[
https://issues.apache.org/jira/browse/YARN-4280?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14971154#comment-14971154
]
Jason Lowe commented on YARN-4280:
----------------------------------
It's a sticky problem. The problem with doing the resource check is that it
can prevent the reservation from being fulfilled indefinitely. For example,
consider a situation like this:
* root queue (near 100% utilization)
** parent queue P (near max capacity)
*** leaf queue A (well under capacity)
*** leaf queue B (almost all of P's utilization)
** leaf queue C (the remainder of root - P)
We have an application X in queue A that needs a large resource. If we do a
limit check against P's max capacity or the root's max capacity, it won't fit.
If we don't make the reservation, then the app in A could be indefinitely
postponed. So let's say we go ahead and let the reservation occur. If the
resource to fill that reservation was freed from within the P queue hierarchy
then we're OK. If it's not, then we cannot fulfill the reservation otherwise
we run over P's max capacity. So in the latter case, do we leave the
reservation? Does this in turn prevent apps in C from making progress because
app X's reservations start locking down the cluster, waiting for the apps in
queue B to free up resources?
Offhand I don't have a great answer for how to tackle the problem. Seems like
either we need to start locking down parts of the cluster and potentially leave
resources fallow, even for other queues outside of P, to make sure app X will
eventually get something or we keep app X from reserving and leave it
vulnerable to indefinite postponement despite containers churning in queue B.
It's like we need to make a reservation _within_ the P queue hierarchy for this
scenario, to make sure queue B isn't allowed to grab more resources while app X
is waiting, but not sure that's right either.
> CapacityScheduler reservations may not prevent indefinite postponement on a
> busy cluster
> ----------------------------------------------------------------------------------------
>
> Key: YARN-4280
> URL: https://issues.apache.org/jira/browse/YARN-4280
> Project: Hadoop YARN
> Issue Type: Bug
> Components: capacity scheduler
> Affects Versions: 2.6.1, 2.8.0, 2.7.1
> Reporter: Kuhu Shukla
> Assignee: Kuhu Shukla
>
> Consider the following scenario:
> There are 2 queues A(25% of the total capacity) and B(75%), both can run at
> total cluster capacity. There are 2 applications, appX that runs on Queue A,
> always asking for 1G containers(non-AM) and appY runs on Queue B asking for 2
> GB containers.
> The user limit is high enough for the application to reach 100% of the
> cluster resource.
> appX is running at total cluster capacity, full with 1G containers releasing
> only one container at a time. appY comes in with a request of 2GB container
> but only 1 GB is free. Ideally, since appY is in the underserved queue, it
> has higher priority and should reserve for its 2 GB request. Since this
> request puts the alloc+reserve above total capacity of the cluster,
> reservation is not made. appX comes in with a 1GB request and since 1GB is
> still available, the request is allocated.
> This can continue indefinitely causing priority inversion.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
