An Efficient Adaptive Meta Learning Model Based VNFs Affinity for Resource Prediction Optimization in Virtualized Networks

In today’s cloud environments, diverse and massive data is growing exponentially. This requires efficient prediction mechanisms to support automatic resource allocation and management. These mechanisms enable anticipating resource consumption variation in complex and distributed systems. To allow the service provider fulfilling the service level objective (SLO), these mechanisms must deal with the heterogeneity and the non-stability of different resource demands. To enable the prediction mechanism adapting to different resource demands, we propose a selector model that chooses among several prediction methods, the best suiTable one to trigger at a given workload situation. The proposed model is a multitask selector based on meta learning strategy MT-MLS. The MT-MLS introduces a novel concept by analyzing similarities of multidimensional resource consumption between virtual network functions (VNFs) of a service function chain (SFC). Attention mechanism is used to identify a weight for each VNF based on the similarity analysis. The MT-MLS is designed as a multitask classifier, that assigns simultaneously and independently for each VNF the corresponding best predictor label to forecast multidimensional resource needs. We developed four deep learning base predictors: a CNN model, a GNN model, an LSTM model, and a hybrid model. These base predictors were used to generate meta data to train the MT-MLS. Then compared the prediction results of each model to the prediction results using the MT-MLS. Various real world SFC resource consumption datasets have been used to evaluate the proposed MT-MLS using three performance metrics for the evaluation of classification and five others for the evaluation of prediction. The performance analysis reveals that the MT-MLS enhance considerably the efficiency of the resource usage prediction.