bounds or reduce jitter, while later priority queuing can reintroduce jitter.Source: B. Delay Segments · p. 5

V. INFORMATION FLOW

  • The information-flow model uses stream and neighboring-device information to calculate segment bounds and is intended to support development of the RAP draft.Resource reservation consumes a latency budget, and admission control rejects reservations that exceed a pre-defined threshold.Source: §V · p. 5

A. A Priori Information

  • Before reservation, direct neighbors share upstream shaper configurations and required parameters; during reservation, switches use pre-configured segment thresholds as advertised guarantees.CBS requires idle slope information for tight bounds, while CQF requires matching cycle timings; changing a threshold after reservations may invalidate them.Source: A. A Priori Information · p. 6

B. Per-Stream Information

Stream reservations carry traffic and path information so each switch can estimate queue occupancy and accumulated latency locally. The tutorial explains these fields, identifies signaling lessons, and outlines latency-bound models for four shaper types.

B. Per-Stream Information

  • Reservation information covers stream identification, traffic specification, accumulated latency, and upstream traffic class, which the downstream switch needs to estimate queue traffic.The upstream class must be signaled explicitly because class support and priority-to-class mappings can differ between switches.Source: B. Per-Stream Information · pp. 6–7
  • RAP identifies streams by sender MAC address plus a unique number, while traffic specifications use token-bucket burst and rate parameters to bound traffic over any interval.The token-bucket representation avoids the overhead and reduced network efficiency of translating legacy frame-count specifications.Source: B. Per-Stream Information · p. 7

C. Lessons learned

  • Decentralized bounds require local state signaling, including original traffic specifications, accumulated maximum latency, upstream class mappings, and—when tighter bounds are sought—the previous ingress link speed.A static pre-configured maximum latency threshold is propagated instead of fluctuating tight accumulated latency; the ingress-speed field is not yet supported by standard signaling protocols.Source: C. Lessons learned · p. 7

VI. LATENCY BOUNDS

  • The latency models provide local calculations for SP, refined SP using Network Calculus, CQF, ATS, and CBS, covering transmission and queuing delays.Propagation and processing delays are assumed to have separate device-specific upper bounds.Source: §VI · pp. 7–8

A. Strict Priority – Simple Bound

The Strict Priority simple bound estimates delay from the effective burstiness of reserved streams, accounting for priority interference and a lower-priority packet already in transmission. Its assumptions include compliant traffic specifications and segment delays kept within configured thresholds by admission control.

A. Strict Priority – Simple Bound

  • Downstream latency jitter can cluster formerly separated packets, so effective burstiness bounds simultaneous queue arrivals using the difference between accumulated maximum and minimum latency.The bound assumes end devices conform to their Traffic Specifications and admission control keeps each segment within its configured latency threshold.Source: A. Strict Priority – Simple Bound · p. 8
  • The effective-burstiness formulation generalizes prior work from inter-arrival-time traffic specifications to standard token-bucket-shaped traffic.The bound assumes a full MTU for lower-priority packet length when Best-Effort traffic is unknown; the example uses ℓmax = 1542 Bytes.Source: A. Strict Priority – Simple Bound · pp. 8–9
  • For the highest class, one lower-class packet can delay service by at most ℓmax/Rseg, while its worst-case burst backlog is bounded using effective burstiness.New higher-class arrivals do not increase this delay when the aggregate rate is below link speed, so the backlog clears faster than it accumulates.Source: A. Strict Priority – Simple Bound · p. 8
  • For lower classes, the bound adds lower-class packet blocking, higher-class burst interference, and higher-class arrivals over the configured delay threshold.Admission control verifies that the resulting worst-case delay remains within that threshold.Source: A. Strict Priority – Simple Bound · pp. 8–9
  • The simple bound supports incremental admission checks because adding a candidate stream leaves existing-stream interference terms and ℓmax unchanged.Cached bounds for existing reservations can reduce computation to one addition, an optimization noted as omitted in this paper.Source: A. Strict Priority – Simple Bound · p. 9

B. Strict Priority – Tighter Bound

The tighter Strict Priority bound models arrivals and service with piecewise linear curves, accounting for ingress-port serialization. It computes worst-case delay from their maximum horizontal deviation, trading tighter bounds for greater computation and reservation-information requirements.

