[
https://issues.apache.org/jira/browse/MAPREDUCE-5873?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14155848#comment-14155848
]
Joep Rottinghuis commented on MAPREDUCE-5873:
---------------------------------------------
[~jlowe] are you guys seeing this behavior as well ?
We're finding that the current way that the shuffle bandwidth is shows is not
only useless, but also confusing for our users.
If a mapper has to be re-run it appears as if the bandwidth is a tiny amount.
This new way of showing bandwidth is a much better indication of what is going
on, and if it is very small, it is actually in indication that there is
something not right in the network.
> Measure bw of a single copy call and display the correct aggregated bw
> ----------------------------------------------------------------------
>
> Key: MAPREDUCE-5873
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-5873
> Project: Hadoop Map/Reduce
> Issue Type: Bug
> Affects Versions: 2.3.0
> Reporter: Siqi Li
> Assignee: Siqi Li
> Attachments: MAPREDUCE-5873.v1.patch, MAPREDUCE-5873.v2.patch,
> MAPREDUCE-5873.v3.patch
>
>
> Currently ShuffleScheduler in ReduceTask JVM status displays bandwidth. Its
> definition however is confusing because it captures the time where there is
> no copying because there is a pause between when new wave of map outputs is
> available.
> current bw is definded as (bytes copied so far) / (total time in the copy
> phase so far)
> It would be more useful
> 1) to measure bandwidth of a single copy call.
> 2) display aggregated bw as long as there is at least one fetcher is in the
> copy call.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
