On Performance VNF Load Prediction Models in Service Function Chaining

One of the fundamental challenges in network function virtualization (NFV) is how to autonomously and dynamically allocate resources to virtualized network functions (VNFs) whose resource requirements frequently changed. To address this issue, several machine learning (ML)-based studies have been proposed, which predict the future loads of VNF in service function chaining (SFC) enabled networks. In this paper, we compare two prediction models: 1) basic long short term memory (LSTM) model that uses historical resource utilization of VNF to predict and 2) context and aspect embedded attentive target dependent LSTM (CAT-LSTM) model that leverages both historical utilization of VNF and its neighbor VNFs. Simulation results demonstrate that performance of basic LSTM is not affected by the number of service chain (SC) allocated in VNF. Meanwhile, the performance of CAT-LSTM is apparently degraded when multiple SCs are allocated in VNFs, which means prediction loss of CAT-LSTM increases from 4% to 6%.