Scenario: vanet

Class: VANET — vehicles on a street grid, corridor-shaped links, buildings that cut line of sight (#301, #488)

← Benchmark index · Metrics · Methodology

A separate family, not a variant of the MANET corpus. Vehicles move on the 1-D streets of a 2-D grid, and the radio is shadowed by the buildings between the streets. Its cells are not comparable with the MANET pages: read them against each other and against the VANET anchors, never against the Broch floor, which is RWP-only.

Re-measured for v2.2.0 (2026-10-16). The v2.2.0 campaign re-ran every cell on this page as its control generation, at a8a16991, with the shipped defaults and the same 20 seeds. Every ##RUN## row is byte-identical to the cells this page cites. No default moved, so the numbers here stand as v2.2.0's. The control cells are results/cells/vanet-*-v220.txt. What the campaign's four knobs do on this family is in network-regimes.md §7.

What it stresses

Vehicular ad hoc networks in a city. Three things change against the open MANET field: - mobility is street-shaped: vehicles follow streets and turn at intersections (the Manhattan model), so relative motion is either head-on/same-direction along one street or across an intersection; - links are corridor-shaped: buildings cut line of sight, so a vehicle hears its own street far further than the parallel one 200 m away; - links break at corners: a vehicle that turns loses line of sight to its former street within metres, so route churn is driven by turns, not only by distance.

The preset is built so that the main cell measures routing, and a second cell measures the partition regime deliberately.

Configuration

--scenario=vanet (added in #488). The preset carries scenario knobs only (ADR-0019): no AntHocNet attribute is changed for this family. Every value below can be overridden on the command line; an explicit flag always wins.

cell	vehicles	field (m)	grid	mobility	speed (m/s)	flows	per-flow rate	time (s)	propagation
vanet (main)	100	800 × 800	4 × 4 blocks of 200 m, 20 m streets	Manhattan	U(10, 20)	20	4 × 64 B/s (2048 bps)	900	urban
vanet + --nNodes=40 (sparse)	40	800 × 800	4 × 4 blocks of 200 m, 20 m streets	Manhattan	U(10, 20)	20	4 × 64 B/s (2048 bps)	900	urban

Flows start uniformly in [0, 180] s, as in the paper preset. The oracle control recomputes its graph every 100 ms under this preset, as under fanet (see the oracle).

The mobility model (--mobility=manhattan)

The Manhattan model of Bai, Sadagopan & Helmy (IMPORTANT, INFOCOM 2003)¹, the model BonnMotion's ManhattanGrid implements: - streets are the block edges, border streets included; - a vehicle starts at a uniform point of the street network (placement is length-weighted over all streets), heading either way along its street; - at each intersection it goes straight with probability 0.5 and turns left or right with 0.25 each, the probabilities of the Camp, Boleng & Davies survey²; a choice that would leave the field is redrawn among the valid ones, and a dead end turns the vehicle back; - each block-to-block leg is driven at a speed drawn U(speedMin, speed); - there are no pauses, so --pause is inert and the preset sets it to 0.

It is built in the harness, as WaypointMobilityModel waypoints from one pinned RNG stream, rather than imported as a BonnMotion trace (#488 scope 1 suggested the import). That needs no Java toolchain in CI and no committed trace files, and a run stays a function of its seed alone (#352): every arm on a seed sees the identical vehicle schedule.

Two simplifications against BonnMotion, recorded rather than hidden: BonnMotion perturbs speed at every update interval, while this model draws one speed per leg; and BonnMotion supports pauses, which this model omits.

The channel (--propagation=urban)

The same two-ray ground path loss as the tworay arm (identical frequency and antenna height, so {tworay, urban} isolates the buildings alone), plus the deterministic building shadowing of Sommer, Eckhoff, German & Dressler (WONS 2011)³, the obstacle model Veins ships:

L_obs = 9 dB × (walls cut) + 0.4 dB/m × (metres inside buildings)

at Veins' shipped defaults. The buildings are the grid's blocks, inset by half the street width from each street centre line.

Provenance

Each value is either sourced (a published setup uses it or brackets it) or chosen (a decision this repo made, with its reason). Nothing is left unlabelled: unsourced constants were the root cause of #88, #169 and #173.

