Menu

Kubernetes DRA (Dynamic Resource Allocation) with NVIDIA GPU Operator

Reading 0:00
15s threshold

Hello guys,

Just a simple question, we are setting up Kubernetes cluster with some GPUs and we were wondering if starting directly with DRA could be a good future proof design.
Instead of starting with device-plugin and then migrate on DRA afterwards.
We will be running workloads (not LLMs) that will make use of different GPUs in a single cluster.

The NVIDIA docs say some DRA features are still alpha full GPU allocation is not officially supported upstream yet, and we’re using Turing, so no MIG.
So for now this cluster is for development, but it could move to production next year.

I’d like to hear from anyone who has run DRA with the GPU Operator: what issues did you encounter, and would you make the same choice today?

Cheers.