Predictive Auto-scaling: LSTM-Based Multi-step Cloud Workload Prediction

Auto-scaling, also known as elasticity, provides the capacity to efficiently allocate computing resources on demand, rendering it beneficial for a wide array of applications, particularly web-based ones. However, the dynamic and unpredictable nature of workloads in web applications poses considerable challenges in designing effective strategies for cloud auto-scaling. Existing research primarily relies on single-step prediction methods or focuses solely on forecasting request arrival rates, thus overlooking the intricate nature of workload characteristics and system dynamics, which significantly affect resource demands in the cloud. In this study, we propose an innovative approach to address this limitation by introducing a multi-step workload prediction method using the Long Short-Term Memory (LSTM) model. By considering workload attributes over a specific time frame, our approach enables accurate predictions of future workloads over designated time intervals through multi-step forecasting. By utilising two real-world web workload datasets, our experiments aim to underscore the significance of using real-world data in delivering a comparative performance analysis between single-step and multi-step predictions. The results demonstrate that our proposed multi-step prediction model outperforms single-step predictions and other baseline models.