[
https://issues.apache.org/jira/browse/SPARK-2774?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14081300#comment-14081300
]
Josh Rosen commented on SPARK-2774:
-----------------------------------
I found two research papers discussing locality-aware reduce task scheduling,
which might provide extra context / motivation:
[Center-of-Gravity Reduce Task Scheduling to Lower MapReduce Network
Traffic|http://www.qatar.cmu.edu/~mhhammou/CoGRS-CLOUD2012.pdf]
and (earlier work by the same group):
[Locality-aware reduce task scheduling for
MapReduce|http://www.qatar.cmu.edu/~mhhammou/mhh_msakr_LARTS_CloudCom2011_paper.pdf]
> Set preferred locations for reduce tasks
> ----------------------------------------
>
> Key: SPARK-2774
> URL: https://issues.apache.org/jira/browse/SPARK-2774
> Project: Spark
> Issue Type: Improvement
> Components: Spark Core
> Reporter: Shivaram Venkataraman
>
> Currently we do not set preferred locations for reduce tasks in Spark. This
> patch proposes setting preferred locations based on the map output sizes and
> locations tracked by the MapOutputTracker. This is useful in two conditions
> 1. When you have a small job in a large cluster it can be useful to co-locate
> map and reduce tasks to avoid going over the network
> 2. If there is a lot of data skew in the map stage outputs, then it is
> beneficial to place the reducer close to the largest output.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
