Satellite suite: the moving constellation (leo-walker)

Regime: a time-varying LEO Walker shell — satellites move, ground stations hand over, ISLs fail and recover. The static ISL torus that came first is isl-grid.md; why satellites need their own suite at all is network-regimes.md. The substrate decision is ADR-0022.

← Benchmark index · Metrics · Methodology

Provenance — measured at 0b5a65ad. Seventeen satellite-benchmark.yml dispatches with harness=leo-walker, one arm per dispatch, on ghcr.io/danieljoppi/ns3:3.48 (default profile), 20 seeds each. Every cell carries its run id on its first line and its full ##CONFIG## line, and every cell passes scenario_check.py (the #297 handover-clock and outage-accounting rules) as committed.

cell	anthocnet	aodv	olsr	geo-greedy	oracle	oracle-delay
walker16	37827387100	37827390737	37827395914	37827410383	37827401516	37827405771
starlink1	37827426356	37827430778		37827443695	37827433965	37827439616
starlink1-storm	37827449496	37827453949		37827467035	37827458721	37827462965
hypatia calibration						37827470132

Cells: docs/benchmarks/cells/leo-<cell>-<arm>.txt. Every leo-walker number measured before 0b5a65ad is superseded — see the pairing bug.

The cells

All cells: 4 ground-to-ground CBR flows (2 kbit/s each) between six cities (New York, London, São Paulo, Johannesburg, Tokyo, Sydney), 300 s, minimum elevation 25°, link delays re-read from ECEF distance every 0.1 s, ISL failure overlay at 1 % (mean outage 2 s), 10 Mbps links. The handover clock fires at second 12 of every 15 s — the Starlink reconfiguration interval measured by Tanveer et al. (WWW 2024), and the phase the LENS trace below confirms.

cell	shell	what it adds
walker16	16 planes × 16 sats, 1150 km, 53°, phasing 1	a small Walker-delta shell: 256 satellites, long hops
starlink1	72 × 22, 550 km, 53°, phasing 39 — Starlink shell 1	the real shell size: 1584 satellites, ~25-hop paths
starlink1-storm	as starlink1	a 10 % mass satellite failure at t = 150 s

Arms:

• anthocnet — this repository, defaults.
• aodv — stock ns-3 AODV, the reactive baseline every other suite uses.
• olsr — stock ns-3 OLSR, the proactive link-state baseline. walker16 only (see below).
• geo-greedy — an idealised greedy geographic comparator: each hop goes to the neighbour nearest the destination's serving satellite, with perfect global position knowledge and no control traffic. It stands for the position-based family the satellite literature proposes, at its best case.
• oracle — the precomputed shortest-hop control (#196), re-solved every 1 s (the oracle module's default RecomputeInterval). A delivery bound, not a delay bound: one hop count covers links of very different lengths.
• oracle-delay — the same oracle with Metric=delay: Dijkstra on current channel delays, re-solved every 0.1 s on walker16 (oracleInterval=0 follows --delayUpdate) and every 1 s on S1 to keep CI time bounded. This is the latency bound, and the Hypatia-style control.

OLSR runs on walker16 only. On S1's 1584 nodes its topology flooding is the practical limit of the harness, not of the substrate (ADR-0022, criterion 5): even the 256-node walker16 cell took 5 h 54 min of CI for 20 seeds.

Headline

• AntHocNet beats AODV on every cell, by more as the shell grows: +15.5 pp PDR on walker16, +28.2 pp on Starlink S1, +32.3 pp under the S1 storm — all 20 of 20 seeds, Wilcoxon p = 1.9e-06 — with lower mean delay on every cell.
• The handover is where the difference lives. AntHocNet loses 9.4–11.2 packets per scheduled handover; AODV 34.5–57.7. The oracles, which re-solve on the new topology instantly, lose ~0.
• The gap to the bound grows with path length. AntHocNet is 7.0 pp under the delay oracle on walker16 (~9 hops) and 15.5 pp under it on S1 (~25 hops). That is the residual a 25-hop ant has to cover between two 15 s reconfigurations.
• Greedy geographic forwarding is not a viable comparator on a Walker grid, even idealised — see the geo-greedy verdict.

