MANET and satellite are not the same problem¶

Both are multi-hop wireless networks without fixed infrastructure. Almost every other property differs — and the differences decide which routing mechanisms make sense, what the benchmark must control for, and what may honestly be claimed.

This page exists because the satellite track (#192) kept producing findings whose real cause was "this regime does not work like the one the protocol was designed for". Rather than re-derive that each time, it is written down once. Every row of the difference table below changed a concrete decision in this repository — a defect, a parameter, a benchmark control, or an architectural call.

1. Where the two regimes sit¶

Infrastructure-free networks, ordered by how knowable the topology is. The two this repository cares about are at opposite ends.

flowchart LR
    subgraph axis [" topology unpredictable  ──────────▶  computable years ahead "]
    direction LR
    A["<b>MANET</b><br/>random waypoint<br/>shared radio"]
    F["FANET<br/>3D, smooth<br/>trajectories"]
    B["VANET<br/>road-constrained"]
    C["WSN / IoT<br/>static, energy-bound"]
    D["GEO<br/>one hop, ~119 ms"]
    E["<b>LEO ISL mesh</b><br/>+Grid, 4 links/node"]
    A ~~~ F ~~~ B ~~~ C ~~~ D ~~~ E
    end
    style A fill:#e2f0ed,stroke:#0f7f70,stroke-width:2px
    style E fill:#e8e6f8,stroke:#5b4fc4,stroke-width:2px

GEO sits near the deterministic end but is a different shape again: a bent-pipe hop to a gateway, with essentially no path to choose. That is why SNS3 — the most mature ns-3 satellite module — is the wrong substrate for routing work (#199).

The intermediate families change the evaluation, not the protocol: the same binary runs everywhere, but each family constrains mobility differently, so each needs its own mobility model, scenario shape, and — sometimes — metrics before a number from it means anything.

Family	Mobility	What it changes	Status in this repo
MANET	unconstrained random (RWP), 1–20 m/s, 2D	nothing — the regime the 2004/2007 sources designed and tuned for	supported: the paper and thesis fields (benchmarks)
FANET	3D smooth trajectories (Gauss-Markov standard), 10–30 m/s, sparse	third dimension, faster link churn, energy budgets that matter	supported (v1.7.0, #300): FANET results. `--scenario=fanet` is a sourced preset (#482): 30 UAVs in 1000 × 1000 × 300 m, 3-D Gauss-Markov at 10–30 m/s. It has its own analytic anchors, and 20-seed results for six arms in a main cell and a partition-bound sparse cell. AntHocNet leads every protocol in both cells, and in the main cell delivers within 2.70 pp of the oracle. No geographic arm runs in 3-D (GPSR is planar). Priority rationale: 2024–2026 FANET surveys evaluate AntHocNet directly and rate it strongest among the classical protocols they test, which makes FANET the family where the protocol's reputation is currently made
VANET	road-constrained (Manhattan, SUMO traces), 10–40 m/s, platooning	churn is fast but street-shaped; density swings block-by-block; RSU/infrastructure hybrids common	supported (v1.9.0, #301): VANET results. `--scenario=vanet` (#488): 100 vehicles on an 800 × 800 m Manhattan grid (Bai & Helmy 2003), 10–20 m/s, two-ray + building shadowing (Sommer et al. 2011), seven arms (GPSR included) and the per-pair oracle, 20 seeds, plus a partition-bound sparse cell. AntHocNet ties AODV on delivery (42.1 vs 42.3 %) with half its `delay99` and +13 NRL; every other arm trails both, and the oracle delivers 81.7 %. Losses are 36 % reconvergence: buildings break links at every corner (#537). The Veins/SUMO confirmation arm (#485) is deferred with the OMNeT++ adapter
Static Wi-Fi mesh	none (community / Freifunk-class networks)	nothing to track: proactive maintenance pays for a topology that never changes, and the Babel / BATMAN-adv baselines are home here	supported: static-mesh results (#484). AntHocNet beats AODV on every metric, and OLSR leads, once the #496 timer fix is in. The earlier `ReconvHoldCap` finding (#494) was that artifact
WSN / IoT	static or near-static, energy-bound	routing problem becomes energy/sleep scheduling, not topology discovery	out of scope (energy-aware `ILinkMetric` is the nearest hook, #145)
LEO ISL mesh	deterministic orbits	the §3 inversion: topology known, traffic unknown	supported as a static +Grid snapshot; dynamics are epic #297

The two supported regimes in detail:

	MANET — paper field	MANET — thesis field	Satellite — ISL +Grid
Harness	`anthocnet-compare` `--scenario=paper`	`anthocnet-compare` `--scenario=thesis`	`isl-grid`
Nodes / field	50 · 1500×300 m (Broch '98 calibration field)	100 · 2400×800 m (Ducatelle 2007 §5.1.3)	rows×cols torus (default 6×6), static snapshot
Mobility	RandomWaypoint, 1–20 m/s, pause 30 s	RandomWaypoint, 1–10 m/s, pause 30 s	none — topology fixed by construction
Medium	802.11b @ 2 Mbit/s, shared broadcast channel (disk or two-ray propagation)	same	point-to-point ISLs, 10 Mbit/s, 5 ms/link, one `/30` subnet each, degree 4
Topology	unknown — discovered by ants	unknown — discovered by ants	deterministic — degree/link count asserted every run
Loss	collisions, retry exhaustion, mobility	same	none on the link; any loss indicts the stack
Traffic	20 CBR flows × 512 bps	20 CBR flows × 2048 bps	4 CBR flows × 4096 bps + adversarial cells (scripted link cut, corridor congestion)
Baselines	AODV / OLSR / DSDV on identical seeds	same	same, plus (planned) precomputed shortest-path control (#216)
Results	benchmarks.md	benchmarks.md	benchmarks/satellite/isl-grid.md (static torus), benchmarks/satellite/leo-walker.md (moving Walker shell)

2. The two topologies¶

A MANET node's degree depends on who happens to be nearby. Neighbours are unknown until a hello arrives, and there is one interface on one shared broadcast domain.

graph LR
    n1((n1)) --- n2((n2))
    n2 --- n3((n3))
    n1 --- n3
    n2 --- n4((n4))
    n2 --- n5((n5))

n2 has degree 4 here, n4 degree 1 — and both change as the nodes move.

A satellite's degree is fixed by construction: two intra-plane links (fore and aft, near-constant length) and two cross-plane links (port and starboard, length varying with latitude).

graph TB
    fore["fore<br/><i>same plane</i>"] --- SAT(("SAT"))
    SAT --- aft["aft<br/><i>same plane</i>"]
    port["port<br/><i>adjacent plane</i>"] --- SAT
    SAT --- stbd["starboard<br/><i>adjacent plane</i>"]

Tiled, that is a +Grid torus — the standard LEO abstraction, and what ns3/anthocnet/examples/isl-grid.cc builds:

        ┌───────────────────────────┐   ← cross-plane wrap
        │                           │
    ────●───────●───────●───────●────┐  ← intra-plane wrap
        │       │       │       │    │
    ────●───────●───────●───────●────┤
        │       │       │       │    │
    ────●───────●───────●───────●────┘
        │                           │
        └───────────────────────────┘

    every node: exactly 4 ISLs, each on its own /30 subnet

Measured on the 4×4 harness: 16 nodes, 32 links — exactly 2 contributed per node, hence degree 4. That count is asserted on every run (#226), because a silently-wrong torus still delivers ~100% of packets and would produce entirely plausible numbers for the wrong network.

3. The inversion that explains everything else¶

	What is unknown	What is given
MANET	the topology — nodes move unpredictably, links appear and vanish, no node can know the graph	traffic demand is usually treated as given
LEO constellation	the traffic — where demand lands (ocean crossings, ground-station clustering, diurnal peaks) does not follow from orbits	the topology, for any second of the next decade, from the orbital elements

Move a routing protocol from one regime to the other and you take away the problem it was designed to solve, then hand it a different one.

