Adaptive LSTM-Based Model for Accurate Forecasting of Workload and Resource Variability in Cloud Computing

Abstract

Cloud/edge computing systems play a crucial role in providing a wide range of services for Internet users. Despite their numerous advantages, providers of these systems face certain challenges, such as accurately predicting large-scale workloads and resource usage traces. The complexity of cloud computing environments makes it difficult for traditional models to accurately predict these traces due to their highly variable nature. Traditional models struggle to handle nonlinear characteristics and long-term memory dependencies. To address this issue, this study proposes an integrated prediction method that combines Bi-directional and Grid Long Short-Term Memory network (BGLSTM) models to predict workload and resource usage traces. The proposed method first smooths the traces using a Savitzky-Golay filter to eliminate extreme points and noise interference. Subsequently, an integrated prediction model is established to achieve accurate predictions for highly variable traces. The effectiveness and adaptability of the BG-LSTM model for different traces are demonstrated through extensive experiments using real-world workload and resource usage traces from Google Cloud data centers. The performance results indicate that BG-LSTM outperforms typical prediction methods in accurately predicting highly variable real-world cloud systems.