MQSeries.net :: View topic

powerlord · Posted: Wed May 10, 2006 7:12 am Post subject:

I've read the help on setting this both on IBM site:

http://publib.boulder.ibm.com/infocenter/wmbhelp/v6r0m0/index.jsp?topic=/com.ibm.etools.mft.doc/au16640_.htm

and on mqseries.net here:

http://www.mqseries.net/phpBB2/viewtopic.php?t=18369&highlight=maxdata

The thing is... I'm not quite sure this is tallying with what I am seeing and want to run it by people:

I am on AIX, on a 2 CPU blade with 2 gig of memory.

I have my flow deployed with 5 additional instances. My flow is very java heavy (lots of java custom nodes, java XSLs, etc).

I see that the DataFlowEngine process defaults to 4 memory segments (from a dump -ov). (0x40000000)

(I have also tried explicited settings LDR_CNTRL)

I start the broker.

I now run up nmon in a different window with top memory by process.

I can see DataFlowEngine using about 100 meg which is the same as on windows.

I start firing messages into it on one thread and memory goes up to about 180ish.

I then try firing in 4 threads of messages and see it go up to 320 meg or so and stabilise.

I try 5 threads, and the broker abends with what looks like a problem with memory allocation. nmon showed memory going up to just over 320 meg.

Now, I stop the broker, set LDR_CNTRL to 0x50000000 and restart broker.

Now, it happily does 5 threads of messages and I see memory stabilise at about 330meg or so....

so, the question is.. why ? 4 segments is 1 gig. That should have been plenty. So why was it abending until I gave it 5 segments (1.25 gig).

It's not the jvm, because memory usage there is minimal, and the default max heap size is -1. Just to be sure I set this to the same as the minimum (128 meg), and it behaved exactly the same way.

Any ideas what is happening here ? Is nmon telling me lies about memory usage ? Or is something weird happening with LDR_CNTL setting ? Or is this affected by being on 2 CPUs (even though DataFlowEngine is only using one CPU obviously).

stu

<nudge>