AWS: Fix nvidia gpu resource name, re-enable ScaleToZero

[discordianfish]

These are two somewhat separate changes, so let me know if I should split this PR:

1. Rename gpu resource name, the same as Fix GPU resource name. #407 but for AWS
2. Set scaleToZero = true. It looks like this was changed (by accident?) in Support autodetection of GCE managed instance groups by name prefix #462

I've tested the changes manually and scaling to and from 0 works on my cluster with the nvidia k8s device plugin.

[aleksandra-malinowska]

Thanks for contributing! @sethpollack @mumoshu can you PTAL?

[sethpollack]

lgtm

@negz was there a reason scaleToZeroSupported was defaulted to false?

[negz]

@sethpollack It's been a while, but I'm fairly sure AWS did not support scale to zero when I added that code.

[negz]

https://github.com/kubernetes/autoscaler/pull/462/files#diff-34ecd32e36ab8898fff1637bb1b39c2cL367

@sethpollack Nope - my bad. I see now that it was enabled and I disabled it. My apologies - completely accidental on my part.

[sethpollack]

@negz thanks

[aleksandra-malinowska]

/lgtm

[bhack]

Is this available in any released image?

[alexnederlof]

@discordianfish would it be possible for you to also PR This on the 1.1.2 release for a 1.1.3 release? Everyone using Kops is still stuck on Kubernetes v1.9, which is bound to autoscaler 1.1.x. Shipping it to the 1.1.x branch would help a lot of people. I'm keen to help, but I've never done anything in Go, and couldn't figure out how to do it myself (just cherry picking didn't work 😉 ) See #903

[alexnederlof]

Or perhaps @aleksandra-malinowska

[kaazoo]

I tried cluster-autoscaler 1.2.2 on a 1.9.9 cluster (created by kops 1.9.1). Allthough it's not not officially supported, scaling up my GPU ASG from 0 seems to work.

[MaciekPytel]

@kaazoo A word of warning: CA may over scale nodes with GPU. This is because the current GPU workflow is:

1. CA notices a pending pod that requires GPU and adds node with GPU.
2. Node becomes ready. At that point GPU drivers are not yet installed and node doesn't have GPU capacity.
3. GPU driver installing daemonset runs on the node.
4. Once the drivers are installed GPU capacity is added to node.

The problem is between steps 2 and 4 CA sees the node has already booted up without GPUs. The pod requesting GPU is pending and there are no upcoming nodes, so CA triggers another scale-up. Due to the details of how CA works this issue will not happen when scaling-up from 0, but otherwise will happen ~100% of times.

There is a GCP/GKE specific hack for this, but we don't have a generic solution yet.