The Future of AI Workloads with Slinky

In this episode, Marlow Warnicke, lead for the Slinky project, and Tim Wickberg, CTO of SchedMD, join us to discuss their work integrating HPC scheduler Slurm with Kubernetes. They provide background on Slurm's origins, its open source nature, and its evolution into Slinky to address Kubernetes's limitations in scheduling AI and HPC workloads. The discussion touches on the unique challenges in the MLOps space, the need for fine-grained resource control, and their collaborative efforts with various communities to enhance Kubernetes's efficiency. They also share the roadmap for Slinky and avenues for community collaboration and contribution.

00:00 Introduction and Guest Introductions

00:39 Overview of Slurm and Its Evolution

01:44 The Fusion of Slurm and Kubernetes: Slinky

04:14 Challenges in Kubernetes Scheduling

09:07 Unique Challenges in MLOps

12:58 Community Collaboration and Future Plans

16:41 Getting Involved and Final Thoughts