CPU Usage Forecasting for Load Balancing in Kubernetes Using LSTM: A Synthetic Traffic Simulation Approach

Kubernetes offers automatic scaling; however, the accuracy of predicting resource requirements remains a challenge in dynamic workload environments. This research proposes LSTM to predict CPU usage in real-time on Kubernetes. The dataset was obtained from e-commerce server logs, and a distribution that identified the beta distribution as the best choice (AIC & BIC). Synthetic data based on the Beta distribution is then simulated using k6, resulting in a time series of CPU usages. The results show that LSTM outperforms ARIMA and GRU with the lowest MSE (0.00001053) and RMSE (0.00324451). The proposed approach can enhance resource allocation efficiency and application stability, as well as provide opportunities to develop real-time workload predictions for more adaptive auto-scaling on Kubernetes.