This paper introduces Kind Slurm Integration, short KSI, as a system for running Kubernetes without root under Slurm.
The implementation is based on Kind and rootless Podman, which enable setting up and running a complete Kubernetes deployment within a Slurm job on a single node.
The paper was written by me but it is based on the master’s thesis of Sören Metje.

As the Slurm job is started, the Kubernetes cluster comes into existence, sets itself up on its assigned node and is then ready to process workloads.
Once the workloads are completed or the Slurm job expires, the cluster is removed.
The provided Kubernetes cluster is feature complete such that any Kubernetes workload can be executed on top of Slurm without changing the Slurm installation as long as a recent rootless Podman installation is provided.

Moreover, the paper provides an extensive evaluation of the performance of KSI compared to other, similar solutions.
Here, two significant short comings of KSI should be noted.
First, KSI only supports operating on a single node as it is based on Kind, which does not support multi-node clusters.
Second, as KSI employs rootless Podman, it employs its default network driver, which was slirp4netns at the time.
Slirp4netns significantly underperformed compared to the alternative approaches, achieving about a tenth of their performance.

To see all details, the complete paper can be found here: https://www.thinkmind.org/library/CLOUD_COMPUTING/CLOUD_COMPUTING_2025/cloud_computing_2025_2_60_20036.html