New Auto-Scaling Method Boosts Serverless Efficiency and Cost Savings

This research introduces a novel autoscaling framework designed for serverless computing environments, addressing challenges like dynamic workloads and cold-start latency. The system models serverless applications as dependency graphs to pinpoint critical functions. It then employs a multi-expert consensus mechanism, combining multiple machine learning models like MLP, LSTM, and CNN, to accurately forecast short-term resource demands. The framework further incorporates cold-start awareness and cost comparison, allowing it to intelligently decide whether to scale up, scale down, or hold resources. Experimental results using real-world data demonstrate that this approach achieves high prediction accuracy and significantly reduces infrastructure costs compared to existing methods, all while maintaining desired performance levels.