That is why a claim about ant-colony routing on a constellation has to be about congestion and disruption, never about finding routes — the conclusion reached independently in satellite-routing-prior-art.md §3, where the published ACO-on-LEO literature turns out to have converged on load balancing for exactly this reason.

4. The differences that bite¶

Property	MANET	LEO satellite	Consequence in this repo
Topology	unknown, stochastic	deterministic from orbital elements	the control to beat is a precomputed shortest path (or OPSPF), not AODV — #216
Medium	shared broadcast radio, 802.11 DCF, contention	point-to-point ISL, one peer per link	no contention ⇒ expected values are analytic, not literature-derived — the satellite anchors in `benchmarks/methodology.md`
Interfaces per node	one	four, each on its own `/30`	exposed a real defect: the data path used interface 0 for every next hop — #203
Node degree	varies with density, 1 to many	exactly 4	degree becomes an assertable invariant — #226
Loss	collisions, retry exhaustion	essentially none on the link itself	anything under 100 % delivery indicts the stack, not the channel
Delay	queueing/contention dominated; `T_hop` = 3 ms	propagation dominated; 3–18 ms per ISL, GEO ~119 ms	delay becomes predictable: 2 hops × 5 ms → measured 10.39 ms. Also means the 802.11-calibrated timers are mis-sized — #205
Neighbour discovery	hello beacons are the only way to know	the peer is fixed and known from geometry	1 Hz hello on a known peer is overhead: NRL 12.18 with nothing to discover — #204
Link failure	random, mobility-driven, constant	mostly scheduled (polar seams, visibility windows)	only unscheduled failure is interesting; the scheduled kind is already in the control's tables
Scale	tens of nodes, diameter ~5	~1584 per shell, diameter 20–40	reactive flooding cost is the open scaling question — #207
Congestion signal	wifi MAC queue occupancy	no wifi MAC exists on an ISL	the A2 metric is inert on satellites until the signal is generalised — #206
Realistic baseline	AODV, OLSR, DSDV	snapshot routing, OPSPF, segment routing	terrestrial protocols become a sanity row, never the claim — §2.2 of the prior-art survey

5. What this means for the algorithm¶

AntHocNet's mechanisms split cleanly along this line:

Reactive discovery — flooding forward ants to find a path nobody knows — is its answer to unknown topology. A constellation does not have that problem: the +Grid is a known graph and every node can compute where a destination is.

Be precise about what that does and does not buy, though. Knowing the topology is not the same as a node already holding a next hop for a given destination: it still has to acquire direction, and on 1584 nodes with four links each, acquiring it by flooding is the expensive way. The correction is not "discovery is free" — it is that discovery here is steerable rather than blind, because the direction is derivable instead of unknown.

That is exactly the gap enableDirectedReactive (configuration.md) probes, and it does so without assuming a constellation: it steers along the diffused virtual gradient, which is pheromone the node already received, not a coordinate. So the same switch is testable on a MANET, where it degrades to "no gradient yet, flood as before" rather than to "wrong answer". - Multipath with delay-weighted pheromone is a different mechanism entirely: it responds to load, and load is precisely what orbits cannot predict.

So the defensible position is not "ACO works on satellites" but "one half of it addresses a problem that exists there, and the other half addresses one that does not". That is narrower, more honest, and testable.

It is also contested. Segment routing already performs congestion-aware traffic engineering on deployed hardware without per-packet stochastic decisions. What remains distinctive is that pheromone needs no central traffic view — which matters exactly when that view is stale or unreachable, i.e. under unpredicted failure and handover churn rather than under steady-state congestion. See satellite-routing-prior-art.md §5.1.

6. What runs where — mechanism × regime¶

Section 5 gives the argument; this table gives the inventory. One attribute set (defaults in configuration.md) serves both regimes — no per-regime build, no per-regime preset. What differs is which mechanisms are live: some bind to the Wi-Fi MAC and physically cannot fire on a point-to-point ISL, and some answer a question the regime doesn't ask. The harnesses set no protocol attributes; every A/B arm goes through explicit --ns3::anthocnet::RoutingProtocol::<Attr>=<v> overrides (#177), so a table row below describes the default run of that regime's harness.

