activemq-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Milton Taylor (JIRA)" <>
Subject [jira] Commented: (AMQ-2239) Stomp wire format doesn't work with tcp or nio transport
Date Fri, 01 May 2009 10:02:38 GMT


Milton Taylor commented on AMQ-2239:

Have analysed this further.  We changed the implementation of the stomp factory to extend
from NIO rather than from tcp.  However unit test code crashed trying to allocate huge amounts
of memory from the heap! Further investigation shows that the NIO implementation is openwire-predicated.
i.e., it has code in it that sets up buffer lengths etc based on openwire packet framing headers
etc. Arggh!  The TCP stuff does not make such assumptions; it works a different way anyway
in terms of buffer management.

So we hacked it a bit more to get beyond this design issue to see how far we could get.  A
lot of unit tests passed. (Did not try them all).  A real-world test rig with a load test
client was then tried. 

Two unsavoury discoveries:
* The performance was totally lousy - much much worse than the original stomp implementation
ever was. There is probably simple explanation for this, but we ran of time to investigate
* It still uses one thread per connection on the server, despite the fact that the NIO code
does have code to create worker threads that look after a whole set of channels.

Also failed to shut down properly after that...I think there may be open issues concerning
NIO on this front?

Would be interested to hear if there are others interested in an NIO implementation of stomp...we
may look at reworking this code. (It also doesn't take advantage of NIO features as well as
it could anyway).

> Stomp wire format doesn't work with tcp or nio transport
> --------------------------------------------------------
>                 Key: AMQ-2239
>                 URL:
>             Project: ActiveMQ
>          Issue Type: Bug
>          Components: Transport
>    Affects Versions: 5.2.0
>            Reporter: Milton Taylor
> It appears to us that there is some sort of architectural problem with how stomp has
been implemented in ActiveMQ. As we understand it, stomp is an application-level wire protocol.
However the implementation seems to combine both wire protocol and it's own transport implementation
based directly on the tcp socket factory. Not sure what the reason was for this...but it seems
like it compromises the layering of wire protocols over transports.
> Why we started looking at this, was we wanted to use NIO with Stomp to get the best scalability.
In particular, we don't want a server thread being consumed for every long-lived stomp connection.
However, although one sees references in mailing lists to the fact that *stomp://localhost:61613*
is equivalent to  *tcp://localhost:61613?wireFormat=stomp*, in fact the latter does not work
and produces this exception:
> {noformat}
> Exception in thread "BrokerService"
> java.lang.ClassCastException: org.apache.activemq.command.BrokerInfo cannot be cast to
> at org.apache.activemq.transport.stomp.StompWireFormat.marshal(
> at org.apache.activemq.transport.tcp.TcpTransport.oneway(
> at org.apache.activemq.transport.InactivityMonitor.oneway(
> at org.apache.activemq.transport.MutexTransport.oneway(
> at
> at
> at
> at org.apache.activemq.thread.PooledTaskRunner.runTask(
> at org.apache.activemq.thread.PooledTaskRunner$
> at java.util.concurrent.ThreadPoolExecutor$Worker.runTask(Unknown Source)
> at java.util.concurrent.ThreadPoolExecutor$ Source)
> at Source)
> {noformat}
> Because of this, it is not possible to specify for example, 
> {noformat}
> <transportConnectors>
>   <transportConnector name="stomp" uri="nio://localhost:61613?wireFormat=stomp"/>
> </transportConnectors>
> {noformat}
> We did try altering the implementation of StompTransportFactory to extend NIOTransportFactory
instead of TcpTransportFactory. It compiled OK, but it didn't seem to actually work in our
tests...the client gets EOF reading the socket for some reason.

This message is automatically generated by JIRA.
You can reply to this email to add a comment to the issue online.

View raw message