knob	value	kind	where it comes from
mobility	Manhattan grid, turns 0.5 / 0.25 / 0.25	sourced	Bai, Sadagopan & Helmy (2003)¹ introduce the Manhattan model for urban movement; the turn probabilities are the Camp, Boleng & Davies survey's².
field	800 × 800 m	chosen, from the sourced setup	Bai & Helmy run their Manhattan patterns on 1000 × 1000 m¹. Under the urban channel that field needs ~145 vehicles to clear the 2·ln(n) connectivity line (#488 preflight), and per-seed cost grows faster than quadratically in nodes (scale sweep: 98.2 min/seed at 162 nodes). 800 × 800 m reaches the same line at 95–100 vehicles.
blocks	4 × 4 of 200 m	assumption	Neither source fixes a block length. 200 m is a city-block order of magnitude, and it keeps the block shorter than the radio's line-of-sight reach (the two-ray decode radius, 423.3 m), so a street is one or two hops long. Recorded as an assumption, not a citation.
street width	20 m	assumption	Sets the building footprint (180 × 180 m per block). Not from a source; recorded as an assumption.
vehicles	100 main, 40 sparse	chosen for density	The #488 preflight's road-network degree (a fixed-seed Monte Carlo over street pairs, applying the two-ray margin against the building loss) is 9.6 at 100 vehicles, clearing 2·ln(100) = 9.2, so the main cell measures routing. At 40 vehicles — Bai & Helmy's count¹ — it is 3.8, right at ln(40) = 3.7, so that cell partitions and is kept as the deliberate sparse cell, like FANET's 250 m cell. The estimate matches the harness: the oracle's measured graph on the main preset averages degree 9.1.
speed	U(10, 20) m/s	chosen, inside the sourced envelope	36–72 km/h, urban arterial speeds, inside the #298 urban envelope (10–40 m/s). Bai & Helmy sweep maxima of 1–60 m/s¹. The 10 m/s floor (--speedMin) keeps vehicles moving, since the model has no stops.
channel	two-ray + Sommer building shadowing	sourced	Sommer et al. (2011)³: 9 dB per wall, 0.4 dB/m inside, Veins' defaults. Threat to validity: those constants were fitted at 5.9 GHz (802.11p), and this harness keeps its 2.4 GHz 802.11b radio (ADR-0019: a family changes the scenario, not the stack). Lower frequency penetrates walls somewhat better, so the buildings here are, if anything, slightly too opaque.
traffic	20 CBR flows, 4 × 64-byte packets/s	sourced	Bai & Helmy (2003)¹: 20 CBR sources, 4 packets/s, 64 bytes. The per-flow rate equals the AntHocNet thesis's (Ducatelle 2007 §5.1.3). Offered load is 41 kbps, about 2 % of the 2 Mbit/s channel, so the cell measures routing, not congestion.
time	900 s	sourced	Bai & Helmy (2003)¹, and the paper/thesis horizon.

1. F. Bai, N. Sadagopan, A. Helmy, "IMPORTANT: a framework to systematically analyze the Impact of Mobility on Performance of RouTing protocols for Adhoc NeTworks", IEEE INFOCOM 2003. The values above are read from the authors' extended chapter, F. Bai and A. Helmy, "A Survey of Mobility Models in Wireless Adhoc Networks", ch. 2 (PDF), §3–§5: 40 nodes, 1000 × 1000 m, 900 s, 250 m range, 20 CBR sources at 4 packets/s of 64 bytes, maximum speeds 1–60 m/s.
2. T. Camp, J. Boleng, V. Davies, "A survey of mobility models for ad hoc network research", Wireless Communications and Mobile Computing 2(5), 483–502 (2002): the Manhattan model's 0.5 straight / 0.25 left / 0.25 right.
3. C. Sommer, D. Eckhoff, R. German, F. Dressler, "A computationally inexpensive empirical model of IEEE 802.11p radio shadowing in urban environments", WONS 2011; the defaults are those of Veins' obstacle configuration (db-per-cut 9, db-per-meter 0.4).

Preflight (#488): both cells pass with WARNs only. The family WARN says the cell is its own family; the sparse cell (--nodes 40, degree 3.8 against ln(40) = 3.7) additionally WARNs that partitions are likely, which is its purpose.

python3 tools/bench/scenario_check.py preflight \
  --nodes 100 --areaX 800 --areaY 800 --blocksX 4 --blocksY 4 \
  --mobility manhattan --propagation urban --speed 20 --pause 0 \
  --flows 20 --cbrBps 2048 --time 900 --pathWindowS 2

The street-grid degree replaces the strip/disk rule, which is wrong for a road network: nodes live on the 1-D edges of a 2-D grid, and a disk around a vehicle mostly covers buildings. Link lifetime at 20 m/s over the 423 m line-of-sight reach is about 10.6 hello periods, so HelloInterval is not a concern by the #481 rule; corner turns break links sooner than that bound, which the route-stability metrics measure directly.

Anchors

The Broch AODV floor is RWP-only and does not transfer. The VANET family has its own analytic anchors, gated in CI on the 3.42 leg (tools/checks/check-anchors.sh, floors in ns3/tools/anchors.yml). Both run the preset shrunk to a single 280 × 280 m block with 20 vehicles, and route only stock AODV and the oracle control.

anchor	field	expected (derived)	measured	what it checks
vanet-single-hop	one block, street width 279 m (a 1 m building)	the 396 m block diagonal is inside the 423.3 m two-ray decode radius and no line of sight is cut, so every flow is one hop: PDR ≈ 100 %, oracle hopsMean ≈ 1.00	AODV 99.5 %, oracle 99.9 %, hopsMean 1.00	the street grid, mobility and channel deliver at all
vanet-building	one block, street width 100 m (a 180 × 180 m building)	vehicles on opposite streets lose line of sight and must relay around a corner: oracle hopsMean > 1.2; a channel that ignored the building reads the single-hop 1.00	AODV 80.4 %, oracle 99.6 %, hopsMean 1.36	the buildings are honoured end to end
vanet-gpcr-vs-gpsr (#709)	the vanet-building block, gpsr,gpcr,oracle	GPCR's own paper trend (Lochert et al. 2005, §V): around a blocked corner the junction-aware arm beats GPSR by ≥ 2 pp on the same seeds; every arm ≥ 75 %	ns-3.42: GPCR 89.3 %, GPSR 84.4 % (+4.9 pp), oracle 99.6 %	the GPCR arm reproduces the trend its paper reports

The harness's own CI smoke (tools/checks/check-manhattan.sh) adds the invariants the anchors do not: every vehicle is on a street at t = 0 and mid-run, every arm delivers on the urban field, the oracle keeps 66 edges against 236 under plain two-ray on the same field, and a same-seed rerun is byte-identical.

The junction-aware comparator: GPCR (#709)

Under ADR-0026 this family is compared against VANET-native routing: - GPSR, position-based; - AODV, the VANET literature's standard reactive baseline; - GPCR, added here as the street grid's own answer to GPSR's failure at corners.

OLSR, DSDV and AOMDV are MANET migrants. Their rows below stay as the dated record of their campaigns, and they leave the plan from the v2.4.0 control campaign.

--protocols=gpcr is Greedy Perimeter Coordinator Routing (Lochert, Mauve, Füßler & Hartenstein, ACM MC2R 2005). It runs as a mode of the vendored GPSR module, with the same beacons, headers and GOD positions. Its rules:

• nodes on a junction ("coordinators") are recognised from the correlation of their neighbours' positions;
• along a street, packets are forwarded toward the next junction and never across it;
• at a junction, the next street is chosen greedily;
• when greedy forwarding fails, the right-hand rule is applied at junctions only.

Two constants are not given by the paper. Both are attributes, recorded as maintainer decisions:

• the junction threshold ε = 0.9 (the paper's digit is illegible);
• a ±30° street cone.

The fidelity sheet is in ns3/baselines/gpsr/README.md.

ε sensitivity, measured on the vanet-building block (ns-3.48, 2 seeds, before the #717 fix): GPCR delivers 90.1 / 90.0 / 90.0 % at ε = 0.8 / 0.9 / 0.95. On the full 100-vehicle grid (1 seed, 120 s) it delivers 24.2 / 23.7 % at ε = 0.8 / 0.95. The threshold moves delivery by under 1 pp on these fields.

GPSR rows measured before #717. The vendored GPSR drew its first hello before the harness pinned its random stream (#717), so each GPSR row depended on the runs before it in the same process. For example, seed 2 on the delivery-smoke field read 27.9 % alone and 18.1 % after seed 1. The GPSR rows on this page are one realisation of that order-dependent run. The v2.4.0 control campaign re-measures them, and GPCR's 20-seed cell is published with it.

The oracle on the street grid

The oracle derives its graph per pair under this channel (decode-los-approx, ns3/baselines/oracle/README.md): inside the two-ray decode disk, a pair is a link iff the deterministic two-ray + building chain reaches the decode floor. Both models are deterministic, so the evaluation draws no random numbers and every arm sees the same channel. It is flagged approx=1 for the same reason as tworay: decoding under load also depends on interference, which no per-pair rule carries.

Recompute cadence: 100 ms. At the 1 s default the oracle lagged the corner turns: on the vanet-building field with 10 vehicles it delivered 48.3 % against AODV's 50.5 %, so it was not a bound. At 100 ms it delivers 99.6 % there. The preset sets 100 ms as fanet does (#506); the oracle is the measuring instrument, not a protocol under test, so this is not the per-family protocol default ADR-0019 forbids.

Restated at v2.4.0: family-native baselines (ADR-0026)

Current numbers: v2.4.0. From v2.4.0 every family is compared against its own protocols plus its oracle (ADR-0026). These cells are the v2.4.0 control campaign at the shipped defaults (no default moved), 20 seeds, mean ± 95 % t-interval. AntHocNet's rows and every kept baseline row are byte-identical to v2.3.0's. The blocks below this one are the dated record of their campaigns; their migrant rows are retired, not deleted.

