Chapter 03 of 07

Ideas.

Working notions, half-arguments, the kind of thinking that ends up in the discussion section of a paper. None of these are settled.

  1. 01 age of information

    Latency is not freshness.

    We chase milliseconds when the right question is often how stale. A control loop with 3ms delivery but unsteady cadence can be worse than 5ms with predictable arrival. Age of Information — the time elapsed since the most recent useful update — captures something that p99 latency hides. The protocol design space looks different once you stop optimizing delay alone.

  2. 02 tsn

    Time is a first-class network primitive.

    TSN treats time the way IP treats addresses: a primitive every node agrees on. That single decision changes everything downstream — scheduling, isolation, recovery, observability. The hard part is not the abstraction. It is the synchronization machinery and the kernel paths underneath that have to be honest about it.

  3. 03 urllc

    The robot does not care about your average.

    URLLC is a worst-case discipline. A robotic arm or a haptic loop tolerates the 99.99th percentile, not the mean. The interesting experiments are the long-tail ones, and most testbeds quietly elide them — because reproducing tail behaviour is expensive and the headlines are easier on the average.

  4. 04 5g–tsn

    5G alone is not deterministic. TSN alone is not wireless.

    Bridging the two is the unglamorous middle of the next decade of industrial communication. The standards are maturing; the implementations are not. Most of what I build lives in that gap — translating between scheduling philosophies that were never designed for each other.

  5. 05 network coding

    Recover. Do not retransmit.

    When the round trip is the budget, retransmission is bankruptcy. Network coding offers a different move: reconstruct what was lost from what survived. Done carefully — with the right history length, the right verification — it buys reliability without the latency tax.

  6. 06 methodology

    Evaluate on what people actually deploy.

    It is tempting to test on bespoke hardware: results are clean, numbers are kind. They rarely transfer. Commodity NICs, mainline Linux, off-the-shelf 5G modems — that is where the gap between theory and practice lives, and where I prefer to look.