Scaling Migrations and Replications of Virtual Network Functions based on Network Traffic Forecasting

TL;DR

This paper explores how traffic forecasting using LSTM models can optimize the migration and replication of virtual network functions, reducing operational costs and resource usage in ISP edge networks.

Contribution

It introduces a traffic prediction-based approach to minimize VNF migrations and replications, demonstrating significant reductions through LSTM forecasting combined with optimization algorithms.

Findings

LSTM-based traffic prediction reduces migrations by up to 45%.

Forecasting decreases reliance on overprovisioning and cloud offloading.

The approach improves resource utilization in dynamic network environments.

Abstract

Migration and replication of virtual network functions (VNFs) are well-known mechanisms to face dynamic resource requests in Internet Service Provider (ISP) edge networks. They are not only used to reallocate resources in carrier networks, but in case of excessive traffic churns also to offloading VNFs to third party cloud providers. We propose to study how traffic forecasting can help to reduce the number of required migrations and replications when the traffic dynamically changes in the network. We analyze and compare three scenarios for the VNF migrations and replications based on: (i) the current observed traffic demands only, (ii) specific maximum traffic demand value observed in the past, or (iii) predictive traffic values. For the prediction of traffic demand values, we use an LSTM model which is proven to be one of the most accurate methods in time series forecasting problems. Based the traffic prediction model, we then use a Mixed-Integer Linear Programming (MILP) model as well as a greedy algorithm to solve this optimization problem that considers migrations and replications of VNFs. The results show that LSTM-based traffic prediction can reduce the number of migrations up to 45\% when there is enough available resources to allocate replicas, while less cloud-based offloading is required compared to overprovisioning.