Skip to content

ADR-0010: Data forwarding — prev-hop-excluded stochastic, flow-stickiness gated

  • Status: Accepted — implementation tracked in item 10 (A1)
  • Date: 2026-06-25

Context

Data packets are forwarded by an independent stochastic draw over regular pheromone at every hop (onDataPacket(dest) → table_.lookup(dest, rng_)). Two consequences:

  • No prev-hop exclusion — a packet can be drawn back toward the neighbour it just arrived from, so transient routing loops are possible and only the IP TTL breaks them.
  • Per-packet reordering — independent draws scatter a single flow across the path mesh, which is harmless for UDP/CBR (the benchmark traffic) but hurts TCP (dup-ACKs, spurious retransmits).

The spec spreads data stochastically over a mesh but assumes the pheromone gradient is effectively loop-free, and real implementations additionally avoid sending a packet back the way it came and damp per-packet flapping. Note the overlap with β_data (item 01): a high greedy β_data already concentrates data on the best path and only spreads when pheromones are comparable, so per-packet reordering is naturally limited in the paper's regime.

Decision

  • Prev-hop exclusion is always on (not gated). Thread the incoming neighbour into the core (onDataPacket(dest, prevHop) → lookup(dest, …, exclude=prevHop)) and skip it when forming the probability sum and selecting — unless it is the only next hop with pheromone, in which case forward to it anyway (no black-holes at leaf nodes). This is pure loop-safety, so it is unconditional.
  • Per-flow stickiness is config-gated, default OFF. When enabled, a bounded (src,dst) → next-hop cache with hysteresis keeps a burst on one path; it is a core concern (routing policy, golden rule #2), bounded like the (src,seq) dedup history, and requires the adapter to pass the packet src (onDataPacket(dest, prevHop, src)). Default off because (1) high β_data already makes data ~sticky, (2) pure per-packet stochastic is the paper-faithful baseline the benchmarks should measure first, and (3) it only helps reorder-sensitive (TCP) traffic, which the current harness doesn't carry. It becomes another benchmark ablation knob (mirrors ADR-0007).

Alternatives considered

  • Stickiness on by default. TCP-friendliness and smoother delay out of the box, but departs from the paper's pure-stochastic spreading and muddies the "canonical AntHocNet" story until a benchmark with reorder-sensitive traffic justifies it. Rejected as a default; available via the flag.
  • No prev-hop exclusion (rely on TTL). Simplest, but accepts transient loops and the wasted transmissions they cause. Rejected — loop-safety is cheap and unconditional.

Consequences

  • The core data entry point gains prevHop (always) and src (when stickiness is compiled/used); both adapters supply them (NS-3 RouteInput idev→sender and origin; NS-2 ch->prev_hop_ and ih->saddr()).
  • Loop-safety no longer depends on TTL; the only-option fallback preserves reachability at leaf nodes.
  • One more benchmark ablation axis (sticky vs pure-stochastic), consistent with the gate-and-measure pattern used for diffusion.
  • No wire-format impact: prev-hop and flow identity are local forwarding inputs, not serialized ant fields (no kWireVersion bump — ADR-0006/0009).