WorksComputer Science
INC-SRv6: In-Network Route Exploration and Fast Failover Using Segment Routing Headers in Dynamic UAV Networks
Abstract
Unmanned aerial vehicle (UAV) networks require a network provisioning mechanism that guarantees delivery and latency quality of service (QoS) under dynamic conditions in next-generation airborne mesh networks by continuously learning and repairing multi-hop paths under rapid link turnover without flooding the control plane. This paper presents INC-SRv6, an in-network computing routing mechanism that repurposes the IPv6 Segment Routing Header (SRH) and its Segments Left (SL) pointer into a self-learning, fast-failover distributed control plane. Rather than treating the SRH as a static source-committed vector, INC-SRv6 uses the active SL pointer to partition the segment list in flight: slots ahead of SL provide capacity for hop-by-hop next-hop recording, while the suffix behind SL serves as an in-band multi-destination reachability vector. Through this dual-zone header discipline, transit routers extract distance-weighted reachability updates from arriving suffixes, gate admission by single-pass confidence with bounded binary probing, advance or reverse SL deterministically, and restore pre-transmission SRH state for local failover upon link loss. Simulations across 200-node 3D UAV networks show that under high-speed flight and repeated 20% relay outages, INC-SRv6 improves packet delivery over reactive routing (92.12% versus 89.12%) while cutting 95th-percentile tail delay from 121.26 to 8.84 ms and control packets by 35.8%, with a 4-byte compact uSID stride model reducing SRv6 control bytes by over 52%.
Introduction
Unmanned aerial vehicle (UAV) swarms provide rapidly deployable, infrastructure-less communication backbones for emergency response, disaster relief, and wide-area sensing [1, 2, 3, 4]. In a large-scale multi-hop aerial mesh, end-to-end connectivity depends on intermediate UAV relays whose geometric positions and operational states change continuously during flight. Even when a source–destination pair remains topologically connected through alternative trajectories, high-speed maneuvers or sudden relay outages can sever active forwarding links on the sub-second timescale. Sustaining reliable low-latency communication and statistical quality-of-service (QoS) guarantees in such dynamic non-terrestrial environments [5] requires a routing architecture that discovers feasible multi-hop paths without broadcast storms, refreshes local forwarding state continuously as contacts evolve, and repairs broken next hops locally without stalling active flows.
Existing routing paradigms struggle to satisfy these requirements simultaneously [6, 7, 8]. Proactive link-state protocols disseminate periodic topology advertisements across the swarm, incurring heavy background overhead while leaving computed paths vulnerable to topological staleness between update intervals. Reactive distance-vector protocols suppress idle overhead by discovering routes on demand, but rely on network-wide request flooding and upstream route-error notifications (RERR) that stall source transmission during link breaks and inflate tail latency into the hundred-millisecond regime. Classical source routing records explicit route vectors during discovery, yet commits each data packet at the source to a fixed end-to-end path that frequently breaks in flight under high mobility. Meanwhile, online reinforcement learning schemes using one-hop delay feedback propagate multi-hop reachability too slowly to track rapid 3D link turnover, whereas continuous multi-agent swarm probing generates prohibitive control packet volumes.
To overcome this fundamental tension, we look to a standardized protocol header already present in modern IPv6 data planes: the Segment Routing Header (SRH, RFC 8754) in Segment Routing over IPv6 (SRv6) [9, 10, 11, 12, 13]. In conventional SRv6 deployments, the SRH is treated as a static, read-only instruction list computed by a centralized controller or link-state database [14, 15]: the source preloads an explicit sequence of 128-bit Segment Identifiers (SIDs) into SRH.segment_list, and transit routers merely decrement the Segments Left (SL) pointer to read the next hop. Under rapid UAV churn, such source-locked segment lists quickly become obsolete.
The central insight of this paper is that the standard RFC 8754 SRH structure, through the interaction between SRH.segment_list and the SL pointer, can be repurposed into a dynamic, in-flight state register for in-network routing computation [16, 17]. While preserving the standard forwarding invariant at every hop, the active SL cursor naturally partitions the segment list into two functional regions during packet traversal. Slots ahead of SL (segment_list[1..SL − 1]) provide pre-allocated capacity where transit routers dynamically record selected next-hop SIDs as they decrement SL, rather than following a rigid source-committed route. Simultaneously, the traversed suffix behind SL (segment_list[SL + 1..LastEntry]) preserves the exact sequence of visited routers. Every arriving packet therefore carries an in-band multi-destination reachability vector: for any suffix index , the receiving router observes that node segment_list[k] is reachable via neighbor segment_list[SL + 1] at an exact hop distance of .
Building on this dual-zone SRH representation, we propose INC-SRv6, an in-network computing mechanism that unifies distributed route exploration, multi-destination learning, loop-free forwarding, and local fast failover within the SRv6 data plane. In INC-SRv6, transit routers continuously extract distance-attenuated reachability updates for all intermediate nodes in arriving packet suffixes, amortizing discovery traffic across intersecting flows and pre-populating backup next hops before links break. When a source lacks a confident route, it launches a bounded binary exploration wave whose branches exclude nodes already present in their SRH suffix to guarantee loop freedom, while the destination reverses the compacted SRH array once so that the acknowledgment retraces the discovered path strictly by decrementing SL without routing lookups. During subsequent data forwarding, routers make fresh hop-by-hop decisions using locally learned weights; if an outbound link fails just before transmission, the detecting router rolls back the packet’s pre-hop SRH state to expunge the untraversed hop and immediately rewrites an alternate loop-free next hop, avoiding upstream route-error propagation and source rediscovery stalls.
The main contributions of this work are summarized as follows:
Dual-zone SRH state representation for in-network routing: We formulate a unified SRH state representation and pointer discipline that transforms
SRH.segment_listand SL from a static source-route header into a dual-zone in-band telemetry and control structure supporting hop-by-hop next-hop recording, multi-destination suffix extraction, loop-free candidate filtering, and lookup-free reverse acknowledgments.
Closed-loop exploration, confidence gating, and local failover: We design a self-learning forwarding mechanism that couples single-pass confidence-calibrated reinforcement (), least-probed binary exploration cloning ( leaves per generation), and pre-transmission SRH snapshot rollback () for immediate loop-free local repair.
Large-scale 3D UAV network evaluation: Across 1,100 simulation runs spanning 10 mobility and fault scenarios in 200-node 3D UAV networks, we benchmark INC-SRv6 against five protocol families (OSPF-SRv6, OLSR-L3, AODV-L3, Q-Routing-L3, and AntHocNet-L3), four mechanism ablations, and a 4-byte compact uSID variant, quantifying the impact of each design component and demonstrating a 52.1%–54.7% control-byte reduction via 4-byte compact uSID stride encoding.
The remainder of this paper is organized as follows. § 2 reviews related work. § 3 defines the 3D UAV network model and programmable SRH state representation. § 4 details the INC-SRv6 processing workflow. § 5 presents the simulation evaluation, ablation study, and CPU microbenchmark. § 6 concludes the paper.
Related Work
Table 1 summarizes the routing state representations, discovery mechanisms, and failure recovery strategies of representative routing protocols compared with INC-SRv6.
| Protocol | Routing state | Discovery / updates | Failure handling |
| OSPF-SRv6 | Local link-state database; source SRH | Periodic LSA dissemination | Local alternate from cached link-state paths |
| OLSR-L3 | Two-hop and MPR topology | HELLO and MPR-relayed TC | Neighbor expiry and route recomputation |
| AODV-L3 | Scalar next-hop and precursor entries | RREQ flood; reverse-hop RREP | RERR and source rediscovery |
| DSR | Cached route vectors; source-route header | RREQ flood; route-carrying RREP | Cached-route packet salvaging |
| Q-Routing-L3 | Per-destination/neighbor delay estimates | One-hop feedback | Next-best eligible neighbor |
| AntHocNet-L3 | Pheromone tables and recorded agent paths | Reactive/proactive probes and diffusion | Local fallback and renewed exploration |
| INC-SRv6 | Multi-destination weights from SRH suffixes | Bounded cloning; reverse-path ACK | Pre-hop restoration and local next-hop selection |
Table 1. Comparison of routing-state representations, discovery mechanisms, and failure handling strategies across protocols. Baseline engineering parameters are specified in § 5.
SRv6 Network Programming and Fast Reroute
Segment Routing over IPv6 (SRv6) encodes forwarding instructions as an ordered list of 128-bit Segment Identifiers (SIDs) inside the Segment Routing Header (SRH, Routing Type 4), using the Segments Left (SL) pointer to index the active segment [9, 10, 11, 18]. Beyond static source-routed traffic engineering and service function chaining embedding [12, 15, 19], SRv6 Network Programming allows transit routers to execute lightweight header inspection, in-band performance monitoring, and state manipulation in eBPF or P4 pipelines [20, 21], while recent IETF standards and micro-SID architectures specify compact SID addressing [22, 23] and compressed segment-list encodings (uSID / NEXT-C-SID) [24] to reduce 128-bit header overhead.
For fault recovery and deterministic low-latency delivery [13], IP Fast Reroute mechanisms such as Loop-Free Alternates (LFA) [25], Topology Independent LFA (TI-LFA) [26], and SDN-assisted UAV architectures [14] precompute backup repair paths from a Link-State Database (LSDB) or centralized controller. However, under rapid 3D UAV mobility and simultaneous multi-relay churn, periodic Link-State Advertisements (LSAs) lag behind sub-second link transitions, causing source-computed SRHs and precomputed TI-LFA repair paths to steer packets into broken links. INC-SRv6 eliminates this dependency on global link-state synchronization by repurposing the SRH itself as an in-flight path-recording and learning structure, updating next-hop weights directly from arriving suffixes and repairing broken links locally.
Routing in Dynamic UAV Networks
Flying Ad Hoc Network (FANET) routing protocols span proactive, reactive, and learning-based designs [1, 2, 3, 6]. Proactive protocols such as OLSR [27] use Multipoint Relays (MPRs) to disseminate Topology Control (TC) messages; while providing immediate lookups in quasi-static topologies, periodic TC snapshots suffer severe blackholing when links change faster than the broadcast interval. Reactive protocols such as AODV [28] discover scalar routes on demand via flooded Route Requests (RREQ) and unicast Route Replies (RREP), using upstream Route Errors (RERR) upon link failure to trigger fresh source flooding. Under repeated relay churn, this drop-invalidate-reflood cycle introduces recurrent discovery stalls that inflate tail latency by an order of magnitude.
Dynamic Source Routing (DSR) [29] records traversed node addresses during discovery for route caching and salvaging, but commits each data packet at the source to a fixed end-to-end route vector that frequently breaks in flight as UAV speed and path length grow. INC-SRv6 differs fundamentally from DSR: (i) rather than locking data packets to source-selected vectors, it converts observed SRH suffixes into normalized per-destination weight tables for fresh hop-by-hop decisions; (ii) it bounds unconfident exploration to a binary tree of at most four leaves rather than network-wide flooding; and (iii) Stage 4 restores the pre-transmission SRH snapshot before local rerouting so failed hops never pollute the recorded suffix.
Reinforcement learning and swarm-intelligence protocols replace rigid tables with adaptive estimates [7, 8, 30, 31, 32, 33]. Q-Routing and recent multi-agent FANET extensions [7, 8, 30, 32, 33] update neighbor quality estimates via one-hop feedback, which bootstraps multi-hop reachability by only a single hop per transmission and converges too slowly under 3D mobility without external geographic beacons. AntHocNet [31] accelerates convergence via reactive/proactive ants and neighbor pheromone diffusion, but incurs millions of control packets in 200-node swarms. INC-SRv6 achieves rapid multi-hop convergence without swarm flooding by extracting up to destination updates per arriving SRH suffix and calibrating its single-pass reinforcement step so that one validated acknowledgment immediately opens the source confidence gate.
In-Network Computing and Path-Carried State
In-network computing (INC) shifts state processing and inference into forwarding elements [16, 17, 34], while In-band Network Telemetry (INT) and its low-overhead variants [35, 36] append device state to packets for passive monitoring. INC-SRv6 turns this in-band state paradigm into an active, closed-loop routing control plane built entirely within standard RFC 8754 SRH fields: every packet’s traversed SRH suffix acts as a multi-destination reachability vector, and each router executes a four-stage pipeline that learns from the suffix, gates or branches outbound traffic, updates the SL pointer, and repairs local link failures in place.
System Model and SRv6 Header State Representation
3D Dynamic UAV Network Pure-L3 Model
We consider a pure network-layer (L3) model of airborne SRv6 routers moving inside a 3D operational volume m m m (Figure 1), capturing rapid 3D aerial mobility and link transitions characteristic of emergency UAV deployments [37]. Each UAV follows a continuous, piecewise-linear 3D random waypoint trajectory under nominal speed profiles m/s (sampling per-leg speeds uniformly from , , , and m/s, respectively).

Figure 1. 3D dynamic UAV mesh network scenario ( UAVs in m airspace, m L3 contact range) experiencing a central L3 relay/link outage zone. Bounded EXPLORE packets (split_rem = 2, solid blue arrows, leaf branches) detour around the outage while recording traversed SIDs into SRH.segment_list[SL]. Destination returns a deterministic unicast SRH-ACK (dashed magenta arrows) along IPv6.DA = SRH.segment_list[SL], and high-confidence DATA packets (solid green arrows) forward along the reinforced trajectory.
Connectivity between operational UAVs is governed by a contact range m:
where denotes operational state. Because flight legs are piecewise-linear, exact link up/down boundary times at are solved analytically via quadratic root finding for L3 link-state queries. Each directed link provides a 10 Mbps rate, 20 s processing delay, distance-dependent propagation delay, and a 128-packet FIFO queue. Link availability is checked at both enqueue and dequeue: failure at either check invokes Stage 4 local repair while the packet remains at the sender, whereas a break after dequeue causes an in-transit loss. In practical airborne wireless interfaces, this sender-side check corresponds to hardware transmit-failure callbacks (e.g., missing link-layer acknowledgment or fast BFD/beacon loss notification) while the pre-transmission header snapshot is retained in the transmit ring buffer.
Programmable SRH Path-State Representation and Byte-Accounting Model
INC-SRv6 operates within a trusted, homogeneous airborne SRv6 domain consistent with RFC 8754 [10], where all participating UAV routers implement the domain-local header conventions defined below. Every INC-SRv6 packet carries an outer IPv6 header (40 bytes) followed by a Segment Routing Header (SRH, Routing Type 4, 8-byte fixed header plus populated SID entries), as illustrated in Figure 2. Throughout forwarding, the Segments Left (SL) pointer satisfies the RFC 8754 invariant [10]:

Figure 2. INC-SRv6 in-network SRH and SL pointer dual-zone state mechanism across a three-hop path. (a) Forward EXPLORE/DATA traversal (): relays decrement SL, write the selected next-hop SID into segment_list[SL], compact unused slots upon direct delivery to (see Table 2 for slot indices), and execute Stage 1 upstream route learning from the accumulated suffix ( at hop distances ). (b) Reverse SRH-ACK traversal (): destination reverses the compacted array once, and each returning hop decrements SL deterministically without routing lookups while extracting multi-destination downstream routes from the suffix.
To record paths up to hops, a router allocates a 17-slot logical segment array. A forward packet originated by source toward target places at index 0, at index 16 (LastEntry = 16), and sets SL = 16. As the packet traverses the network, the suffix [SL..16] records visited routers in order, while unpopulated slots [1..SL − 1] provide local capacity for subsequent hops. At each transit relay, the router decrements SL and writes the selected next-hop SID into segment_list[SL]; when the selected next hop is the final target , unused intermediate slots are compacted out and SL is set to 0. Packet mode (EXPLORE, ACK, or DATA) and remaining split budget () occupy domain-local bits of the 8-bit SRH Flags field, while the generation identifier (tracked as a monotonic 32-bit integer in the simulator) maps on the wire via modulo- reduction into the 16-bit SRH.Tag field (65,536 active generations far exceeding in-flight lifetime).
