Convolutional Neural Network-based Workload and Resource-aware Task Management in Fog Networks

As latency-sensitive IoT applications proliferate, fog computing must meet stringent real-time requirements by performing computation near the edge. However, existing task assignment and offloading approaches are predominantly reactive, relying on current system states without anticipating future load fluctuations—leading to frequent congestion, suboptimal resource utilization, and missed task deadlines under bursty workloads. Although a few studies have explored workload prediction, most employ simplistic statistical models or heavy recurrent networks (e.g., LSTM) that are unsuitable for real-time deployment on resource-constrained fog nodes. To address these limitations, this paper introduces Convolutional Assignment Strategy for Tasks (CAST), a Convolutional Neural Network (CNN)-based, predictive task assignment framework designed for resource-constrained fog networks, enabling proactive, deadline-aware, and load-balanced decision-making. CAST operates within a three-layer architecture consisting of IoT devices, fog nodes, and a centralized controller. At the IoT layer, tasks are executed locally or offloaded to fog nodes based on task size, deadline, and device constraints. Fog nodes queue tasks using Earliest Deadline First (EDF) scheduling while periodically reporting workload to the controller. The centralized controller predicts future workload using a CNN trained on time-series workload matrices representing all fog nodes. Unlike Recurrent Neural Networks (RNNs) or Transformers, which incur higher computational latency due to sequential processing or quadratic self-attention complexity, the CNN leverages the spatio-temporal grid of workload matrices to efficiently extract hierarchical features capturing both temporal dynamics and spatial correlations between nodes. Its computational efficiency and parallelizable nature make it suitable for real-time execution on resource-constrained networks. Predicted workloads guide the controller in classifying nodes as crowded or uncrowded, enabling deadline-aware task assignment and dynamic load balancing. Tasks are then offloaded among fog nodes or, if necessary, to the cloud to prevent congestion, minimize queuing delays, and maximize the number of tasks completed within their deadlines. CAST is a robust mechanism, because the CNN model can be periodically retrained to adapt to evolving workload patterns. Simulation results show that CAST reduces average task response time by 30%, improves the deadline-hit ratio by 48%, and achieves more balanced workload distribution compared to state-of-the-art heuristic and modern Deep Reinforcement Learning (DRL) baselines, while maintaining low prediction and decision overhead. Furthermore, Pareto efficiency analysis confirms that CAST represents the optimal solution for latency-critical applications, occupying the high-performance frontier compared to energy-conservative baselines.