Abstract
<title>Abstract</title> <p>The provisioning of cloud resources must achieve a balance between cost-effectiveness, performance, and responsiveness in increasingly dynamic environments centered on human interaction. Conventional static methods like round-robin, threshold-based allocation, and greedy packing frequently struggle to adjust to fluctuating workloads, resulting in ineffective scaling and compromised SLA (Service Level Agreement) compliance. This study provides an empirical analysis of four provisioning methods utilizing authentic workload traces from the Google Cluster dataset. The investigations evaluate baseline efficiency, resource utilization, cost-performance trade-offs, and responsiveness to simulated workload spikes. Findings indicate that while threshold and greedy packing enhance utilization, they still lead to SLA breaches when subjected to cost limitations and sudden load increases. Importantly, the static system documented 3,297 SLA breaches at a decreased cost of ₹298.65, while spike simulation tests highlighted the limitations of the system's reactive capabilities. The results reveal the limitations of heuristic provisioning and encourage a transition to adaptive models. We suggest viewing cloud provisioning as a sequential decision-making issue centered around human interaction and validate the use of Deep Reinforcement Learning (DRL) as a scalable and agile approach. This research establishes a data-informed basis for incorporating DRL-driven agents into real-time orchestration systems to improve utility, adaptability, and user assistance in managing cloud resources.</p>