[
https://issues.apache.org/jira/browse/HADOOP-3285?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12590939#action_12590939
]
Amar Kamat commented on HADOOP-3285:
------------------------------------
The namenode (via {{Namenode.getBlockLocations()}}) returns different block
location information than the one in {{split.getLocations()}}. Note that
{{split.getLocations()}} is used for task cache creation in JobInProgress while
{{Namenode.getBlockLocations()}} is used by the DFS client for pulling the
split before starting the actual map phase. One possible problem is 2027. We
are investigating.
> map tasks with node local splits do not always read from local nodes
> --------------------------------------------------------------------
>
> Key: HADOOP-3285
> URL: https://issues.apache.org/jira/browse/HADOOP-3285
> Project: Hadoop Core
> Issue Type: Bug
> Components: mapred
> Reporter: Runping Qi
>
> I ran a simple map/reduce job counting the number of records in the input
> data.
> The number of reducers was set to 1.
> I did not set the number of mappers. Thus by default, all splits except the
> last split of a file contain one dfs block (128MB in my case).
> The web gui indicated that 99% of map tasks were with local splits.
> Thus I expected that most of the dfs reads should have come from the local
> data nodes.
> However, when I examine the traffic of the ethernet interfaces,
> I found about 50% traffic of each node were through the loopback interface
> and other 50% were through the ethernet card!
> Also, the switch monitoring indicated that a lot of traffic went through the
> links and cross racks!
> This indicated that the data locality feature does not work as expected.
> To confirm that, I set the number of map tasks to a very high number so that
> it forced the split size down to about 27MB.
> The web gui indicated that 99% of map tasks were with local splits, as
> expected.
> The ethernet interface monitor showed that almost 100% traffic went through
> the loopback interface, as it should be.
> I found about 50% traffic of each node were through the loopback interface
> and other 50% were through the ethernet card!
> Also, the switch monitoring indicated that there were very little traffic
> through the links and cross racks.
> This implies that some corner cases are not handled properly.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
