DRL ‐Based Virtual Resource Allocation for Web 3‐Tier Containers Using Low‐Layer Metrics

Container‐based virtualization is widely adopted as an execution substrate for cloud‐native systems. However, when multiple containers share CPU and memory on a single host, workload imbalance can easily degrade tail latency. This paper proposes a dynamic resource allocation method for a Web 3‐tier container system (Nginx/Flask/MySQL) using Deep Reinforcement Learning (DRL) with observations augmented by low‐layer metrics. In experiments using a real hardware deployment, the policy improved p95 response time and showed reward convergence while maintaining throughput and avoiding excessive resource cost increase.