[
https://issues.apache.org/jira/browse/HBASE-11297?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14025402#comment-14025402
]
Nicolas Liochon commented on HBASE-11297:
-----------------------------------------
Hum, it's difficult to have something obviously better. The attach patch seems
to bring some improvements. It's very difficult to test. Some points:
- with the current code, if we stay under the socket buffer size, we just queue
on the lock. The responder is not really involved. That's also why we're seeing
the lock on responseQueue. When the buffer size is reached, then the behavior
changes. This depends a lot on the network or client (mis)behavior. The patch
changes this.
- I've used ??sudo sysctl net.ipv4.tcp_wmem="1024 1024 1024"?? to change the
buffer size and test the correctness.
- the performances are very sensitive to warming.
- I've also changed PerformanceEvaluation to share the connection.
I haven't run all the unit test, I'm going to use hadoop qa for this...
> Remove some synchros in the rpcServer responder
> -----------------------------------------------
>
> Key: HBASE-11297
> URL: https://issues.apache.org/jira/browse/HBASE-11297
> Project: HBase
> Issue Type: Bug
> Components: regionserver
> Affects Versions: 0.99.0
> Reporter: Nicolas Liochon
> Assignee: Nicolas Liochon
> Fix For: 0.99.0
>
> Attachments: 11297.v1.patch, 11297.v2.patch
>
>
> This is on top of another patch that I'm going to put into another jira.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
