Radio MACs

How radio MAC models map onto Florete's control architecture. Companion to the Link Contract — this page covers its resource-contract requirement R9 and the RRA extension; R8's design lives in Time Synchronization. Terms: Glossary.

Two MAC families

• Contention (CSMA) — e.g. 802.11. Access is decided locally by carrier sensing; link establishment is fully local; no allocator exists. The ctrl-plane resource contract is thin, not empty: a logically centralized optimizer can still improve a contention medium — per-node transmit-power caps that shrink interference domains and mitigate hidden terminals, channel assignment, spatial reuse of links. This is the B2 reference target.
• Scheduled (TDMA/OFDMA/…) — resource blocks (slots, subcarriers, power) are allocated by an entity with a wider-than-local view. That allocator must live somewhere — and Florete's control model already has the right slots for it.

RRA vs. scheduler. RRA is the whole radio-resource function, across all timescales and both MAC families. A scheduler is the scheduled-MAC mechanism implementing RRA's fast-timescale slice (per-frame grants). CSMA has RRA without a scheduler.

RRA placement: the timescale hierarchy

Radio resource allocation splits across Florete's existing decision-authority layers:

Regulatory constraints being operator-signed bounds is a natural fit: an out-of-bounds allocation is rejected locally, and a hijacked control plane stays DoS-class.

Control-plane roadmap: C0 — the Coordinator only delivers mgmt decisions; C1 — adds the minimal ctrl stream; B2 — adds the first RRA extension: the minimal-but-operational CSMA RRA (working candidate: centrally computed per-node TX-power caps within regulatory bounds). Scheduled-MAC RRA (semi-persistent grants) lands only with a concrete vendor integration.

Node-side RRA placement

The Coordinator extension is only the computing end. On each node, RRA decisions travel a chain down to the node fast-path:

1. flor agent — receives and verifies RRA ctrl artifacts exactly like any other ctrl state; makes no radio decisions itself.
2. Link vertex (via its medium adapter) — applies semi-persistent decisions to the radio driver through the host interface (R10): sets power caps, channels, slot maps; owns queues and per-link policy. For WiFi this is a UDP socket + nl80211; for a custom radio, whatever control interface R10 documents.
3. Radio driver over the radio module (vendor) — executes the fast path: the fast-path scheduler selects slots locally among the options the slow path provisioned, plus timed transmissions, backoff, rate adaptation, link-level retransmission. Timing this hard lives in hardware/firmware (often an FPGA); the radio driver itself may be split kernel/userspace — link-layer fragmentation or ARQ can live in userspace as long as timing-critical parts stay below.

The fast-path MAC is always vendor-side. Florete's medium adapter is ordinary userspace software on commodity boards — it runs no per-slot loop and reconstructs no medium access. The timing-critical fast path (medium-access arbitration, and any sub-ms/HARQ retransmit) must live in the radio driver/module, below the host interface (R10); a PHY-only radio is not integrable until its vendor supplies that MAC. What Florete builds above R10 is the non-real-time upper MAC — per-neighbor addressing, discovery, coarse (ms-scale) ARQ, telemetry, and applying the slow-path options.

What remains radio-specific is the finer boundary above that line — how much provisioned-option state the medium adapter holds versus what the radio driver caches. The mesh layer follows the same pattern at its own level: path failover on liveness loss is a mesh-vertex decision, not an agent one.

No infrastructure asymmetry

A hard constraint on MAC choice (contract R9) — but a precise one: it is not a demand for a distributed scheduler. The split:

• The fast path is local. Per-slot selection runs in the radio driver/module, choosing among semi-persistently provisioned options. This mirrors label forwarding exactly: the mesh-vertex picks labels from tables it did not compute.
• The slow path is logically centralized — and that is fine. The options themselves are provisioned by the Coordinator's RRA, just as label allocation is PCE-computed rather than negotiated hop-by-hop. Florete deliberately took the PCE route over distributed label-distribution protocols, and the same choice applies to slot provisioning: distributed, hop-by-hop scheduling protocols (6TiSCH's MSF and kin) are an implementation shape, not something R9 demands — they too are just slow-path producers filling local tables.
• What must never be required: fixed infrastructure with fast-timescale authority. The Coordinator is logically centralized but relocatable and partition-tolerant — it splits, moves, and re-emerges in a disconnected group (see Coordinator evolution), and the emergent one plays the RRA role too; meanwhile nodes keep operating on their provisioned options. A MAC fails R9 only if it pins scheduling authority to fixed infrastructure at per-slot timescale, so that losing the controller stops the medium — the LTE eNB/UE asymmetry. Florete is a single-rank network: every node runs the same stack, and coordination roles are movable.