Native set: AntHocNet, AODV (the VANET literature's standard reactive baseline), GPSR, GPCR (#709), the oracle. Retired: OLSR, DSDV, AOMDV. GPSR's rows differ from v2.3.0's: the #717 fix moved its first hello after stream pinning.

main (100 vehicles)

protocol	PDR % ± 95 % CI	delay99 ms	NRL
AntHocNet	42.09 ± 0.66	1849.3	73.61
AODV	42.26 ± 0.60	3748.5	60.45
GPSR	25.82 ± 0.50	790.6	5.42
GPCR	28.73 ± 0.52	673.6	4.87
oracle	81.70 ± 0.45	609.0	0.00

AntHocNet − oracle -39.61 [-40.20, -39.01] pp; against the best native baseline (aodv, 42.26 %): dPDR -0.16 [-0.59, +0.26] pp, d delay99 -1899.15 [-1949.00, -1852.55] ms, d NRL +13.16 [+11.40, +14.91] (PAIRED-MIXED).

sparse (40 vehicles)

protocol	PDR % ± 95 % CI	delay99 ms	NRL
AntHocNet	23.93 ± 0.59	649.5	27.19
AODV	23.79 ± 0.65	8583.1	7.06
GPSR	16.30 ± 0.52	1685.5	3.44
GPCR	16.90 ± 0.53	1686.8	3.33
oracle	30.79 ± 0.52	26.3	0.00

AntHocNet − oracle -6.85 [-7.02, -6.68] pp; against the best native baseline (aodv, 23.79 %): dPDR +0.15 [-0.18, +0.47] pp, d delay99 -7933.60 [-8041.95, -7829.40] ms, d NRL +20.13 [+19.35, +20.90] (PAIRED-MIXED).

AntHocNet ties AODV on delivery in both cells, with a far shorter tail at more overhead. GPCR beats GPSR on the same seeds in both cells (+2.91 pp on main, +0.59 pp on sparse, paired-improved), as its own paper predicts, but both geographic arms trail AODV by 7–16 pp on this street grid.

Restated at v2.3.0 defaults (quiet mode on)

Dated record: v2.3.0 defaults (superseded by the v2.4.0 block above). v2.3.0 turns quiet mode on by default (#571, the v2.3.0 campaign). AntHocNet's rows below are the campaign's quiet arm. Every baseline row is the campaign's control, and the attribution control found those identical to the v2.2.0 corpus. 20 seeds, paired per seed; a bold difference is one whose 95 % t-interval excludes zero.

Quiet mode's effect here, paired against v2.2.0's AntHocNet: noise on PDR, delay99 and NRL in both cells: corner breaks reset the back-off.

Provenance. The campaign's runs at main, 54cdfae7..cc230126. Cells: results/cells/<cell>-v230-published.txt, composed by tools/bench/compose_cell.py from the campaign's <cell>-v230.txt (control) and <cell>-v230-quiet.txt. Tables recomputed from the cells' ##RUN## rows (t-intervals; delay99 by bootstrap).

main (100 vehicles)

protocol	PDR %	delay99 (ms, bootstrap)	NRL
anthocnet	42.10 ± 0.66	1849.3 [1811.4, 1892.2]	73.61 ± 1.94
aodv	42.26 ± 0.60	3748.5 [3706.7, 3793.4]	60.45 ± 1.27
aomdv	18.87 ± 0.48	9028.1 [8982.8, 9073.8]	73.60 ± 2.07
olsr	9.19 ± 0.38	183.8 [72.4, 314.6]	11.23 ± 0.52
dsdv	11.81 ± 0.38	833.6 [796.5, 876.9]	66.64 ± 2.48
gpsr	25.95 ± 0.50	764.0 [727.0, 803.3]	5.39 ± 0.13
oracle	81.70 ± 0.45	609.0 [541.5, 669.6]	0.00 ± 0.00

AntHocNet minus	ΔPDR (pp)	Δdelay99 (ms)	ΔNRL
aodv	-0.16 [-0.59, +0.26]	-1899.2 [-1951.8, -1846.5]	+13.16 [+11.40, +14.91]
aomdv	+23.22 [+22.77, +23.67]	-7178.8 [-7250.0, -7107.5]	+0.01 [-2.31, +2.33]
olsr	+32.90 [+32.47, +33.33]	+1665.6 [+1539.5, +1791.7]	+62.37 [+60.62, +64.13]
dsdv	+30.28 [+29.83, +30.73]	+1015.8 [+940.6, +1090.9]	+6.97 [+4.59, +9.34]
gpsr	+16.14 [+15.77, +16.51]	+1085.3 [+1023.2, +1147.5]	+68.21 [+66.33, +70.10]

sparse (40 vehicles)

protocol	PDR %	delay99 (ms, bootstrap)	NRL
anthocnet	23.93 ± 0.59	649.5 [608.6, 690.1]	27.19 ± 0.91
aodv	23.79 ± 0.65	8583.1 [8490.3, 8674.6]	7.06 ± 0.19
aomdv	18.09 ± 0.52	9301.6 [9257.5, 9346.6]	10.39 ± 0.30
olsr	8.68 ± 0.39	15.6 [14.6, 16.6]	4.19 ± 0.19
dsdv	10.90 ± 0.45	69.4 [41.8, 119.2]	8.71 ± 0.36
gpsr	16.29 ± 0.46	1712.2 [1586.1, 1846.3]	3.44 ± 0.10
oracle	30.79 ± 0.52	26.3 [25.7, 26.9]	0.00 ± 0.00

