[ 
https://issues.apache.org/jira/browse/YARN-7461?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16406047#comment-16406047
 ] 

Tao Yang commented on YARN-7461:
--------------------------------

Thanks [~sunilg] for your comments.

{quote}

But my concern is that such a scenario may not compute accurate ratio and could 
impact higher level api uses like DRC#lessThanOrEquals which internally calls a 
RC#compare and eventually comes to ratio. Ideally we want to see whether 
resourceA is less than resourceB and in such case its possible that resourceA 
may have 5 as resource value for resource type X and resourceB may have 0 for 
same type X.

{quote}

DRC#lessThanOrEquals perhaps is Resources#lessThanOrEquals, right? I can't see 
it comes to 

ResourceCalculator#ratio, I think this patch will not affect 
DRC#lessThanOrEquals, please tell me if I missed something. Thanks.

> DominantResourceCalculator#ratio calculation problem when right resource 
> contains zero value
> --------------------------------------------------------------------------------------------
>
>                 Key: YARN-7461
>                 URL: https://issues.apache.org/jira/browse/YARN-7461
>             Project: Hadoop YARN
>          Issue Type: Bug
>    Affects Versions: 3.0.0-alpha4
>            Reporter: Tao Yang
>            Assignee: Tao Yang
>            Priority: Minor
>         Attachments: YARN-7461.001.patch, YARN-7461.002.patch, 
> YARN-7461.003.patch, YARN-7461.004.patch
>
>
> Currently DominantResourceCalculator#ratio may return wrong result when right 
> resource contains zero value. For example, there are three resource types 
> such as <memory-mb, vcores, vssd>, leftResource=<5, 5, 0> and 
> rightResource=<10, 10, 0>, we expect the result of 
> DominantResourceCalculator#ratio(leftResource, rightResource) is 0.5 but 
> currently is NaN.
> There should be a verification before divide calculation to ensure that 
> dividend is not zero.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to