B. Strict Priority – Tighter Bound

  • Aggregating arrivals per ingress port captures physical limits on simultaneous bursts and yields a tighter arrival curve than summing individual stream curves.Each port’s curve is constrained by source production and ingress link speed; the full curve sums the port-level curves.Source: B. Strict Priority – Tighter Bound · pp. 9–10
  • Worst-case delay is the maximum horizontal distance between the arrival and service curves, with service represented by rate-latency curves.The service latency accounts for waiting imposed by higher-priority traffic; arrival and service curves are based on signaling information.Source: B. Strict Priority – Tighter Bound · pp. 9–11
  • For the highest class, service proceeds at link speed once selected, but an already-started lower-class transmission can delay its start by up to one MTU.For lower classes, higher-priority bursts and newly arriving higher-priority traffic reduce the service guarantee.Source: B. Strict Priority – Tighter Bound · p. 11
  • The ingress-port arrival-curve implementation requires stream-to-port mapping and upstream link speeds to be supplied during reservation.A port curve uses a burst based on the maximum frame length; stream burstiness affects the bend time rather than that immediate-arrival burst.Source: B. Strict Priority – Tighter Bound · p. 11
  • The tighter Strict Priority bound is tighter than the simple bound, but more complex and slower to compute, and requires more reservation information.The paper describes numerical computation by iterating through arrival-curve pieces until their slope is no greater than the service rate.Source: B. Strict Priority – Tighter Bound · pp. 11–12

C. Cyclic Queuing and Forwarding

CQF is modeled by adapting the Strict Priority latency model to account for alternating queues and synchronized gating. Its fixed hop-to-hop timing depends on admission control keeping delay below the configured window threshold, while its token-bucket model can overestimate interference for periodic traffic.

C. Cyclic Queuing and Forwarding

  • CQF’s out-of-phase PSFP and TAS gates make each hop between shapers take exactly T time units when admission control keeps delay below δ ≤ T − dmisc.Here dmisc accounts for synchronization inaccuracies, and the downstream shaper treats a frame as eligible exactly T after the upstream shaper did.Source: C. Cyclic Queuing and Forwarding · p. 12
  • CQF halves usable traffic classes because each external class needs two internal queues, and consolidating classes into higher priorities increases interference and worst-case delay.The evaluation uses eight available classes, with classes 4–7 allocated to reserved traffic.Source: C. Cyclic Queuing and Forwarding · p. 12
  • CQF adapts the simple Strict Priority model because gating prevents the faster-than-arrival backlog clearing assumed by Strict Priority, allowing traffic to accumulate for up to the delay threshold.The paper therefore uses the simple Strict Priority formula; the serialization-based tighter bound does not apply, and the final TAS stage would need a different service curve.Source: C. Cyclic Queuing and Forwarding · p. 13
  • A token-bucket model can assign at least 240 bytes of worst-case interference to a 120-byte packet sent every 2 ms when the CQF window is 2 ms.The model permits a second packet before queue alternation; the authors identify specialized periodic-stream models as beneficial for such cases.Source: C. Cyclic Queuing and Forwarding · p. 13

D. Asynchronous Traffic Shaping

ATS enforces each stream’s token-bucket constraints with per-stream shaping and virtual queues to limit head-of-line blocking. Its latency bound uses original traffic specifications without accumulated-latency fields or configured thresholds, but its queue allocation rules require more complex hardware.

