Elastic Virtual Network Function Orchestration Policy Based on Workload Prediction

Yunjie Gu,Yuxiang Hu,Yuehang Ding,Jie Lu,Jichao Xie

doi:10.1109/access.2019.2929260

Abstract

By separating network functions from hardware-dependent middleboxes, network function virtualization (NFV) is expected to lead to significant cost reduction and the flexibility improvement in network management. Elastic orchestration of virtual network functions (VNF) is a key factor to achieve NFV goals. However, most existing VNF orchestration researches are limited to offline policy, ignoring the dynamic characteristics of the workload. To reduce the operational expenditure of NFV providers, this paper proposes an Elastic Virtual Network Function Orchestration (EVNFO) policy based on workload prediction. We adapt the online learning algorithm for predicting the flows rate of service function chains (SFC), which can help to obtain the VNF scaling decision. We further design the online instance provisioning strategy (OIPS) to accomplish the deployment of VNF instances according to the decision. The simulation proves that EVNFO can provide good performance with dynamic resource provision. The throughput of VNF is improved by 11.1%–22.9%, and the total operational expenditure can be reduced by 13.8% compared with other online approaches.

Highlights

INTRODUCTIONFor Network Function Virtualization (NFV) providers that rent virtual resource in datacenter, we aim to minimize their operational expenditure by provision and scaling the instances of virtual network functions (VNF) in a proactive way
Enterprise network ubiquitously deploys hardware dedicated middleboxes, such as firewalls, network address translators and proxies to offer network service
This study focuses on online elastic virtual network function orchestration

Summary

INTRODUCTION

For NFV providers that rent virtual resource in datacenter, we aim to minimize their operational expenditure by provision and scaling the instances of VNF in a proactive way. To help NFV providers make proactive scaling decisions, we introduce online learning into the Evolution Prediction algorithm of Flow Rate (EPFR), which can predict the upcoming workload with the minimized error. An overview of our contributions is presented as follows: i) We introduce an online learning algorithm to predict the upcoming aggregated flow rate, which performs better than other algorithms according to the simulation; ii) We design an online instance provision strategy to orchestrate VNF instances; iii) We introduce a release mechanism for redundant instances to cut down the running cost and the deployment cost.

RELATED WORK

EFFECTIVENESS OF FLOW RATE EVOLUTION

Findings

CONCLUSION