LEARNING-BASED DYNAMIC RESOURCE ALLOCATION FOR SERVERLESS COMPUTING WITH GRAPH NEURAL NETWORKS

Serverless computing has emerged as a transformative paradigm in cloud infrastructure, offering dynamic resource provisioning and pay-per-use economics that significantly reduce operational overhead for application developers. However, the inherent challenges of serverless architectures, including unpredictable workload patterns, heterogeneous resource demands, and stringent quality-of-service requirements, necessitate intelligent resource allocation mechanisms that can adapt to rapidly changing conditions. This paper proposes a novel learning-based approach that leverages Graph Neural Networks (GNNs) to model the complex dependencies and resource relationships in serverless computing environments. Our framework captures the intricate structural patterns of function invocations, resource utilization, and inter-function dependencies through graph representations, enabling more effective resource allocation decisions. The GNN-based model employs a deep reinforcement learning architecture where an intelligent agent learns optimal policies through continuous interaction with the serverless environment. Through comprehensive experimental evaluation, we demonstrate that our approach achieves superior performance compared to traditional heuristic-based methods including Shortest Job First (SJF), Tetris, and Packer algorithms, reducing average job slowdown by approximately 40% under high load conditions while maintaining robust performance across varying workload intensities. The proposed system exhibits strong scalability with efficient training on graphs containing up to 10 million edges and demonstrates excellent generalization capabilities across diverse workload patterns.