DIMA: machine learning based dynamic infrastructure management for containerized applications

DIMA: machine learning based dynamic infrastructure management for containerized applications

연구 분야: Infrastructure

학회: Computing

Microservices transform application architecture using containerized, loosely connected components that can be deployed, reused, and maintained independently. These components scale efficiently to manage dynamic workloads. Traditional autoscaling methods, which include reactive, proactive, and rule-based strategies, focus on horizontal scaling by adding or removing container resources based on a pre-configured number of virtual machines. They often rely on single-metric predictions like CPU utilization, memory usage, or response time. However, these approaches overlook the challenge of multi-metric dynamic adjustment of container and VM resources during fluctuating workloads. As a result, it can lead to overprovisioning during low-demand periods or underprovisioning during traffic spikes, negatively impacting performance and cost. This paper introduces a machine learning-based dynamic infrastructure management approach (DIMA) for containerized applications. DIMA predicts and adjusts real-time infrastructure requirements using a multi-layered perceptron (MLP) model to meet service-level agreements (SLAs). It reduces the response time violations and operational costs. We evaluate DIMA using a benchmark microservices application deployed on a Kubernetes cluster with diverse real-world workloads and compare it to state-of-the-art autoscaling methods. The results show that DIMA reduces response time violations by 15–20% and lowers operational costs by 1.5x to 2.5x compared to existing methods.