[
https://issues.apache.org/jira/browse/HADOOP-2568?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12565291#action_12565291
]
Amar Kamat commented on HADOOP-2568:
------------------------------------
I guess we are talking of following things here :
1) _Bundler_ : Bundles map outputs from different maps on a node for a reducer
at the destination node.
2) _Separator_ : Separates a single map's output for multiple reducers on the
destination node.
3) _Bundler_ and _Separator_ : 1 + 2
The _Bundler_ assumes that multiple maps on the source node gets completed at
the same time so that their outputs for one reducer can be bundled and shipped
across the network. This will require a common _BUNDLER_ at the map side. This
might not be true since maps might finish at different times or might fail and
get re-executed.
The _Separator_ assumes that there is a common _SHUFFLER_ at the destination
node and the reducers are pinned to each other every time. But reducers might
fail and get re-executed. Taking care of these dynamics will add a lot of
complexity. Not sure how much this will help. Corrections/clarifications?
> Pin reduces with consecutive IDs to nodes and have a single shuffle task per
> job per node
> -----------------------------------------------------------------------------------------
>
> Key: HADOOP-2568
> URL: https://issues.apache.org/jira/browse/HADOOP-2568
> Project: Hadoop Core
> Issue Type: Improvement
> Components: mapred
> Reporter: Devaraj Das
> Assignee: Devaraj Das
> Fix For: 0.17.0
>
>
> The idea is to reduce disk seeks while fetching the map outputs. If we
> opportunistically pin reduces with consecutive IDs (like 5, 6, 7 ..
> max-reduce-tasks on that node) on a node, and have a single shuffle task, we
> should benefit, if for every fetch, that shuffle task fetches all the outputs
> for the reduces it is shuffling for. In the case where we have 2 reduces per
> node, we will decrease the #seeks in the map output files on the map nodes by
> 50%. Memory usage by that shuffle task would be proportional to the number of
> reduces it is shuffling for (to account for the number of ramfs instances,
> one per reduce). But overall it should help.
> Thoughts?
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
