Satellite suite: the ISL grid¶

Regime: satellite / inter-satellite-link mesh — the second of the two regimes this project measures. The MANET suite (the original) lives under scenarios/; why the two regimes are different enough to need separate suites is network-regimes.md.

← Benchmark index · Metrics · Methodology

Provenance — measured at 0f7a3ab, with the upper-bound control. These are the first satellite numbers this project has produced against a control, from the v1.5.0 campaign's phase 3 (campaign plan, #415): run 31830581632, ref=main, commit=0f7a3ab9f3b7ba837b43bbd2d3fd7e04f7090c1c, image=ghcr.io/danieljoppi/ns3:3.42-opt, profile=release, harness=isl-grid. Cell: 4×4 +Grid torus — 16 satellites / 32 ISLs (mean degree 4.00) at 5.00 ms / 10 Mbps — 8 flows, 900 s, 20 seeds, protocols=anthocnet,aodv,olsr,oracle. They supersede nothing, because nothing comparative was published from this suite before. Reproduce the base-torus results at commit 0f7a3ab; the #432 adversarial cells carry their own provenance (corridor/failcell 820f5cf, seam cell 5220bd0).

The headline result: the base torus does not discriminate¶

The #216 control now exists, it ran on this cell, and its first finding is a negative one — all three real protocols already sit exactly on the upper bound, so this cell cannot carry a comparative delivery or latency claim in either direction.

protocol	PDR%	delay (ms)	delay99 (ms)	thrput (kbps)	NRL	NRL bytes	jitter (ms)
anthocnet	100.0	10.2	11.0	23.35	2.16	3.63	0.00
aodv	100.0	10.2	11.0	23.34	2.03	1.06	0.01
olsr	100.0	10.2	11.0	23.29	1.93	4.76	0.00
oracle	100.0	10.2	11.0	23.36	0.00	0.00	0.00

The control is exact here¶

On this topology the oracle arm (#419) is not an approximation of an upper bound — it is one. Every seed self-describes as

##ORACLE## <seed> oracle mode=wired approx=0 nodes=16 edges=64 recomputes=900 changes=1 range=-1.0 noRoute=0 nrl=0.00

mode=wired approx=0 — adjacency comes from ns-3 Channel co-membership, so on point-to-point ISLs the graph is the wiring; there is no radius, no propagation model and nothing to guess (range=-1.0).
edges=64 directed edges = the 32 ISLs of the 4×4 torus, both directions.
recomputes=900 changes=1 — the edge rebuild runs every second for the whole 900 s run, but the topology is static, so one Dijkstra solve covers the entire run.
noRoute=0 and nrl=0.00 — never without a route, and not one routing packet on the wire.

That is the contrast with the MANET suite, where the same arm runs approx=1 against a geometric disk and is therefore a reference point rather than a proven bound — see the grid page for what that costs there (do not carry its numbers across regimes).

What the tie means¶

PDR, mean delay and delay99 are identical to the bound on all three real protocols. The 10.2 ms is simply two ISL hops at 5.00 ms — the analytic floor this page already gates on — plus the small unattributed excess tracked in #250, which the oracle now shows is not a protocol artefact: the control pays it too. On a static, lossless, uncongested 4-regular torus every protocol finds a shortest path; there is no routing difficulty here to be better or worse at.

So no delivery or latency claim can be made from this cell in either direction: a protocol cannot beat 100 % / 10.2 ms, and failing to separate the protocols here says nothing about them.

This is exactly the service the control provides, and nothing else could have provided it. Before the oracle arm existed, "all three protocols deliver 100 %" was indistinguishable from "all three protocols are excellent here." The oracle shows the ceiling was touching the floor.

The only separating axis is overhead¶

Routing overhead is the only axis that separates the arms at all. (Throughput spans 23.29–23.36 kbps and jitter 0.00–0.01 ms across all four rows — that is the arms agreeing, not differing.)

NRL — olsr 1.93 < aodv 2.03 < anthocnet 2.16, against the oracle's exact 0.00.

But nrl_bytes reorders that ranking — aodv 1.06 < anthocnet 3.63 < olsr 4.76 — which is its own caution: OLSR sends the fewest control packets and the most control bytes. Quote both, or neither; never NRL alone. (Both metrics are defined in metrics.)

