ScaleIP: A hybrid autoscaling of VoIP services based on deep reinforcement learning

Abstract

Adaptive resource provisioning has become crucial for cloud-based applications, especially those managing real-time traffic like Voice over IP (VoIP), which experience rapidly fluctuating workloads. Traditional static provisioning methods often fall short in these dynamic environments, leading to inefficiencies and potential service disruptions. Existing solutions struggle to maintain performance under varying traffic conditions, particularly for time-sensitive applications. This paper introduces ScaleIP, a hybrid autoscaling solution for containerized VoIP services that offers real-time adaptability and efficient resource management. ScaleIP leverages Deep Reinforcement Learning to make dynamic and efficient scaling decisions, improving call latency, increasing the number of successfully routed calls, and maximizing resource utilization. We evaluated ScaleIP through extensive experiments conducted on a real testbed utilizing the customer Call Detail Record (CDR) from 2023 provided by World Direct, encompassing over 89 million calls. The results show that ScaleIP consistently maintains call latency below 2 s, increases the number of successfully routed calls by 3.26 ×, and increases the resource utilization up to 60 % compared to state-of-the-art autoscaling methods.