Minimum stack size..

[Thomas Stüfe]

Hi,

Steve, did you maybe set LDR_CNTRL in your environment? If yes, and it
forces C-heap pages to 64K, this may be the reason for the high default
(1920k are 30 64K pages). A workaround may be to switch off 64K pages for
the C-Heap (LDR_CNTRL="..DATAPSIZE=4K..")

Stacks for java threads on AIX are allocated by the pthread library from
C-heap and are therefore subject to the C-heap data page size, which can be
either 4K or 64K. This depends on the setting you are running in - it can
be changed via the LDR_CNTRL environment variable (eg
LDR_CNTRL="..DATAPSIZE=64K..", see
http://publib.boulder.ibm.com/infocenter/javasdk/v6r0/index.jsp?topic=%2Fcom.ibm.java.doc.user.aix32.60%2Fuser%2Faix_ldr_cntrl_page_sizes.html)
or in the XCOFF itself (with ldedit, I think).

I think the SAPJVM is able to cope with 64K pages in the C-Heap - we put
some work into this - but the usual mode of operation is a C Heap page size
of 4K. I think the wasteful stacks are the reason for that. For stacks, you
want small pages.

Note that all this stuff is independent from -XX:+UseLargePages, which
changes the page size of the java heap but leaves the C heap page size
alone.

Kind Regards, Thomas



On Wed, Apr 17, 2013 at 7:40 PM, Volker Simonis <volker.simonis at gmail.com>wrote:

> Hi Steve,
>
> I suspected you'll come back with this issue, especially after I
> pretended that it will be easy to fix :)
>
> You're actually right, the minimum stack size is computed at VM
> startup from various other parameters one of which is the OS page
> size.
> It's a long standing problem in HotSpot that at various places it is
> assumed that the default OS page size is 4k. That's obviously not true
> anymore (on Linux/PPC it's usually 64k) and that leads to such kind of
> errors (if you'd divide 1920k by 16 you'd get 120k which would be
> reasonable).
>
> We've already changes this behavior in our SAP JVM  but that's quite a
> large shared code change which we've omitted from the port so far.
> Please let me discuss this with Goetz tomorrow to see haw we can fix
> this for our port - but I think it would be reasonable to fix it in
> the HotSpot and not in the test.
>
> Regards,
> Volker
>
> On Wed, Apr 17, 2013 at 6:30 PM, Steve Poole <spoole at linux.vnet.ibm.com>
> wrote:
> >
> > hi.  we are  working though the last remaining jtreg test failures for
> > PPCAIX.
> >
> > I have some that are failing on Linux PPC as the  test explicitly sets
> the
> > minimum stack size  to "-Xss256k" . The VM then fails to launch with the
> > usual error
> >
> > The stack size specified is too small, Specify at least 1920k
> >   Error: Could not create the Java Virtual Machine.
> >   Error: A fatal exception has occurred. Program will exit.
> >
> >
> >
> > So I can adjust the stack size in the testcases to suit this number or
> > remove it altogether.  But...
> >
> > 1 - This number of 1920k seems a trifle  excessive  - is it something I'm
> > doing wrong in building hotspot or is this actually a reasonable number?
> > 2 - Since it seems from reading the hotspot code that this number is
> > calculated is there any Java API I can use to get this number so I can
> make
> > the tests more flexible?
> >
> >
> > Cheers
> >
> >
> >
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.openjdk.java.net/pipermail/ppc-aix-port-dev/attachments/20130417/76f8e985/attachment.html