Benchmarking and revisiting time series forecasting methods in cloud workload prediction

Over the past decades, cloud computing has become a cornerstone of modern infrastructure. Accurate cloud workload prediction is crucial for assessing quality of service and ensuring efficient resource utilization in cloud center. Despite significant advances in cloud workload prediction research, we find that there remains a substantial performance gap between existing specialized models and cutting-edge Long-term Time Series Forecasting (LTSF) models. In this paper, we conduct the first comprehensive evaluation of advanced LTSF models compared to existing cloud workload prediction models, examining both predictive accuracy and computational efficiency. Our in-depth analysis reveals that the presence of distribution shifts in cloud workload data and the inadequate robustness of current cloud workload prediction models contribute to their underperformance relative to leading LTSF models. Our further research demonstrates that advanced techniques proposed in the LTSF field, such as Channel Independence (CI) and Instance Normalization (IN), can mitigate these challenges and significantly enhance model performance. Our findings not only highlight the potential of integrating LTSF techniques into cloud workload prediction models but also provide valuable insights for advancing research in this domain.