Coordinator DFS (Dynamic Frequency Selection) — Result
Status: Stage 2 design experiment, quantifying issue #65 (coordinator DFS). Date: 2026-06-22.
Reproduce: cd sim && PYTHONPATH=. python3 -m experiments.dfs_channel_selection
(writes docs/figures/dfs_channel_selection.png).
Occupancy data: sim/data/ng_channel_occupancy.json, derived from the live North-Georgia
Meshtastic firehose /exchange/mtnme-mqtt/packets_log.csv (see Occupancy source, below).
What this answers
Loomwave coordinators currently use a fixed operating channel. Issue #65 reframes this as
coordinator DFS: scan the candidate band, pick the least-occupied channel, announce it in the
beacon (frequency-agility-multisector-note.md §3). We hit the cost of not doing this live —
two Loomwave coordinators both on 911.5 collided (co-channel contention). This experiment
quantifies, on the existing PHY/collision engine, the DFS benefit in two cases:
- Single cell — a fixed coordinator that lands on an occupied channel (the survey-flagged 915.0 / loud-Meshtastic case) vs a DFS coordinator that measures and picks the quietest candidate.
- Two co-located coordinators — fixed (both may pick the same channel → the 911.5 collision) vs DFS (they spread to different clear channels → frequency separation).
Occupancy source — real mtnme.sh data (not synthetic)
The per-channel busy fractions are anchored to real data, the North-Georgia Meshtastic firehose
at /exchange/mtnme-mqtt/packets_log.csv (753,979 rows, 9.65-day window 2026-06-12 → 06-22,
712k GA-region packets). Processing (sim/data/ng_channel_occupancy.json):
- The CSV
channelfield is a Meshtastic modem preset (MediumFast / LongFast / ShortFast …), which sets both the SF/BW and the default US915 frequency slot (channel_num = xorHash(name) % numChannels, freq = 902 + bw/2 + ch·bw). Computed slots: MediumFast → 909.375 MHz, LongFast → 904.625 MHz. - Airtime per packet from the SX126x time-on-air formula (the project's
phy.time_on_air_ms), using the realpayload_len. Packets de-duplicated by(mesh_packet_id, channel)so each over-air transmission is counted once (the firehose is a multi-gateway aggregate). - Busy fraction = Σ ToA / window, bucketed per hour.
Measured per-channel occupancy (GA, busy-hour p90 of the deduped floor):
| Meshtastic channel | SF/BW | default freq | busy-frac median | mean | p90 | max |
|---|---|---|---|---|---|---|
| MediumFast (loud slot) | SF9/250 | 909.375 | 1.14% | 1.24% | 1.67% | 2.12% |
| LongFast (quiet slot) | SF11/250 | 904.625 | 0.34% | 0.36% | 0.49% | 0.74% |
| ShortFast | SF7/250 | 904.625 | ~0% | ~0% | ~0% | ~0% |
Key real ratio used: the MediumFast slot runs ≈3.4× the airtime of the LongFast slot — i.e.
the spectrum is genuinely not uniform, the premise the band survey asserted
(frequency-agility-multisector-note.md §2: 915.0 OCCUPIED, 911.5/902.5/921 CLEAR).
Caveat (and how we handle it): the firehose is gateway-deduped, so it is a floor of the
true on-channel occupancy a node senses — a single node hears more local transmissions than the
gateway-merged trace, and the NG validation gate measured real per-node channel_utilization at
median 5.7%, p90 14.8%, max 37.3% vs this deduped ~1.2%. We therefore do not treat the
deduped floor as the absolute busy level. We take the relative structure (loud vs quiet, the
3.4× ratio) as real, and sweep an interference scale that lifts the loud channel from the floor
(1.7%) up to ×20 (33%), which brackets the gate's measured busy-hour reality (≈37% max). All
results are reported across that range.
So: real data sets the channel structure and ratio; the absolute interference level is swept across the real-data-bracketed range rather than asserted from the deduped floor.
