[
https://issues.apache.org/jira/browse/YARN-2026?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14011523#comment-14011523
]
Sandy Ryza commented on YARN-2026:
----------------------------------
The nice thing about fair share currently is that it's interpretable as an
amount of resources that, as long as you stay under, you won't get preempted.
Changing it to depend on the running apps in the cluster severely complicates
this. It used to be that each app and queue's fair share was min'd with its
resource usage+demand, which is sort of a continuous analog to what you're
suggesting, but we moved to the current definition when we added multi-resource
scheduling.
I'm wondering if the right way to solve this problem is to allow preemption to
be triggered at higher levels in the queue hierarchy. I.e. suppose we have the
following situation:
* root has two children - parentA and parentB
* each of root's children has two children - childA1, childA2, childB1, and
childB2
* the parent queues' minShares are each set to half of the cluster resources
* the child queue' minShares are each set to a quarter of the cluster resources
* childA1 has a third of the cluster resources
* childB1 and childB2 each have a third of the cluster resources
Even though childA1 is above its fair/minShare, We would see that parentA is
below its minShare, so we would preempt resources on its behalf. Once we have
YARN-596 in, these resources would end up coming from parentB, and end up going
to childA1.
> Fair scheduler : Fair share for inactive queues causes unfair allocation in
> some scenarios
> ------------------------------------------------------------------------------------------
>
> Key: YARN-2026
> URL: https://issues.apache.org/jira/browse/YARN-2026
> Project: Hadoop YARN
> Issue Type: Bug
> Components: scheduler
> Reporter: Ashwin Shankar
> Assignee: Ashwin Shankar
> Labels: scheduler
> Attachments: YARN-2026-v1.txt
>
>
> While using hierarchical queues in fair scheduler,there are few scenarios
> where we have seen a leaf queue with least fair share can take majority of
> the cluster and starve a sibling parent queue which has greater weight/fair
> share and preemption doesn’t kick in to reclaim resources.
> The root cause seems to be that fair share of a parent queue is distributed
> to all its children irrespective of whether its an active or an inactive(no
> apps running) queue. Preemption based on fair share kicks in only if the
> usage of a queue is less than 50% of its fair share and if it has demands
> greater than that. When there are many queues under a parent queue(with high
> fair share),the child queue’s fair share becomes really low. As a result when
> only few of these child queues have apps running,they reach their *tiny* fair
> share quickly and preemption doesn’t happen even if other leaf
> queues(non-sibling) are hogging the cluster.
> This can be solved by dividing fair share of parent queue only to active
> child queues.
> Here is an example describing the problem and proposed solution:
> root.lowPriorityQueue is a leaf queue with weight 2
> root.HighPriorityQueue is parent queue with weight 8
> root.HighPriorityQueue has 10 child leaf queues :
> root.HighPriorityQueue.childQ(1..10)
> Above config,results in root.HighPriorityQueue having 80% fair share
> and each of its ten child queue would have 8% fair share. Preemption would
> happen only if the child queue is <4% (0.5*8=4).
> Lets say at the moment no apps are running in any of the
> root.HighPriorityQueue.childQ(1..10) and few apps are running in
> root.lowPriorityQueue which is taking up 95% of the cluster.
> Up till this point,the behavior of FS is correct.
> Now,lets say root.HighPriorityQueue.childQ1 got a big job which requires 30%
> of the cluster. It would get only the available 5% in the cluster and
> preemption wouldn't kick in since its above 4%(half fair share).This is bad
> considering childQ1 is under a highPriority parent queue which has *80% fair
> share*.
> Until root.lowPriorityQueue starts relinquishing containers,we would see the
> following allocation on the scheduler page:
> *root.lowPriorityQueue = 95%*
> *root.HighPriorityQueue.childQ1=5%*
> This can be solved by distributing a parent’s fair share only to active
> queues.
> So in the example above,since childQ1 is the only active queue
> under root.HighPriorityQueue, it would get all its parent’s fair share i.e.
> 80%.
> This would cause preemption to reclaim the 30% needed by childQ1 from
> root.lowPriorityQueue after fairSharePreemptionTimeout seconds.
> Also note that similar situation can happen between
> root.HighPriorityQueue.childQ1 and root.HighPriorityQueue.childQ2,if childQ2
> hogs the cluster. childQ2 can take up 95% cluster and childQ1 would be stuck
> at 5%,until childQ2 starts relinquishing containers. We would like each of
> childQ1 and childQ2 to get half of root.HighPriorityQueue fair share ie
> 40%,which would ensure childQ1 gets upto 40% resource if needed through
> preemption.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
