aws-samples / aws-efa-eks

Deploying EFA in EKS utilizing GPUDirectRDMA where supported
MIT No Attribution
35 stars 19 forks source link

Conflicting setup in launch template and EKS node group when not using cloudformation #19

Open ziyuanchen-db opened 9 months ago

ziyuanchen-db commented 9 months ago

For some internal reasons we cannot use cloudformation or terraform to set this up. When I tried to set this up in the UI it seems that:

  1. In the Launch Template setup, we cannot create multiple network interfaces, all without subnet specified, for p4de.24xlarge using Ubuntu images.
  2. If we try to specify the subnets for network interfaces 1-3, then we got this error "the launch template version can only have a max of one network interface" when trying to create an EKS node pool from this launch template.

So it seems (1) EKS requires all launch templates to have at most 1 network interface and without subnet(s) specified, while this tutorial requires us to set up 4 EFA network interfaces, which is rejected by the AWS UI.

Was there anything wrong with the steps I took, or would this setup only work with cloudformation / terraform, which somehow call the APIs differently than the UI?