Proactive Auto-Scaling for Service Function Chains in Cloud Computing Based on Deep Learning

Auto-scaler system enables high Quality of Service (QoS) with low cost to survive in a competitive market. Indeed, the auto-scaling of Virtual Network Functionality (VNFs) can adaptively allocate the Cloud resources for various VNFs based on workload demands at any time. However, the intensity of workload is dynamically changed because of the variation in service demand over time. The predominant auto-scaling approaches use scaling rules (threshold-based reactive approach) or scaling policies (schedule-based proactive approach) to adapt resources and meet the performance requirements of each VNF. The reactive approaches can significantly degrade the VNF performance for improper reconfiguration or variation of auto-scaling rules. Conversely, the proactive approaches dynamically adjust the scaling policies according to the workload variation. These approaches rely on accurate workload predictive models (e.g., time-series models). This paper proposes a real-time proactive auto-scalar system based on a deep learning model that can efficiently predict the future values of CPU, Memory, and Bandwidth for VNFs for a Service Function Chain (SFC) to proactively auto-scale the resources allocated to each VNF in a Cloud platform. A hybrid model of MLP-LSTM is used to forecast the values of different features. Auto-correlation is used to identify the abnormal events of instances in the Cloud platform by measuring the repeated pattern for each identified impact feature. Moreover, the auto-scalar system enables to predict the abnormal values for some features during the online stage using the Auto-regression model to meet the QoS requirements of an SFC.