[
https://issues.apache.org/jira/browse/YARN-4314?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14982340#comment-14982340
]
Raju Bairishetti commented on YARN-4314:
----------------------------------------
I feel adding timestamp to each resource request will be costly and all the
existing applications will need to migrate to use this metric.
Had a discussion with [~sriksun] earlier about this approach. Resource request
is prepared by AM. In future if we want to use this timestamp as priority for
allocating resources then there is a chance that user/AM can misuse the system
by saying they have older time stamps.
Thinking about this approach:
AppSchedulingInfo has all the scheduling info about an application. When RM
receives first resource a request from AM then RM can note down the system time
as resource request time. Whenever new request comes(i.e.
UpdateResourceRequest() in AppSchedulingInfo or allocate()) then we can measure
how many containers were waited till this time from the last request time. We
can mostly listen on the container request & allocate events.
I will put up a detailed doc with all my thoughts & approaches.
> Adding container wait time as a metric at queue level and application level.
> ----------------------------------------------------------------------------
>
> Key: YARN-4314
> URL: https://issues.apache.org/jira/browse/YARN-4314
> Project: Hadoop YARN
> Issue Type: New Feature
> Reporter: Lavkesh Lahngir
> Assignee: Lavkesh Lahngir
>
> There is a need for adding the container wait-time which can be tracked at
> the queue and application level.
> An application can have two kinds of wait times. One is AM wait time after
> submission and another is total container wait time between AM asking for
> containers and getting them.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