AntHocNet minus	ΔPDR (pp)	Δdelay99 (ms)	ΔNRL
aodv	+0.15 [-0.18, +0.47]	-7933.6 [-8050.3, -7816.9]	+20.13 [+19.35, +20.90]
aomdv	+5.85 [+5.46, +6.23]	-8652.0 [-8717.7, -8586.4]	+16.80 [+16.11, +17.49]
olsr	+15.25 [+14.86, +15.64]	+634.0 [+589.4, +678.6]	+23.00 [+22.23, +23.78]
dsdv	+13.04 [+12.73, +13.35]	+580.1 [+521.3, +639.0]	+18.49 [+17.82, +19.15]
gpsr	+7.64 [+7.36, +7.92]	-1062.7 [-1194.3, -931.1]	+23.75 [+22.92, +24.58]

The blocks below are the dated v2.2.0 record (quiet mode off). Their baseline rows are unchanged; where AntHocNet's numbers disagree with this block, this block is current. Diagnostic tables below (drop causes, paths, handovers, churn) describe the v2.2.0 runs.

Results

Provenance. paper-benchmark.yml at main @ 1a923576 (the #536 merge), image 3.42-opt, 900 s, seven arms (anthocnet, aodv, olsr, dsdv, aomdv, gpsr, oracle), seeds 1–20 per cell, --scenario=vanet with the inputs nNodes, areaX=800, areaY=800, pause=0, speed=20, propagation=urban, mobility=manhattan; seeds split with extraArgs=--firstRun=N.

cell	runs
main (100 vehicles, 4 seeds per job)	37638802518 (1–4), 37638813699 (5–8), 37638825144 (9–12), 37638836620 (13–16), 37638847225 (17–20)
sparse (40 vehicles, 10 seeds per job)	37638857489 (1–10), 37638868141 (11–20)

Cells: results/cells/vanet-{main,sparse}.txt, 140 ##RUN## rows each (20 seeds × 7 arms). A main-cell job took 3.5–4 h (about 50 min per seed, seven arms); a sparse-cell job 1.2–2 h.

scenario_check.py results: WARN only, 0 FAIL, on both cells. The WARNs are the oracle's approx=1 and partition notes, the #230 path-diversity caveat, and the matched-hop note below. One rule did FAIL on the first read and is now scoped, with its evidence:

The oracle matched-hop check is not a bound on a moving street grid. On the packets every arm delivered, the oracle took 2.1–3.0 hops against every arm's ~1.0–1.2 (sparse cell, 20/20 seeds). That is the #506 mechanism #519 scoped off 3-D fields: the "same" packet is forwarded at different instants in each arm (AODV's matched delay99 is 8.8 s), and the oracle re-routes per hop over a graph whose links flip at every corner. Control: the same cell nearly static (0.01–0.02 m/s, 3 seeds) gives oracle 2.400 against AntHocNet 2.400, AODV 2.405, OLSR 2.411 — no excess, so the per-pair building rule is not missing links. scenario_check.py now reports the excess once per file as a WARN when ##CONFIG## names mobility=manhattan above 1 m/s; a near-static grid keeps the FAIL as its control.

Reproduce: make reproduce FIG=family-vanet re-renders it from the committed data and diffs the PNG; MODE=rerun NS3DIR=… re-measures a cell first (#621).

Bars: mean of 20 seeds per arm, AntHocNet highlighted; whiskers: 95 % CI (t for PDR and NRL, bootstrap for delay99). The vertical line is the oracle control (not drawn for NRL, where it is 0 by construction). Drawn by ns3/tools/family-charts.py from the committed cells vanet-{main,sparse}-v230-published.txt: the v2.3.0 generation, whose tables are in the block above. The tables below are the dated v2.2.0 record.

Means ± 95 % t-CI half-width over 20 seeds; paired deltas are per seed (t-CI, Wilcoxon p). Energy is per delivered bit.

Main cell (vanet, 100 vehicles)

protocol	PDR %	mean delay (ms)	delay99 (ms)	NRL	energy (mJ/bit)
anthocnet	42.10 ± 0.66	166.5 ± 3.7	1848 ± 43	73.59 ± 1.94	5.52 ± 0.11
aodv	42.26 ± 0.60	318.9 ± 11.2	3749 ± 48	60.45 ± 1.27	5.43 ± 0.09
gpsr	25.95 ± 0.50	34.5 ± 1.4	764 ± 42	5.39 ± 0.13	8.80 ± 0.21
aomdv	18.87 ± 0.48	1060.4 ± 28.9	9028 ± 49	73.60 ± 2.07	11.97 ± 0.35
dsdv	11.81 ± 0.38	23.5 ± 1.3	834 ± 44	66.64 ± 2.48	19.22 ± 0.72
olsr	9.19 ± 0.38	15.4 ± 1.7	184 ± 134	11.23 ± 0.52	24.53 ± 1.14
oracle (100 ms)	81.70 ± 0.45	18.8 ± 1.1	609 ± 71	0.00	2.74 ± 0.03

