[
https://issues.apache.org/jira/browse/MAPREDUCE-2729?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13079451#comment-13079451
]
Sherry Chen commented on MAPREDUCE-2729:
----------------------------------------
I manually verified this fix on the 10 nodes cluster.
Verification steps:
1. Replace hadoop-capacity-scheduler.jar with the fix on the cluster gateway
2. Modify the capacity-scheduler.xml to ensure a queue have multiple map &
reduce task slots
3. restart mapred
4. Submit jobs for a user which start reduces when 5% (default) maps complete,
submit jobs for 2nd user (same queue as 1st user) which start reduces when 50%
maps complete.
5. Verify that 1st user got all queue reduce capacity whatever the 2nd user
hasn't used yet, it is greater than user-limit.
> Reducers are always counted having "pending tasks" even if they can't be
> scheduled yet because not enough of their mappers have completed
> -----------------------------------------------------------------------------------------------------------------------------------------
>
> Key: MAPREDUCE-2729
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-2729
> Project: Hadoop Map/Reduce
> Issue Type: Improvement
> Affects Versions: 0.20.205.0
> Environment: 0.20.1xx-Secondary
> Reporter: Sherry Chen
> Assignee: Sherry Chen
> Fix For: 0.20.205.0
>
> Attachments: MAPREDUCE-2729.patch
>
>
> In capacity scheduler, number of users in a queue needing slots are
> calculated based on whether users' jobs have any pending tasks.
> This works fine for map tasks. However, for reduce tasks, jobs do not need
> reduce slots until the minimum number of map tasks have been completed.
> Here, we add checking whether reduce is ready to schedule (i.e. if a job has
> completed enough map tasks) when we increment number of users in a queue
> needing reduce slots.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