Tables below are the output of python3 ns3/tools/leo-summary.py walker16 starlink1 starlink1-storm, pasted unedited; rerun it on the committed cells to reproduce every number.

walker16

arm	seeds	PDR %	mean delay (ms)	delay99 (ms, boot)	NRL	refused at source %
anthocnet	20	92.84 ± 0.40	67.33 ± 1.18	140.5 [131.4, 150.9]	97.63 ± 1.69	0.00
aodv	20	77.34 ± 2.58	71.97 ± 0.95	125.4 [123.0, 128.6]	92.84 ± 3.24	0.00
olsr	20	83.88 ± 0.53	67.37 ± 0.47	124.7 [123.5, 125.8]	306.59 ± 1.87	2.91
geo-greedy	20	41.35 ± 0.19	76.91 ± 2.07	417.3 [396.9, 432.2]	0.00 ± 0.00	0.00
oracle (bound)	20	98.35 ± 0.12	69.85 ± 0.09	123.2 [123.0, 123.5]	0.00 ± 0.00	0.00
oracle-delay (bound)	20	99.79 ± 0.03	61.56 ± 0.05	103.2 [103.0, 103.6]	0.00 ± 0.00	0.00

AntHocNet − arm, paired per seed (95 % t-CI, Wilcoxon p):

vs	n	Δ PDR (pp)	Δ mean delay (ms)
aodv	20	+15.50 [+12.91, +18.09], p=1.9e-06	-4.64 [-6.20, -3.07], p=1.3e-05
olsr	20	+8.96 [+8.36, +9.56], p=9.6e-05	-0.04 [-1.34, +1.26], p=0.94
geo-greedy	20	+51.50 [+51.06, +51.94], p=9.6e-05	-9.57 [-12.09, -7.05], p=3.8e-06

Handover metric family (sums over seeds; outage = ≥ 3 consecutive lost send attempts):

arm	startup n / lost	scheduled n / lost	unplanned n / lost	massfail n / lost	other n / lost	lost per scheduled handover	median outage, scheduled (s)	median outage, unplanned (s)
anthocnet	1 / 3	257 / 2062	649 / 4329	0 / 0	10 / 74	9.37	2.00	1.75
aodv	2 / 17	232 / 12695	509 / 7256	0 / 0	106 / 1223	57.70	3.25	2.25
olsr	80 / 2789	259 / 6009	551 / 5685	0 / 0	55 / 543	27.31	5.25	2.12
geo-greedy	80 / 41274	41 / 13236	75 / 998	0 / 0	0 / 0	60.16	92.00	2.12
oracle	0 / 0	1 / 4	252 / 830	0 / 0	4 / 15	0.02	1.00	0.75
oracle-delay	0 / 0	0 / 0	0 / 0	0 / 0	0 / 0	0.00	—	—

arm	hop changes / flow / min	mean hops (delivered)
anthocnet	2.07	9.29
aodv	0.76	9.05
olsr	0.83	8.65
geo-greedy	0.84	6.33
oracle	1.02	8.72
oracle-delay	1.08	8.76

AODV's and OLSR's lower delay99 on walker16 (125 vs 140 ms) is survivorship: they deliver 15.5 and 9.0 pp fewer packets, and the ones they drop are the ones a detour would have made late.

OLSR, the proactive baseline, sits between AODV and AntHocNet. AntHocNet delivers +8.96 pp more (20 of 20 seeds, p = 9.6e-05) at the same mean delay (−0.04 ms, p = 0.94), with a third of OLSR's routing overhead (NRL 97.6 vs 306.6). OLSR loses 27.3 packets per scheduled handover — half of AODV's 57.7, three times AntHocNet's 9.4 — consistent with its topology view lagging each 15 s reconfiguration by its TC interval. It is also the only arm that refuses packets at the source (2.91 %: no route yet), and every one of its 80 flows (4 × 20 seeds) opens with a startup outage while the first topology floods converge.

