Enabling secure HPC resource utilization for cloud and HPC workloads

Serving AI-models to users requires access to GPUs, which are available in HPC centers of universities and research institutes. However, these GPUs are already part of existing batch-scheduling systems and require a strategy for utilizing them for both serving regular HPC workloads and AI inference. To this end, we have explored multiple approaches, one of which forms the foundation of the GWDG AI service offerings. In this article we discuss two other approaches that we have published about that enable the operation of Kubernetes in HPC environments along with regular HPC workloads.