On the wire, only populated SID entries () are serialized and charged bandwidth within the airborne SR domain:
which maps back for forward packets at the receiving router to the logical pointer SL = when (intermediate relay hop) and SL = 0 when (final destination hop). A reverse ACK reverses the compacted segment array once at and decrements SL in place along the return path. The simulator maintains the logical SRH state transitions and charges exact serialized wire bytes and transmission times according to
with bytes for DATA and 0 bytes for EXPLORE/ACK. Standard INC-SRv6 uses 128-bit SIDs ( bytes), whereas INC-SRv6-uSID models a 4-byte compact locator stride ( bytes, abstracting 128-bit carrier boundary padding to isolate per-segment identifier width) [22, 23, 24]. Non-SRv6 baseline data packets carry the 40-byte IPv6 header plus 512-byte payload (552 bytes). Table 2 traces the logical and wire SRH states across a three-hop traversal and its returning acknowledgment .
| Hop | Logical SL | Wire SL | Outbound segment list | Split budget | Bytes 16B / 4B |
| 15 | 1 | 1 | 96 / 60 | ||
| 14 | 1 | 0 | 112 / 64 | ||
| 0 | 0 | 0 | 112 / 64 | ||
| (ACK) | 2 | 2 | 0 | 112 / 64 | |
| (ACK) | 1 | 1 | 0 | 112 / 64 | |
| (ACK) | 0 | 0 | 0 | 112 / 64 |
Table 2. Forward exploration and reverse acknowledgment header states. Byte columns give 128-bit / 4-byte-stride sizes with zero payload.
Proposed INC-SRv6 Four-Stage Packet and SL Pointer Workflow
Overview of the Four-Stage Processing Pipeline
Each INC-SRv6 router maintains a local neighbor set (discovered via periodic link-local multicast HELLO beacons every 1.0 s with a 3.2 s expiration timeout evaluated on maintenance ticks) and a local normalized routing weight table for each known destination and active neighbor , satisfying . In addition, router tracks the learned hop distance , last reinforcement timestamp , and local repair status per destination.
Algorithm 1 gives normal packet processing; Algorithm 2 handles failed egress attempts. Stage 1 runs on arrival, Stage 2 selects forward next hops, Stage 3 writes the selected hop or advances an ACK, and Stage 4 restores and repairs a failed forward transmission. The predicate confident requires and .
Algorithm 1. Packet processing at router
Require: Packet , target , mode , and local routing state
if arrives from a neighbor then
Reconcile neighbors, age accessed rows, and learn suffix (Stage 1)
Release non-expired pending DATA whose routes are confident
end if
if then
if then
Deliver payload
else if and first arrival of this generation then
Reverse path; initialize ACK at SL =
Enqueue ACK to its active hop with no rollback snapshot
end if
return
end if
if then
Decrement SL; enqueue to the indexed hop; return ▷ Stage 3
end if
if and generation already forwarded then
return
end if
If EXPLORE, mark generation forwarded; compute candidates (Stage 2)
if then
Buffer source DATA; drop transit packet; return
end if
if then
if is source and route is unconfident then
Buffer DATA; request cooldown-gated exploration; return
end if
Select best candidate; refresh confident-route activity
If relay is unconfident, request exploration (250 ms cooldown)
else if route is confident and is a transit relay then
Select best candidate without consuming split budget
else
Rank by probe count, rounded weight, and seeded random tie-break
Select two candidates if budget permits, otherwise one
Increment selected probe counts; clone and reduce budget if splitting
end if
for all selected packet/next-hop pairs do
If HopLimit or ( and SL ), drop ; continue
Save pre-hop snapshot ; advance header toward (Stage 3)
Enqueue ; check link at enqueue and dequeue
end for
Algorithm 2. Egress-failure callback (Stage 4)
Require: Failed next hop , packet , target , pre-hop snapshot
Remove ; age and renormalize weights; mark target for repair
if is ACK or is absent then
Drop ; return
end if
Restore the complete snapshot
Recompute loop-free candidates from the restored history
if is EXPLORE then
Set split budget to zero and retain generation identity
if then
Drop ; return
end if
Select best candidate ; if Stage 3 hop/SL bound fails, drop and return
Save snapshot, advance toward (Stage 3), and enqueue checked transmission
else if then
Buffer source DATA; drop relay DATA
else if is the DATA source then
Buffer DATA; originate exploration with cooldown bypass
else
Originate exploration with cooldown bypass; select best candidate
If Stage 3 hop/SL bound fails, drop and return
Save snapshot, advance relay DATA toward (Stage 3), and enqueue transmission
end if
Stage 1: Ingress SRH Suffix Learning (Multi-Hop Path Extraction)
Before looking up or updating destination , router reconciles weight row with its active neighbor set and applies lazy exponential aging toward the uniform prior () over idle interval :
with s, followed by normalization . If has forwarded DATA toward on a confident route within the preceding 250 ms activity window, aging on is suspended and is refreshed. When new neighbors join while holds a confident non-repair route to ( and ), their combined prior share is capped at to preserve the confidence margin.
For each arriving suffix index , let denote the learned destination, denote the immediate upstream sender, and denote the exact hop distance from to along that suffix. Forward EXPLORE and DATA suffixes teach routes toward upstream origins and relays, whereas returning ACK suffixes teach validated routes toward the target destination and downstream relays. Each tuple (generation, mode, ) is learned at most once per router, and an observation is ignored if already holds a confident, non-repair route to with a strictly smaller recorded hop distance than .
When passes deduplication and short-route protection, router reinforces neighbor using a distance-attenuated step size :
and updates the routing weights via convex combination (clearing from once ):
Because for all , a single validated observation elevates strictly above in one pass. For example, from a four-neighbor uniform prior (), a two-hop ACK () applies and raises the reinforced neighbor’s weight to in one RTT (Figure 3b).

Figure 3. Stage 2: Bounded Exploration Cloning and Confidence-Gated Next-Hop Selection. (a) Least-probed bounded binary exploration cloning tree (), restricting each unconfident exploration wave to at most 4 leaf branches while excluding visited suffix nodes. (b) Single-pass confidence gate () crossed in one RTT (illustrated for a representative returning SRH-ACK with , elevating from 0.250 to 0.764; a three-hop return similarly yields and ) to release buffered DATA.
Stage 2: Confidence-Gated Next-Hop Selection and Bounded Cloning
Figure 3 illustrates bounded exploration cloning and the single-pass confidence gate. For any transit packet (SL , target ), Stage 2 forms the loop-free candidate set by excluding all routers already recorded in the traversed suffix :
Router is confident for when and . For DATA packets, an unconfident source enqueues the payload in a per-destination FIFO buffer (bounded at 64 packets with 1.0-s expiration) and triggers an EXPLORE wave (50-ms source cooldown, retried every 250 ms on maintenance ticks); any Stage 1 update that renders confident immediately releases all buffered DATA. Transit relays forward DATA greedily along the highest-weight candidate in and trigger a background EXPLORE probe (250-ms relay cooldown) if locally unconfident.
For EXPLORE packets, each generation is forwarded at most once per relay. A confident transit relay steers the probe along its single best candidate without decrementing the split budget . At the source or an unconfident relay, Stage 2 ranks by ascending probe count , descending weight, and a seeded tie-breaker. When and , the router clones the packet across the top two candidates and decrements the child budget:
With at the source, binary splitting occurs at most twice along any path, bounding each discovery wave to at most leaf branches (Figure 3a).
Stage 3: SL Pointer Advancement and Deterministic SRH-ACK Reversal
For each outbound forward branch with positive HopLimit, Stage 3 saves a pre-transmission snapshot . If , Stage 3 compacts unused slots into and sets . Otherwise, if and (meaning the branch has exhausted the intermediate segment capacity without reaching $d), or if , the router drops the branch to enforce the strict path-depth bound. For a valid intermediate hop ( and ), Stage 3 advances the pointer, decrements HopLimit, and writes the next-hop SID:
When the first EXPLORE branch of a generation reaches destination () with compacted segment array , reverses the array once into
sets and , and transmits the unicast ACK. Each returning hop executes Stage 1 suffix learning and decrements the pointer in place:
until reaching source at (Table 2).
Stage 4: Local Link-Down Failover and Table Aging
When the enqueue or dequeue check at router detects that next hop is unreachable, Stage 4 executes immediate local failover (Algorithm 2 and Figure 4). First, removes from , marks , and renormalizes remaining neighbor weights:

Figure 4. Stage 4: Local Link-Down Failover and Pre-Transmission SRH Snapshot Restoration. (a) Relay detects a broken egress link to primary next hop , removes from , renormalizes surviving weights, and reroutes immediately to loop-free alternate while triggering a background probe (with idle entries aging exponentially at s). (b) Three-step header rollback and rewrite along the logical traversal chain: the failed attempt at had decremented SL from 14 to 13 (Step 1); restoring snapshot resets SL = 14 and expunges unsent hop (Step 2) before rewriting alternate at SL = 13 (Step 3).
