GRAPH-NEURAL-NETWORK-DRIVEN PROACTIVE RESOURCE ALLOCATION FOR CLOUD WORKLOADS: A PREDICTIVE GRAPHMODELING FRAMEWORK

Abstract

Cloud infrastructures face high variability in workload demands, making static or reactive resource allocation inefficient: either resources are over-provisioned (wasting cost) or under-provisioned (violating SLAs). This paper proposes a novel framework that models the cloud infrastructure and workload interactions as a graph, and employs a Graph Neural Network (GNN) to forecast future resource demands for nodes and edges in the graph. By leveraging this prediction, the framework drives a proactive resource allocator that adjusts provisioning and workload placements ahead of time. Experimental results based on realistic workload traces demonstrate that the GNN-based predictor achieves significantly higher accuracy compared to conventional time-series methods, enabling the allocator to reduce cost by X% and SLA violation rate by Y%. The contributions include (1) a graph abstraction of the cloud system and workloads; (2) a spatio-temporal GNN architecture for demand prediction; and (3) a resource allocation engine integrating prediction with provisioning decisions.