Where discrimination has to come from instead¶

The harness is now validated end to end — anchors, determinism, and the control — but the base torus cannot carry a comparative result. Discrimination must come from the adversarial cells this suite already anticipates:

the asymmetric-congestion corridor (#280, instrument described below) — now measured with the oracle arm, below;
scripted ISL breaks (#260) — now measured with the oracle arm, below;
larger or less regular constellations — this one now has a designed, measured answer: the #432 item 3 seam cell (a deliberately irregular 6×6 torus with the rows-axis wrap seam cut, PR #453) was measured at 20 seeds with the oracle arm, below — and static irregularity also cannot discriminate on this substrate. Headline discrimination in the static suite comes only from load (the corridor) or an event instrument (the failcell); the residue of item 3 (dynamics/churn) is subsumed by #297.

The recommendation this section used to close on — add the oracle arm to those cells before any satellite comparison is published — was followed: both instrumented cells ran with the control beside them, and the control stayed exact on the cut-ISL topology too (a downed interface simply drops out of the Channel adjacency).

The adversarial cells: corridor and failcell (#432)¶

Provenance — measured at 820f5cf, 20 seeds per cell, with the exact control. Both cells are from the #432 dispatch (dispatch record + pre-registered expectations): ref=main, commit=820f5cfa2295baf8f7b734ece862bd93747b2521, image=ghcr.io/danieljoppi/ns3:3.42-opt, profile=release, harness=isl-grid, 900 s, 20 seeds (1–20), protocols=anthocnet,aodv,olsr,oracle (the pinned canonical order). Corridor cell: run 31922945507. Failcell: run 31922957019. Every number below is a 20-seed mean computed by script from the per-seed ##RUN## / # corridor / # failcell rows (bench_parse.py plus a diagnostic-line reducer — nothing eyeballed or hand-averaged); ± is a 95 % t-CI half-width, boot[lo,hi] a percentile-bootstrap 95 % interval (delay99), and [lo,hi] after a paired difference is the paired per-seed 95 % t-CI. Both cells passed scenario_check.py results (0 fail, 0 warn). Reproduce this section's results at commit 820f5cf.

These are the first two adversarial regimes, each re-run with the oracle arm beside it as recommended above — and they are the suite's first discriminating satellite results. In one sentence each: the corridor cell separates the arms on the headline delay metric, and AntHocNet beats the congestion-blind bound on it (with a mechanism caveat that belongs in every quote of that claim); the failcell still cannot separate anyone on headline metrics, but its reconvergence instrument now discriminates — OLSR reconverges 5× slower than the oracle's floor while AntHocNet and AODV sit statistically on it. A third cell — the #432 item 3 seam cell, measured after these two — closes the adversarial-design question for the static suite: even genuine topological irregularity does not move the headline metrics off the bound.

Cell: 6×6 torus (36 satellites, 72 ISLs), 5 ms / 10 Mbps, --corridorLoad=12Mbps --corridorLoadAt=15 (ρ = 1.2 — the documented deliberate overload of the east corridor's second link), --flows=0, so the headline ##RUN## row is the probe flow (0,0)→(0,3) over two equal 3-hop corridors (the #280 zero-new-parsing contract). The AntHocNet arm runs at shipped defaults (EnableMacMetric off): this cell's question is where the shipped protocols sit against the exact bound; the mac-metric/gate mechanism ladder stays on #216/#180.

protocol	probe PDR %	delay (ms)	delay99 (ms)	NRL	clean-corridor seeds
anthocnet	100.00 ±0.00	55.6 ±19.4	57.5 boot[40.9, 74.1]	36.62 ±0.05	10/20
aodv	99.93 ±0.01	97.3 ±0.1	99.0 boot[99.0, 99.0]	36.38 ±0.04	0/20
olsr	100.00 ±0.00	68.7 ±18.9	70.0 boot[53.4, 86.5]	39.44 ±0.05	7/20
oracle	100.00 ±0.00	97.2 ±0.1	99.0 boot[99.0, 99.0]	0.00	0/20

("Clean-corridor seeds" = seeds whose # corridor line shows the majority of post-loadStart probe packets leaving on the clean west corridor.)

The cell discriminates, and the pre-registered headline happened: an arm beat the bound. The oracle routes the probe through the loaded corridor in all 20 seeds (viaLoaded≈3536 viaClean=0, probe delay 97.19 ±0.13 ms): its Dijkstra tie-break picks east in every seed, and being blind to congestion by construction it stays there while the 12 Mbps background saturates the link. AntHocNet's probe delay is lower by a paired per-seed difference of −41.6 ms, 95 % CI [−60.9, −22.2] at the same 100.00 % probe PDR, and its delay99 (57.5 boot[40.9, 74.1]) sits far below the bound's 99.0.

