[jboss-jira] [JBoss JIRA] Commented: (JBCLUSTER-131) Evaluate impact of message bundling in HttpSession replication performance

[Brian Stansberry (JIRA)]

    [ http://jira.jboss.com/jira/browse/JBCLUSTER-131?page=comments#action_12341141 ] 
            
Brian Stansberry commented on JBCLUSTER-131:
--------------------------------------------

Ran tests on AS 4.0.4.GA, JBC 1.4.0.SP1, JGroups 2.2.9.2, testing UDP, TCP and TCP_NIO with bundling on and off, REPL_ASYNC and REPL_SYNC.  4 node cluster, no buddy replication.  Results are attached. The basic JBC config files are also attached.   The versions used for no-bundling tests looked like the attached, except "enabled_bundling" was "false".  The versions used for REPL_SYNC tests looked like the attached except for CacheMode being REPL_SYNC and the FC protocol was commented out of the protocol stack.

Test app basically puts an 8Kb payload in a single attribute for each request.

Both SESSION and ATTRIBUTE granularity were tested.  Since the session only has a single attribute, the payload size difference between the two granularities is not great.  A big difference is SESSION does not use a TX during replication (it just does a single put, so no need), while ATTRIBUTE does use a TX.

Interesting points:

1) TCP_NIO is by far the fastest transport, consistent with what I've seen previously.
2) Disabling bundling actually *hurts* TCP_NIO performance.
3) For UDP and TCP, disabling bundling has only a minor effect ( < 1.5%), in most cases negative.  This is actually good news, since with the multiplexer a REPL_ASYNC cache may share a channel with a synchrononous HAPartition. It is important that an optimization for one does not harm the other.
4) With SESSION + SYNC, the benefit of turning of bundling is largest.  This is interesting, as this combination uses no TX, so the costly 2PC protocol doesn't get involved.
5) Turning off bundling actually hurts a bit in the ATTR + SYNC case.  This is odd, as this case uses the full 2PC protocol.

> Evaluate impact of message bundling in HttpSession replication performance
> --------------------------------------------------------------------------
>
>                 Key: JBCLUSTER-131
>                 URL: http://jira.jboss.com/jira/browse/JBCLUSTER-131
>             Project: JBoss Clustering
>          Issue Type: Task
>      Security Level: Public(Everyone can see) 
>            Reporter: Brian Stansberry
>         Assigned To: Brian Stansberry
>             Fix For: Q3Y6
>
>         Attachments: bundleeffect.xls, tc5-cluster-service.xml.nio, tc5-cluster-service.xml.tcp, tc5-cluster-service.xml.udp
>
>
> As described in http://wiki.jboss.org/wiki/Wiki.jsp?page=ProblemAndClusterRPCs the use of message bundling is not optimal for RPC calls.  Essentially all clustering traffic in the AS is RPC calls, so we should evaluate the impact of disabling message bundling.
> I'm going to evaluate this using both REPL_SYNC and REPL_ASYNC.  Any negative effect of bundling with REPL_ASYNC should be less, as the caller does not wait for the response.

-- 
This message is automatically generated by JIRA.
-
If you think it was sent incorrectly contact one of the administrators: http://jira.jboss.com/jira/secure/Administrators.jspa
-
For more information on JIRA, see: http://www.atlassian.com/software/jira