[benchmark][ns3] ISL-grid scenario — routable satellite topology with no orbital mobility

[danieljoppi]

Child of #196. The cheapest useful step in the whole satellite track (#192) — it needs no substrate decision and no orbital mechanics.

Context

The standard abstraction of a LEO ISL mesh is a +Grid torus: each satellite holds two intra-plane links (fore/aft, near-constant length) and two cross-plane links (port/starboard, length varying with latitude). Frozen at one instant, that is just a static torus of point-to-point links — buildable in stock ns-3 today.

That is enough to answer most of the questions this track is actually blocked on:

• does AntHocNet route a 4-degree, high-diameter, non-random topology at all?
• does the multi-interface fix ([ns3] Data path always uses the first interface's output device — breaks any multi-interface node #203) hold on a node with exactly the 4 ISLs it was written for?
• what does control-plane cost look like as n grows ([spike][protocol] Control-plane cost at constellation scale — does reactive flooding survive 1584 nodes? #207)?

None of that needs a satellite to move. Doing this first means the substrate spikes (#193) are evaluating options against a scenario we already understand, instead of the other way round.

Scope

An ns-3 example building an M × N +Grid torus of p2p links with configurable per-link delay, running the existing traffic/metric plumbing (anthocnet-compare-style FlowMonitor metrics, --diag, CSV output). Configurable: grid dimensions, ISL delay (uniform to start; per-link later), flow count and placement, and which protocols to run.

Keep it a scenario, not a module — no new devices, no mobility. If it grows either, it belongs in #195.

Acceptance

A 10×10 torus (100 nodes, 4 ISLs each) delivers end-to-end under AntHocNet and under at least one baseline; the run emits a standard CSV row; scaling the grid is a command-line flag. Run through the benchmark-results skill so the numbers get a noise verdict rather than an eyeball.

Related

#196 (parent), #203 (this is its natural test bed), #207 (consumes this for the scale numbers), #200 (its two-node delay-mutation script extends into per-link ISL delay here), #192.

[danieljoppi]

Implemented in 427c0d0 on claude/satellite-networking-simulation-dsjhay. CI: run 30190452648 — 13/13 jobs green, and the new ISL-grid smoke ran on every NS-3 leg (3.36, 3.41, 3.42, 3.47, 3.48).

Result (4×4 torus, 3.42 leg)

##RUN## 1 anthocnet 100.00 10.39 11.00 4.10 12.18 19.48 0.00
protocol,runs,rows,cols,nodes,links,isl_delay_ms,flows,pdr_pct,delay_ms,delay99_ms,throughput_kbps,nrl,nrl_bytes,jitter_ms
anthocnet,1,4,4,16,32,5.0,2,100.0,10.4,11.0,4.10,12.180,19.480,0.00
AntHocNet ISL-grid PDR% = 100.0

PDR 100.0% — AntHocNet routes a 4-degree torus where every node carries 4 ISLs on 4 separate /30s. That is the first of the three questions on this ticket answered, and the first evidence at real satellite shape for the #203 fix (its own test case only has a 2-interface relay).

Two independent sanity checks that the scenario is measuring what it claims:

• Topology. 16 nodes, 32 links — exactly 2 links contributed per node, i.e. degree 4, which is what a +Grid torus must produce. A construction bug would show here.
• Delay matches the analytic value. Both flows (0→15 and 1→14) are 2 hops on a 4×4 torus — the wrap makes opposite corners near neighbours, distance min(3, 4−3) = 1 per dimension. At 5 ms per ISL that predicts 10 ms, and the measured mean is 10.39 ms, the remainder being serialisation. Jitter 0.00 as expected on a static topology with no contention.

What this hands to the sibling tickets

NRL 12.18 — twelve routing-control packets per delivered data packet, on a topology with no churn whatsoever to discover. That is the hello cost predicted in #204: 16 nodes × 4 point-to-point interfaces × 1 Hz, each hello reaching exactly one known, permanently-present peer. It is a first concrete datapoint for #207, and it is measured at n=16 — the interesting question is the slope, which is now a loop over --rows/--cols.

Note this is a 30 s run at low offered load, so NRL is flattered against the control plane by the small data denominator; do not read it as a constellation-scale figure. #207 should set the load deliberately.

Scope notes

• Separate example rather than a --topology flag on anthocnet-compare. Its RunOne is the 420-line wifi/mobility function producing the numbers in [epic] First fully paper-compliant version (v1.0): ledger, ordering, and open gaps #91's ledger; branching it for a p2p topology risks perturbing those for no benefit here. Cost is ~60 duplicated lines of FlowMonitor accumulation — [benchmark] Precomputed-shortest-path control + satellite scenario class in the taxonomy #216 adds the precomputed-shortest-path control to this same harness and is the natural point to extract the shared half.
• dOff50/dOff90 are not computed here, deliberately. They exist to expose survivorship bias under mobility-driven route loss, which a static topology does not have. Omitted rather than zero-filled so nothing downstream mistakes a placeholder for a measurement.
• A CI smoke step was added although this ticket did not ask for one: without it the acceptance criterion ("delivers end-to-end") would be unverifiable, since an NS-3 build needs a simulator tree this repo does not vendor. It reuses the existing e2e gate's loose "must deliver something" assertion and costs ~1 s per leg. point-to-point was already in CI's module list, so no configure change was needed.

Acceptance status

• a torus delivers end-to-end under AntHocNet
• the run emits a standard CSV row
• scaling the grid is a command-line flag (--rows, --cols)
• under at least one baseline — not yet run. AODV/OLSR/DSDV are wired up (--protocols=anthocnet,aodv) but only AntHocNet is exercised in the smoke, to keep per-leg CI cost down. The baseline comparison belongs with the control row in [benchmark] Precomputed-shortest-path control + satellite scenario class in the taxonomy #216 rather than as a CI gate.

Remaining on this ticket is therefore the baseline row; everything else is done and green.

---

Generated by Claude Code