DeepVRM: Deep Learning Based Virtual Resource Management for Energy Efficiency

• Published in: Journal of network and systems management 2023-10, Vol.31 (4), p.66, Article 66
• Main Authors: Zaman, Zakia, Rahman, Sabidur, Rafsani, Fazle, Rahman, Ishraq R., Naznin, Mahmuda
• Format: Article
• Language: English
• Subjects: Algorithms; Artificial neural networks; Communications Engineering; Communications traffic; Computer Communication Networks; Computer Science; Computer Systems Organization and Communication Networks; Deep learning; Energy consumption; Energy efficiency; Forecasting; Hardware; Information Systems and Communication Service; Machine learning; Mathematical models; Multivariate analysis; Network function virtualization; Networks; Operations Research/Decision Theory; Resource allocation; Resource management; Virtual networks
• ISSN: 1064-7570; 1573-7705
• DOI: 10.1007/s10922-023-09752-1

Network function virtualization (NFV) provides dynamic, energy-efficient, and cost-effective network services by segregating network services from expensive and energy-hungry hardware components. Traditional hardware-based network functions are always ON with 100% capacity, resulting in massive energy consumption. In contrast, NFV enables different network services (NS) through virtual network functions (VNFs), which can reduce energy consumption, by automatic ON/OFF (auto-scaling) of the VNFs depending on the traffic load. However, auto-scaling will introduce delays in the network service management process. This delay can be minimized by estimating future resource requirements accurately and taking proactive steps ahead of time. In this research, we propose the use of deep learning based forecasting algorithms to predict the required VNFs and virtual CPU for each VNF to complete a network service, often deployed as service function chain (SFC). We use long short term memory (LSTM), bidirectional-LSTM, and convolutional neural network (CNN) based algorithms to forecast resources from univariate and multivariate network traffic data. Experimental results show that CNN and bidirectional-LSTM-based forecasting models can forecast more accurately than the prior studies. Based on the forecasting models, we propose a resource allocation algorithm (DeepVRM) that can efficiently allocate resources to complete SFC.