Queue cache or persistance I/O slowdown

[jaywright00]

I'm running performance tests with a JMS application on both 2.4.9 and 3.0.3 on solaris. Same code.

In 3.0.3 I am having performance issues where the io wait is around 20% and we can see that about 300kb are being written to disk each second. But I can't figure out where. I don't see any files growing. I hope it's just an ill-configured cache or persistence manager.

In 2.4.9 it runs without problem, no io wait, so it's not any logging or application induced i/o.

For our application, we do NOT want to persist any messages.

Has anyone seen this problem in porting to 3.x from 2.x?

Thanks,
Jay

[alwyn]

Look in db/jbossmq/file. Do a lot of fast subsequent ls -la's and you'll see a bunch of small files being created and destroyed.

I solved it with a ramdisk. Alternative is rollinglogged PersistenceManager, but it kills performance by about 300 messages/second.

Strange things also happen when you actually want persistence...

[jaywright00]

Actually i don't see that, not in db/jbossmq/file or db/jbossmq/file/QUEUE.myQueue. Do you know what is being written and by which class?

Most importantly, can it be turned off? Either through configs or extending a class (or editting a class) and rebuilding jboss?

Why would this disk io be necessary?

[kelbahi]

you may want to check if the in memory queue is being written to disk. This is governed by
org.jboss.mq.server.MessageCache

and
org.jboss.mq.pm.file.CacheStore

looks like 1 message per file once the vm is the size specifiec by messageCache.
/Kel

[alwyn]

The message cache writes to tmp/jbossmq I think...

In my tests the Cache values were set in the region of 1.5Gb
and I never came close to reach that so I suspect it is the Persistence Manager.

I'm toying with the idea of writing my own PM that does nothing, but that is a bit of a low priority idea at the moment.