Recently, universities and research institutions have been building systems that provide independent virtual practice environments to large numbers of students based on Kubernetes. While this architecture improves learning efficiency and accessibility...
Recently, universities and research institutions have been building systems that provide independent virtual practice environments to large numbers of students based on Kubernetes. While this architecture improves learning efficiency and accessibility, it introduces a problem in which pods remain in the cluster and continue to consume resources for extended periods if users do not explicitly terminate them after completing their practice sessions. Kubernetes default garbage collection mechanism is limited to dependency-based policies that statically delete subordinate resources when their parent resources are removed, and therefore cannot automatically release or delete pods whose actual usage has ended. As a result, educational clusters frequently experience resource exhaustion and failures in provisioning new practice environments. To address these limitations, this thesis proposes a dynamic pod reclamation algorithm based on user activity and process-level activity states. The proposed approach analyzes activity metrics within pods—such as CPU utilization, context switching, and page faults—to detect inactive pods and automatically reclaim resources while taking user activity into account. Experimental results demonstrate that the proposed command and process-based state classification
algorithm improves both inactive pod detection accuracy and resource reclamation efficiency by jointly considering user interaction and system activity. These findings confirm that stable resource circulation and efficient cluster operation can be achieved even in large-scale hands-on practice environments.