[ 
https://issues.apache.org/jira/browse/YARN-2022?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14043104#comment-14043104
 ] 

Sunil G commented on YARN-2022:
-------------------------------

Thank You [~mayank_bansal] for the review. I understood the scenario which you 
explained and split-up of 50% should be 5% + 45%. 
bq.Resource maxAMCapacity = 
Resources.multiply(Resources.multiply(clusterResource,  
qT.leafQueue.getAbsoluteMaximumCapacity()),   
qT.leafQueue.getMaxAMResourcePerQueuePercent());
Yes, I used *getMaxAMResourcePerQueuePercent* with *getAbsoluteMaximumCapacity* 
to get the maxAMCapacity in a Queue. *getAbsoluteCapacity* was a choice instead 
of *getAbsoluteMaximumCapacity*.
But when I checked LeafQueue for *maxActiveApplications* calculation, I found 
that getAbsoluteMaximumCapacity was used there.

We can use *getAbsoluteCapacity* here for calculating the AMCapacity for a 
Queue. With this approach, there are chances that more AM will be killed under 
scenarios like below.
For eg, Queue A with Capacity 25%, Max capacity 100%, Queue B with capacity 
75%, AM Percent is 0.5. 
If QueueA is using 100% of cluster with 50% of AM (it is possible with 
different users) then with a big demand from QueueB, 37.5% of AMs from QueueA 
also can be preempted. I feel this is fine as you suggested. 

Could I go along by changing maxAMCapacity w.r.t getAbsoluteCapacity?

> Preempting an Application Master container can be kept as least priority when 
> multiple applications are marked for preemption by 
> ProportionalCapacityPreemptionPolicy
> ---------------------------------------------------------------------------------------------------------------------------------------------------------------------
>
>                 Key: YARN-2022
>                 URL: https://issues.apache.org/jira/browse/YARN-2022
>             Project: Hadoop YARN
>          Issue Type: Sub-task
>          Components: resourcemanager
>    Affects Versions: 2.4.0
>            Reporter: Sunil G
>            Assignee: Sunil G
>         Attachments: YARN-2022-DesignDraft.docx, YARN-2022.2.patch, 
> YARN-2022.3.patch, YARN-2022.4.patch, YARN-2022.5.patch, YARN-2022.6.patch, 
> YARN-2022.7.patch, Yarn-2022.1.patch
>
>
> Cluster Size = 16GB [2NM's]
> Queue A Capacity = 50%
> Queue B Capacity = 50%
> Consider there are 3 applications running in Queue A which has taken the full 
> cluster capacity. 
> J1 = 2GB AM + 1GB * 4 Maps
> J2 = 2GB AM + 1GB * 4 Maps
> J3 = 2GB AM + 1GB * 2 Maps
> Another Job J4 is submitted in Queue B [J4 needs a 2GB AM + 1GB * 2 Maps ].
> Currently in this scenario, Jobs J3 will get killed including its AM.
> It is better if AM can be given least priority among multiple applications. 
> In this same scenario, map tasks from J3 and J2 can be preempted.
> Later when cluster is free, maps can be allocated to these Jobs.



--
This message was sent by Atlassian JIRA
(v6.2#6252)

Reply via email to