Mechanism (attribute)	Default	MANET (Wi-Fi broadcast)	Satellite ISL (p2p grid)
Reactive forward-ant flood (`EnableReactive`)	on	live, essential — the only way to learn a topology nobody knows	live, but steerable-not-blind is the honest framing (§5): the flood re-derives direction the geometry already gives
Proactive ants (`EnableProactive`, `ProactiveInterval` 10 s) + diffusion (`EnableDiffusion`)	on	live — path maintenance and improvement during a session	live — diffusion carries the virtual gradient across the grid; it is what `EnableDirectedReactive` would steer along
Proactive emission gate (`ProactiveVirtualMargin`)	0 = off	off deliberately — the thesis's 10% gate measured harmful (#180)	same
Quiet mode (`EnableQuietMode`: proactive back-off on a stable route, #571)	on since v2.3.0	live but rarely engaged: routes change before the back-off grows (often byte-identical to off); on the static mesh it cuts NRL 8.6 % (section 8)	live — the torus's routes hold, so the back-off engages; NRL −0.04 to −0.11
Hello beacons (`HelloInterval` 1 Hz)	on	live, essential — sole neighbour-discovery mechanism	live but redundant: one fixed, known peer per link; measured cost NRL 12.18 on a churn-free grid (#204)
Multipath acceptance (`EnableMultipath`, a1 `AntAcceptanceFactor` 0.9, a2 `AntAcceptanceFactorNewHop` 2.0)	on	live — disjoint paths when density allows	live — the torus has equal-length corridors by construction; this is the half of the protocol with a real satellite claim (§5)
Local repair ants (`EnableRepair`, `RepairWaitFactor` 5, `RepairTimeout` 1 s)	on	live, constantly exercised — mobility breaks links	live, exercised only by unscheduled failure — the `--breakLink` cell (#260); scheduled failure belongs to the control's tables
Link-failure notifications (`EnableLinkFail`, cooldown `LinkfailNotifyInterval` 5 s)	on	live	live
Failure detector A — hello timeout	always on	live	live (and sufficient: `Ipv4::SetDown` on a cut link also raises the interface event)
Failure detector D — Wi-Fi MAC transmit failures (`EnableMacFailureDetector`, `TxFailureThreshold` 3)	on	live — the fast detector (ADR-0008)	inert — subscription requires a `WifiNetDevice`; a p2p ISL has none
A2 congestion metric (`EnableMacMetric`, `MacServiceAlpha` 0.7)	off	available — reads `WifiMacQueue` (AC_BE_NQOS) occupancy	inert even if enabled — no Wi-Fi MAC queue to read; needs a generalised congestion signal (#206); the corridor experiments (#216) exposed the deeper attribution gap (#292)
Directed reactive discovery (`EnableDirectedReactive`)	off	A/B arm — no gradient yet ⇒ degrades to the stock flood	A/B arm — the regime this switch exists to probe (§5, #245)
Pending-queue timing (`QueueTimeout` 3 s, `ReconvHoldCap` 200 ms, `RepairHoldCap` 0, `ReactiveRetryInterval` 0.25 s)	#21/#103 frontier values; `ReconvHoldCap` re-picked at 200 ms by #371	live — tuned on this regime's delay tail	live but calibrated against 802.11 contention delays, not 5–18 ms propagation floors (#205)
Goodness timing reference (`HopTime` 3 ms)	thesis value (#88)	live — matches an unloaded 802.11 hop	mis-sized — an ISL hop is 5 ms of pure propagation before queueing (#205)
Pheromone dynamics (`Alpha` 0.7, `Gamma` 0.7, `BetaAnts` 20, `BetaData` 20)	shipped values (the β exponents are the thesis's, adopted on a measured A/B — #179)	live	live — and on a static grid the closed-form orbit of these constants is exactly reproducible (#216 derivation)

Reading the table column-wise gives each regime's honest summary. MANET: everything live; the protocol is in its design regime. Satellite ISL: the discovery half runs but answers a solved problem, two Wi-Fi-coupled mechanisms (detector D, A2 metric) are inert, hello pays overhead for nothing, and the timing constants are calibrated for a medium that isn't there — while the multipath/load-response half faces exactly the problem the regime does have. That asymmetry is the satellite research programme (#192), not a defect list.