This is not a paradox, and it must never be quoted as "faster than optimal". The oracle is a shortest-path bound: exact on topology, blind to load by construction — which is the reason this cell exists (the #202 survey §6's "congestion the precomputed control cannot see"). On a saturated corridor the shortest path is not the fastest path, so the bound's delay is beatable here precisely because the cell was built to make it beatable; its PDR and its NRL = 0.00 remain the bounds they always were.

The mechanism caveat that belongs in every quote of the claim. The per-seed # corridor lines are all-or-nothing: every arm's every seed sends essentially 100 % of post-loadStart probe packets down a single corridor. AntHocNet locks clean-west in 10/20 seeds (15.2–15.5 ms) and loaded-east in 10/20 (92.3–97.4 ms), with no seed shifting corridors after the load arrives — the # pher traces show both corridors hold pheromone at load onset, then the unused corridor's entry evaporates to zero and never re-forms at the source. The #216 round-2 lock-in observation therefore persists at 900 s; session length did not break it (follow-up (a) of the round-2 readout, now answered). What beats the bound is not measured congestion adaptation but initial-choice diversity: the stochastic ant choice lands half the seeds on the corridor that will stay clean, and lock-in keeps them there. OLSR makes the same point from the other side: a purely load-blind hop-count protocol also "beats the bound" (−28.4 ms, 95 % CI [−47.2, −9.6]) on nothing but seed-dependent tie-breaking (7/20 clean seeds, at the identical 15.2 ms), and AntHocNet vs OLSR is statistically indistinguishable on this cell (paired −13.1 ms, 95 % CI [−41.6, +15.3]). AODV, which commits its discovery-time route on the quiet net and keeps it, ties the bound exactly (+0.1 ms, 95 % CI [−0.0, +0.3]) and is the only arm below 100 % probe PDR (99.93 ±0.01 — tail-drops on the saturated ISL).

The quotable claims, in full: (1) this is the suite's first cell where headline metrics separate the arms, at 20 seeds with CIs; (2) AntHocNet at shipped defaults beats the congestion-blind shortest-path bound on probe delay with a paired CI excluding zero — the first legitimate satellite "beats-the-bound" claim, valid only with the shortest-path qualifier attached; (3) at shipped defaults the advantage is corridor lottery plus lock-in, not observed load-shifting — AntHocNet does not separate from load-blind OLSR here. Whether EnableMacMetric=true turns 10/20 clean seeds into 20/20 is exactly the #216/#180 mechanism ladder, deliberately not part of this dispatch.

The failcell: reconvergence at the oracle floor (#260)¶

Cell: the published base-torus cell exactly (4×4, 8 flows) plus one scripted break — --breakLink=0,0,3,0 --breakAt=450 cuts the ISL (0,0)–(3,0), which preflight verified lies on a shortest path of flows 0→15 and 3→12; equal-cost alternates survive on the torus, and 450 s of post-break run remain. The oracle's recompute cadence is 1 s, and every seed self-reports changes=2 noRoute=0 — the initial solve plus exactly one post-break recompute, never without a route.

protocol	PDR %	delay (ms)	delay99 (ms)	NRL	tReconverge (s)
anthocnet	100.00 ±0.00	10.16 ±0.00	11.0 boot[11.0, 11.0]	2.12 ±0.00	0.91 ±0.10
aodv	99.95 ±0.00	10.18 ±0.01	11.0 boot[11.0, 11.0]	2.00 ±0.00	0.95 ±0.11
olsr	99.93 ±0.02	10.15 ±0.00	11.0 boot[11.0, 11.0]	1.89 ±0.02	4.49 ±1.03
oracle	100.00 ±0.00	10.15 ±0.00	11.0 boot[11.0, 11.0]	0.00	0.86 ±0.11

On headline metrics the break cell still does not discriminate — that is the finding, and it is worth stating plainly. A single cut ISL with equal-cost alternates on a static lossless torus is absorbed at ~100 % PDR / 10.2 ms by every arm; the visible cost of the entire event is ≤ 0.07 pp of PDR (aodv 99.95, olsr 99.93 — the packets lost inside each arm's reconvergence window), far below any materiality threshold. This is the measured confirmation of the prediction that kept #432 item 3 undispatched: a topology event the routing can absorb does not move headline numbers on this substrate; only an instrument aimed at the event window sees it.

