Date of Publication :5th July 2025

Abstract: Serverless computing has revolutionized cloud computing by allowing developers to build and deploy applications without managing the underlying infrastructure. However, efficiently allocating resources to handle dynamic workloads remains a significant challenge. This paper presents an approach for auto-scaling in serverless environments using Proximal Policy Optimization (PPO), a reinforcement learning technique that optimizes resource allocation in real-time. Unlike previous methods that relied on Deep Q-Learning (DQL) or Q-Learning (QL), PPO enhances stability and scalability by directly learning optimal policies. A synthetic workload dataset is used to simulate realistic traffic patterns for model training. Experimental results on AWS Lambda demonstrate that PPO reduces average response time by 35% compared to QL and 20% compared to DQL, ensuring faster job execution. Energy consumption is lowered by 25% and 15%, respectively, improving efficiency. Additionally, throughput increases by 18% over QL and 10% over DQL, while success rate improves by 12% and 8%, ensuring more reliable task execution. These findings highlight PPO’s superior effectiveness in reinforcement learning- based resource management, making it a promising solution for autoscaling in serverless computing.

Reference :