starlink1

arm	seeds	PDR %	mean delay (ms)	delay99 (ms, boot)	NRL	refused at source %
anthocnet	20	79.43 ± 0.41	84.61 ± 1.18	210.2 [205.1, 215.7]	955.45 ± 17.11	0.00
aodv	20	51.23 ± 1.18	108.48 ± 5.40	1644.1 [1461.1, 1785.0]	904.16 ± 24.48	0.00
geo-greedy	20	3.81 ± 0.01	25.29 ± 0.45	27.9 [27.0, 28.9]	0.00 ± 0.00	0.00
oracle (bound)	20	95.27 ± 0.34	78.66 ± 0.11	132.1 [132.0, 132.2]	0.00 ± 0.00	0.00
oracle-delay (bound)	20	94.92 ± 0.22	74.89 ± 0.07	117.5 [116.7, 118.5]	0.00 ± 0.00	0.00

AntHocNet − arm, paired per seed (95 % t-CI, Wilcoxon p):

vs	n	Δ PDR (pp)	Δ mean delay (ms)
aodv	20	+28.20 [+26.90, +29.50], p=1.9e-06	-23.86 [-28.97, -18.75], p=1.9e-06
geo-greedy	20	+75.62 [+75.21, +76.03], p=9.6e-05	+59.32 [+57.97, +60.68], p=1.9e-06

Handover metric family (sums over seeds; outage = ≥ 3 consecutive lost send attempts):

arm	startup n / lost	scheduled n / lost	unplanned n / lost	massfail n / lost	other n / lost	lost per scheduled handover	median outage, scheduled (s)	median outage, unplanned (s)
anthocnet	0 / 0	592 / 5958	1573 / 11779	0 / 0	100 / 538	9.93	2.25	1.75
aodv	6 / 820	399 / 20705	874 / 16701	0 / 0	276 / 7453	34.51	5.12	2.25
geo-greedy	80 / 83762	40 / 7353	0 / 0	0 / 0	0 / 0	12.26	47.00	—
oracle	2 / 7	7 / 27	693 / 2354	0 / 0	17 / 66	0.04	1.00	0.75
oracle-delay	1 / 4	7 / 25	782 / 2671	0 / 0	22 / 77	0.04	1.00	0.75

arm	hop changes / flow / min	mean hops (delivered)
anthocnet	8.71	25.60
aodv	1.86	20.70
geo-greedy	1.37	4.14
oracle	2.82	23.75
oracle-delay	2.83	24.23

On S1, AntHocNet's control overhead (NRL 955) is in the same range as AODV's (904), on a 1584-node shell — the cost question #207 asked. It pays for that with 4.7× the hop-change rate of AODV: ants keep moving traffic to the currently-shortest path, which is what the delay result shows. The oracle arms' "unplanned" outages are the 1 s re-solve interval on S1 meeting the 1 % ISL failure overlay — a cost of the bound's sampling, not of routing.

starlink1-storm

arm	seeds	PDR %	mean delay (ms)	delay99 (ms, boot)	NRL	refused at source %
anthocnet	20	78.79 ± 1.04	90.04 ± 1.95	212.8 [208.1, 217.4]	957.42 ± 63.39	0.00
aodv	20	46.49 ± 2.39	111.82 ± 5.71	1704.0 [1546.8, 1841.3]	1018.31 ± 63.81	0.00
geo-greedy	20	3.60 ± 0.48	25.48 ± 0.63	27.4 [26.6, 28.4]	0.00 ± 0.00	0.00
oracle (bound)	20	95.08 ± 0.29	81.75 ± 1.34	134.2 [133.2, 135.7]	0.00 ± 0.00	0.10
oracle-delay (bound)	20	94.95 ± 0.26	78.51 ± 1.29	126.8 [125.3, 128.4]	0.00 ± 0.00	0.10

AntHocNet − arm, paired per seed (95 % t-CI, Wilcoxon p):

vs	n	Δ PDR (pp)	Δ mean delay (ms)
aodv	20	+32.30 [+30.03, +34.57], p=1.9e-06	-21.78 [-27.29, -16.27], p=1.9e-06
geo-greedy	20	+75.19 [+73.94, +76.44], p=1.9e-06	+64.56 [+62.35, +66.77], p=1.9e-06

