Graph Neural Networks for Dynamic Resource Allocation in Cloud-Edge Continuum

Abstract

The infrastructure running modern applications increasingly spans a continuum from hyperscale cloud regions through
regional edge sites to constrained devices, with workloads that must be placed, and repeatedly re-placed, across tiers whose
latency, bandwidth, and load change on timescales of seconds. This article synthesizes the literature on using graph neural
networks (GNNs) to learn dynamic resource allocation policies for that continuum, arguing that the infrastructure's topology,
workload dependency structure, and traffic patterns are naturally graph-shaped in a way flat, fixed-dimensional
representations cannot capture. We trace the technique from foundational message-passing and attention-based GNN
architectures through spatio-temporal and heterogeneous variants built for evolving, mixed-type graphs, to concrete systems,
cluster job schedulers, wireless power-control policies, network performance digital twins, and edge and vehicular
task-offloading controllers, that combine GNN encoders with reinforcement learning to allocate resources across cloud and
edge infrastructure. Read together, the evidence, including double-digit percentage reductions in job completion time and
single-digit-percentage error against packet-level network simulation reported in the cited sources, supports GNN-based
resource allocation as an architecturally well-motivated and increasingly well-evidenced pattern, while leaving full multi-tier
continuum evaluation, continuous-time topology change, and heterogeneous node/edge modeling as a substantially open
research agenda.