[
https://issues.apache.org/jira/browse/MAPREDUCE-4893?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Bikas Saha updated MAPREDUCE-4893:
----------------------------------
Attachment: MAPREDUCE-4893.1.patch
Attaching patch that fixes the issue by making a node-local pass followed by a
rack-local pass for the maps. The performance should not be different from
earlier since the actual operations on a container are still identical. The
only overhead is an extra pass over the list of containers. New test for the
specific fix. Assuming existing TestRMContainerAllocator tests will verify
errors in refactoring. The refactoring keeps the ealier logic intact except for
a 2 pass assignment cycle for maps.
> MR AppMaster can do sub-optimal assignment of containers to map tasks leading
> to poor node locality
> ---------------------------------------------------------------------------------------------------
>
> Key: MAPREDUCE-4893
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-4893
> Project: Hadoop Map/Reduce
> Issue Type: Bug
> Reporter: Bikas Saha
> Assignee: Bikas Saha
> Fix For: 3.0.0
>
> Attachments: MAPREDUCE-4893.1.patch
>
>
> Say the MR AppMaster asks the RM for 3 containers on nodes n1, n2 and n3.
> There are 10 node n1-n10 in the same rack. The RM can give it allocated
> containers in the list order n5, n2, n1. The way AM map->container assignment
> happens, the AM will try to assign node local maps to n5, failing which it
> will assign rack local maps to n5. These rack local maps could be node local
> on n2 and n1 and would have been assigned to containers on n1 and n2 if the
> AM had not made an early rack local match for them on n5. This can lead to
> poor locality.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
