[
https://issues.apache.org/jira/browse/YARN-9580?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16851623#comment-16851623
]
Tao Yang commented on YARN-9580:
--------------------------------
Found another problem about reservation when multi-node enabled, scheduler
iterates all nodes and try to allocate for reserved container in
LeafQueue#allocateFromReservedContainer. Here there are two problems:
# The node of reserved container should be taken as candidates instead of all
nodes when calling FiCaSchedulerApp#assignContainers, otherwise later scheduler
may generate a reservation-fulfilled proposal on another node, which will
always be rejected when committing.
# Assignment returned by FiCaSchedulerApp#assignContainers could never be null
even if it's just skipped, it will break the normal scheduling process for this
leaf queue because of the if clause in LeafQueue#assignContainers: "if (null !=
assignment) \{ return assignment;}"
> Fulfilled reservation information in assignment is lost when transferring in
> ParentQueue#assignContainers
> ---------------------------------------------------------------------------------------------------------
>
> Key: YARN-9580
> URL: https://issues.apache.org/jira/browse/YARN-9580
> Project: Hadoop YARN
> Issue Type: Bug
> Components: capacityscheduler
> Reporter: Tao Yang
> Assignee: Tao Yang
> Priority: Major
> Attachments: YARN-9580.001.patch
>
>
> When transferring assignment from child queue to parent queue, fulfilled
> reservation information including fulfilledReservation and
> fulfilledReservedContainer in assignment is lost.
> When multi-nodes enabled, this lost can raise a problem that allocation
> proposal is generated but can't be accepted because there is a check for
> fulfilled reservation information in
> FiCaSchedulerApp#commonCheckContainerAllocation, this endless loop will
> always be there and the resource of the node can't be used anymore.
> In HB-driven scheduling mode, fulfilled reservation can be allocated via
> another calling stack: CapacityScheduler#allocateContainersToNode -->
> CapacityScheduler#allocateContainerOnSingleNode -->
> CapacityScheduler#allocateFromReservedContainer, in this way assignment can
> be generated by leaf queue and directly submitted, I think that's why we
> hardly find this problem before.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
