[infinispan-issues] [JBoss JIRA] Commented: (ISPN-1015) Optimize HotRod client to match Memcached performance

[Dan Berindei (JIRA)]

    [ https://issues.jboss.org/browse/ISPN-1015?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12597916#comment-12597916 ] 

Dan Berindei commented on ISPN-1015:
------------------------------------

I added a new pull request, https://github.com/infinispan/infinispan/pull/271
Should still keep the issue open, there should be some more optimizations coming.

I changed lots of things (although it seems the only ones with a significant impact were client-side buffering and server-side the number of worker threads):

I found a few other places where we were building the trace/debug message even if the debug level was set to info or higher.
I changed this on the "happy" flows, even though none of them looks like a performance hotspot.

Improved ConsistentHashPerfTest, now it reports standard deviation, average abs deviation and median abs deviation.
Changed PreInvocationContext to stop allocating new EnumSets.
Made some fields in AbstractConsistentHash non-volatile.
Added SKIP_CACHE_LOAD flag when the client doesn't need the return value in the hotrod server.

Use BufferedOutputStream to coalesce socket writes in as few packets as possible.
Also use BufferedInputStream to reduce calls into the network stack.

Use a specialized ClassResolver to avoid looking at the context class loader completely if serialization doesn't require it.

I changed the default number of worker threads for the Infinispan servers to 20 * number of processors.
The setting for the number of master threads wasn't really doing anything so I removed it.

Added VNodesCHPerfTest to test the TopologyAwareConsistentHash with virtual nodes.

> Optimize HotRod client to match Memcached performance
> -----------------------------------------------------
>
>                 Key: ISPN-1015
>                 URL: https://issues.jboss.org/browse/ISPN-1015
>             Project: Infinispan
>          Issue Type: Task
>            Reporter: Dan Berindei
>            Assignee: Dan Berindei
>             Fix For: 4.2.2.BETA1, 4.2.2.FINAL, 5.0.0.BETA1, 5.0.0.FINAL
>
>
> Load tests show the HotRod protocol performing worse than the Memcached client.
> Since the HotRod protocol should be easier to parse and we also have the extra topology information, this is quite surprising.
> We have analyzed the servers and seem to perform about the same, so we need to optimize the client.

--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira