Auto Scaling of Service Function Chain based on the predicted network traffic using Temporal Fusion Transformer

Abstract

Efficient network operations are crucial for Internet service providers, and they necessitate swift management actions like VNF deployment, VNF auto-scaling, and VNF live migration. These actions need to be performed in response to user demands and changes in network virtualization environments. Sole reliance on manual interventions can lead to longer response times and additional costs when problems arise. This is due to several factors, including the reactive nature of manual operations compared to the anticipatory nature of proactive management, the limitations of human speed and decision-making capabilities compared to automated processes, and the increasing difficulty of manually managing complex and growing networks. To address these challenges and shift towards more proactive management strategies, network traffic analysis and future traffic prediction become essential. In this thesis, we employ the Temporal Fusion Transformer (TFT), a cutting-edge model for time series data prediction, to forecast real network and application traffic, particularly Abilene and Wiki traffic. The Mean Squared Error (MSE) of TFT is compared with that of Long Short Term Memory (LSTM) and Gated Recurrent Unit (GRU), and a threshold-based horizontal auto-scaling mechanism is deployed that operates based on the predicted traffic. Our experiments show that the use of the Temporal Fusion Transformer (TFT) for network traffic prediction leads to a significant decrease in Mean Squared Error (MSE), by as much as 94\% compared to traditional methods such as Long Short Term Memory (LSTM) and Gated Recurrent Unit (GRU). This clearly underscores the superior accuracy of TFT predictions. Further, applying auto-scaling based on TFT-predicted traffic notably reduces both under-provisioning and over-provisioning, especially when handling burst-heavy traffic like Wiki traffic. When compared to the LSTM and GRU predictions, the TFT prediction method leads to a decrease in under-provisioning by 47.1\% and 24.77\% respectively, and a decrease in over-provisioning by 43.21\% and 35.19\% respectively. Moreover, across both Abilene and Wiki traffic, there was a reduction in the number of scaling events by up to 76\% when these were based on TFT-predicted traffic. This outcome highlights how a proactive, TFT-based auto-scaling approach can considerably reduce the frequency of scaling changes, providing potential savings in terms of both cost and time.