Optimization of Kubernetes for the High-Performance Computing with Kubernetes, Performance Analysis, and Dynamic Workload Placement toward the Enhancement of Cloud Computing

The rapid growth of cloud computing has created a demand for efficient resource utilization and high-performance computing (HPC) solutions. An essential component for overseeing cloud workloads is Kubernetes, a top-tier container orchestration platform. However, challenges such as resource underutilization, workload inefficiencies, and performance bottlenecks persist in dynamic cloud environments. This paper presents a comprehensive evaluation of optimizing Kubernetes for high-performance computing (HPC) in hybrid cloud-edge environments. Kubernetes is configured with components such as Etcd, Kube-APIServer, Kube-controller-manager, and Kube-scheduler to enable efficient resource management and workload orchestration. Kubeflow operators are integrated to automate machine learning workflows, including distributed training, hyperparameter tuning, and model serving. Performance metrics were analyzed for two methods, KFT and KFL. KFT achieved a deployment time of 173 seconds, a task completion time of 4.62 hours, CPU utilization of 3.47%, and RAM utilization of 3708 MB. Conversely, KFL demonstrated a faster deployment time of 51.29 seconds, a task completion time of 5.31 hours, CPU utilization of 13.77%, and RAM utilization of 2725 MB. Furthermore, KFT outperformed KFL in accuracy, reaching 0.55 at epoch 10, compared to 0.50 for KFL. These findings highlight a trade-off between resource utilization and performance, offering key insights into optimizing Kubernetes for scalable HPC systems in cloud-native environments.