Second, for a forward packet carrying snapshot , Stage 4 restores (segment_list, SL, IPv6.DA, and HopLimit) to its exact pre-hop arrival state. Because Stage 3 had already decremented SL and written SID() into the header prior to the link check, restoring prevents the untraversed node from polluting the suffix or consuming an extra segment slot. Third, recomputes the loop-free candidates . A failed EXPLORE branch sets its split budget to zero and advances to the best surviving candidate; a failed DATA packet at a transit relay triggers a cooldown-bypassed EXPLORE probe and immediately re-advances the restored packet toward the highest-weight surviving candidate in (e.g., rerouting from failed hop to alternate in Figure 4).
Performance Evaluation
Pure-L3 Simulation Setup, Baselines, and Timing Definitions
We implement the pure-L3 discrete-event simulator, the proposed INC-SRv6 mechanism, all baseline protocols, and the evaluation pipeline in Python 3.10 (verified by 64 deterministic unit tests; open-source repository details are provided in the Data Availability Statement). Table 3 summarizes the physical, network-layer, and protocol parameters used across all experiments. Each simulation run spans 50.0 s (40.0 s of active traffic generation followed by a 10.0 s drain window) across UAVs. Traffic consists of 40 concurrent constant-bit-rate (CBR) flows between randomly sampled distinct source-destination pairs, each generating 512-byte UDP payloads at 8 packets/s (32.768 kbps payload rate per flow, 125 ms inter-packet interval), with flow start times staggered uniformly in [2.0, 4.0] s. All periodic control messages (HELLO, LSA, TC, and RREQ) are transmitted via link-local L3 multicast, charging one serialization event at the sender’s multicast FIFO and delivering independent copies to all currently connected L3 neighbors. Note that this collision-free pure-L3 multicast abstraction is intentionally conservative in favor of broadcast-heavy baselines (AODV-L3, OLSR-L3, OSPF-SRv6, and AntHocNet-L3): under realistic CSMA/CA wireless contention, network-wide RREQ and LSA/TC broadcasts suffer severe broadcast-storm collisions and contention backoffs, whereas INC-SRv6 relies on bounded unicast exploration ( leaf branches) and passive suffix extraction.
| Category | Parameter | Value / Engineering Description |
| 3D UAV Topology | Number of UAV routers () | 200 airborne SRv6 routers |
| 3D UAV Topology | Operational airspace () | 1500 m 1500 m 300 m (3D Random Waypoint) |
| 3D UAV Topology | L3 communication range () | 220 m (mean geometric node degree –15.2) |
| 3D UAV Topology | Simulation horizon | 40.0 s active traffic + 10.0 s drain (50.0 s total) |
| 3D UAV Topology | Paired random seeds | 10 paired seeds per scenario (seed ) |
| Pure-L3 Forwarding | Directed link capacity () | 10 Mbps point-to-point L3 pipe (zero MAC retry) |
| Pure-L3 Forwarding | Per-hop SRH processing () | 20 s per forwarded packet |
| Pure-L3 Forwarding | Interface FIFO queue () | 128 packets per directed interface |
| Pure-L3 Forwarding | Offered CBR workload | 40 concurrent flows, 8 pkts/s/flow, 512 B payload |
| Pure-L3 Forwarding | Modeled byte accounting | 40 B IPv6 + 8 B SRH base + (16 B SID / 4 B uSID model) |
| INC-SRv6 Parameters | Confidence gate () | 0.51 (strict majority confidence at source) |
| INC-SRv6 Parameters | Reinforcement step () | , (Equation (6)) |
| INC-SRv6 Parameters | Bounded cloning budget () | split_rem = 2 ( leaf branches per exploration wave) |
| INC-SRv6 Parameters | Maximum SRH path depth () | 16 hops (17 logical segment slots in SRH.segment_list) |
| INC-SRv6 Parameters | Table aging & active refresh | s lazy exponential decay; 250 ms active DATA window |
| INC-SRv6 Parameters | Source pending buffer | Max 64 pkts/destination, 1.0 s wait expiration |
Table 3. Pure-L3 3D UAV network simulation parameters and INC-SRv6 configuration.
Five pure-L3 engineering baselines share the common L3 forwarding engine (establishing controlled algorithmic reference baselines within the shared L3 discrete-event engine): (i) OSPF-SRv6 [38] disseminates neighbor LSAs every 2.0 s (batched over a 10-ms window), computes shortest-path source SRH segment lists from its local LSDB, and performs local LFA/TI-LFA style alternate repair upon egress link loss using a cached reverse shortest-path tree excluding the repair node; (ii) OLSR-L3 transmits 1.0-s HELLO beacons and 2.0-s MPR-relayed TC advertisements to maintain hop-by-hop shortest paths; (iii) AODV-L3 discovers scalar next-hop routes via flooded RREQ and unicast RREP messages (8.0-s active route lifetime, 500-ms per-destination request cooldown) and propagates RERR notifications to upstream precursors upon link failure to trigger source rediscovery; (iv) Q-Routing-L3 updates per-destination/neighbor delivery delay estimates via charged one-hop feedback packets (, exploration); and (v) AntHocNet-L3 combines reactive forward/backward ants with 2.0-s proactive path probing, periodic pheromone evaporation (0.9 factor), and neighbor pheromone diffusion.
To isolate the contribution of each design element in our workflow, four mechanism ablations modify exactly one component of full INC-SRv6: INC-SRv6-OriginOnly restricts Stage 1 suffix learning to the packet’s origin node () rather than the full traversed suffix; INC-SRv6-NoClone disables Stage 2 binary exploration cloning (, single-branch probing); INC-SRv6-SmallStep replaces the single-pass reinforcement rule in Equation (6) with a weak step size ; and INC-SRv6-NoFastFailover disables Stage 4 immediate neighbor invalidation, snapshot restoration, and local rerouting upon link-down events, relying instead on periodic HELLO expiration. In addition, INC-SRv6-uSID evaluates the 4-byte compact SID stride ( bytes).
End-to-end packet delivery ratio (PDR) is defined as unique delivered DATA payloads divided by total generated DATA payloads across the 40 flows. End-to-end delay is measured over all delivered DATA packets and includes source pending-buffer waiting time, interface FIFO queuing, 20 s per-hop processing, serialization, and propagation delay. Each table entry reports the arithmetic mean across the 10 paired seeds, where delay is computed per seed and averaged across seeds. Control overhead counts all on-wire control hop transmissions (HELLO, EXPLORE, ACK, LSA, TC, RREQ, RREP, RERR, FANT, BANT, DIFFUSE, and QFEEDBACK) over the 50.0-s simulation horizon.
Table 4 defines the ten evaluation scenarios (S1–S10), covering pure mobility scaling from 15 to 45 m/s (S1–S4), central high-degree relay hub failures (S5–S6), contiguous spatial area blackouts (S7), repeated multi-wave relay churn (S8–S9), and combined high-speed flight with repeated 20% relay churn (S10). Relay failures affect only intermediate transit nodes while keeping all flow endpoints operational. Every protocol and variant executes on the exact same precomputed 3D trajectories, flow schedules, and failure cohorts for each seed, yielding 1, 100 paired simulation runs.
| Scenario ID | Category | Speed Range (m/s) | Relay Outage | Operational / Fault Injection Description |
| S1 | Mobility | 15 ([10, 20]) | 0% | Normal patrol cruise speed, zero node failures |
| S2 | Mobility | 25 ([20, 30]) | 0% | Fast transit cruise speed, zero node failures |
| S3 | Mobility | 35 ([30, ?]) | 0% | High-speed dash maneuver, rapid link turnover |
| S4 | Mobility | 45 ([35, ?]) | 0% | Extreme aerial speed, severe link churn |
| S5 | Hub Outage | 25 ([20, 30]) | 10% (20 relays) | Top-20 highest-degree central relays fail at s |
| S6 | Hub Outage | 25 ([20, 30]) | 20% (40 relays) | Top-40 highest-degree central relays fail at s |
| S7 | Spatial Blackout | 25 ([20, 30]) | 20% (40 relays) | 40 relays nearest a seeded reference UAV position fail at s |
| S8 | Churn | 25 ([20, 30]) | Three 4-s relay outage waves (20 relays each) at s | |
| S9 | Churn | 25 ([20, 30]) | Three 4-s relay outage waves (40 relays each) at s | |
| S10 | Combined | 35 ([30, 40]) | High-speed flight (35 m/s) + three 4-s, 40-relay outage waves |
Table 4. Experimental scenario matrix (S1–S10) for the 200-node 3D UAV network campaign.
