Hi,
marc fleury wrote:
> Rickard addressed something similar the first time we has speed issues.
>
> ( will try to dig, ok found it forwarded it )
Yes, at least some of this was similar to
what Rickard saw.
> Timpl WAS in the claspath of the RMI subsystem yours is not,
Cannot see that. At least not in the
jboss-client.jar file. Maybe it should
have, but never was. A performance
problem due to this is currently only
seen on non-optimized transactional calls
within the same VM, which would only be
set up by people that don't care about
performance.
I added GlobalId to jboss-client.jar,
removed it from jboss.jar, made it public
and changed a few other things to avoid
class loader problems.
This brings the non-optimized invocation
time down to about half.
But there is still some way down to the
non-optimized invocation time we saw
before the JTA changes.
Optimized invocation time seems pretty
much the same as before.
Also, I don't like that I don't understand
why this happens to GlobalId, but not to
the TransactionImpl we used before. After
all, they are both simple serializable
objects.
Key to understanding the problem here is
probably understanding the difference
between these two.
So I took a pre-JTA-change snapshot and
ran the same speed test (non-optimized
bean-2-bean call with a transaction
context) on that one to check for
differences. To my big surprise this was
just as slow as current CVS, so it would
probably not help to profile for
differences here.
Maybe the change described above fixes an
unrelated performance problem, and maybe
it doesn't matter when the real problem
is found and fixed.
Some more work is needed on this, but that
will have to wait until tomorrow.
Best Regards,
Ole Husgaard.
