[
https://issues.apache.org/jira/browse/YARN-6193?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15870838#comment-15870838
]
Daniel Templeton commented on YARN-6193:
----------------------------------------
Thanks, [~kasha]. A couple more comments:
* {{SchedulingPolicy.instances}} should be {{INSTANCES}}. Not your fault, but
while you're in there...
* Why don't the want the comparator to be static?
* You've changed the behavior of {{FifoPolicy.checkIfUsageOverFairShare()}}.
Are you sure that's the right thing to do?
* You should explain in this JIRA thread why you're upping the test iterations
from 100 back to 1000.
* On a philosophical note, {{checkIfUsageOverFairShare()}} and
{{checkIfUsageUnderFairShare()}} don't actually have anything to do with the
fair share. They're just resource comparisons. Should the {{fairShare}} arg
maybe be replaced with a call to {{getFairShare()}}?
* I took some time to stare at the code for a bit, and I don't see where this
{code}- Resource currentUsage = getResourceUsage();
- Resource fairshare = getFairShare();
- Resource overFairShareBy = Resources.subtract(currentUsage, fairshare);
-
- return (Resources.fitsIn(container.getAllocatedResource(),
- overFairShareBy));
+ Resource usageAfterPreemption = Resources.subtract(
+ getResourceUsage(), container.getAllocatedResource());
+ return !fsQueue.getPolicy().checkIfUsageUnderFairShare(
+ usageAfterPreemption, getFairShare());{code} actually makes any
difference. Seems to be the same operation to me, assuming that the app is
over its fair share.
> FairScheduler might not trigger preemption when using DRF
> ---------------------------------------------------------
>
> Key: YARN-6193
> URL: https://issues.apache.org/jira/browse/YARN-6193
> Project: Hadoop YARN
> Issue Type: Sub-task
> Components: fairscheduler
> Affects Versions: 2.8.0
> Reporter: Karthik Kambatla
> Assignee: Karthik Kambatla
> Attachments: YARN-6193.000.patch, YARN-6193.001.patch,
> YARN-6193.002.patch
>
>
> {{FSAppAttempt#canContainerBePreempted}} verifies preempting a container
> doesn't lead to new starvation ({{Resources.fitsIn}}). When using DRF, this
> leads to verifying both resources instead of just the dominant resource.
> Ideally, the check should be based on policy.
> Note that current implementation of
> {{DominantResourceFairnessPolicy#checkIfUsageOverFairShare}} is broken.
--
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