Mobility Scaling Analysis (S1–S4)
Figure 5 and Table 5 present the delivery ratio, mean delay, 95th-percentile () tail latency, and control overhead as nominal UAV speed scales from 15 m/s (S1) to 45 m/s (S4). Key performance figures are underlined in Table 5 for ease of comparison. Across high-delivery regimes in Table 5 and Table 6, arithmetic mean delay can slightly exceed delay (e.g., INC-SRv6 in S3–S4 and AODV-L3 in S6–S7): over 95% of delivered payloads traverse established routes within a few milliseconds, whereas fewer than 5% of initial or post-break packets wait hundreds of milliseconds (up to the 1.0-s buffer cap) in the source pending queue during route discovery, pulling up the arithmetic mean without shifting the 95th percentile.
Figure 5. Mobility scaling performance across S1 (15 m/s) to S4 (45 m/s). (a) Mean packet delivery ratio (PDR, %) with 95% confidence intervals across 10 paired seeds. (b) 95th-percentile () end-to-end tail delay (ms, log scale). (c) Total on-wire control packet transmissions (, log scale).
| Scenario | Protocol | PDR (%) | Mean Delay (ms) | Delay (ms) | Control Pkts | Control MB |
| S1 | INC-SRv6 | 94.40 | 8.11 | 8.68 | 81,930 | 13.54 |
| S1 | INC-SRv6-uSID | 94.70 | 7.47 | 7.29 | 77,856 | 6.44 |
| S1 | AODV-L3 | 94.63 | 4.86 | 3.68 | 134,072 | 8.02 |
| S1 | AntHocNet-L3 | 81.25 | 19.19 | 13.58 | 1,723,589 | 159.69 |
| S1 | OSPF-SRv6 | 79.67 | 2.74 | 5.85 | 45,404 | 78.61 |
| S1 | OLSR-L3 | 49.14 | 1.57 | 3.24 | 38,148 | 59.50 |
| S1 | Q-Routing-L3 | 30.60 | 2.50 | 6.15 | 120,672 | 7.68 |
| S3 | INC-SRv6 | 92.28 | 9.89 | 8.85 | 162,155 | 27.61 |
| S3 | INC-SRv6-uSID | 92.33 | 9.17 | 7.49 | 153,329 | 12.51 |
| S3 | AODV-L3 | 89.59 | 14.47 | 113.07 | 260,248 | 15.19 |
| S3 | AntHocNet-L3 | 74.31 | 39.74 | 263.81 | 2,509,042 | 230.87 |
| S3 | OSPF-SRv6 | 41.42 | 2.24 | 5.36 | 45,799 | 88.15 |
| S3 | OLSR-L3 | 21.22 | 1.00 | 1.94 | 39,808 | 67.05 |
| S3 | Q-Routing-L3 | 17.29 | 2.04 | 5.68 | 93,126 | 6.09 |
| S4 | INC-SRv6 | 90.63 | 11.55 | 9.08 | 191,806 | 33.07 |
| S4 | INC-SRv6-uSID | 90.39 | 11.19 | 7.59 | 189,864 | 15.49 |
| S4 | AODV-L3 | 87.81 | 18.79 | 129.37 | 298,515 | 17.37 |
| S4 | AntHocNet-L3 | 71.95 | 46.36 | 328.47 | 2,784,461 | 257.34 |
| S4 | OSPF-SRv6 | 34.51 | 2.05 | 5.20 | 45,997 | 91.01 |
| S4 | OLSR-L3 | 17.39 | 0.86 | 1.71 | 40,236 | 69.13 |
| S4 | Q-Routing-L3 | 15.88 | 1.79 | 5.31 | 83,018 | 5.51 |
Table 5. Mobility scaling performance comparison across S1 (15 m/s), S3 (35 m/s), and S4 (45 m/s) (10-seed arithmetic means; key comparison values are underlined).
| Scenario | Protocol | PDR (%) | Mean Delay (ms) | Delay (ms) | Control Pkts | Control MB | Recovery s (cens./all) |
| S6 | INC-SRv6 | 92.37 | 9.60 | 8.80 | 116,707 | 19.51 | 2.30 s (0/10) |
| S6 | INC-SRv6-uSID | 92.31 | 9.38 | 7.28 | 115,671 | 9.30 | 2.00 s (0/10) |
| S6 | AODV-L3 | 91.25 | 10.44 | 4.22 | 177,524 | 10.31 | 2.10 s (0/10) |
| S6 | AntHocNet-L3 | 76.97 | 30.75 | 185.38 | 1,827,881 | 166.43 | 2.40 s (0/10) |
| S6 | OSPF-SRv6 | 56.72 | 2.50 | 5.71 | 39,811 | 52.75 | 5.56 s (1/10) |
| S6 | OLSR-L3 | 29.91 | 1.20 | 2.54 | 33,317 | 38.57 | 5.00 s (1/10) |
| S6 | Q-Routing-L3 | 25.38 | 2.36 | 6.05 | 104,814 | 6.58 | 5.78 s (1/10) |
| S7 | INC-SRv6 | 91.55 | 10.24 | 8.84 | 121,127 | 20.27 | 2.20 s (0/10) |
| S7 | INC-SRv6-uSID | 91.21 | 9.90 | 7.30 | 120,354 | 9.69 | 2.40 s (0/10) |
| S7 | AODV-L3 | 90.25 | 11.32 | 4.35 | 179,885 | 10.47 | 2.30 s (0/10) |
| S7 | AntHocNet-L3 | 75.64 | 32.93 | 208.07 | 1,799,634 | 163.93 | 3.70 s (0/10) |
| S7 | OSPF-SRv6 | 55.67 | 2.46 | 5.62 | 39,512 | 55.00 | 5.00 s (1/10) |
| S7 | OLSR-L3 | 30.01 | 1.20 | 2.59 | 33,100 | 39.92 | 3.57 s (3/10) |
| S7 | Q-Routing-L3 | 24.26 | 2.26 | 5.87 | 103,373 | 6.52 | 5.75 s (2/10) |
| S10 | INC-SRv6 | 92.12 | 10.57 | 8.84 | 158,619 | 26.87 | 2.17 s (0/30) |
| S10 | INC-SRv6-uSID | 91.78 | 9.88 | 7.46 | 154,342 | 12.51 | 2.30 s (0/30) |
| S10 | AODV-L3 | 89.12 | 16.02 | 121.26 | 247,062 | 14.40 | 2.07 s (0/30) |
| S10 | AntHocNet-L3 | 74.72 | 40.98 | 284.32 | 2,426,164 | 222.77 | 2.17 s (0/30) |
| S10 | OSPF-SRv6 | 39.28 | 2.19 | 5.33 | 43,597 | 76.43 | 5.05 s (11/30) |
| S10 | OLSR-L3 | 20.32 | 0.98 | 1.94 | 37,515 | 57.63 | 3.95 s (9/30) |
| S10 | Q-Routing-L3 | 16.80 | 2.02 | 5.73 | 91,011 | 5.91 | 3.95 s (9/30) |
Table 6. Table 6. Fault resilience and post-outage recovery comparison across 20% hub fault (S6), 20% spatial blackout (S7), and 35 m/s high-speed repeated 20% relay churn (S10) (10-seed arithmetic means; key comparison values are underlined).
The mobility scaling results in Figure 5 and Table 5 highlight two distinct operating regimes:
Moderate Cruise Regime (S1, 15 m/s): Both INC-SRv6-uSID (94.70% PDR; 94.40% for 128-bit INC-SRv6) and reactive AODV-L3 (94.63% PDR) achieve near-ceiling delivery. However, whereas AODV-L3 floods RREQ broadcasts across all 200 nodes upon route expiration (134,072 control packets, 8.02 MB), INC-SRv6 reuses suffixes learned from intersecting flows and restricts unconfident probing to at most four leaf branches, reducing control packets by 41.9% (77,856) and control bytes by 19.7% (6.44 MB) under uSID. Because AODV-L3 pins scalar shortest-hop routes and drops packets upon link loss rather than rerouting locally, its surviving payloads exhibit lower delay (3.68 ms versus 7.29 ms for INC-SRv6-uSID).
