Optimizing resource allocation for federated LLM training via workload forecasting and autoscaling in edge environments

Abstract The growing deployment of applications in real-world edge environments is driving an increasing demand for large language model (LLM) training in distributed systems. This demand creates significant challenges for resource management, particularly in resource-constrained edge cloud environments. Federated learning (FL) enables privacy-preserving LLM fine-tuning across multiple clients, but it also introduces dynamic computational and communication loads that can strain system resources. To address these challenges, we propose an integrated framework that combines LLM fine-tuning using federated adaptive local low-rank adaptation (FedALoRA) with forecasting-driven autoscaling policies. Differential privacy (DP) is incorporated to ensure secure aggregation, balancing privacy guarantees with system performance. Using the WikiText-2 dataset for FL-based LLM training and the Bitbrains dataset for CPU usage forecasting, we evaluate deep learning models (LSTM, GRU, BiLSTM, CNN, CNN-LSTM, Transformer). Results demonstrate that FedALoRA enables fast federated convergence, reducing perplexity from 199.3 to 85.2 without DP and from 211.6 to 92.2 with DP, incurring only a 7–13% privacy overhead. Among forecasting models, CNN-LSTM achieves the lowest MSE (0.25) and consistently requires the fewest pods (as low as 11 replicas), highlighting its superior accuracy and resource efficiency for proactive autoscaling. Comparative analysis demonstrates that proactive autoscaling consistently outperforms reactive autoscaling, with non-DP settings achieving slightly faster adaptation and DP settings offering stronger privacy guarantees. Overall, the proposed framework balances performance, adaptability, and privacy, making it well-suited for federated-LLM Autoscaling in dynamic edge environments.