Cost-Based Efficient Resource Allocation for Edge Computing Using Graph Neural Networks: A FlaskBased Scheduler

Abstract

Edge computing has emerged as a critical paradigm for processing latency-sensitive applications at the network periphery, reducing bandwidth consumption and improving response times for billions of connected devices. However, efficient resource allocation in heterogeneous edge environments remains a fundamental challenge due to the dynamic nature of workloads, varying computational capacities of edge servers, and stringent cost constraints. Traditional scheduling algorithms, including heuristics and classical optimization techniques, struggle to capture the complex interdependencies between tasks and the topological structure of edge networks, leading to suboptimal resource utilization and increased operational costs. This paper introduces a novel Graph Neural Network (GNN)-based resource allocation framework implemented as a Flask web application that intelligently schedules computational tasks across distributed edge servers to minimize execution costs while maintaining load balance and ensuring reliable performance. The system models the edge computing infrastructure as a heterogeneous graph where nodes represent computational resources (edge servers, cloud nodes, network switches) with attributes including processing capacity (2-64 cores), memory availability (4-256 GB), storage capacity (100 GB - 10 TB), network bandwidth (100 Mbps - 10 Gbps), energy consumption profiles (50-500 W), and operational costs ($0.02-0.50 per hour). Edges in this graph capture communication latency (1-50 ms), bandwidth constraints, and data transfer costs between infrastructure components. Incoming tasks, whether single requests or bulk workloads, are represented as computational graphs with nodes representing subtasks and edges representing data dependencies and communication requirements. A multi-layer Graph Attention Network (GAT) processes these graphs to learn optimal task-toserver mappings, with attention mechanisms identifying critical resource constraints and task dependencies. The GNN architecture comprises three graph convolutional layers with 64, 128, and 256 hidden units respectively, incorporating residual connections and layer normalization to stabilize training. The scheduler operates in two modes: online scheduling for single tasks with sub-second decision latency, and batch scheduling for bulk workloads with optimization over multiple tasks simultaneously. Experimental evaluation on a simulated edge infrastructure with 500 servers across 50 geographical locations demonstrates that the GNN-based scheduler reduces average task execution cost by 34.7% compared to round-robin scheduling, 28.3% compared to random allocation, 21.5% compared to least-loaded-first, 18.2% compared to genetic algorithms, and 12.4% compared to reinforcement learning approaches. Load balancing metrics show a 42.3% reduction in server utilization variance, indicating more even distribution of workloads across available resources. The system maintains 99.2% scheduling success rate under peak loads of 10,000 tasks per second, with average decision latency of 47 ms for single tasks and 2.8 seconds for batches of 1,000 tasks. The Flask-based web application provides real-time monitoring dashboards, RESTful APIs for task submission, and visualization of resource utilization patterns. This work represents the first productionready integration of graph neural networks for costaware resource allocation in edge computing, demonstrating significant improvements in operational efficiency and resource utilization.