High-Speed Maneuver Regime (S2–S4, 25–45 m/s): Rapid link turnover continuously breaks active paths. Under AODV-L3, intermediate link breaks drop in-flight packets, propagate RERR messages upstream, and stall sources behind discovery cooldowns until a new RREQ/RREP flood completes, degrading PDR to 87.81% in S4 while inflating tail delay by over 30 from 3.68 ms (S1) to 113.07 ms (S3) and 129.37 ms (S4). Conversely, INC-SRv6 executes Stage 4 snapshot rollback and local alternate forwarding
at the detecting relay, sustaining 92.28% (S3) and 90.63% (S4) PDR while bounding delay tightly at 8.85–9.08 ms (7.49–7.59 ms for uSID), achieving a > 92% tail-latency reduction alongside 36.4%–41.1% fewer control packets than AODV-L3.
Proactive, Learning, and Swarm Baselines: Proactive OSPF-SRv6 and OLSR-L3 fall from 79.67% and 49.14% PDR in S1 to 34.51% and 17.39% in S4 while consuming 59.50–91.01 MB of control traffic, as 2-s topology advertisements lag behind sub-second link changes. Q-Routing-L3 delivers only 15.88%–30.60% because one-hop feedback propagates too slowly across multi-hop trajectories, whereas AntHocNet-L3 achieves 71.95%–81.25% PDR at the cost of 13.58–328.47 ms delay (168.19–328.47 ms across 25–45 m/s) and 1.72–2.78 million control packets (159.69–257.34 MB).
Resilience Under Relay Hub Faults, Spatial Blackouts, and Repeated Churn (S5–S10)
Figure 6 and Table 6 evaluate protocol resilience across the relay failure scenarios (S5–S10). For each outage event, cohort recovery time measures the time required for two consecutive 1-s payload generation cohorts over a 10-s post-fault start window to recover at least 90% of the pre-fault 4-s baseline PDR (minimum recovery window 2.00 s, observed completion up to 11.0 s after single outages or 8.0 s for the final s wave before the 40-s generation cutoff; reported values are arithmetic means over uncensored events, and the censoring count denotes non-recovery within that window).
Figure 6. Fault resilience and post-outage recovery across S5–S10. (a) Mean packet delivery ratio (PDR, %) with 95% confidence intervals under single hub faults (S5–S6), spatial area blackout (S7), repeated relay churn (S8–S9), and high-speed churn (S10). (b) Post-outage cohort recovery time (s) and fractions of events right-censored at the observation limit across the 30 outage events in S10.
Across all six fault-injection scenarios (S5–S10), INC-SRv6 achieves the highest delivery ratio and recovers from 100% of outage events (0 censored events across 10 outage events per scenario in S5–S7 and 30 outage events per scenario in S8–S10):
Central Hub Outages (S5–S6) and Spatial Blackout (S7): When 20–40 central hubs or 40 contiguous core relays fail simultaneously at s, INC-SRv6 sustains 92.90% (S5), 92.37% (S6), and 91.55% (S7) PDR, outperforming AODV-L3 by +1.12 to +1.54 pp with 33.1%–36.9% fewer control packets. Because border routers already hold alternative next-hop weights learned from prior suffixes, Stage 4 detours in-flight packets around the void immediately and achieves 2.00–2.30 s cohort recovery.
Repeated Relay Churn and High-Speed Stress (S8–S10): Under three 4-s outage waves at 25 m/s (S8–S9), INC-SRv6 sustains 92.79% and 92.74% PDR with 8.81 ms delay. In the combined stress scenario (S10: 35 m/s flight plus three 40-relay outage waves), standard INC-SRv6 improves PDR from 89.12% (AODV-L3) to 92.12% (+3.00 ± 0.90 pp), reduces mean delay from 16.02 to 10.57 ms (9.88 ms for uSID), and suppresses tail delay by an order of magnitude from 121.26 ms (AODV-L3) and 284.32 ms (AntHocNet-L3) to 8.84 ms (7.46 ms for uSID), while recovering across all 30 outage waves (2.17 s mean recovery versus 9–11 censored events for OSPF-SRv6, OLSR-L3, and Q-Routing-L3).
Control Overhead and 4-Byte Compact uSID Identifier-Width Experiment
Figure 7a compares the modeled control byte volume across S1–S4 and S10, while Figure 7b illustrates the three-dimensional Pareto trade-off between delivery ratio, tail latency, and control byte volume in S10.
Figure 7. Control overhead and 4-byte compact uSID stride model efficiency. (a) Modeled control byte volume (MB) across S1–S4 and S10, highlighting the 53.4% byte reduction under the B identifier-width model. (b) Pareto trade-off in S10 between PDR (%), tail delay (ms, log scale), and modeled control byte volume (marker area).
In S10, 128-bit INC-SRv6 sends 35.8% fewer control packets than AODV-L3 (158, 619 versus 247, 062), but incurs higher control byte volume (26.87 versus 14.40 MB) due to 16-byte SID entries. Adopting the 4-byte compact stride (INC-SRv6-uSID, bytes) reduces control volume in S10 by 53.4% to 12.51 MB (13.1% below AODV-L3). Across all ten scenarios, INC-SRv6-uSID cuts SRv6 control bytes by 52.1%–54.7% and trims delay to 7.28–7.59 ms via shorter serialization times. For a four-entry path (Table 2) on a 552-byte IPv6 DATA packet, SRH encapsulation adds 72 bytes (13.0%) with 128-bit SIDs and only 24 bytes (4.35%) with 4-byte uSID, placing INC-SRv6-uSID on the Pareto frontier of high delivery, sub-10-ms tail delay, and low overhead (Figure 7b).
Ablation Study of the Four-Stage Workflow
Figure 8 and Table 7 report paired differences ( with 95% Student- CIs and Control Bytes) between full INC-SRv6 and its four mechanism ablations plus the compact-stride uSID variant.
Figure 8. Ablation study of the four-stage INC-SRv6 workflow. (a) Paired PDR advantage (percentage points 95% CI) of full INC-SRv6 over the four mechanism ablations across S1, S3, S4, S7, and S10. (b) Absolute PDR (%) and tail delay (ms) across all six INC-SRv6 variants in S10.
| Scenario | Ablated Variant | Variant PDR (%) | ΔPDR pp ± 95% CI | Variant (ms) | ΔControl Bytes (MB) |
| S1 | OriginOnly | 88.68 | +5.72 ± 0.67 | 9.45 | −7.80 |
| S1 | NoClone | 90.75 | +3.65 ± 0.74 | 9.15 | +1.96 |
| S1 | SmallStep | 31.55 | +62.85 ± 3.18 | 513.10 | −44.56 |
| S1 | NoFastFailover | 45.05 | +49.36 ± 1.03 | 32.77 | +6.90 |
| S1 | uSID | 94.70 | −0.30 ± 0.81 | 7.29 | +7.09 |
| S3 | OriginOnly | 85.75 | +6.53 ± 0.60 | 9.63 | −14.64 |
| S3 | NoClone | 87.53 | +4.75 ± 0.39 | 9.58 | +4.09 |
| S3 | SmallStep | 27.26 | +65.02 ± 2.44 | 655.25 | −34.19 |
| S3 | NoFastFailover | 26.25 | +66.03 ± 1.20 | 605.31 | +21.42 |
| S3 | uSID | 92.33 | −0.05 ± 0.45 | 7.49 | +15.11 |
| S4 | OriginOnly | 83.59 | +7.04 ± 0.67 | 10.07 | −15.78 |
| S4 | NoClone | 85.75 | +4.88 ± 0.68 | 9.77 | +4.97 |
| S4 | SmallStep | 27.96 | +62.66 ± 2.58 | 694.39 | −30.02 |
| S4 | NoFastFailover | 23.96 | +66.67 ± 1.00 | 668.53 | +26.94 |
| S4 | uSID | 90.39 | +0.24 ± 0.52 | 7.59 | +17.58 |
| S7 | OriginOnly | 85.74 | +5.81 ± 0.84 | 9.53 | −9.95 |
| S7 | NoClone | 87.39 | +4.16 ± 0.53 | 9.39 | +3.16 |
| S7 | SmallStep | 32.89 | +58.66 ± 2.83 | 617.53 | −35.19 |
| S7 | NoFastFailover | 33.36 | +58.19 ± 2.08 | 390.20 | +14.10 |
| S7 | uSID | 91.21 | +0.34 ± 0.50 | 7.30 | +10.58 |
| S10 | OriginOnly | 85.72 | +6.40 ± 0.82 | 9.62 | −14.49 |
| S10 | NoClone | 87.64 | +4.48 ± 0.46 | 9.64 | +3.39 |
| S10 | SmallStep | 29.16 | +62.95 ± 2.28 | 675.39 | −33.63 |
| S10 | NoFastFailover | 25.80 | +66.32 ± 1.35 | 627.01 | +20.87 |
| S10 | uSID | 91.78 | +0.33 ± 0.30 | 7.46 | +14.36 |
Table 7. Paired ablation study across representative scenarios (ΔPDR and ΔControl Bytes report paired INC-SRv6 (Full) − Comparator differences across 10 paired seeds; negative ΔControl Bytes indicates that full INC-SRv6 incurs lower control overhead than the ablated variant; key ablation impacts are underlined).