Scheduler ownership. The slow-path scheduler is the RRA extension — Florete-side by default, which is the natural split with a PHY-focused vendor (their radio driver runs the fast path). A vendor- or third-party slow-path scheduler is admissible as a rete app, but its decisions must flow through the Coordinator's ctrl chain — bounded and agent-verified like any ctrl state — never through a side channel to the radio, or the bounds model breaks.

Evolution

B2 is deliberately the degenerate radio setup: one computing host per node, directly attached omnidirectional non-beam-steering radios, one link vertex per radio. Two real-world axes it must not foreclose:

• Multi-host, multi-radio carriers (supernodes). A real vehicle/robot/drone carries several computing hosts and radio modules over an internal network. The outer mesh then spans the carrier's border hosts, the on-board network is itself a (recursive) Florete network, and RRA must eventually allocate across all radio modules of one carrier, not per host.
• Directional antennas: sectoral, beam-steering (phased-array/AESA), MIMO. A steered beam is naturally a per-neighbor link — the pairwise star model of R1 fits directional radios even better than omni. Multi-beam operation (one module forming several simultaneous beams) appears as parallel links from a single module. What changes is RRA — beams become a first-class resource dimension (a beam schedule alongside power/channels/slots) and topology computation gains geometry — and neighbor discovery, which needs beam sweeping/alignment (part of R6's open design). Structurally this converges with the supernode case: several spatial channels behind one carrier.

What B2 does about this now: keeps the ctrl-artifact resource vocabulary open-ended (a new resource dimension is an additive change), keeps telemetry per-link rather than per-radio, and keeps discovery an explicit design surface instead of assuming omni broadcast.

Case studies

Each studied against three fixed questions: where does the allocator live, what is its interface, what timescales does it run at?

Open questions

• Exact scope of B2's minimal CSMA RRA: power caps only, or also channel assignment / link spatial separation?
• Node fast-path split: link vertex vs radio driver — who owns the fast-path scheduler state for a scheduled MAC?
• Vendor/third-party slow-path schedulers as rete apps: their interface to the Coordinator, and how their decisions enter the ctrl chain.
• RRA ↔ TCE/PCE coupling: one joint optimization or staged?
• The ctrl-artifact schema for radio decisions (power caps now; grants, slot maps, beam schedules later): a vendor-agnostic radio section in the link-vertex's config — the medium adapter's counterpart to the mesh-vertex's forwarding table, each delivered to its own vertex, not one payload beside the other.
• Time: the two-clock-domain split is fixed in Time Synchronization, with the designs in Slot Clock and Network Clock; open questions live there.

Footnotes

1. IEEE 802.11s — the 802.11 amendment for link-layer mesh networking (path selection at L2). ↩

2. RFC 9030 — An Architecture for 6TiSCH; the distributed Minimal Scheduling Function is RFC 9033. ↩

3. O-RAN Alliance specifications; the (n)FAPI PHY↔MAC interface family is published by the Small Cell Forum. ↩

4. 3GPP TS 38.300 (NR overall description) covers sidelink resource-allocation Modes 1 and 2; the LTE ancestors are V2X Modes 3 and 4 (sidelink itself dates to Rel-12 ProSe). Practical availability (2026): sidelink ships almost exclusively in C-V2X automotive chipsets; commodity LTE/5G modems do not expose it — so sidelink is a future medium behind this same contract, while cellular today enters Florete only as an infrastructure WAN uplink. ↩

5. No open standard exists — tactical waveforms are proprietary, documented mostly in vendor whitepapers (TrellisWare TSM, Persistent Systems Wave Relay, Silvus). The most readable technical entry point is TrellisWare's published work on barrage relay networks (searchable term; Halford & Chugg's papers), plus the general academic MANET literature. ↩