Scheduling HPC and AI Workloads the Cloud-Native Way

High performance computing (HPC) practitioners and AI researchers have long relied on traditional schedulers such as Slurm, PBS, and LSF to manage workloads. While these systems remain foundational, they can struggle to accommodate the growing convergence of simulation, data analytics, and AI/ML workflows — particularly when heterogeneous accelerators and cloud environments come into play.

This tutorial explores cloud-native approaches to workload scheduling using Kubernetes, focusing on how modern tools can complement and extend traditional HPC schedulers. We will cover both conceptual and practical dimensions:
- The state of scheduling today: where traditional HPC tools succeed and where they fall short for modern AI/ML workflows.
- Running Slurm on Kubernetes with the Slinky project: extending familiar HPC scheduling concepts into a container-native environment.
- Using Kueue with Kubernetes: a purpose-built cloud-native job scheduler for AI and HPC-style batch workloads.
- Hybrid scheduling models: how to integrate simulation pipelines, GPU-heavy ML training, and inference workloads into a unified Kubernetes platform.
- Real-world use cases: lessons from deployments in scientific, aerospace, and space computing environments.