Prediction and Dynamic Adjustment of Resources for Latency-Sensitive Virtual Network Functions

In this paper, we propose a scheme to employ predictive multiple regression models for dynamic resource adjustment of latency-sensitive virtual network functions (VNFs). The objective is to optimize resource allocation with agile control so that both latency and resource utilization meet the target performance requirements, despite the fluctuations in the workload. As a use case of latency-sensitive VNF, we select the Internet-of-things directory service (IoT-DS). IoT-DS is capable of storing up to a billion IoT devices records and provide fast database lookup with few milliseconds of latency. Our experimental results with an operational VNF serving as the IoT-DS show that dynamic adjustment of CPU by the proposed scheme reduces CPU resource requirements by 21.9% and avoids lookup latency requirement violations by 58.2 % compared to a threshold rule-based conventional algorithm. Besides, our scheme can provide agile vertical CPU scaling within a one-second window, which is five times faster than a related machine learning-based prior work.

I. Introduction

Virtual network functions (VNFs) placed in the cloud may not satisfy latency requirements due to their remote location from the end-users. Cloud computing facility is introduced at the end-users' proximity, known as the edge cloud with limited resources, to satisfy latency requirements of latency-sensitive end-user services, such as augmented reality, telemedicine, online multiparty video games, and self-driving vehicles. Dynamic fine-tuning of VNF resource allocation at the edge is necessary to optimize resource utilization and satisfy performance requirements. In this paper, we present a prediction and dynamic resource adjustment scheme for latency-sensitive VNFs to optimize resource utilization and satisfy latency requirements simultaneously.