The Coronavirus pandemic and the work-from-home have drastically changed the working style and forced us to rapidly shift towards cloud-based platforms & services for seamless functioning. The pandemic has accelerated a permanent shift in cloud migration. It is estimated that over 95% of digital workloads will reside in cloud-native platforms. Real-time workload forecasting and efficient resource management are two critical challenges for cloud service providers. As cloud workloads are highly volatile and chaotic due to their time-varying nature; thus classical machine learning-based prediction models failed to acquire accurate forecasting. Recent advances in deep learning have gained massive popularity in forecasting highly nonlinear cloud workloads; however, they failed to achieve excellent forecasting outcomes. Consequently, demands for designing more accurate forecasting algorithms exist. Therefore, in this work, we propose 'MAG-D', a Multivariate Attention and Gated recurrent unit based Deep learning approach for Cloud workload forecasting in data centers. We performed an extensive set of experiments on the Google cluster traces, and we confirm that MAG-DL exploits the long-range nonlinear dependencies of cloud workload and improves the prediction accuracy on average compared to the recent techniques applying hybrid methods using Long Short Term Memory Network (LSTM), Convolutional Neural Network (CNN), Gated Recurrent Units (GRU), and Bidirectional Long Short Term Memory Network (BiLSTM).
Keywords: Cloud data centers; Deep learning; Energy-efficiency; Prediction approaches; Resources’ utilization; Time-series.
© 2023 Elsevier B.V. All rights reserved.