Model
On the existing PHY/collision engine (phy.py / network.py), unchanged:
- Cell: an INFRA coordinator (TDMA beacon authority,
InfraTdmaMac) + K clients (ClientTdmaMac, one slot each), MediumFast (SF9/BW250), 32 B uplinks, 30 s per-client cadence, K ∈ {10, 20}. Clients are mutually hidden (cell geometry); the hub hears all. - Ambient occupancy is a Poisson interferer co-located with the cell on its channel
(
AmbientInterferer), sized so its airtime/run = the channel's busy fraction. It is audible to the cell, so it genuinely collides with member uplinks via the same capture-aware rule (6 dB). - Candidate set: 915.0 (OCCUPIED, the loud channel), 911.5 / 902.5 / 921.0 (CLEAR), with busy fractions from the table above scaled by the swept interference level.
- Fixed picks its configured default (modelled as the occupied named slot, 915.0 — the realistic worst case of an operator hard-coding the popular channel). DFS picks the least-occupied candidate (921.0 here).
- 3 seeds × 1200 s.
Result 1 — single cell: fixed vs DFS
| interf level | loud busy% | K | fixed ch | fixed deliv | DFS ch | DFS busy% | DFS deliv | deliv gain |
|---|---|---|---|---|---|---|---|---|
| floor | 1.7 | 10 | 915.0 | 0.966 | 921.0 | 0.24 | 0.997 | +0.030 |
| floor | 1.7 | 20 | 915.0 | 0.966 | 921.0 | 0.24 | 0.995 | +0.029 |
| ×5 | 8.3 | 10 | 915.0 | 0.882 | 921.0 | 0.24 | 0.997 | +0.115 |
| ×5 | 8.3 | 20 | 915.0 | 0.874 | 921.0 | 0.24 | 0.995 | +0.121 |
| ×10 | 16.7 | 10 | 915.0 | 0.750 | 921.0 | 0.24 | 0.997 | +0.247 |
| ×10 | 16.7 | 20 | 915.0 | 0.767 | 921.0 | 0.24 | 0.995 | +0.228 |
| ×20 | 33.4 | 10 | 915.0 | 0.542 | 921.0 | 0.24 | 0.997 | +0.455 |
| ×20 | 33.4 | 20 | 915.0 | 0.594 | 921.0 | 0.24 | 0.995 | +0.401 |
Reading it:
- At the deduped firehose floor (1.7% busy), the benefit is small (~+0.03 delivery) — i.e.
on a genuinely-clear band, a fixed coordinator is fine, exactly the note's prediction that
single-radio DFS is marginal when the default is already CLEAR
(
frequency-agility-multisector-note.md§3 open question). - As the occupied channel approaches busy-hour reality (the gate's measured 15–37%), the fixed coordinator's delivery falls to 0.77 (×10) and 0.54–0.59 (×20), while DFS holds ≈0.99 by sitting on a clear channel. The DFS win grows to +0.23 (×10) and +0.40–0.46 (×20).
- Headline: on a quiet band DFS earns little; on a contested band (the 915-style ISM pile-up the survey found) DFS recovers 23–45 points of delivery. The benefit is conditional on there actually being a busy channel to avoid — which the real data confirms exists (3.4× ratio).
Result 2 — two co-located coordinators (the live 911.5 collision)
Two coordinators within RF range. Fixed: both configured to the same channel → their TDMA grids are referenced to different beacons, uncoordinated, so cell B's uplinks land in cell A's slots and the beacons collide (one shared medium). DFS: the two cells spread to different clear channels → frequency separation means they share no medium (modelled as each cell alone on its own quiet channel — the physically-correct consequence of different frequencies).
