AWS: GPU node autoscaling doesn't work
cyrildiagne opened this issue · comments
Cyril Diagne commented
GPU nodes don't autoscale with the AWS provider.
Because Knative doesn't support nodeSelectors and tolerations (knative/serving#1816) we can't rely on node taints to limit the workload being assigned to the GPU nodeGroup.
An immediate workaround could be to use namespace-level annotations: knative/serving#1816 (comment)
This doesn't seem to be a problem on GKE so it's also worth investigating how they do it.
Cyril Diagne commented
Out of scope for now