AntHocNet minus	ΔPDR (pp)	ΔNRL	Δdelay99 (ms)	Δenergy (mJ/bit)
aodv	−0.16 [−0.58, +0.26], p = 0.34	+13.14 [+11.39, +14.89]	−1901 [−1953, −1849]	+0.092 [+0.033, +0.151]
gpsr	+16.15 [+15.77, +16.52]	+68.19 [+66.31, +70.08]	+1084 [+1023, +1145]	−3.277 [−3.401, −3.153]
aomdv	+23.23 [+22.77, +23.68]	−0.01 [−2.32, +2.30], p = 0.99	−7180 [−7250, −7110]	−6.444 [−6.712, −6.175]
dsdv	+30.28 [+29.83, +30.74]	+6.95 [+4.58, +9.32]	+1014 [+939, +1089]	−13.698 [−14.337, −13.059]
olsr	+32.91 [+32.47, +33.34]	+62.35 [+60.60, +64.11]	+1664 [+1538, +1790]	−19.003 [−20.051, −17.955]
oracle	−39.60 [−40.19, −39.01]	+73.59	+1239 [+1151, +1326]	+2.780 [+2.688, +2.871]

Every interval that excludes zero has Wilcoxon p ≤ 9.6 × 10⁻⁵.

Sparse cell (--nNodes=40, the partition regime)

protocol	PDR %	mean delay (ms)	delay99 (ms)	NRL	energy (mJ/bit)
anthocnet	23.90 ± 0.56	35.1 ± 1.2	632 ± 35	27.22 ± 0.93	3.79 ± 0.09
aodv	23.79 ± 0.65	906.0 ± 22.9	8583 ± 101	7.06 ± 0.19	3.78 ± 0.10
aomdv	18.09 ± 0.52	1162.5 ± 32.2	9302 ± 50	10.39 ± 0.30	4.95 ± 0.14
gpsr	16.29 ± 0.46	99.4 ± 6.8	1712 ± 143	3.44 ± 0.10	5.80 ± 0.16
dsdv	10.90 ± 0.45	13.4 ± 2.0	69 ± 49	8.71 ± 0.36	8.31 ± 0.34
olsr	8.68 ± 0.39	6.0 ± 1.3	16 ± 1	4.19 ± 0.19	10.38 ± 0.46
oracle (100 ms)	30.79 ± 0.52	6.1 ± 0.3	26 ± 1	0.00	2.90 ± 0.05

AntHocNet minus AODV: ΔPDR +0.11 [−0.20, +0.42] (p = 0.32), ΔNRL +20.16 [+19.36, +20.95], Δdelay99 −7951 ms [−8058, −7844]. AntHocNet leads AOMDV, GPSR, DSDV and OLSR by +5.81, +7.60, +13.01 and +15.21 pp (all 20/20 seeds). Even the oracle delivers only 30.8 %: 40 vehicles on this grid are partitioned most of the time, which is the cell's purpose.

What the numbers say

• AntHocNet ties AODV on delivery in both cells, and every other arm trails both. This is the first family where AntHocNet does not lead delivery outright (see the four-family statement below).
• The tail is AntHocNet's clear win over AODV: half of AODV's delay99 in the main cell and a thirteenth of it in the sparse cell. AODV buffers packets for seconds during route discovery; AntHocNet delivers or drops them sooner.
• Overhead is AntHocNet's cost. Its NRL is the highest of the arms that deliver comparably: +13.1 over AODV in the main cell, 3.9× AODV's in the sparse cell. It ties AOMDV's in the main cell. Energy per delivered bit is level with AODV (+0.09 mJ/bit) and far below every other arm.
• The gap to the oracle is the widest of any family: 39.6 pp. The oracle delivers 81.7 % on the same field, so the network is connected most of the time; the protocols lose the difference.

Where AntHocNet's losses go (drop decomposition, % of offered, mean over the jobs' compact blocks):

main cell	route (reconv / repair)	queue	MAC	channel	TTL
anthocnet	39.9 (36.2 / 3.9)	0.0	7.1	10.3	0.7
aodv	13.2	5.6	21.4	16.7	0.7
oracle	7.5 (partition)	0.0	6.1	4.7	0.0

AntHocNet loses 36 % of its traffic to reconvergence: packets that arrive while a route is being re-established and outlive the 200 ms ReconvHoldCap. The route-stability metrics (##ROUTE##, #294) show why:

main cell	median route setup (s)	pathChg	median path lifetime (s)	mean hops
anthocnet	1.17 ± 0.18	0.38	0.75	4.83
aodv	1.76 ± 0.33	0.23	1.32	3.56
oracle	0.27	0.42	0.50	3.72

Read the single-path arms for churn: AntHocNet's pathChg and path lifetime also count its deliberate multipath spreading (metrics.md). The oracle's shortest path changes every 0.50 s (median) and AODV's routes live 1.32 s: the topology turns over in about one hello period. Buildings cut line of sight the moment a vehicle turns a corner, so links die far sooner than distance alone predicts. The #488 preflight's link-lifetime rule (range / 2·speed = 10.6 hello periods) reasons from distance and overestimates link life on this channel: it passed a cell whose shortest paths change twice per hello period. AntHocNet's routes are also longer than AODV's (4.83 against 3.56 hops), so each crosses more of those breaks.