D. Asynchronous Traffic Shaping

  • ATS assigns each stream a virtual token bucket and separates traffic into virtual queues by ingress port, egress port, or traffic class to mitigate head-of-line blocking.A frame transmits only when its size in tokens can be subtracted without making the bucket negative.Source: D. Asynchronous Traffic Shaping · p. 13
  • The ATS delay bound uses the end devices’ original burstiness and does not require a preconfigured latency threshold.Its queueing rules require more complex hardware, and limited head-of-line blocking makes the serialization-based tighter bound inapplicable.Source: D. Asynchronous Traffic Shaping · p. 13
  • The ATS model is independent of accumulated latency fields and latency thresholds, making its implementation simpler than the Strict Priority Algorithms 1 and 2.The paper notes that its analysis was previously presented in.Source: D. Asynchronous Traffic Shaping · p. 13

E. Credit-Based Shaping

CBS shapes traffic by class using credit state, so aggregation can add latency jitter and burstiness. Its local latency model adapts arrival and service curves to CBS behavior, while the credit-reset interaction complicates separation of shaping from priority selection.

E. Credit-Based Shaping

  • Class-based CBS aggregates streams under one shaper, and that aggregation introduces latency jitter and increased effective burstiness.Credit rises at the idle slope and decreases during transmission; positive credit accumulates only while the queue is non-empty.Source: E. Credit-Based Shaping · p. 14
  • CBS complicates strict separation of shaping and priority selection because credit resets depend on whether same-class frames remain queued.The shaping step completes only after a frame is transmitted, while priority queuing is treated as empty; the downstream switch accounts for its own queuing delay.Source: E. Credit-Based Shaping · p. 14
  • The CBS latency analysis adapts the SP model by assigning queuing delay downstream and incorporating the upstream idle slope into its piecewise arrival curve.The section assumes the upstream switch also uses CBS and has already accounted for its own queuing delay.Source: E. Credit-Based Shaping · p. 14
  • CBS arrival curves use the infimum of three upper bounds, reflecting line-rate transmission while credit is positive and idle-slope limits afterward.The curves have intersection points, and ingress-port arrival curves are summed for the class-level arrival curve.Source: E. Credit-Based Shaping · pp. 14–15
  • The service curve uses the CBS idle slope and omits temporary higher-rate phases from accumulated positive credit.The text notes those high-rate phases can be interrupted by other events; service-curve offsets are computed in two steps.Source: E. Credit-Based Shaping · p. 15
  • Algorithm 7 represents CBS arrival curves with two bend positions and computes service offsets recursively because each class offset depends on the next higher class.Streams are grouped by ingress port and kept separate when upstream traffic-class mappings—and therefore ingress idle slopes—differ.Source: E. Credit-Based Shaping · p. 15

F. Lessons learned

The paper derives decentralized worst-case delay models for SP, CBS, CQF, and ATS, with pseudocode implementations. More detailed SP and CBS Network Calculus models can tighten bounds, but require more information and computation.

F. Lessons learned

  • Additional Network Calculus models for SP and CBS provide tighter bounds at the cost of more information and more complex computation.They exploit serialization: additional workload arrives over time while previously received workload can already be served.Source: F. Lessons learned · pp. 15–16
  • All four shapers follow a shared pattern: combine traffic specifications and accumulated latencies to bound workload, then use link speed or shaper rate to compute service time.The section reports formal proofs for worst-case delays in decentralized architectures and suggests pseudocode implementations.Source: F. Lessons learned · pp. 15–16
  • The approach constructs local arrival curves from accumulated maximum latency and traffic specifications instead of repeatedly recalculating Network Calculus theorems when streams change elsewhere.The avoided theorems include leftover service and output characterization.Source: F. Lessons learned · p. 16

VII. EVALUATION

The evaluation compares five latency models across three scenarios and reports performance through selected KPIs. The scenarios examine stream acceptance under fixed thresholds, end-to-end guarantees when all streams are accepted, and individual parameter effects.

VII. EVALUATION

  • The evaluation compares five latency models using KPIs and scenario parameters summarized in Table IV.The study also provides its implementation on GitHub.Source: §VII · p. 16
  • The fixed-threshold scenario analyzes stream acceptance, while the auto-configured scenario accepts all streams to evaluate end-to-end guarantees.These are two of the three evaluation scenarios described.Source: §VII · p. 16
  • A single-switch parameter study observes the effects of individual parameters.The three scenarios are summarized in Table IV.Source: §VII · p. 16

A. KPIs

The KPI section treats admitted streams as the direct reservation-performance indicator and also tracks traffic volume, latency-related measures, and computation time. It cautions that threshold-based bounds are valid only while configured limits hold, and that computation timings are rough comparisons.