The ablation results verify that all four stages are essential and mutually reinforcing:
Stage 1 Full-Suffix Learning (OriginOnly): Learning only the origin node reduces PDR by 5.10–7.04 pp (+6.40 ± 0.82 pp in S10) while increasing control bytes by 7.80–15.78 MB (saving 14.49 MB in S10, i.e., ΔControl Bytes = −14.49 MB), proving that intermediate suffix extraction both pre-populates backup routes and amortizes discovery across flows.
Stage 2 Bounded Cloning (NoClone): Disabling binary splitting () reduces PDR by 3.59–4.88 pp (+4.48 ± 0.46 pp in S10), as single-branch probes stall in local topological pockets whereas ≤ 4-leaf cloning explores diverse corridors at modest cost (1.96–4.97 MB).
Single-Pass Confidence Calibration (SmallStep): Using collapses PDR to 27.26%–32.89% (+62.95 ± 2.28 pp gap in S10) and inflates delay to 513.10–694.39 ms because a single multi-hop ACK (, ) cannot open the source admission gate from a uniform prior.
Stage 4 Fast Failover (NoFastFailover): Disabling immediate egress-failure recovery drops PDR to 45.05% in S1 and 23.96%–25.80% in S4/S10 (+66.32 ± 1.35 pp in S10, delay 627.01 ms), confirming that snapshot-restored local rerouting is the primary driver of high-churn resilience.
SRH Processing Pipeline CPU Microbenchmark
Table 8 reports per-packet CPU execution times on an x86-64 server (48 logical CPUs, Python 3.10.13, 7 × 20,000 operations via time.process_time_ns()). Median times are 10.84 s (92.26 kpps) for ACK processing (Stages 1 and 3), 33.87 s (29.52 kpps) for exploratory forwarding with Stage 2 cloning (Stages 1–3), and 52.24 s (19.14 kpps) for link-down failover with snapshot rollback (Stages 1–4), confirming that per-packet SRH processing requires only tens of microseconds even in pure Python.
| Operation | Stages | Median CPU Time (ns/pkt) | Pure-Python Rate (kpps) |
| ACK processing | 1 + 3 | 10,839.4 | 92.26 |
| Forwarding | 1–3 | 33,870.9 | 29.52 |
| Failure branch | 1–4 | 52,242.3 | 19.14 |
Table 8. Per-packet CPU execution microbenchmark of the INC-SRv6 processing pipeline on the benchmark host (pure Python 3.10 reference implementation, 7 repeats × 20,000 packets).
Discussion and Conclusions
This paper presented INC-SRv6, an in-network computing mechanism that repurposes the standard IPv6 Segment Routing Header (SRH) and its SL pointer into a distributed route exploration, multi-destination learning, and fast-failover control plane for dynamic UAV networks. By partitioning SRH.segment_list around the active SL pointer into pre-allocated forward slots and a traversed path suffix, transit routers extract multiple destination/next-hop observations from every arriving packet, gate admission by single-pass confidence () with bounded binary probing ( leaves), reverse the SRH array at the destination for lookup-free ACK return, and restore pre-transmission SRH snapshots () for immediate loop-free local repair upon link failure.
Across 1,100 pure-L3 simulation runs in 200-node 3D UAV networks, INC-SRv6-uSID matches the delivery of reactive AODV-L3 at 15 m/s cruise speed (94.70% versus 94.63%) while cutting control packets by 41.9% and control bytes by 19.7%. Under 35 m/s high-speed flight with repeated 20% relay churn (S10), INC-SRv6 improves delivery from 89.12% to 92.12%, reduces tail delay by an order of magnitude from 121.26 to 8.84 ms (7.46 ms with uSID), and achieves 100% post-outage cohort recovery, while the 4-byte uSID encoding reduces SRv6 control byte volume by 52.1%–54.7% across all scenarios. In practical wireless deployments, Stage 4 local failover can be gated by a small link-layer retry threshold (e.g., triggering neighbor removal and SRH snapshot rollback after consecutive short MAC retry failures rather than a single lost frame) to filter transient multipath fading without sacrificing millisecond-scale repair. Future work will integrate cross-layer IEEE 802.11ax/be MAC link-quality hints, implement C-SID wire serialization in eBPF/XDP, and validate INC-SRv6 on a hardware-in-the-loop UAV testbed.
Author Contributions: Conceptualization, Z.R.; methodology, Z.R. and Z.L.; software, Z.R.; validation, Z.R. and Z.L.; formal analysis, Z.R.; investigation, Z.R.; data curation, Z.R.; writing—original draft preparation, Z.R.; writing—review and editing, Z.R.; visualization, Z.R. All authors have read and agreed to the published version of the manuscript.
Funding: This work was supported by the National Key Research and Development Program of China (No. 2024YFE0200302).
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: All source code for the pure-L3 discrete-event simulator, the INC-SRv6 protocol implementation, the 64 deterministic unit tests, the CPU microbenchmark, and the 1,100-run simulation dataset are openly available at https://github.com/Computing-in-Network/INC-SRv6.
Conflicts of Interest: The authors declare no conflicts of interest.
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
References
- [1]Gupta, L.; Jain, R.; Vaszkun, G. Survey of Important Issues in UAV Communication Networks. IEEE Communications Surveys & Tutorials 2016, 18, 1123–1152. https://doi.org/10.1109/COMST.2015.2495297.
- [2]Jiang, J.; Han, G. Routing Protocols for Unmanned Aerial Vehicles. IEEE Communications Magazine 2018, 56, 58–63. https://doi.org/10.1109/MCOM.2017.1700326.
- [3]Lakew, D.S.; Sa’ad, U.; Dao, N.N.; Na, W.; Cho, S. Routing in Flying Ad Hoc Networks: A Comprehensive Survey. IEEE Communications Surveys & Tutorials 2020, 22, 1071–1120. https://doi.org/10.1109/COMST.2020.2982452.
- [4]Yao, Z.; Cheng, W.; Zhang, W.; Zhang, T.; Zhang, H. The Rise of Next-Generation Wireless Technologies for Emergency Wireless Communications in Harsh Environments. IEEE Network 2022, 36, 28–37. https://doi.org/10.1109/MNET.001.2100691.
- [5]Wang, J.; Cheng, W.; Zhang, W.; Liang, H. Statistical QoS Provisioning Architecture for 6G Satellite-Terrestrial Integrated Networks. Journal of Communications and Information Networks 2024, 9, 34–42. https://doi.org/10.23919/JCIN.2024.10494937.
- [6]Khan, M.F.; Yau, K.L.A.; Noor, R.M.; Imran, M.A. Routing Schemes in FANETs: A Survey. Sensors 2020, 20, 38. https://doi.org/10.3390/s20010038.
- [7]Alam, M.M.; Moh, S. Survey on Q-Learning-Based Position-Aware Routing Protocols in Flying Ad Hoc Networks. Electronics 2022, 11, 1099. https://doi.org/10.3390/electronics11071099.
- [8]Zhang, L.; Ma, X.; Zhuang, Z.; Xu, H.; Sharma, V.; Han, Z. Q-Learning Aided Intelligent Routing With Maximum Utility in Cognitive UAV Swarm for Emergency Communications. IEEE Transactions on Vehicular Technology 2023, 72, 3707–3723. https://doi.org/10.1109/TVT.2022.3221538.
- [9]Filsfils, C.; Previdi, S.; Ginsberg, L.; Decraene, B.; Litkowski, S.; Shakir, R. Segment Routing Architecture. RFC 8402, Internet Engineering Task Force (IETF), 2018. https://doi.org/10.17487/RFC8402.
- [10]Filsfils, C.; Dukes, D.; Previdi, S.; Leddy, J.; Matsushima, S.; Voyer, D. IPv6 Segment Routing Header (SRH). RFC 8754, Internet Engineering Task Force (IETF), 2020. https://doi.org/10.17487/RFC8754.