7. Measured: the v2.2.0 knobs per regime¶

Section 6 says which mechanisms are live in each regime. The v2.2.0 campaign measured four pending-queue and link knobs in every family: - 18 cells; - 20 seeds; - paired per seed against the shipped defaults.

None moved a default: each loses delivery somewhere (ADR-0019). This table is what each one does where, so a reader who runs one regime can still use it. Each entry shows AntHocNet's paired change in PDR, delay99 and NRL; "≡" marks byte-identical runs, where the knob never binds.

knob (attribute)	MANET grid	static mesh	FANET	VANET	satellite ISL grid	LEO shells
`RepairHoldCap=200ms` (#433)	two-ray: PDR −0.3 to −1.0 pp, d99 −3 to −13 %; Nakagami: PDR noise, d99 −22 to −33 %, NRL −4 to −9 %	≡	main: d99 −10 %; sparse: PDR −0.4 pp, d99 −22 %	main: PDR +1.1 pp, d99 −1.1 s, NRL −11 (recommended); sparse: PDR −0.2 pp, d99 −340 ms	≡	≡
`LinkMetric=sinr` (#181, ADR-0024)	two-ray: PDR +0.1 to +0.4 pp; Nakagami: d99 up to +98 ms, NRL +1 to +2.4	d99 +0.4 ms	main: PDR +0.4 pp, d99 +18 ms; sparse: d99 +535 ms	main: PDR −1.6 pp, d99 +1 s, NRL +13 (avoid); sparse: PDR +0.4 pp, d99 +1.5 s	PDR and delay identical (no PHY SINR on a p2p ISL: falls back to hop count)	starlink1: PDR +0.7 pp at NRL +49 (the access links do report SINR)
`HelloInterval=500ms` (#537)	PDR −0.5 to −8.0 pp, NRL +3.8 to +20 (avoid)	PDR −0.4 pp, NRL +3.3	PDR +0.5 to +0.6 pp, NRL +0.9 to +1.5	main: PDR +1.9 pp, d99 −15 % (recommended); sparse: PDR +0.5 pp, NRL +1.7	PDR ≡, NRL +1 to +18 (pure beacon cost)	PDR +2.6 to +6.5 pp, NRL +76 to +550 (load roughly doubles)
`ReconvHoldCap=1s` (#537)	two-ray: PDR +0.3 to +1.2 pp, d99 +6 to +36 %; Nakagami: PDR −1.4 to −6.0 pp, d99 +310 to +470 ms (avoid)	noise	main: PDR +1.3 pp; sparse: PDR +4.4 pp; d99 +180 to +590 ms	main: PDR +5.3 pp; sparse: PDR +6.2 pp; d99 +210 to +360 ms (recommended where delay is tolerable)	≡	walker16: noise; Starlink: PDR +0.5 to +0.8 pp, d99 +100 to +120 ms

Read the table by the regime's link-break signature:

Breaks that are clean and frequent (VANET corners, sparse FANET) reward every knob that waits or re-detects faster: the hold caps and the faster hello.
Breaks that are transient fades (Nakagami) punish waiting. Held packets queue behind ones the next attempt would have delivered, and faster beacons add contention.
Breaks that cannot happen to the routing (the ISL torus, the static mesh) leave the caps idle. The hello pays its beacons and buys nothing.

These are the same three signatures section 4 derives from the topology. The campaign's contribution is the measured size of each effect.

