Published Paper Details:

Cloud operating systems face challenges in managing dynamic workloads for applications like e-commerce and streaming, processing over 2.5 quintillion bytes daily. Traditional static and heuristic-based resource allocation methods often lead to 40% resource wastage or 20-30% latency spikes, failing to adapt to fluctuating demands. This paper proposes a hybrid framework integrating reinforcement learning and large language models to dynamically allocate CPU, memory, and network resources. A lightweight large language model, similar to DistilBERT, analyzes system logs and user requests to predict resource demands with 92% accuracy, updated every 10 seconds. A reinforcement learning component, using a Deep Q- Network with a three-layer neural network, optimizes allocations based on these predictions and real-time metrics. Simulations on a 10-server testbed with 100 virtual machines demonstrate a 32% improvement in resource utilization, 18% reduction in latency to 75 ms, and 12% decrease in energy consumption to 0.70 kWh compared to heuristic methods. Throughput increased by 31%, handling 12,500 requests per second. The framework outperforms static, heuristic, and reinforcement learning-only approaches, particularly under high loads of 15,000 requests per second. Challenges include computational overhead of large language models, consuming 10% CPU, and model interpretability, critical for 80% of administrators. Future work focuses on lightweight algorithms and multi-cloud scalability to enhance efficiency and practicality in dynamic cloud environments.