- [11]Filsfils, C.; Camarillo, P.; Leddy, J.; Voyer, D.; Matsushima, S.; Li, Z. Segment Routing over IPv6 (SRv6) Network Programming. RFC 8986, Internet Engineering Task Force (IETF), 2021. https://doi.org/10.17487/RFC8986.
- [12]Ventre, P.L.; Salsano, S.; Polverini, M.; Cianfrani, A.; Abdelsalam, A.; Filsfils, C.; Camarillo, P.; Clad, F. Segment Routing: A Comprehensive Survey of Research Activities, Standardization Efforts, and Implementation Results. IEEE Communications Surveys & Tutorials 2021, 23, 182–221. https://doi.org/10.1109/COMST.2020.3036826.
- [13]Polverini, M.; Cianfrani, A.; Caiazzi, T.; Scazzariello, M. SRv6 Meets DetNet: A New Behavior for Low Latency and High Reliability. IEEE Journal on Selected Areas in Communications 2025, 43, 448–458. https://doi.org/10.1109/JSAC.2025.3528806.
- [14]Secinti, G.; Darian, P.B.; Canberk, B.; Chowdhury, K.R. SDNs in the Sky: Robust End-to-End Connectivity for Aerial Vehicular Networks. IEEE Communications Magazine 2018, 56, 16–21. https://doi.org/10.1109/MCOM.2017.1700456.
- [15]Cianfrani, A.; Listanti, M.; Polverini, M. Incremental Deployment of Segment Routing into an ISP Network: A Traffic Engineering Perspective. IEEE/ACM Transactions on Networking 2017, 25, 3146–3160. https://doi.org/10.1109/TNET.2017.2731419.
- [16]Sapio, A.; Abdelaziz, I.; Aldilaijan, A.; Canini, M.; Kalnis, P. In-Network Computation is a Dumb Idea Whose Time Has Come. In Proceedings of the 16th ACM Workshop on Hot Topics in Networks (HotNets), 2017, pp. 150–156. https://doi.org/10.1145/3152434.3152461.
- [17]Ports, D.R.K.; Nelson, J. When Should the Network Be the Computer? In Proceedings of the 17th Workshop on Hot Topics in Operating Systems (HotOS), 2019, pp. 209–215. https://doi.org/10.1145/3317550.3321439.
- [18]Abdelsalam, A.; Clad, F.; Filsfils, C.; Salsano, S.; Siracusano, G.; Veltri, L. Implementation of Virtual Network Function Chaining Through Segment Routing in a Linux-Based NFV Infrastructure. In Proceedings of the 2017 IEEE Conference on Network Softwarization (NetSoft), 2017, pp. 1–5. https://doi.org/10.1109/NETSOFT.2017.8004208.
- [19]Zheng, S.; Ren, Z.; Cheng, W.; Zhang, H. Performance Analysis for Subgraph Isomorphism Based Embedding Service Function Chains. IEEE Transactions on Network Science and Engineering 2023, 10, 3099–3113. https://doi.org/10.1109/TNSE.2023.3247625.
- [20]Xhonneux, M.; Duchene, F.; Bonaventure, O. Leveraging eBPF for Programmable eNetwork Functions with IPv6 Segment Routing. In Proceedings of the 14th International Conference on emerging Networking Experiments and Technologies (CoNEXT), 2018, pp. 67–72. https://doi.org/10.1145/3281411.3281426.
- [21]Scarpitta, C.; Sidoretti, G.; Mayer, A.; Salsano, S.; Abdelsalam, A.; Filsfils, C. High Performance Delay Monitoring for SRv6-Based SD-WANs. IEEE Transactions on Network and Service Management 2024, 21, 1067–1081. https://doi.org/10.1109/TNSM.2023.3300151.
- [22]Krishnan, S. Segment Routing over IPv6 (SRv6) Segment Identifiers in the IPv6 Addressing Architecture. RFC 9602, Internet Engineering Task Force (IETF), 2024. https://doi.org/10.17487/RFC9602.
- [23]Tulumello, A.; Mayer, A.; Bonola, M.; Lungaroni, P.; Scarpitta, C.; Salsano, S.; Abdelsalam, A.; Camarillo, P.; Dukes, D.; Clad, F.; et al. Micro SIDs: A Solution for Efficient Representation of Segment IDs in SRv6 Networks. IEEE Transactions on Network and Service Management 2023, 20, 774–786. https://doi.org/10.1109/TNSM.2022.3205265.
- [24]Cheng, W.; Filsfils, C.; Li, Z.; Decraene, B.; Clad, F. Compressed SRv6 Segment List Encoding. RFC 9800, Internet Engineering Task Force (IETF), 2025. https://doi.org/10.17487/RFC9800.
- [25]Atlas, A.; Zinin, A. Basic Specification for IP Fast Reroute: Loop-Free Alternates. RFC 5286, Internet Engineering Task Force (IETF), 2008. https://doi.org/10.17487/RFC5286.
- [26]Bashandy, A.; Litkowski, S.; Filsfils, C.; Francois, P.; Decraene, B.; Voyer, D. Topology Independent Fast Reroute Using Segment Routing. RFC 9855, Internet Engineering Task Force (IETF), 2025. https://doi.org/10.17487/RFC9855.
- [27]Clausen, T.; Jacquet, P. Optimized Link State Routing Protocol (OLSR). RFC 3626, Internet Engineering Task Force (IETF), 2003. https://doi.org/10.17487/RFC3626.
- [28]Perkins, C.; Belding-Royer, E.; Das, S. Ad hoc On-Demand Distance Vector (AODV) Routing. RFC 3561, Internet Engineering Task Force (IETF), 2003. https://doi.org/10.17487/RFC3561.
- [29]Johnson, D.; Hu, Y.; Maltz, D. The Dynamic Source Routing Protocol (DSR) for Mobile Ad Hoc Networks for IPv4. RFC 4728, Internet Engineering Task Force (IETF), 2007. https://doi.org/10.17487/RFC4728.
- [30]Boyan, J.A.; Littman, M.L. Packet Routing in Dynamically Changing Networks: A Reinforcement Learning Approach. In Proceedings of the Advances in Neural Information Processing Systems 6 (NIPS), 1993, pp. 671–678.
- [31]Di Caro, G.; Ducatelle, F.; Gambardella, L.M. AntHocNet: An Adaptive Nature-Inspired Algorithm for Routing in Mobile Ad Hoc Networks. European Transactions on Telecommunications 2005, 16, 443–455. https://doi.org/10.1002/ett.1062.
- [32]da Costa, L.A.L.F.; Kunst, R.; de Freitas, E.P. Q-FANET: Improved Q-learning based routing protocol for FANETs. Computer Networks 2021, 198, 108379. https://doi.org/10.1016/j.comnet.2021.108379.
- [33]Ke, Y.; Huang, K.; Qiu, X.; Song, B.; Xu, L.; Yin, J.; Yang, Y. Distributed Routing Optimization Algorithm for FANET Based on Multiagent Reinforcement Learning. IEEE Sensors Journal 2024, 24, 24851–24864. https://doi.org/10.1109/JSEN.2024.3415127.
- [34]Zhang, X.; Cui, L.; Tso, F.P.; Li, W.; Jia, W. IN3: A Framework for In-Network Computation of Neural Networks in the Programmable Data Plane. IEEE Communications Magazine 2024, 62, 96–102. https://doi.org/10.1109/MCOM.001.2300587.
- [35]Kim, C.; Sivారaman, A.; Katta, N.; Bas, A.; Dixit, A.; Wobker, L.J. In-Band Network Telemetry via Programmable Dataplanes. In Proceedings of the ACM SIGCOMM 2015 Conference (Industrial Demo), 2015, pp. 1–2.
- [36]Qian, M.; Cui, L.; Tso, F.P.; Deng, Y.; Jia, W. OffsetINT: Achieving High Accuracy and Low Bandwidth for In-Band Network Telemetry. IEEE Transactions on Services Computing 2024, 17, 1072–1083. https://doi.org/10.1109/TSC.2023.3323697.
- [37]Chen, Y.; Cheng, W.; Zhang, W. Reconfigurable Intelligent Surface Equipped UAV in Emergency Wireless Communications: A New Fading-Shadowing Model and Performance Analysis. IEEE Transactions on Communications 2024, 72, 1821–1834. https://doi.org/10.1109/TCOMM.2023.3336223.
- [38]Coltun, R.; Ferguson, D.; Moy, J.; Lindem, A. OSPF for IPv6. RFC 5340, Internet Engineering Task Force (IETF), 2008. https://doi.org/10.17487/RFC5340.



