CNCF-hosted Co-located Events Europe 2025

Kubernetes is widely adopted for inference workloads, but distributed ML training still presents challenges, such as dynamic resource scaling, GPU scheduling, and efficient inter-node communication. Recent advancements, including KubeRay, Kubeflow, and Slurm integration, have expanded Kubernetes' capabilities for training workloads, making it a more viable option for complex, large-scale ML tasks.

This session focuses on the next step: benchmarking these tools to evaluate and optimize their performance for distributed ML training. We’ll review existing solutions, discuss the design and implementation of our benchmarking platform, and demonstrate how it provides actionable insights to improve throughput, scalability, and efficiency.