Handover metric family (sums over seeds; outage = ≥ 3 consecutive lost send attempts):

arm	startup n / lost	scheduled n / lost	unplanned n / lost	massfail n / lost	other n / lost	lost per scheduled handover	median outage, scheduled (s)	median outage, unplanned (s)
anthocnet	0 / 0	612 / 6610	1528 / 11347	42 / 354	85 / 510	11.15	2.25	1.75
aodv	6 / 410	378 / 21513	716 / 14792	18 / 5404	297 / 8116	36.28	5.62	2.50
geo-greedy	80 / 84112	38 / 7184	2 / 14	0 / 0	0 / 0	12.11	47.00	1.75
oracle	3 / 10	28 / 124	714 / 2428	0 / 0	11 / 42	0.21	1.00	0.75
oracle-delay	1 / 4	23 / 99	776 / 2613	0 / 0	19 / 68	0.17	1.00	0.75

arm	hop changes / flow / min	mean hops (delivered)
anthocnet	9.22	26.64
aodv	1.79	20.37
geo-greedy	0.94	5.16
oracle	3.39	24.47
oracle-delay	3.59	24.81

The mass failure costs AntHocNet 0.6 pp of PDR and costs AODV 4.7 pp. The outage table says why: AntHocNet's 42 mass-failure outages lost 354 packets in total (≈ 8 each, a few seconds of traffic); AODV's 18 lost 5404 (≈ 300 each — the flow stays black until a route error finally reaches the source and a fresh discovery succeeds).

Calibration

The model is circular orbits on a spherical Earth with propagation-only links (ADR-0022), so calibration asks one thing: are the delays and the reconfiguration timing the right size? Two external references answer it.

Against Hypatia: Paris–Luanda RTT on Starlink S1

Hypatia (Kassing et al., IMC 2020, Fig. 13) publishes the Paris–Luanda RTT over 200 s on Starlink S1 (72 × 22, 550 km, 53°, minimum elevation 25°, path re-solved every 100 ms): 85–117 ms. The hypatia cell reproduces that setup — --pairs=hypatia, the delay oracle re-solved every 0.1 s, 1000 s, one flow.

	min	p25	p50	p75	p95	max
leo-walker RTT (ms)	54.8	77.2	101.0	140.3	186.4	207.5

• The median sits inside Hypatia's band (101.0 ms), and 33.5 % of the seconds fall inside 85–117 ms. Mean path length is 20.8 hops.
• The spread is wider than Hypatia's, and the cause is known: leo-walker gives each ground station one serving satellite (the highest one) and hands over on the 15 s clock; Hypatia lets the station use any visible satellite and picks whichever minimises the end-to-end path. The plateaus in the chart are exactly the 15 s-to-minutes stretches where the serving satellite on one end is a poor entry point into the grid (high plateaus) or an unusually good one (low plateaus, under Hypatia's floor). The minimum (54.8 ms) stays above the 43.4 ms great-circle-at-c line; nothing is faster than physics.
• What this means for the routing results: the single-serving-satellite GSL makes every arm's paths longer than an any-visible-satellite model would, by the same geometry for every arm. It moves absolute delays; it does not favour an arm.

Against LENS: one day of a real Starlink dish

LENS (Zhao & Pan, MMSys 2024; dataset CC BY-SA 4.0) publishes raw 10 ms ping traces from Starlink dishes to their point of presence. One full day of the bruhl dish (Brühl → Frankfurt PoP, 2026-10-06, 24 hourly traces, 5.96 M pings) summarised by ns3/tools/lens-calibration.py; derived numbers only, in docs/benchmarks/cells/lens-bruhl-20261006.txt:

	min	p5	p50	p95	p99	loss
LENS bruhl RTT (ms)	11.4	16.0	20.8	33.6	47.2	0.186 %

• Delay: a dish-to-PoP RTT is one bent-pipe hop (dish → satellite → gateway, and back). At 550 km and 25° minimum elevation the propagation-only floor for that is 7.3–15.0 ms; LENS's 11.4 ms minimum and 20.8 ms median are that floor plus scheduling and queueing that leo-walker does not model. The model's propagation is the right size; its RTTs are lower bounds.
• Timing — the 15 s clock is confirmed in the wild: binning loss by UTC second mod 15, LENS loss peaks at phases 11 and 12 (0.399 % and 0.230 % against a ~0.16 % baseline on every other phase), and median RTT steps up at phase 12 (21.2 ms against 20.6–20.9). leo-walker's handover clock fires at phase 12 (clock=15 clockOffset=12). The scheduled-handover class in the tables above is placed where the real network puts it.

