[
https://issues.apache.org/jira/browse/YUNIKORN-580?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17305846#comment-17305846
]
Wilfred Spiegelenburg commented on YUNIKORN-580:
------------------------------------------------
This is important but could be a complex part to implement. The first choice is
where do we implement this. Based on the fact that the scheduler is the one
that tracks we could do it in the core. On the other hand if we want to reject
early we could do this in the shim but that must be done in every shim then.
Both are possible and have their pros and cons.
We then need to decide on what the behaviour is that we want to allow or
implement. If we do not just want to say they must be the same, and enforce
that, we need to look at the how and what. There are two cases that we could
distinguish if we don't enforce equality:
# adjust the real ask resource request
# schedule as a non task group member
For case 1 we can easily adjust what gets accounted for in the real allocation.
We adjust the resource ask for the real allocation to the placeholder ask. That
could be simple to implement. The real use on K8s will be different than what
the scheduler knows this might have side effects like auto scaling decisions.
In case 2 we can schedule the real task as a non gang task if the resources are
not equal. We then release the placeholder early. That keeps everything in
check nicely too. The side effect could be that you get an issue with quotas on
the queues if the placeholders are smaller than the real allocation.
So both ways of fixing have their drawbacks. Completely rejecting the ask is
also possible but that has other implications.
> Check if the placeholders and real resources are the same
> ---------------------------------------------------------
>
> Key: YUNIKORN-580
> URL: https://issues.apache.org/jira/browse/YUNIKORN-580
> Project: Apache YuniKorn
> Issue Type: Sub-task
> Reporter: Kinga Marton
> Priority: Critical
>
> The design doc stated that the placeholder resources and the real resource
> ask should be the same. However we are not checking it, so it is possible to
> ask for more or less resources, what might cause some unexpected issues.
> Also the resource cleanup is not working properly in this case.
> For example if we have higher resources set for the placeholders than for the
> real pods, after the cleanup the resource difference still be there in the
> queue and node.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