The instrument, however, now discriminates — and no arm beats the floor. The oracle's 1 s recompute cadence puts its reconvergence floor at 0.86 ±0.11 s, inside the pre-registered ≤ 1 s. AntHocNet (0.91 ±0.10 s) and AODV (0.95 ±0.11 s) sit statistically on that floor — paired differences +0.05 s [−0.13, +0.23] and +0.09 s [−0.08, +0.27], both CIs spanning zero — while OLSR reconverges 5× slower: 4.49 ±1.03 s, paired +3.63 s [+2.55, +4.71] above the floor, the cost of waiting out its periodic HELLO/TC machinery instead of reacting to the loss event. Pre-registered expectation 2 is confirmed: nobody beats the topology-change floor; the comparison is who reaches it, and two of the three real arms do.

Read tDetect/tReconverge with the documented caveats (above): tDetect is 0.00 in all 20 AntHocNet seeds by construction for a scripted break (the interface-down fast path is the detection; the baselines expose no detection trace — theirs is nan), and tReconverge is a proxy whose per-seed minima (0.22–0.38 s across arms) include the ~125 ms CBR-gap contribution of unaffected flows.

The seam cell: static irregularity also ties the floor (#432 item 3)¶

Provenance — measured at 5220bd0, 20 seeds, with the exact control. Run 31985601548, ref=main, commit=5220bd0d6275413d58b9701d944b45d5496462d3, image=ghcr.io/danieljoppi/ns3:3.42-opt, profile=release, harness=isl-grid. Cell: 6×6 +Grid torus with the rows-axis wrap seam removed at columns 1–5 (--removeLinks=5,1,0,1;5,2,0,2;5,3,0,3;5,4,0,4;5,5,0,5 — 67 ISLs, degree min 3 / max 4, mean 3.72, connected), 5 ms / 10 Mbps, 8 standard flows (i → 35−i), cbrBps 4096, 900 s, seeds 1–20, protocols=anthocnet,aodv,olsr,oracle (the pinned canonical order). Design and pre-registration: PR #453 and the issue design record, both written before any 20-seed data existed. Every number below is a 20-seed mean computed by script from the per-seed ##RUN## rows (bench_parse.py; paired per-seed deltas against the oracle rows via the skill's stats_util — nothing eyeballed or hand-averaged); ± is a 95 % t-CI half-width, boot[lo,hi] a percentile-bootstrap 95 % interval (delay99), and paired [lo,hi] a per-seed paired 95 % CI (t for PDR/delay, bootstrap for delay99), n = 20 seeds. scenario_check.py results: OK (0 fail, 0 warn). Reproduce this subsection at commit 5220bd0.

The cell is the designed answer to "larger or less regular constellations" (#432 item 3). The base torus tied because on a 4-regular torus every alternative is an equal-cost shortest path, so no routing choice has a price. Cutting the rows-axis wrap seam at every column but 0 makes choices priced while staying static, lossless and uncongested: non-uniform degree (3–4), genuine path-length asymmetry (flow shortest paths 2/4/6/6/4/2/4/6 hops, mean 4.25 — analytic oracle floor 21.25 ms mean, 31.0 delay99 in the 1 ms bins), and the surviving wrap ISL (0,0)–(5,0) — "the funnel" — on a shortest path of all 8 flows, strictly so for two of them (the detour around the open rows axis costs +4 hops = +20 ms).

protocol	PDR %	delay (ms)	delay99 (ms)	NRL
anthocnet	100.00 ±0.00	21.58 ±0.00	31.0 boot[31.0, 31.0]	4.56 ±0.01
aodv	99.91 ±0.01	21.71 ±0.04	31.0 boot[31.0, 31.0]	4.27 ±0.00
olsr	100.00 ±0.00	21.57 ±0.00	31.0 boot[31.0, 31.0]	4.68 ±0.00
oracle	100.00 ±0.00	21.57 ±0.00	31.0 boot[31.0, 31.0]	0.00 ±0.00

Pre-registered expectation 1 — the exact bound — held. Every seed self-reports ##ORACLE## … mode=wired approx=0 nodes=36 edges=134 recomputes=900 changes=1 noRoute=0 nrl=0.00 (134 directed edges = the 67 post-removal ISLs, both directions — the oracle proves it solved the seam graph, not the full torus). Oracle PDR 100.00, delay99 exactly 31.0 in all 20 seeds, and mean delay 21.57 ±0.00 = the 21.25 ms analytic floor plus a 0.32 ms excess (~0.075 ms/hop at 4.25 mean hops — the same #250-class per-hop excess the base cell pays).

