[
https://issues.apache.org/jira/browse/YARN-6210?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Karthik Kambatla updated YARN-6210:
-----------------------------------
Attachment: YARN-6210.1.patch
Patch (v1) does the following:
# Only starved apps can reserve nodes
# The starvation check considers a single resource. For instance, in the case
of DRF, this corresponds to the dominant resource: (1) the app is considered
starved if it is starved on the dominant resource, (2) an app can be preempted
as long as it is not starved on the dominant resource.
# FSAppAttempt#toString for easier debugging of tests
# DominantResourceFairnessComparator#compare can return 0 if the submit time is
the same. This leads to issues while allocating containers. Changed this to be
in line with FairShareComparator and return either -1 or 1.
# Added the test from YARN-6151. FYI. [~yufeigu]
> FS: Node reservations can interfere with preemption
> ---------------------------------------------------
>
> Key: YARN-6210
> URL: https://issues.apache.org/jira/browse/YARN-6210
> Project: Hadoop YARN
> Issue Type: Bug
> Components: fairscheduler
> Affects Versions: 2.9.0
> Reporter: Karthik Kambatla
> Assignee: Karthik Kambatla
> Attachments: YARN-6210.1.patch
>
>
> Today, on a saturated cluster, apps with pending demand reserve nodes. A new
> app might not be able to preempt resources because these nodes are already
> reserved. This can be reproduced by the example in YARN-6151.
> Since node reservations are to prevent starvation of apps requesting large
> containers, triggering these reservations only on starved applications would
> avoid this situation.
--
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
