DMV-DAB: a multivariate model for cloud resource forecasting

Cloud Computing provides numerous benefits to the users but, it grapples with the problem of inefficient resource usage. Although, resource forecasting is a prominent solution for better resource management, but high dynamicity in the workload traces poses a significant predictive hurdle. Moreover, performance metrics related to one resource of the workload traces may depend on utilization of other resources also. This highlights the need of a model which considers multiple input parameters in resource forecasting decision. This work proposes a multi-variate deep learning model named Deep Multi Variate-Dual Attention BiGRU. The model incorporates a Bi-directional GRU layer and dual attention mechanism for CPU resource forecasting. These layers enable the model to capture long-range dependencies in the data and helps to capture both temporal and spatial dependencies which enhances the model's ability to understand and predict complex interrelationships in resource dynamics. To evaluate the proposed model, experiments are conducted using data traces of Google and Bitbrains. Experimental results demonstrate that the proposed model improves prediction accuracy in terms of RMSE by an average of 63.48%, 29.74%, and 14.26% compared to statistical, machine learning, and deep learning methods respectively on Google workload traces, and 62.90%, 25.46%, and 17.81% on Bitbrains fastStorage dataset.