Pre-registered expectation 2 — hop-stretch discrimination — failed: the predicted mechanism did not occur. The registered signature was AntHocNet mean delay above the oracle floor by evaporation lock-in onto longer-than-shortest funnel-flow paths. The smallest version of that effect is quantized at the hop: a wrong first hop on a strictly-shortest funnel flow costs +20 ms on that flow, i.e. +2.5 ms on that seed's 8-flow mean if held for the run, +0.25 ms if held for a tenth of it. Measured: AntHocNet ties the floor — PDR 100.00 in every seed (paired dPDR all-zero), delay99 exactly 31.0 in every seed (paired d_delay99 all-zero), and paired mean delay +0.0165 ms [+0.0102, +0.0228] — fifteen times smaller than even the tenth-of-a-run signature, four orders below the +20 ms detour. No seed in 160 flow-seed combinations shows a held longer path: the stochastic ant choice plus lock-in landed on true shortest paths every time. OLSR ties the floor outright (paired dPDR −0.001 pp [−0.002, +0.000], d_delay +0.0025 ms [−0.002, +0.007], both spanning zero). The secondary NRL prediction is not attributable: packet-NRL ordering did invert vs the base cell (olsr 4.68 > anthocnet 4.56 > aodv 4.27 here, vs olsr cheapest on the 4×4), but the grid size changed with the seam and no full 6×6 reference cell exists, so the shift cannot be pinned on the seam — noted, not claimed.

Three paired CIs exclude zero, and none of them is discrimination — the letter-level deviation from the pre-registration, disclosed. The outcome-3 clause said "all paired CIs spanning zero"; strictly, three do not: AntHocNet's +0.0165 ms mean-delay offset, and AODV's dPDR −0.0925 pp [−0.1036, −0.0814] with d_delay +0.143 ms [+0.108, +0.178]. All three are constant sub-hop-quantum offsets, 30–300× below the 5 ms hop price the cell was built to charge, resolvable only because this substrate's seed-to-seed dispersion is ~0.01 ms — at that dispersion a paired CI resolves fixed per-protocol processing cost, not path choice. AODV's PDR deficit is the flood-discovery startup transient, not a seam effect: ≈ 53 of 57 600 offered packets (≈ 7 per flow) dropped before first routes exist, present in every prior cell (corridor 99.93, failcell 99.95, and ≈ 0.9 pp at the PR #453 120 s smoke — a fixed packet count amortized over run length), with delay99 untouched at 31.0. The methodological finding travels forward: on a near-deterministic substrate, "paired CI excludes zero" is not a sufficient discrimination criterion — the next static-cell pre-registration must pin an effect-size floor to the hop quantum (material fraction of 5 ms; ≥ 1 pp PDR).

Verdict — pre-registered outcome 3 obtains: static irregularity also cannot discriminate, and #432 item 3 closes on its falsification branch. All three real arms sit at or indistinguishably near the oracle bound at headline scale — no arm pays a hop anywhere on a deliberately irregular graph where wrong choices finally had a price. With the corridor and failcell results above, the static satellite suite's discrimination map is now complete and consistent: headline metrics move only under load (the corridor's instrument); events and now irregularity are absorbed without headline trace (the failcell needed its reconvergence instrument; the seam cell has no instrument left to fall back on — that is its verdict). The remaining scope of item 3 — dynamics, churn, handover — is subsumed by the #297 satellite evaluation-credibility epic, per the closing criterion pre-registered on the issue. Honest caveats: the cell is lossless and uncongested by design (that isolation is the point, and also the reason nothing separates); the 1 ms delay99 binning makes 31.0 a coarse ceiling check, not a tail measurement; and the tie says the arms all find shortest paths here — it says nothing about what they do under the dynamics this cell deliberately excludes.

What this suite is, and is not¶

One harness, ns3/examples/isl-grid.cc: an R×C +Grid torus of point-to-point ISLs — one /30 subnet per link, so every satellite holds one interface per neighbour (degree 4 on the torus). That multi-interface shape is the point: it is what the MANET suite can never exercise, and it is what #203 broke on before #224 fixed next-hop resolution per interface. Metrics mirror anthocnet-compare's definitions at the same IP-layer counting point, so a number here is comparable in kind (never in regime) to a MANET number.