This is the knob-watchlist item #301 named (HelloInterval vs link lifetime, hold caps under corridor-shaped reconvergence). ADR-0019 forbids retuning per family before a measured A/B, so it is recorded as a follow-up, #537, not changed here: a paired HelloInterval / ReconvHoldCap A/B on this cell, and a corner-aware link-lifetime rule for the preflight. Both have since landed: the preflight rule in v2.2.0's phase 0, and the A/B in the v2.2.0 campaign (the next section).

Tuning for urban VANET (v2.2.0 campaign)

The v2.2.0 campaign ran #537's A/B on this page's two cells, together with the hold-cap and link-metric arms, on the same 20 seeds as the control. Each knob below helps on this regime. None became a default, because each costs delivery in another family (ADR-0019; network-regimes.md §7). If you run urban vehicular networks, these are measured settings to start from:

attribute	main cell (100 vehicles)	sparse cell (40)	costs elsewhere
ReconvHoldCap=1s	PDR +5.3 pp, delay99 +210 to +340 ms	PDR +6.2 pp, delay99 +360 ms, NRL −3 to −4	grid on Nakagami: PDR −1.4 to −6.0 pp
HelloInterval=500ms	PDR +1.9 pp, delay99 −15 %	PDR +0.5 pp, delay99 −100 ms, NRL +1.7	MANET grid: PDR −0.5 to −8 pp
RepairHoldCap=200ms	PDR +1.1 pp, delay99 −1.1 s, NRL −11	PDR −0.2 pp, delay99 −340 ms	grid two-ray, sparse FANET: PDR −0.3 to −1.0 pp
LinkMetric=sinr	avoid: PDR −1.6 pp, delay99 +1 s, NRL +13	PDR +0.4 pp, delay99 +1.5 s	—

The ns-3 attribute prefix is --ns3::anthocnet::RoutingProtocol::.

The knobs were measured one at a time, never combined. The pre-registered combined arm runs only when two defaults move, and none did. Their gains are not known to add.

None of them closes the reconvergence gap. At best a knob claws back 5–6 pp of the 39.6 pp to the oracle. The loss is set by link lifetime at corners (above), which no timer fixes. That is why v2.3.0's link-lifetime mechanism (#574) keeps its place.

Threats to validity

• Radio. 802.11b at 2.4 GHz with Sommer constants fitted at 5.9 GHz (Provenance). Buildings here are, if anything, slightly too opaque.
• Mobility. One speed per block leg and no stops (traffic lights, queues): real urban traffic clusters at intersections, which this model does not.
• The oracle. approx=1, and its matched-hop numbers are not a bound on a moving grid (above). Its delivery is a bound here: it leads every arm on every seed in both cells.
• No Veins confirmation. The Veins/SUMO arm (#485) is deferred with the OMNeT++ adapter; these numbers are ns-3 only.
• GPCR is a reimplementation. No reference code was used. Its paper leaves two constants open (ε, the street cone); they are set here and disclosed above. The neighbour-table junction test and the paper's proposed street-connectivity probing are not implemented.

Ranking stability across the families

ADR-0019 keeps AntHocNet's configuration identical across families, so a ranking that changes between them is a property of the network. From v2.4.0 the statement is read in two terms that need no migrated protocol (#711, ADR-0026): AntHocNet's gap to each cell's oracle, and AntHocNet against each cell's best native baseline (highest mean PDR), both paired per seed. The numbers are the v2.4.0 control campaign, 20 seeds per cell, and the disaster and SAGIN families' own v2.4.0 campaigns (disaster, SAGIN), also 20 seeds. Eight families in all.

Each dot is the mean over 20 paired seeds with its 95 % paired t-CI; filled = the CI excludes 0. Drawn by ns3/tools/family-charts.py from the committed v2.4.0 cells; the numbers come from tools/bench/native_view.py.

