Terminate on migrate for instances with GPU's attached. about google-compute-engine-plugin HOT 9 CLOSED

One thing I see so far as a block is how to emulate live migration in integration testing.

from google-compute-engine-plugin.

Decided to look into this once you mentioned it. I think this should be usable, but the documentation is pretty sparse: https://cloud.google.com/compute/docs/reference/rest/v1/instances/simulateMaintenanceEvent

It's sparse in the sense that the only explanation given both on that page and the gcloud docs page is "simulates a maintenance event". I'm 90% sure that it means your instance would behave as if a real maintenance event happened (i.e. terminate if required). The other 10% is that it may be a dry run and generate logs without actually doing anything to the instance. That should be easy to test manually at least.

from google-compute-engine-plugin.

@RaphC @stephenashank The problem is not with simulating host maintanance
The problem is that we are trying to set wrong value for OnHostMaintenance field when instance gave GPU added..
By default that value is set to migrate but with GPU it needs to be switched to terminate see
https://cloud.google.com/compute/docs/gpus/#restrictions

Now using instances with GPU is available via templates.
My personal opinion is that we should remove support for GPUs in plugin and leave it only via templates (it never worked anyway).

from google-compute-engine-plugin.

@rachely3n Why you have copied #63??
It would be much easier to have 1 thread..

from google-compute-engine-plugin.

@ingwarsw simulating the host maintenance in this case would just be for unit testing after I make it so that instances with GPU's terminate on migrate.

what do you mean by

My personal opinion is that we should remove support for GPUs in plugin and leave it only via templates (it never worked anyway).

If it's available in the templates, shouldn't I still support GPU's in the plugin?

from google-compute-engine-plugin.

Plugin dont need to support GPU (and never actually did)..
If plugin supports templates most things can be done via them..

Rarely used things that are supported via plugin but could be only supported via templates (and they would made plugin bit simpler)

• GPU
• Minimum Cpu Platform

from google-compute-engine-plugin.

@ingwarsw simulating the host maintenance in this case would just be for unit testing after I make it so that instances with GPU's terminate on migrate.

But we are never testing migration..
And Im not sure we should, its GCE feature not ours..

from google-compute-engine-plugin.

@ingwarsw
I'm not sure I understand what's going on here. If GPU's can be added via templates, it can be used in the plugin then. I would then want to fix the migration problem for GPU instances by terminating them.

from google-compute-engine-plugin.

Going to keep this at #63 since yes it is easier to just have 1 thread to track this issue.

from google-compute-engine-plugin.