It is a static snapshot grid: no orbital mechanics, no GSL handover, no link churn. That makes it the quiet-cell instrument — and the measured answer is sharper than the expectation this page used to record. It said the #216 precomputed shortest-path control was expected to win here and AntHocNet expected to lose; with the control measured, neither happened — every protocol ties the bound exactly, and the cell discriminates nothing. Of the adversarial cells that could separate them (unpredicted ISL loss, asymmetric congestion, handover churn), the first two have instruments:

Unpredicted ISL loss — #260 added a scripted single-ISL break (--breakLink=r1,c1,r2,c2 --breakAt=<s>, cut via Ipv4::SetDown on both endpoint interfaces) that reports a per-run # failcell … tDetect=<s> tReconverge=<s> line.
Asymmetric congestion (#216 cell 1, unblocked by #206) — --corridorLoad=<rate> offers background load over one of two equal-length corridors and reports a per-run # corridor line; see the congestion cell below.

The handover cell is still #216's scope and does not exist yet. The #216 control row has now run beside both instruments — the #432 adversarial cells above — so the failcell and corridor lines carry an upper bound and support protocol-level claims, with the caveats recorded there.

Read the failcell numbers honestly. tDetect (break → the protocol's first neighbour-loss event for the severed peer, from the RouteChanged trace; AntHocNet only — the baselines expose no such trace) is ~0 by construction for a scripted break: the adapter's #260 fast path is the Ipv4 interface-down notification itself, the only failure signal a PointToPointNetDevice offers (no retry-limit trace; IP drops packets to a down interface before any device trace can fire). That models an ISL terminal reporting loss-of-light locally within ms. A silent failure — the interface stays up but frames stop arriving (e.g. a receive-side error model) — is not covered by the fast path and still waits the full hello timeout, helloInterval × allowedHelloLoss = 2 s at defaults; a failure cell built on silent loss lower-bounds the protocol, exactly the pre-#260 caveat. tReconverge (break → the last of the per-flow first deliveries after the break) is a proxy: the harness does not know which flows crossed the broken ISL, so unaffected flows contribute ~one CBR inter-packet gap (~125 ms at the 64 B / 4096 bps defaults) and only a value clearly above that gap measures re-convergence.

Configuration¶

knob	default	meaning
`rows` × `cols`	4 × 4	orbital planes × satellites per plane
`torus`	true	wrap the edges (+Grid); false = open grid
`islDelayMs`	5	one-way ISL propagation delay (LEO ISLs are ~3–13 ms)
`islRate`	10Mbps	ISL data rate
`flows` / `cbrBps`	8 / 4096	CBR load
`protocols`	anthocnet,aodv,olsr,oracle	`anthocnet,aodv` is the #250 hop-delay discriminator pair; `oracle` is the #419 exact upper bound. `dsdv` is not available in this suite — see below
`breakLink` / `breakAt`	off	#260 scripted single-ISL break: endpoints `r1,c1,r2,c2` + cut time (s); emits `# failcell` detect/reconverge lines
`corridorLoad` / `corridorLoadAt`	off / 15	#216 cell 1: background rate (e.g. `12Mbps`) loading one of two equal-length corridors, switched on at `corridorLoadAt` (s); emits `# corridor` path-shift lines. Needs `torus=true` and even `cols` ≥ 4
`removeLinks`	off	#432 item 3: static ISL removals — semicolon-separated `r1,c1,r2,c2` endpoint quadruples, each adjacent on the grid (same rules as `breakLink`), never built (absent from t=0, both directions). Makes the constellation deliberately irregular: non-uniform degree, unequal-length alternatives. The remaining graph must stay connected (harness abort; the preflight proves it by BFS)

Why `dsdv` is not one of the arms (#420)¶

This table used to list anthocnet,aodv,olsr,dsdv as the default protocol list. That list never ran. DSDV has been an arm of isl-grid.cc since the file was added, and it has aborted on every multi-ISL grid since — the suite's own documented default was never once exercised end to end. It surfaced only when the #415 campaign asked for all five arms and the process died, with an empty stderr, straight after the olsr rows. This is a latent defect, not a regression; the narrowing above is a correction of the documentation to what the harness can actually measure, not a quiet retreat from a working configuration.

The cause is in ns-3's dsdv::RoutingProtocol, and it is structural rather than a tuning problem. DSDV there is written for a node with exactly one non-loopback interface: when it advertises itself it hardcodes m_ipv4->GetAddress(1, 0), so a satellite holding four ISLs announces only the address of the first one. Its peers, however, learn next hops from the source address of the update, which is the address of the /30 the update arrived on — three of the four are addresses DSDV never advertised. The next-hop lookup in LookForQueuedPackets() therefore misses, its return value is not checked, and the packet is forwarded on a default-constructed Ipv4Route whose output device is null. Ipv4L3Protocol::SendRealOut asserts on that in a debug build (cond="interface >= 0") and indexes the interface list with -1 in the optimised profile the campaign runs — hence a SIGSEGV with nothing on stderr.

