hadoop-yarn-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Zoltan Siegl (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (YARN-8707) It's not reasonable to decide whether app is starved by fairShare
Date Fri, 05 Oct 2018 21:41:00 GMT

    [ https://issues.apache.org/jira/browse/YARN-8707?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16640377#comment-16640377

Zoltan Siegl commented on YARN-8707:

Hi [~uranus],

I have been trying to understand this deeper. 

One thing is that to me it seems that this is a bug indeed, but it should not cause any real
problems, as {{Resource fairShareStarvation()}} uses the correct calculation, so it will go
- Are you starved?
- Hell yes!
- And how many resources do you need?
- Amm.. None.

So no preemption is done in favour for the app in question.
However this should still be fixed.

According to the patch:

Resource fairShareStarvation() {
    long now = scheduler.getClock().getTime();
    Resource threshold = Resources.multiply(
        getFairShare(), getQueue().getFairSharePreemptionThreshold());
    Resource fairDemand = Resources.componentwiseMin(threshold, demand);

This code path here considers {{getQueue().getFairSharePreemptionThreshold()}}, unlike the
current patch, and I believe that should be the correct approach to this. 
Actually the whole code around here is a bit fuzzy. 
What we probably should do here is wise up the {{isStarvedForFairShare()}} method so it takes
care of:
- calculating fair demand
- being starved for fair demand
- being in or out of preemption time treshold
while {{Resource fairShareStarvation()}} should call {{isStarvedForFairShare()}} and if so
it should return the amount, else empty resource.

Also we definitely need some unit tests around this area, as using the word "obvious" seems
a bit brave in the world of fair scheduler. :)

> It's not reasonable to decide whether app is starved by fairShare
> -----------------------------------------------------------------
>                 Key: YARN-8707
>                 URL: https://issues.apache.org/jira/browse/YARN-8707
>             Project: Hadoop YARN
>          Issue Type: Bug
>          Components: fairscheduler
>    Affects Versions: 3.0.0-alpha3
>            Reporter: Zhaohui Xin
>            Assignee: Zhaohui Xin
>            Priority: Minor
>         Attachments: YARN-8707.patch
> When app's usage reached demand, it's still be considered fairShare starved. Obviously,
that's not reasonable!
> {code:java}
> boolean isStarvedForFairShare() {
> return isUsageBelowShare(getResourceUsage(), getFairShare());
> }
> {code}

This message was sent by Atlassian JIRA

To unsubscribe, e-mail: yarn-issues-unsubscribe@hadoop.apache.org
For additional commands, e-mail: yarn-issues-help@hadoop.apache.org

View raw message