family	best native baseline	AntHocNet − best native, PDR	AntHocNet − oracle, PDR
MANET grid (6 cells)	OLSR (AODV on Gauss-Markov/two-ray)	+5.8 … +11.3 pp, every cell	+0.5 … +0.6 pp on two-ray; −4.2 … −8.0 pp on Nakagami
static mesh	OLSR	−0.65 pp	−0.60 pp
FANET	3-D GPSR (main), AODV (sparse)	+6.20 pp (main), +2.52 pp (sparse)	−2.41 pp (main), −7.70 pp (sparse)
VANET	AODV	tie (−0.16 and +0.15 pp, CIs straddle 0)	−39.6 pp (main), −6.85 pp (sparse)
ISL grid (4 cells)	none measured (OPSPF deferred)	—	0.00 pp, every cell
LEO shells (3)	OPSPF	−6.17 pp (walker16), −17.48 / −16.94 pp (Starlink S1, storm)	−5.62 pp (walker16), −15.9 pp (S1); the 1 s hop oracle is not a bound there (#720)
disaster	AODV	tie (+0.58 pp, CI straddles 0)	−3.77 pp; the oracle itself loses 18.5 % to partitions
SAGIN (3 shells + HAPS)	OPSPF	−5.95 pp (walker16), −17.11 / −17.74 pp (Starlink S1, storm)	−5.61 pp (walker16), −14.9 / −16.3 pp (S1); the 1 s oracles are not bounds there (#720)

What holds across families: - AntHocNet leads delivery on open mobile fields: every MANET cell and both FANET cells, against each family's strongest native protocol. - Its tail is shorter than the best native baseline's wherever links break abruptly: −1.3 to −1.6 s on the Nakagami grid cells, −1.9 and −7.9 s on VANET, −4.7 s on sparse FANET, −0.79 s on the disaster scene.

What does not hold: - On stable topologies, link state wins. OLSR leads the static mesh and OPSPF leads every LEO shell, with or without the SAGIN air layer. The OPSPF lead is large on Starlink S1 (−17.5 pp), bought with 20× AntHocNet's control. - On the street grid nothing native beats AODV, and neither does AntHocNet. It ties AODV, and the geographic comparators (GPSR, GPCR) trail both. VANET main's 39.6 pp gap to the oracle is still the largest in the corpus. - On the disaster scene it ties AODV too, 77.70 % against 77.12 % (+0.58 pp, CI [−0.22, +1.38]), with OLSR 11 pp behind both. AntHocNet is 3.8 pp from the oracle, which itself loses 18.5 % of the packets to the partition: no path exists when they are sent (#727). - The overhead ordering is family-dependent. AntHocNet sends less control than AODV on the MANET grid (−23.9 NRL on Gauss-Markov/two-ray) but more than OLSR in every family OLSR runs in, and more than AODV on VANET (+13.2 / +20.1) and on the disaster scene (+1.84).

Dated record: AntHocNet − AODV across four families (v2.3.0)

The statement below is the v2.3.0 generation, read against AODV in every family. AODV is no longer a baseline on the static mesh or the satellite cells (ADR-0026), so this view is retired. Its numbers stand as measured; the chart it read (families-vs-aodv.png) is replaced by the one above.

	MANET grid (6 cells)	static mesh	FANET main	VANET main
delivery leader	AntHocNet, every cell	OLSR (+0.49 pp over AntHocNet), DSDV +0.32	AntHocNet (+8.32 pp over AODV)	AntHocNet = AODV (−0.16 pp, p = 0.34)
AntHocNet vs AODV, PDR	+10.8 … +21.0 pp	+2.05 pp	+8.32 pp	tie
AntHocNet vs AODV, NRL	−23.9 … −28.0	−13.0	−0.36	+13.1
AntHocNet vs AODV, delay99	−164 … −372 ms	−205 ms	−293 ms	−1901 ms
gap to the oracle (PDR)	AntHocNet above it on two-ray; 4.2–8.0 pp below on Nakagami	0.6 pp	2.45 pp	39.6 pp
OLSR	78–89 %	leads	47.9 %	9.2 % (last)

What holds in every family: - AntHocNet beats AODV on the tail. delay99 is lower in every cell of every family, from −164 ms (MANET) to −1901 ms (VANET). - AntHocNet delivers at least as much as AODV everywhere, and strictly more everywhere except VANET.

What does not hold: - "AntHocNet delivers most" is a claim about open mobile fields. It holds on every MANET cell and both FANET cells. It fails on the static mesh (OLSR and DSDV lead) and becomes a tie with AODV on the street grid. - AntHocNet's overhead advantage over AODV vanishes and reverses as links shorten. −24 to −28 NRL on MANET, −13 on the static mesh, −0.4 on FANET, +13.1 on VANET: its proactive ants and repair keep sampling links that live under one hello period. - OLSR is the family-sensitive protocol. It leads the static mesh, delivers 78–89 % on MANET, 47.9 % on FANET and 9.2 % on VANET: its topology tables hold next hops that a corner or a fast flight has already broken. - The distance to the oracle grows with how abruptly links break: none on two-ray MANET, a few points on Nakagami and FANET, 39.6 pp when buildings cut links at every corner. VANET is where a better protocol has the most room, and where AntHocNet's reconvergence (36 % of its losses) is the measured place to look.

Provenance of the numbers

Per-seed rows: results/cells/vanet-{main,sparse}.txt. Statistics from stats_util (t-CI, Wilcoxon signed-rank). Drop decomposition from the # drops compact blocks, ##ROUTE## from the per-seed rows. Reproduce with paper-benchmark.yml at 1a923576 and the inputs in the provenance block above.