There is no fix on this side of the boundary. One interface per link is the ISL mesh (it is the shape #203 exists for), so the harness rejects the combination up front with an explanatory message instead of dying mid-campaign. The rejection is keyed on the topology, not on the arm: the 1×2 single-ISL grid used by the #237 anchors gives every node one interface, and DSDV is correct there. DSDV remains a full baseline in the MANET suite, where every node has a single wifi interface and the assumption holds.

ns3/tools/check-sat-arms.sh runs every supported arm on a small torus per PR so a defect of this class cannot reach a campaign dispatch again.

dsdv is therefore absent by necessity, not by choice, and its absence from the result table above is not a gap in the comparison.

The congestion cell (#216 cell 1)¶

"Congestion the precomputed control cannot see" — the adversarial regime the #202 survey §6 names first, runnable since #206 gave EnableMacMetric a real per-next-hop signal on ISLs (ADR-0017).

Construction (fixed, derived from the grid): probe flow from satellite (0,0) to (0,cols/2) at the standard cbrBps/64 B. On the torus row ring there are exactly two shortest paths, each cols/2 hops — east (0,0)→(0,1)→…→(0,cols/2) and west (0,0)→(0,cols−1)→…→(0,cols/2); any path leaving row 0 is ≥ 2 hops longer. A background OnOff flow (0,1)→(0,2) (1000 B packets, rate corridorLoad, UDP port 10, constant duty — the OnOff default of 1 s on/1 s off would oscillate the queue empty→full→empty every 2 s, and an ant sampling the off-phase reads zero backlog; the first dispatches measured exactly that) loads the east corridor's second link from corridorLoadAt onward — after discovery has settled on the quiet net, so a reactive baseline has already committed a route. Hop count cannot distinguish the corridors; only a congestion signal can. The background load is offered load, not traffic under measurement: port 10 is excluded from the data metrics and from the NRL control counter.

Output, one line per (protocol, seed) next to the ##RUN## row:

# corridor <proto> seed=<s> loadStart=<s> viaLoaded=<n> viaClean=<n> viaOther=<n> probePdr=<pct> probeDelayMs=<ms>

viaLoaded/viaClean/viaOther count probe data packets by the interface they leave the source on (east / west / off-row), from loadStart onward; probePdr/probeDelayMs are the probe flow's own whole-run FlowMonitor numbers. AntHocNet runs additionally emit a # pher line every 30 s from loadStart — the probe source's regular/virtual pheromone toward each corridor's first hop for the probe destination (# pher <proto> seed=<s> t=<t> eastR=<> eastV=<> westR=<> westV=<>) — the mechanism trace for arms that fail to shift: it shows whether a clean-corridor gradient ever forms at the source and whether diffusion (the V columns) feeds it. Pass criterion: the congestion-aware arm (--ns3::anthocnet::RoutingProtocol::EnableMacMetric=true) moves probe traffic off the loaded corridor (viaClean dominates, or at least the loaded share drops materially vs the blind arms) and its probePdr/probeDelayMs beat the blind control's on the same seeds.

Read it honestly: the background flow is itself routed by the protocol under test — real cross-traffic is — so an adaptive arm may spread the load across both corridors rather than leave it east; and AntHocNet's default wall-clock ant metric also feels queueing delay, so the mac-metric-OFF arm is not fully blind. The truly load-blind references are the hop-count baseline (OLSR — not DSDV, which cannot run here at all, above) and the #216 precomputed control, which has now run beside this cell — see the measured corridor result, where the OLSR reference did real work: it showed that landing on the clean corridor does not require congestion awareness. Judge the cell on the probe's counters and QoS across arms, not on the background's path. With --flows=0 the headline ##RUN##/table row is the probe flow, so the standard pipeline compares the cell without any new parsing.

How to run it¶

Dispatch the manual Satellite benchmark workflow (satellite-benchmark.yml) — the regime's counterpart of the MANET paper-benchmark.yml. Same contract: results file + time -v artifact, a compact tail block cheap to fetch from the job log, extraArgs for ns-3 attribute overrides (e.g. --ns3::anthocnet::RoutingProtocol::EnableDirectedReactive=true for the #244 satellite arm).
Locally, with an ns-3 tree carrying the module: ./ns3 run "isl-grid --rows=4 --cols=4 --runs=3 --protocols=anthocnet,aodv".