A. KPIs

  • The number of successfully admitted stream reservations is the most direct performance indicator for deterministic-network resource reservation.Reservations are accepted or rejected along their path.Source: A. KPIs · p. 16
  • Accepted data rate reflects traffic volume, while calculated latency bounds indicate admission potential because lower predictions permit more admissions.The passage also identifies latency jitter and effective burstiness as possible use-case-specific measures.Source: A. KPIs · p. 16
  • Most decentralized models require preconfigured latency thresholds, and a bound remains valid only while every network link stays below its threshold.Excessive admissions that push a link past its threshold invalidate the model assumptions and calculated bounds.Source: A. KPIs · p. 16
  • When reservations are rejected, the resulting latency bounds do not properly reflect model performance because they represent the predefined configuration.The passage states that the underlying latency-model assumptions fail in this situation.Source: A. KPIs · p. 16
  • Computation time offers only a rough comparison because Python timings vary with garbage collection and omit optimizations such as caching intermediate results.The passage nevertheless presents the timings as a rough overview of relative algorithm complexity.Source: A. KPIs · p. 16

B. Network Topologies

The evaluation combines several topology types with pre-configured admission scenarios to compare shapers. ATS generally leads on accepted streams, throughput, and bursts in industrial topologies, while topology shape and priority affect the comparisons.

B. Network Topologies

  • The evaluation spans linear, synthetic industrial, and single-switch topologies; the linear baseline uses 25 switches, while industrial profiles scale from small to big.Small, medium, and big profiles combine chains, rings, and branches, with the big profile adding a fat-tree section; the single-switch study varies sender count.Source: B. Network Topologies · pp. 16–17

C. Pre-configured Scenario & Accepted Streams

  • ATS generally accepts the most streams, throughput, and burst in industrial scenarios; CQF favors higher rates, while SP favors higher bursts.CQF, SP SIMPLE, and SP NC accept similar stream counts, with little difference between the two SP models.Source: Figure 18 · p. 17, C. Pre-configured Scenario & Accepted… · p. 18
  • Pre-configured reservations are simulated sequentially across five models, with 100 generated topologies and 100 shuffled stream orders per scenario.The setup uses per-hop latency thresholds and a 20% per-class bandwidth threshold for CBSA NC.Source: C. Pre-configured Scenario & Accepted… · pp. 17–18
  • Industrial topologies show similar KPI trends across small, medium, and big sizes, but network shape can change which latency model is preferred.The relevant structural differences include path lengths, ingress-port counts, and fan-in/fan-out.Source: C. Pre-configured Scenario & Accepted… · p. 18
  • In the linear topology, CQF appears best because its higher latency threshold increases accepted streams, while the all-priority-7 traffic neutralizes its consolidation penalty.This advantage does not occur in other topologies.Source: C. Pre-configured Scenario & Accepted… · p. 18
  • In the linear topology, ATS, CBSA NC, and SP NC perform similarly across KPIs, while SP SIMPLE accepts only half as many streams.The comparison attributes the capacity difference to improvements in the other models.Source: C. Pre-configured Scenario & Accepted… · p. 18
  • In industrial topologies, CBSA performs worst across KPIs, particularly for rates, and its 75% utilization quantile remains below other shapers despite a 20% bandwidth cap.The lower performance is visible before the cap is reached.Source: Figure 20 · p. 18, C. Pre-configured Scenario & Accepted… · p. 18

D. Auto-configured Scenario & E2E Guarantees

The section evaluates auto-configured per-hop bounds and compares them with global configuration. In industrial topologies, CQF and CBSA have much higher latency than other shapers, while the relative performance of ATS and SP NC depends on the metric and scenario.

D. Auto-configured Scenario & E2E Guarantees

  • Auto-configuration iteratively adjusts per-hop latency thresholds until calculated delay bounds fit, but CQF instead uses a shared maximum-delay threshold because it cannot optimize per link.The study accepts all streams for this comparison and records latency, effective burstiness, and computation time.Source: D. Auto-configured Scenario & E2E… · pp. 18–19
  • In industrial topologies, CQF and CBSA latencies can differ by a factor of 100 from other shapers, while SP NC ranks second across KPIs.CQF’s global threshold spreads bottleneck-link delays to other links; CBSA NC’s reduced service rate contributes to its poor results in these scenarios.Source: Figure 21 · p. 19, D. Auto-configured Scenario & E2E… · p. 19
  • Big topologies show a low-latency step for ATS and both SP models: roughly 18% of end-to-end latencies fall between 20 µs and 100 µs.The step reflects short paths in the generated 1 Gbit/s fat-tree subnet; it is not visible for CQF or CBSA.Source: D. Auto-configured Scenario & E2E… · p. 19
  • Linear topology latency curves have 25 steps; CBSA and CQF do better without fan-out, while SP NC outperforms ATS, likely through serialization gains.All streams share an egress port, and the delay-segment calculations have five ingress ports.Source: D. Auto-configured Scenario & E2E… · p. 19
  • ATS and CQF provide constant receiver bursts: ATS shapes to source traffic specifications, while CQF’s constant-delay shaping prevents burstiness cascading.CQF’s jitter bound is also constant across streams in the linear topology because the streams share the same last hop.Source: D. Auto-configured Scenario & E2E… · p. 19
  • In medium topologies, ATS, SP NC, and SP SIMPLE show lower delay and jitter at higher priorities, while ATS effective bursts remain uniform across priorities.CBSA’s service rate limits dominate its priority trend; CQF’s consolidated priority pairs perform similarly.Source: Figure 22 · p. 20, D. Auto-configured Scenario & E2E… · p. 20
  • ATS matches or outperforms SP NC on effective bursts, SP NC outperforms SP SIMPLE on latency, and ATS competes closely with SP NC, especially at high priorities.For priority 7, 25% of ATS streams outperform SP NC, while 30% of SP NC streams outperform ATS; the choice depends on per-hop shaping versus physical serialization.Source: D. Auto-configured Scenario & E2E… · p. 20
  • CBSA NC is the most computationally demanding model for large topologies, whereas SP SIMPLE and ATS are highly efficient.The timing results are a naive baseline and could shift with algorithm optimization.Source: D. Auto-configured Scenario & E2E… · p. 20

E. Single Switch Parameter Study

The single-switch main-effects study varies parameters individually to isolate why latency-model behavior differs across topologies and traffic settings. CBSA NC’s idle-slope ratio is a key discriminator, alongside model-specific effects from other parameters.

E. Single Switch Parameter Study

  • Figure 26 summarizes a one-parameter-at-a-time study with other factors held at defaults and parameters scaled to avoid confounding; CBSA NC is an unexpected outlier.Source: E. Single Switch Parameter Study · pp. 20–21, Figure 26 · p. 21
  • Higher bandwidth reduces latency across all shapers as queues empty faster.Source: E. Single Switch Parameter Study · p. 21
  • More ingress ports increase SP NC and CBSA NC arrival burstiness and slopes; beyond roughly 17 ports, concurrency outweighs ingress shaping and CBSA NC delays exceed SP NC.CQF, ATS, and SP SIMPLE are unaffected by port count in these models; SP NC converges toward SP SIMPLE as ports exceed stream count.Source: E. Single Switch Parameter Study · p. 21
  • Beyond five streams, SP NC delay decreases slightly as burstiness is split across streams; CBSA NC shows a similar trend on a larger scale, while 1,000 streams trigger a burst correction.For SP NC, delay rises through five streams, matching the five default ingress ports; at 1,000 streams, individual bursts fall below 64 bytes and are corrected, increasing total delay.Source: E. Single Switch Parameter Study · pp. 21–22
  • Higher total burst increases latency for all models except CBSA NC, whose delay plateaus after roughly 128 kB.The paper attributes the CBSA NC plateau presumably to the burst-to-maximum-credit ratio making a linear arrival-curve segment insignificant.Source: E. Single Switch Parameter Study · p. 22
  • Higher total data rates increase SP SIMPLE, SP NC, and CQF delays, while ATS is unaffected within the same class and CBSA NC delay decreases.CQF shows the largest increase, partly due to consolidated priority interference and partly to the main-effects setup; CBSA NC’s increased egress idle slope dominates ingress-slope effects in fan-in scenarios.Source: E. Single Switch Parameter Study · p. 22
  • Higher accumulated jitter increases SP NC, SP SIMPLE, and CQF delays, while higher-priority streams raise delay similarly across shapers except CBSA NC.CBSA NC incorporates jitter but appears dominated by other arrival-curve segments; possible effects on its accumulated positive credit and service-curve latency are insignificant in this case.Source: E. Single Switch Parameter Study · p. 22
  • For idle-slope ratios below 1.0, CBSA NC delay rises to nearly 6 times other shapers’ values; at ratios of 1.0 or more, it is smaller, making the ratio a primary factor behind its topology-dependent performance.The paper says CBSA can be a good option when the egress-to-ingress ratio stays above 1.0, but feasibility depends on topology; it also says SP NC can be used at a switch without additional delay penalties if the criteria cannot be met.Source: E. Single Switch Parameter Study · p. 22

