[
https://issues.apache.org/jira/browse/YARN-5829?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15904453#comment-15904453
]
Miklos Szegedi commented on YARN-5829:
--------------------------------------
[~Tao Jie], I like the idea that you described in YARN-5636.
I did some testing and prototyping on this Jira. I am open to a wider solution
as well.
1. What I have learned so far is that preemption has a conflict with the
following code:
{code}
isReservable(capability) &&
reserve(pendingAsk.getPerAllocationResource(), node, reservedContainer,
type, schedulerKey)
{code}
Basically the preempted application will have an excess demand and it will
reserve the lost resources. They will be assigned back to it, when they become
free.
2. We also need to be careful with prioritizing reservations. For example how
it works now is that a reservation takes priority before any other request.
What happens, if I have a preemption from a lower priority request but there is
a demand from a higher priority application?
3. It is a great idea to have a timeout. We also need to take into
consideration proper release of the reservation. If the app is killed, all
reservations should be released.
I do not see such code in the current reservation code for FS.
> FS preemption should reserve a node before considering containers on it for
> preemption
> --------------------------------------------------------------------------------------
>
> Key: YARN-5829
> URL: https://issues.apache.org/jira/browse/YARN-5829
> Project: Hadoop YARN
> Issue Type: Sub-task
> Components: fairscheduler
> Reporter: Karthik Kambatla
> Assignee: Karthik Kambatla
>
> FS preemption evaluates nodes for preemption, and subsequently preempts
> identified containers. If this node is not reserved for a specific
> application, any other application could be allocated resources on this node.
> Reserving the node for the starved application before preempting containers
> would help avoid this.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: yarn-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: yarn-issues-h...@hadoop.apache.org
