Hey
marc fleury wrote:
> > Ok, now I looked more at the CPU-load results. Much time is actually
> > spent marshalling stubs, which is why sending EJBObjects is slow. I
> > think I can speed this up by doing some simple things such as
> > implementing my own serialization and stuff like that.
>
> I wish Norbert would talk more...
>
> The guy optimized that last week (mail on spyderMQ) and went from 500
> messages / second with normal RMI to 8000 messages/second. The key is that
> we ALWAYS send the same structure except for the actual payload of the
> message (the byte[] inside).
>
> In other words, we can now move the MessageInvocation almost ONLY on
> metaData. If we only deserialize the meta data information of the MI we can
> move that really fast in the server and definitely out of the invocation
> layer. Only the "routing" information (EJB name, type, key, method name
> etc) is needed the actual "payload" (args) is not needed. We end up with a
> message structure that is always the same and we can greatly speed that up
> with custom serialization. Of course the stub sending is the same we can
> speed that up greatly.
This would only be interesting for routed messages/invocations. For the
plain case (client<->server) there's not much to do about it.
> We can also write JNI serializers ;-) for Linux and windows... but that
> would suck I think...
Duh.
/Rickard
--
Rickard �berg
@home: +46 13 177937
Email: [EMAIL PROTECTED]
http://www.telkel.com
http://www.jboss.org
http://www.dreambean.com
