Slinky: Slurm in Kubernetes, Performant AI and HPC Workload Management in Kubernetes - Tim Wickberg

Tim Wickberg introduces Slinky, an integration of the Slurm workload manager with Kubernetes to support high-performance AI and HPC workloads.

Don't miss out! Join us at our next Flagship Conference: KubeCon + CloudNativeCon events in Hong Kong, China (June 10-11); Tokyo, Japan (June 16-17); Hyderabad, India (August 6-7); Atlanta, US (November 10-13). Connect with our current graduated, incubating, and sandbox projects as the community gathers to further the education and advancement of cloud native computing. Learn more at kubecon.io Slinky: Slurm in Kubernetes, Performant AI and HPC Workload Management in Kubernetes - Tim Wickberg, SchedMD Kubernetes was designed for microservices. With AI rapidly advancing, Kubernetes must adapt to also support both AI training and multi-node inference. It needs to improve not only at scheduling these workloads within the cluster, but also at fine-grained resource assignment on the nodes. High Performance Computing (HPC) systems use workload managers such as Slurm. Slurm, the most used HPC workload manager with over two decades of development, excels at gang scheduling, fair usage, job planning, and batch scheduling. We will show the current state of Slinky, a fully open-source toolset designed to integrate Slurm with Kubernetes and to solve the difficulties of getting AI clusters working more performantly and efficiently. Slinky includes a Slurm operator, a Slurm client library, and a metrics exporter. Here, we will outline our architecture and discuss the challenges of achieving the fine-grained control needed in Kubernetes for full functionality for AI and HPC workloads.