F. Lessons learned

The evaluations favor SP NC overall, while the best shaper depends on topology and KPI. The discussion identifies modeling, signaling, and deployment boundaries that shape how these decentralized models can be used.

F. Lessons learned

  • Shaper comparisons vary with configuration: default global settings showed little difference, while optimal per-hop settings revealed orders-of-magnitude gaps and a factor-of-2.0 difference between previously similar points.CBS, CQF, and ATS each have situational advantages; ATS performs well on receiver burstiness and worst-case latency, while its hardware cost and availability are disadvantages.Source: F. Lessons learned · pp. 22–23
  • SP NC outperforms SP SIMPLE across scenarios and KPIs except computation time; SP SIMPLE often fails to accept all streams under global configuration because of cyclic dependencies.The authors recommend SP NC when signaling and algorithmic complexity are acceptable.Source: F. Lessons learned · pp. 22–23

VIII. DISCUSSION AND FUTURE WORK

  • Hybrid shaper deployments benefit from shared delay segments, but current models assume homogeneous upstream shaping and require adaptations for mixed shapers and traffic classes.The discussion proposes propagating both effective and original Talker TSpecs to communicate burstiness changes downstream.Source: §VIII · p. 23
  • CBS differs from the other models because its shaping step includes queuing and transmission, so hybrid transitions can require downstream bridges to calculate both upstream and downstream queuing delays.With redundancy, a disrupted faster path can leave older frames arriving from the slower path and temporarily increase burstiness after merging.Source: §VIII · pp. 23–24
  • The work supplies a theoretical foundation rather than the full reservation framework, and assumes stream paths and priorities are known in advance.Assigning streams to priorities remains an optimization problem; propagation and internal processing delays also need separate bounds.Source: §VIII · p. 24
  • The models cover transmission and queuing delays, not propagation or internal processing delays, and adopt the ATS delay-segment model for all mechanisms, which enables decentralized CQF with tradeoffs.TAS is not covered except for CQF because it relies on central orchestration.Source: §VIII · pp. 23–24

Lessons learned

  • The shared delay-segment framework and signaling proposals support decentralized resource reservation and shaper transitions, targeting heterogeneous deployments with hard real-time guarantees.The conclusion identifies mixed-shaper latency models and self-configuring decentralized thresholds as future directions.Source: Lessons learned · p. 24, §IX · pp. 24–25

IX. CONCLUSION

  • SP NC consistently ranked among the top approaches; it is preferred over SP SIMPLE when its computational complexity is acceptable, while results depend on topology and stream configuration.CBSA NC degrades rapidly when the idle-slope ratio falls below 1.0 on any link, a condition found in almost all tested industrial topologies.Source: §IX · p. 24
Loading 2610.11999v1…