[
https://issues.apache.org/jira/browse/MAPREDUCE-284?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12743141#action_12743141
]
Arun C Murthy commented on MAPREDUCE-284:
-----------------------------------------
bq. I still don't see a strong reason for reducing the ping interval.
The reasoning behind reducing rpc ping intervals was to allow the child tasks
to 'fail fast' when sockets to either the TaskTracker or the DataNode were
unexpectedly shutdown.
bq. Also, I am worried about the approach followed in the patch. In particular,
i don't like the way that ping interval is hardcoded to 5 seconds in the Child
process (overriding the configuration passed). I also think that in the current
implementation, it affects the task->namenode ping interval which is not
desirable.
Fair enough, any other ideas?
bq. I propose that for this patch, we just fix the case of TCP_NODELAY being ON
by default.
+1 if we can't come up with better heuristics. Maybe in that case we can open
another jira to track it?
> Improvements to RPC between Child and TaskTracker
> -------------------------------------------------
>
> Key: MAPREDUCE-284
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-284
> Project: Hadoop Map/Reduce
> Issue Type: Improvement
> Reporter: Arun C Murthy
> Assignee: Ravi Gummadi
> Fix For: 0.21.0
>
> Attachments: MR-284.patch
>
>
> We could improve the RPC between the Child and TaskTracker:
> * Set ping interval lower by default to 5s
> * Disable nagle's algorithm (tcp no-delay)
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
