WANE: Workload Adaptive Neuro-genetic Engine for Container Usage Prediction
Soyeon Park, Hyokyung Bahn · 2024
Microservice architectures in Kubernetes-based container environments often face challenges in optimally allocating resources due to fluctuating workload usage. Traditional methods, which rely on heuristic computing resource requests, typically result in overprovisioning and inefficient utilization. Predicting the actual runtime resource demand of containerized applications accurately is challenging. To address this issue, the authors introduce a novel hybrid cloud resource prediction engine that combines the strengths of genetic algorithms and Bayesian neural networks. Trained with real-world trace data, the proposed model outperforms existing techniques like ARIMA and exponential smoothing, particularly in reducing the risk of underprediction. Notably, these enhancements significantly improve CPU prediction accuracy, demonstrating potential for optimizing resource allocation and enhancing cost efficiency in cloud environments.