A Deep Learning Prediction Approach for Machine Workload in Cloud Computing
Yiming Bai, Lei Chen, Ying Lei, Hongjuan Xie · 2023
Accurate workload prediction for cloud computing clusters is essential to ensure Quality of Service (QoS), meet Service Level Agreements (SLAs), and minimize energy consumption. In a cloud computing environment, cloud servers collect and store large sets of time series data, including metrics such as CPU usage, network traffic, and disk I/O at each point in time. The fluctuations in these metrics show noticeable temporal correlations. Therefore, these data can be used in time series data prediction models to predict upcoming workload scenarios. However, traditional statistical methods have limitations such as requiring manual feature extraction, high data requirements, and limited generalizability, leading to inaccurate predictions. In addition, most standard deep learning models ignore the problem of sequence noise. To overcome these hurdles, we have created a hybrid deep learning model utilizing advanced techniques. This model integrates Time Convolutional Networks (TCN), Gated Recurrent Units (GRU), and self-attention mechanism to achieve a more accurate workload prediction. Additionally, we perform a comparative analysis of different methods during the preprocessing stage and select the optimal one to effectively remove noise from the original sequences. Experimental results demonstrate that our proposed model outperforms other workload prediction algorithms in terms of prediction accuracy.