MAG-D: A multivariate attention network based approach for cloud workload forecasting

Area of Science:

• Cloud Computing
• Artificial Intelligence
• Data Science

Background:

• The COVID-19 pandemic accelerated cloud migration, with over 95% of digital workloads expected in cloud-native platforms.
• Accurate real-time workload forecasting and resource management are critical challenges for cloud service providers.
• Classical machine learning and existing deep learning models struggle with the volatile and nonlinear nature of cloud workloads.

Purpose of the Study:

• To address the limitations of current forecasting methods.
• To propose a novel deep learning approach for enhanced cloud workload prediction.

Main Methods:

• Developed MAG-DL (Multivariate Attention and Gated Recurrent Unit based Deep Learning) approach.
• Utilized Google cluster traces for extensive experimental validation.
• Compared MAG-DL against hybrid methods including LSTM, CNN, GRU, and BiLSTM.

Main Results:

• MAG-DL effectively captures long-range nonlinear dependencies in cloud workload data.
• The proposed model demonstrates improved prediction accuracy compared to existing state-of-the-art techniques.
• Experimental results confirm the efficacy of MAG-DL on real-world cloud workload traces.

Conclusions:

• MAG-DL provides a more accurate solution for cloud workload forecasting.
• The model's ability to handle complex workload dynamics offers significant advantages for cloud service providers.
• This research contributes to more efficient resource management in cloud data centers.