<scp>AI</scp> ‐Driven Adaptive Pod Co‐Location in Kubernetes Using Real‐Time Multi‐Objective Optimization

ABSTRACT As microservices and containerized architectures gain prominence, Kubernetes has become the foundation for orchestrating distributed workloads. However, its default scheduling strategy is typically static and resource‐centric, lacking responsiveness to changing workload patterns, inter‐service communication needs, or energy efficiency goals. This paper introduces AICoLoc‐K8s, an adaptive scheduling enhancement framework that augments Kubernetes with intelligent, real‐time decision‐making. AICoLoc‐K8s leverages live system metrics collected from Prometheus and a lightweight AI model to inform pod placement decisions. Integrating a custom scheduler extender, the system dynamically evaluates candidate nodes based on multiple criteria, including CPU load, memory availability, and service tier affinity. The AI model is trained to optimize pod co‐location, especially for workloads categorized by frontend, backend, and database roles. We implemented AICoLoc‐K8s on an RKE2 cluster running inside KubeVirt virtual machines and validated its behavior through controlled experiments. The framework consistently demonstrated better placement alignment with workload characteristics than the default scheduler. Although quantifiable gains like latency and energy savings are reserved for future evaluation, initial results confirm the effectiveness of this real‐time adaptive model.