CNCF-hosted Co-located Events Europe 2025

Dynamic Resource Allocation (DRA) is quickly gathering momentum to become the go-to way of advertising GPUs on Kubernetes clusters. In this talk, we will present the current state of the project, the latest implementation updates, and feature additions. We will walk through how to get started with DRA, and why this is relevant for any engineer trying to improve the GPU offering on their clusters. We continue with configuring time-slicing, MPS, and MIG, and explain how to build more custom layouts on top.

Next, we will show how DRA is used at CERN to colocate machine learning workloads on the same GPU. We start by explaining how to choose the best-fitted sharing mechanism depending on the performance requirements. We present extensive training and inference benchmarking results, and how DRA comes into play to make the system flexible and easy to use. Lastly, we go through GPU sharing tradeoffs, and how in the end this approach can help save resources.