8. Measured: the v2.3.0 mechanisms per regime¶

The v2.3.0 campaign measured the six regime-gated mechanisms the same way: 18 cells, 20 seeds, paired per seed against the shipped defaults at the campaign commit. Each entry shows AntHocNet's paired change in PDR, delay99 and NRL where the 95 % interval excludes zero; "noise" means none does, and "≡" marks byte-identical runs.

mechanism (attribute)	MANET grid	static mesh	FANET	VANET	satellite ISL grid	LEO shells
link lifetime `EnableLinkLifetime` (#574)	two-ray: PDR +0.3 to +0.7 pp, d99 −28 to −37 ms; Nakagami: PDR +2.1 to +3.9 pp, d99 −159 to −215 ms, NRL −1.1 to −1.8 (recommended)	d99 +0.7 ms	main: PDR +0.5 pp, d99 −11 ms; sparse: noise	main: NRL +1.7, PDR noise; sparse: noise	≡ (nothing moves, so no link is predicted to end)	walker16, storm: noise; Starlink S1: PDR −0.5 pp, NRL +37
quiet mode `EnableQuietMode` (#571)	noise; ≡ on the three Nakagami cells	NRL −0.40 (−8.6 %), d99 −0.7 ms, PDR noise (default on from v2.3.0)	noise	noise	NRL −0.04 to −0.11, PDR and delay ≡	noise
adaptive evaporation `EnableAdaptiveEvaporation` (#572)	two-ray: noise; Nakagami: d99 +51 to +59 ms, NRL +1.3 to +2.0, PDR −1.1 pp on Gauss-Markov	NRL −0.14	main: PDR +0.4 pp; sparse: d99 +13 ms	main: d99 +180 ms, NRL +1.5; sparse: d99 +108 ms	noise (seam: NRL −0.01)	walker16: PDR −0.5 pp; Starlink: PDR −5.4 to −5.5 pp, NRL +164 to +236 (avoid)
energy metric `LinkMetric=energy` (#145)	two-ray: d99 +8 to +28 ms, NRL +0.2; Nakagami: d99 +126 to +139 ms, NRL +1.4 to +2.3	noise	main: PDR +0.5 pp, d99 +20 ms; sparse: d99 +564 ms	main: PDR −1.5 pp, d99 +1.0 s, NRL +13 (avoid); sparse: PDR +0.5 pp, d99 +1.6 s	noise (no energy probe: hop count)	walker16: NRL +2.0; Starlink S1: PDR +0.7 pp, NRL +49
propagation timing `--timingProfile=propagation` (#205)	—	—	—	—	noise	noise
p2p hello throttle `P2pHelloEvery=5` (#204)	—	—	—	—	base, failcell, corridor: NRL −0.5 to −10.5, PDR ≡; seam: PDR −0.12 pp, d99 +9 ms	—

Five of the six cannot move a default: each loses delivery somewhere, or (timing) shows no effect at all. Quiet mode is the one mechanism with no regression in any cell, so the pre-registered rule moves its default: it ships on from v2.3.0. Three readings carry over from section 7:

Lookahead pays where breaks are fades, not where they are corners. Link lifetime's largest gain is on the Nakagami grid, where it steers off links that are about to fade. Its target, the VANET corner break, is a topology change that kinematics alone does not predict.
The fixed-topology regimes reward doing less. Quiet mode and the hello throttle cut overhead only where the routes cannot change: the static mesh and the ISL torus. The seam, where links do fail, is where throttled hellos lose packets.
A non-hop-count metric at full batteries behaves like sinr did. The energy metric prices every hop at about 1.0 to 1.2. Its losses match v2.2.0's LinkMetric=sinr row cell for cell in sign and size: what costs is leaving hop-count goodness, not the energy signal.

Provenance of the numbers¶

Measurements are from this repository's own CI, not from the literature: isl-grid on a 4×4 torus gives 10.39 ms mean delay against a 10 ms analytic floor (2 hops × 5 ms) and 5.1 ms at one hop, with 32 links across 16 nodes and NRL 12.18 on a churn-free grid (#214, #237).

T_hop = 3 ms is the 2007 Ducatelle thesis value (#88). The protocol landscape and the ~119 ms GEO figure are sourced in satellite-routing-prior-art.md, which also records which of its claims have been checked against a full text and which have not.