Improving scalability, energy efficiency, and cost-effectiveness in Kubernetes clusters using a nonlinear regression-based predictive replica model and ORLE algorithm

Abstract

Container-based deployments have transformed how modern applications are packaged, deployed, and scaled. They bring increasing benefits in terms of development agility, testing, and collaboration. Kubernetes is a famous container orchestration engine that manages the life cycle of containerized applications by automatically scaling the containers and load balancing among them. In this paper, we present an application-level leader election method known as ORLE(Optimal Replica Leader Election). After conducting a detailed performance analysis of ORLE, we also developed a nonlinear regression Predictive Replica Model that predicts the throughput in real time, which helps in the early identification of conditions that require replicas’ scaling up or down. We integrated the proposed ORLE with this nonlinear regression Predictive Replica Model to improve the performance of Kubernetes clusters. Our model autoscales replicas concerning real-time traffic measurements to maximize resource utilization and system scalability. ORLE selects the most suitable replica as the leader to optimize the load distribution and reduce the overall latency of the request. Our experimental results show that our proposed solution outperforms the original leader election mechanism (Default RAFT) of Kubernetes and an existing state-of-the-art algorithm Balanced Leader Distribution (BLD) by up to 25% throughput improvement, 40% latency reduction, optimization of energy consumed, and cost efficiency per working replica under real-time conditions. The proposed model is more beneficial for stateful applications and microservices architectures since consistent performance and fast leader election are crucial in ensuring the Kubernetes cluster reliability and performance and are highly suitable for Kubernetes clusters deployed in a cloud computing environment, which demands high scalability, low latency, and efficient resource management.