The geo-greedy verdict

Greedy geographic forwarding — the family behind most "position-based LEO routing" proposals — is given every advantage here: perfect global positions, zero control traffic, a forced uplink at the ground station, and the destination's serving satellite (not the city) as the target. It still delivers 41 % on walker16 and under 4 % on Starlink S1.

The failure is structural, not tuning. On a Walker +grid, a satellite's four neighbours sit fore/aft in its own plane and left/right in the adjacent planes. Reaching a destination across planes needs, at some point, a hop that does not reduce Euclidean distance to the target — the grid is a lattice wrapped on a sphere, and the greedy-progress invariant has local minima everywhere planes cross. The 80 "startup" outages per cell are every flow on every seed (4 × 20) beginning inside a local minimum — stuck for a median ~200 s on walker16 and for essentially the whole 300 s run on S1, until satellite motion happens to carry the flow's entry point out of it. Face/perimeter routing, the planar-graph repair that makes greedy forwarding complete in 2-D, has no 3-D equivalent with the same guarantee. geo-greedy's low delay on S1 (25 ms) is survivorship: the only flows it delivers are the short ones.

The cross-plane pairing bug and the superseded runs

The first leo-walker paired each satellite with the same slot in the neighbouring plane. With Walker phasing, the same slot in the next plane is not the nearest satellite there: on S1, same-slot cross links averaged 1470.7 km against 636.9 km for the best constant shift, and the Hypatia calibration came out near 274 ms — almost 3× the published range. An independent Python model of the shell reproduced the 3×, and the fix (0b5a65ad) picks one constant slot shift for the whole shell — the one that minimises the mean cross-link length at t = 0 — with the seam plane pairing to its nearest satellite.

Each cell's # anchor isl line records the choice, and the CI smoke (ns3/tools/check-leo-walker.sh) asserts it is no longer on average than same-slot pairing and shorter than the in-plane chord:

shell	chosen shift (slots)	cross mean	same-slot mean	in-plane chord
walker16	15 (≡ −1)	1838.4 km	2527.3 km	2934.5 km
starlink1	21 (≡ −1)	636.9 km	1470.7 km	1969.9 km

Every leo-walker cell measured before 0b5a65ad was re-run and is superseded; none was published. The longer cross links had made every arm's paths longer by the same geometry, so the arm ranking did not change — but the absolute delays and the calibration did, and only the re-run numbers are on this page.

Threats to validity

• Circular orbits, spherical Earth. The stock ns-3.48 model; no J2, no eccentricity, no TLEs. Results are about routing over Walker geometry and delay, not orbital precision (ADR-0022).
• Propagation-only links. ISL and GSL delays are distance / c; no PHY, no weather, no queueing at the gateway. RTTs are lower bounds (the LENS comparison above).
• Single serving satellite per station. The main cause of the wider Hypatia spread; it lengthens every arm's paths alike.
• The oracle bounds are sampled, not continuous. The hop oracle re-solves every 1 s everywhere; the delay oracle every 0.1 s on walker16 and every 1 s on S1, to keep the 20-seed CI run bounded. Their "unplanned" outages are that sampling meeting the ISL failure overlay.
• AODV is order-dependent in ns-3 (#362): stock AODV keys socket maps on heap addresses. Each arm runs alone in its own dispatch here, so no arm's numbers depend on another's.
• 10 Mbps links, 2 kbit/s flows. The load is light on purpose — this is a reachability-under-motion test, not a capacity one. Congestion on ISLs is #206's question.
• No OLSR on S1. Proactive link-state flooding on 1584 nodes exceeds the CI budget; it is measured on walker16 only, where 20 seeds already took 5 h 54 min.

Reproduce

# one arm of one cell, locally in the 3.48 image (S1, 20 seeds)
./ns3 run "leo-walker --csv --protocols=anthocnet --planes=72 --sats=22 --phasing=39 \
  --altitude=550 --inclination=53 --minElevation=25 --time=300 --runs=20 \
  --islDown=0.01 --oracleInterval=1"

# the tables on this page, from the committed cells
python3 ns3/tools/leo-summary.py walker16 starlink1 starlink1-storm

# the charts
python3 ns3/tools/family-charts.py

In CI: satellite-benchmark.yml, harness=leo-walker, one arm per dispatch, with the cell's flags in extraArgs.