Validation gates (run on every PR, 3.42 CI leg)¶

The satellite suite's anchors are analytic, and therefore stricter than the MANET suite's literature-derived ones (methodology): a point-to-point grid with fixed per-link delay has no stochastic channel, so

single-isl: delivery over one ISL must be ≥ sat_single_isl_pdr_min (99.0 — physics says 100);
hop-delay: mean delay must sit within sat_hop_delay_slack_ms (1.5 ms) of the analytic floor hops × islDelayMs;
determinism: the same seed twice must be byte-identical (check-determinism.sh … isl-grid);
arms actually run: check-sat-arms.sh exercises every arm this page advertises, so an unrunnable default cannot stay latent again (#422; the DSDV case above is why it exists).

Values live in ns3/tools/anchors.yml; the gate is ns3/tools/check-sat-anchors.sh (#237/#238). Known open question against the hop-delay floor: the ~0.25 ms excess tracked in #250 — inside the slack, not yet attributed.

Results¶

The suite's committed results are the v1.5.0 phase-3 base-torus cell and the three #432 adversarial cells above — all 20 seeds with the exact oracle control, so above the statistical policy's ≥ 10-run bar. The base cell is a negative result: it does not discriminate, so nothing comparative may be quoted from it — which is also why no interval is attached to a column where four arms report the same number. The corridor cell is the first positive comparative result (quote it only with its mechanism caveat), the failcell separates the arms on its reconvergence instrument while confirming the headline-metric tie, and the seam cell is the suite's second designed negative: static irregularity also ties the floor, closing the static discrimination map. The pipeline that lands and gates results is real (#259) — the same dispatch → rescue → validate → parse loop the MANET suite runs, so no satellite number is ever eyeball-only. When results do land here they fall under the statistical policy: published points need ≥ 10 runs with 95% CIs, except the analytic anchor cells (single-isl, hop-delay), which are derivation checks on deterministic point-to-point links — those stay low-run and are read as diagnostic identities, not estimates (#318 wording pass):

Dispatch satellite-benchmark.yml (above). The results file (satellite-results.txt: ##RUN## per-seed rows plus the summary table) is uploaded as the satellite-benchmark artifact (30-day retention).
Rescue it past expiry with the rescue-artifacts workflow (sat_run_ids input); it is committed as docs/benchmarks/campaign/<runid>-sat.txt.
Validate before reading: python3 .claude/skills/benchmark-results/scenario_check.py results FILE understands both the --csv schema and the human ##RUN##/table output, runs the generic plausibility rules (PDR bounds, delay99 ≥ mean, negatives, dead cells) and adds the satellite invariants — mean delay at or above the one-ISL propagation floor (isl_delay_ms), and the sat_single_isl_pdr_min floor from ns3/tools/anchors.yml on any AODV row whose topology is the 2-node/1-link anchor. A FAIL is a harness bug: do not compare, publish, or quote the numbers.
Parse / A/B with python3 .claude/skills/benchmark-results/bench_parse.py OFF ON — both the text output and --csv rows are accepted; the #244 directed-arm and #250 comparisons are exactly this. Record the verdict + run IDs on the driving issue (ADR-0013).

This suite has no per-merge refresh (the MANET quick taxonomy keeps that job). Do not quote the CI smoke numbers — they are delivery gates at tiny scale, not measurements.

What the suite is waiting on, in dependency order. The control is no longer one of them — #419 built it, #415 measured it, and it runs approx=0 here. What the suite waits on now is a cell where the control's answer and the protocols' answers can differ at all:

Discriminating cells — the adversarial regimes under #216, each re-run with the oracle arm beside it. Three are now measured: asymmetric congestion (#280), scripted ISL breaks (#260), and static irregularity (the #432 item 3 seam cell — measured, and it ties: of the three static instruments, only load moves headline metrics). Still open: the handover cell (no instrument yet) and dynamics/churn generally — subsumed by #297. A satellite comparison published from the base torus would be a comparison of four ties.
#206 — per-next-hop congestion signal, precondition for the congestion cell.
#244 / #245 — the directed reactive A/B and its open steering hazard.
The strategy frame for reading any of it: #192 analysis (detect+reconverge race, path-stretch cost, bootstrap uniformity).