[
https://issues.apache.org/jira/browse/YUNIKORN-3092?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18004297#comment-18004297
]
John Daciuk commented on YUNIKORN-3092:
---------------------------------------
[~wwei] sorry for the late response, I've been away for 1.5 weeks. Thanks for
your questions and to answer them:
We observed a linear increase in memory until the OOM. Not sure if that would
eventually flatten out if we increased memory. But don't think we will try that
at the moment as we still observed preemption misses even with the env var.
Unfortunately I didn't see any telling logs during that time, so I don't think
sharing would help. When we get a chance we can try to repro the OOM and then
share more info and try to get to the bottom of it.
[~mani] Thanks for taking up this issue! Your PR LGTM and looking forward to
upgrading to that release. Personally we'd like to also cancel reservations of
equal priority, but I understand that wouldn't appeal to everyone. Since
reservations cause complications, I wonder if disabling them should be
officially supported and a configmap option? Also what do you think about more
visibility around reservations? If we had a metric to tell us what nodes are
currently reserved by what pods, that would resolve a lot of user questions we
get about why pods are pending when we have available GPUs (potentially after
preemption).
Another question for you both: how might queue set up affect preemption? We
actually ONLY set priority-class on our pods and not at all on our queue
config. Is this a mistake? Would also setting priority on our queues
potentially improve behavior?
> Reservations can permanently block nodes, leading to preemption failure and a
> stuck scheduler state
> ---------------------------------------------------------------------------------------------------
>
> Key: YUNIKORN-3092
> URL: https://issues.apache.org/jira/browse/YUNIKORN-3092
> Project: Apache YuniKorn
> Issue Type: Bug
> Components: core - scheduler
> Affects Versions: 1.6.3
> Reporter: John Daciuk
> Assignee: Manikandan R
> Priority: Minor
> Labels: preemption, pull-request-available
> Attachments: Screenshot 2025-06-23 at 8.28.55 PM.png
>
> Original Estimate: 336h
> Remaining Estimate: 336h
>
> h2. Context
> Since deploying Yunikorn back in October 2024 we've encountered occasional
> preemption misses. We find a high priority pod pending for hours, manually
> delete a low priority pod, then see the high priority pod schedule.
> We dug into this earlier this year and found the upgrade from 1.5.2 to 1.6.2
> helpful. In particular [this
> PR|https://github.com/apache/yunikorn-core/pull/1001] achieved 100% expected
> preemption in our testing due to it's reservation removal logic. However we
> still find that 1.6.2 is not reliable with respect to preemption in practice.
> And we can repro preemption misses by scaling up our original preemption load
> test by 4x.
> h2. Repro
> With Yunikorn 1.6.3 schedule ~400 low priority pods that live forever and
> fill up all node capacity. Once they are running, schedule the same number of
> high priority pods to a different queue. Use the same resources for all the
> pods.
> We expect that all the high priority pods will eventually schedule. However
> we find about 10% of them stuck pending. This can be seen in the screenshot
> attached, where the high priority pods are tier0.
> If we add logging like in [this diff from
> branch-1.6|https://github.com/apache/yunikorn-core/compare/branch-1.6...jdaciuk:yunikorn-core:jdaciuk-1.6]
> we see
> {quote}{{{}2025-06-23T04:54:26.776Z INFO core.scheduler.preemption
> objects/preemption.go:93 Removing node ip-100-76-60-239.ec2.internal from
> consideration. node.IsReserved: true, node reservations:
> map[847f05e1-f74c-403a-8033-154cd76d89c0:ip-100-76-60-239.ec2.internal ->
> tier0-1-395-157140|847f05e1-f74c-403a-8033-154cd76d89c0], node fits ask: true
> {"applicationID": "tier0-1-406-328120", "allocationKey":
> "e589c683-faf1-4793-97b8-c5f3b3bc34b5", "author": "MLP"{}}}}
> {quote}
> A node (with tons of potential victims) ip-100-76-60-239.ec2.internal is
> removed from consideration for preemption because it's reserved. Looking at
> the reservation map above, we see that pod tier0-1-395 has the reservation.
> The pod tier0-1-395 is blocking the entire node. Why can't it schedule and
> release the reservation?
> {quote}{{{}2025-06-23T04:43:45.942Z INFO core.scheduler.application
> objects/application.go:1008 tryAllocate did not find a candidate
> allocation in the node iterator, allowPreemption: true,
> preemptAttemptsRemaining: 0 {"applicationID": "tier0-1-395-157140",
> "author": "MLP"{}}}}
> {quote}
> Because there's no more preemption attempts allowed for the particular queue
> this cycle. And unfortunately this situation repeats itself every cycle since
> pod tier0-1-395 is not among the first in the queue to ever tryAllocate.
> h2. Thoughts
> This is one example, but there are a number of ways we can get stuck in such
> a cycle. Particular to the preemption failure here, it seems like we need
> some way to either remove the dead reservation or ignore it while considering
> preemption victims.
> So for example, when we iterate through the nodes in [this
> code|https://github.com/apache/yunikorn-core/blob/master/pkg/scheduler/objects/preemption.go#L163],
> perhaps we could first try with filtering out reserved nodes (as the code
> is) then try another loop ignoring and/or breaking reservations if we find
> victims. Would ignoring the reservation be enough, or do we have to delete it
> for the preemption to then result in scheduling?
> We'd love to get feedback as to the following
> * Is passing a test like described above even a goal of Yunikorn preemption?
> * If so, how can we be more strategic about releasing reservations that
> become major blockers, esp. in the preemption context?
> * We don't suppose there's a simple way to opt out of the reservation
> feature altogether is there? We don't ever want a reservation to block a
> node. If the pod can't schedule in the current cycle, we'd like it to wait
> without a reservation (in our case a full node will always free up at some
> point all at once). Or is there something we're misunderstanding that makes
> us need reservations?
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
