When Least-Loaded Placement Backfires: Dispatch Granularity and Deadline Attainment in Multi-RSU Vehicular Edge Computing
Year: 2026
Type: preprint
Abstract: Roadside units (RSUs) that accept offloaded vehicular tasks must decide where each admitted task executes. "Send it to the least-loaded RSU" is a common heuristic in vehicular edge computing (VEC), yet how the rule is implemented is rarely specified. We show that the implementation can reverse the conclusion. In a trace-driven evaluator of a count-matched Manchester road network with nine to twelve RSUs, a frozen multi-agent reinforcement-learning offloading policy and a backlog-based deadline gate, selecting one least-workload RSU per scheduling substep lowers offered-task deadline attainment relative to executing at the radio-ingress RSU, whereas reselecting after every reservation raises it. We prove that the common-target rule feeds at most min(K,R) of R RSUs per batch of K substeps, concentrates work on a deterministic low-index prefix, and caps the work the gate can admit ahead of short-deadline tasks; the per-task rule keeps workloads within one task of each other and gate-rejects only when every RSU fails the gate. A reduced-order model with no fitted parameters reproduces the evaluator's common-target destination shares to within 0.2 percentage points (pp). Across 248 full-horizon runs in five scenarios, with seed-level pairing and adjusted intervals, per-task placement exceeded ingress by 3.7–4.4 pp in four scenarios and 0.53 pp in a saturated incident hour, while common-target placement lost 2.1–3.6 pp. Causal round-robin recovered about 85% of the gain; workload awareness added 0.43–0.82 pp that grew when service was halved. Dispatch granularity and reservation visibility should be specified as part of any VEC placement algorithm.
Institution: University of Manchester
Language: English
Topic: IoT and Edge/Fog Computing
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236654481
