STOWP: A light-weight deep residual network integrated windowing strategy for storage workload prediction in cloud systems

Accurate storage workload forecasting of big data applications is a constructive approach to improve the job scheduling and fine-grained load balancing in real-time cluster systems. However, despite the recent advances of deep learning architectures, demands for more accurate workload time series forecasting algorithm exist. Therefore, we propose a light-weight STOrage Workload time series Prediction method named as 'STOWP' integrating Neural Basis Expansion Analysis (N-BEATS) deep model with windowing strategy. The STOWP approach implements a multi-input-multi-output (MIMO) window strategy for capturing the historical storage variation patterns of the workload data. Furthermore, a within window scaling strategy is adopted to effectively estimate the diversity of the workload requests during different time horizons. For experimental evaluation, we used Web-Search dataset containing Search Engine's I/O real-time workload traces. To improve the performance of STOWP, the hyper-parameters' sensitivity is well investigated. Through results, we observed that the 'STOWP' improves the RMSE by 3.33% and MAE by 3.44% atleast in comparison with the existing benchmarks storage workload forecasting techniques.