| interf | K | fixed deliv | fixed p50 (ms) | fixed p90 (ms) | DFS deliv | DFS p50 (ms) | DFS p90 (ms) | deliv gain |
|---|---|---|---|---|---|---|---|---|
| floor | 10 | 0.437 | 1580 | 2492 | 0.990 | 1597 | 2417 | +0.553 |
| floor | 20 | 0.464 | 3163 | 4970 | 0.988 | 3143 | 4853 | +0.524 |
| ×10 | 10 | 0.357 | 1646 | 2653 | 0.990 | 1597 | 2417 | +0.634 |
| ×10 | 20 | 0.367 | 3318 | 5680 | 0.988 | 3143 | 4853 | +0.621 |
Reading it:
- Two coordinators on the same channel deliver only 0.36–0.46 — roughly half their traffic is lost to co-channel contention, with the loss worsening as the shared channel also carries ambient load (0.46 → 0.37 from floor → ×10). This is the live failure quantified.
- DFS frequency separation restores delivery to ≈0.99 — a +0.52 to +0.63 delivery gain, i.e. the co-channel case loses >2× the traffic the separated case does.
- Latency p50/p90 are similar between the two (latency here is dominated by the beacon/slot frame period, ~1.6 s at K=10 / ~3.2 s at K=20); the separation win is overwhelmingly a delivery win, not a latency one — fixed loses packets, it doesn't merely delay them.
Headline numbers
- Single cell, DFS vs fixed: +0.03 delivery on a clear band (firehose floor), rising to +0.23 at ~17% occupancy and +0.40–0.46 at ~33% occupancy (fixed falls 0.97 → 0.54 as the occupied channel loads; DFS holds ≈0.99).
- Two co-located coordinators: fixed co-channel delivers 0.36–0.46; DFS separation delivers ≈0.99 — a +0.52 to +0.63 delivery gain (the 911.5 collision is the headline case for DFS).
- Real data: the channel structure (one loud slot ≈3.4× busier than the quiet slots) is
derived from real mtnme.sh firehose data; the absolute interference level is swept across
a range bracketed by the NG gate's measured busy-hour
channel_utilization(15–37%), because the firehose is gateway-deduped (a floor).
Assumptions and limits (explicit)
- Deduped firehose = floor. Absolute occupancy is swept, not asserted; the loud-channel level is the variable, anchored to the real relative ratio and the gate's real busy-hour band.
- Fixed = occupied named slot. We model the fixed coordinator on the channel the survey flagged OCCUPIED (915.0), i.e. the realistic worst case of hard-coding the popular default. A fixed coordinator that happens to be configured on a clear channel matches the DFS curve — which is exactly the point: DFS removes the dependence on a lucky/expert static configuration.
- DFS = perfect, static pick. We model a clean bandscan that picks the truly-quietest candidate
once at bring-up. We do not model re-selection cost (stranding registered clients), scan
error, or hidden occupancy the coordinator can't sense — those are the conservative caveats in
frequency-agility-multisector-note.md§3 and bound the upper end of the benefit. - Two-cell DFS = full separation. Different frequencies are modelled as separate media (no cross-cell interference). In-enclosure desense at small frequency spacing (§4 of the note) is a separate bench-RF concern, not modelled here.
- Propagation is the standard assumed log-distance + 6 dB shadowing model (uncalibrated absolute range), per the project's standing discipline; these are relative A/B comparisons on identical geometry, which the NG gate licenses.
Bottom line
DFS's value is conditional and the real data tells us the condition is met: the North-Georgia spectrum is non-uniform (3.4× loud/quiet ratio), so a fixed coordinator that lands on the loud slot loses 23–45 points of delivery that DFS recovers — and two coordinators sharing a channel (the live 911.5 failure) lose >half their traffic, which DFS frequency-separation restores to ≈0.99 (+0.52–0.63 delivery). This supports the note's recommendation: the single-cell win is marginal on an already-clear default, so the strong case for DFS is multi-coordinator co-channel avoidance and contested-band sites — exactly the cases the live failure and band survey identified.
Files: experiment sim/experiments/dfs_channel_selection.py; occupancy data
sim/data/ng_channel_occupancy.json; figure docs/figures/dfs_channel_selection.png.