Why deterministic guarantees? From TSN to worst-case bounds
Time-Sensitive Networking (TSN) promises that a standard Ethernet network can carry control traffic with bounded latency. This first post explains what "bounded" means, why it cannot be established by measurement alone, and why deterministic network calculus is the analytical tool of choice for establishing it.
1. The promise of TSN
TSN is a family of IEEE 802.1 amendments that add deterministic behaviour to bridged Ethernet. Industrial guides typically single out four of them [3]:
| Standard | Mechanism | Role in the guarantee |
|---|---|---|
| 802.1AS | Generalised Precision Time Protocol (gPTP) | Common time base for all bridges and end stations |
| 802.1Qbv | Time-Aware Shaper (TAS) | Gate control lists open and close queues at scheduled instants |
| 802.1Qci | Per-Stream Filtering and Policing (PSFP) | Enforces the traffic contract declared by each stream |
| 802.1CB | Frame Replication and Elimination (FRER) | Seamless redundancy over disjoint paths |
A shorter glossary is kept in the TSN glossary.
The promise is converged traffic: control loops, video and best-effort traffic share one physical network, and the control traffic still meets its deadlines. That promise is only as strong as the argument that supports it.
2. Average-case versus worst-case performance
Classical queueing theory answers questions such as "what is the mean delay of a packet?" or "what is the probability that the delay exceeds 1 ms?". Such answers are indispensable for dimensioning best-effort networks, but they are the wrong kind of answer for a safety-critical control loop, for two reasons.
- Rare events are exactly the events that matter. A braking command delayed once in frames is acceptable in a statistical sense and unacceptable in an engineering sense.
- Stochastic assumptions are hard to justify. Poisson arrivals and independence between flows are modelling conveniences; certification authorities ask for arguments that hold for every admissible traffic pattern.
Deterministic network calculus replaces probabilistic assumptions with envelopes: each flow is assumed only to respect an upper bound on the amount of data it can emit in any interval, and each network element is assumed only to guarantee a lower bound on the service it provides. From these two envelopes it derives bounds that hold for every trajectory compatible with them.
3. Why measurement does not suffice
The practical TSN migration guide by TSF Technology proposes a five-phase method: audit the existing network and identify critical flows, design the TSN topology and synchronisation domains, pilot on a segment, configure scheduled traffic with 802.1Qbv, and finally validate latency and jitter against requirements [3]. The last phase is usually carried out by measurement.
Measurement establishes a lower bound on the worst-case delay: the largest observed value. A certificate needs an upper bound. The gap between the two can be large, because the worst case arises from a precise alignment of bursts on every hop of a path, which a test campaign is unlikely to provoke. In the companion repository on certified admission control, packet-level simulation of 4,915 admitted flows observed a median worst delay of only 19 % (control traffic) and 10 % (stream traffic) of the analytical bound [4]. These figures are a sanity check, not a tightness measure: they illustrate that the observed maximum and the provable maximum are different quantities.
A measured maximum answers "what happened?". A network calculus bound answers "what can happen, under the stated assumptions?". Only the second can be used as a guarantee.
4. Static configuration and the reconfiguration problem
TSN guarantees are usually established for a configuration that is computed and verified offline, then frozen. In practice, networks change: flows are added and removed, links fail, and production lines are reconfigured. Two recent lines of work, discussed in the reconfiguration post, illustrate the tension:
- Sertbaş Bülbül, Ergenç and Fischer use a software-defined networking (SDN) controller to migrate time-sensitive flows to new paths at run time [5];
- Boyer and Henia design a Time-Aware Shaper schedule that can absorb rescue flows after a permanent link fault without modifying existing windows [6].
Both raise the same question: after the change, does every admitted flow still meet its deadline? Answering it quickly and soundly is the motivation for this series.
5. Plan of the series
The posts are grouped by topic. Topics are meant to be read in this order, and posts in the order listed.
Recommended viewing
- Networks and Calculus, lecture by Prof. Jean-Yves Le Boudec (EPFL School of Computer and Communication Sciences). YouTube.
References
- J.-Y. Le Boudec, P. Thiran, Network Calculus: A Theory of Deterministic Queuing Systems for the Internet, LNCS 2050, Springer, 2001. Online edition
- A. Bouillard, M. Boyer, E. Le Corronc, Deterministic Network Calculus: From Theory to Practical Implementation, Wiley-ISTE, 2018.
- TSF Technology, "Migration vers TSN : guide pratique," technical guide. Link
- CNAM-ECE, Certified Admission Control for TSN, companion repository. GitHub
- N. Sertbaş Bülbül, D. Ergenç, M. Fischer, "Towards SDN-based Dynamic Path Reconfiguration for Time Sensitive Networking," IEEE/IFIP NOMS 2022. IEEE Xplore
- M. Boyer, R. Henia, "Designing a Resilient Time-Aware Shaper Configuration for TSN," Ethernet & IP @ Automotive Technology Day, Toulouse, October 2025. Slides