This thesis describes the implementation of a GPU-cluster using Rancher Kubernetes and KAI Scheduler. After conducting thorough research and preparations related to require-ments, tools and circumstances, a functional cluster was implemented as a proof of con-cept. The cluster is designed to enable students and researchers at University West to train AI models using Machine Learning.

Implementation included configuring Kubernetes to optimize GPU utilization and integrat-ing KAI Scheduler to enable several jobs to simultaneously run on the same GPU. NVIDIA GPU Operator was also used to automate the installation of necessary software components, simplifying GPU availability and enabling easier scalability when adding more nodes to the cluster.

To align with University West’s vision of the cluster, several important features were imple-mented, including checkpoints for saving progress in case of unexpected interruptions, and job splitting to run segments in parallel instead of in sequence, potentially reducing training times for large jobs. While these features are implemented and tested, their usability could be improved in future work.

The findings of this thesis highlight the cluster's potential for enhancing AI model training at University West. Although the cluster is not usable at scale, the objective of developing a proof of concept was achieved as all specified requirements were met, laying a solid foun-dation for future work to improve the cluster further. Date: