Fwd: Problem with MPE with IBM MPI

[Anthony Chan <chan@…>]

A reminder that I need to look at AC_C_BIGENDIAN on AIX.

----- "Anthony Chan" <chan@mcs.anl.gov> wrote:

> > > 1) checking whether byte ordering is bigendian... universal
> > >   POWER5 is a big endian machine, looks like a bug here.
> >
> > This is a new feature in autoconf - see the MPICH2 configure for
> > what I had to add to deal with some of this.
>
> I will look at what you did in MPICH2 configure.
> I have been under the impression that universal byteordering
> is only relevant for MacOSX, i.e. I assume __BIG_ENDIAN__ is
> only available on gcc for MacOSX.  I am apparently wrong.
>
> I recall jumpshot was working fine on icrunch when we switched to
> autoconf 2.62.
>

[Anthony Chan]

This message has 0 attachment(s)

[chan]

autoconf-2.63 contains an updated AC_C_BIGENDIAN that seems to fix the problem. Need to find an AIX machine to confirm this.

[chan]

autoconf-2.62's AC_C_BIGENDIAN mistakenly detected BlueGene?/P and AIX6 as "universal" instead of "big endian". autoconf-2.63 fixes the bug (autoconf-2.59 which does not have universal binary support does not have the bug). MPE logging in SVN works OK on AIX6 both in terms of endianness and large file support even with the buggy 2.62 AC_C_BIGENDIAN. The option here is if MPICH2 should update to autoconf-2.63 for the correct AC_C_BIGENDIAN or overload autoconf-2.62's AC_C_BIGENDIAN with the bugfixed version ? The latest MCS linux machine, e.g. crank.mcs.anl.gov, has autoconf-2.63 comes preinstalled. The latest Fedora-11 comes with autoconf-2.63 as well.

[balaji]

If autoconf-2.63 already fixes it, I see no reason to copy the code ourselves to a macro, instead of just updating the version requirement. Of course, there might be other bugs in autoconf-2.63, but then, we don't know what all bugs are there in autoconf-2.62 either :-).

[chan]

From our perspective, updating to autoconf-2.63 is the easiest route plus 2.63 has been out since last sept. Since we just updated to 2.62 not too long ago, updating to 2.63 may seem fast and may cause extra burden to some of our collabrators ?

[balaji]

Replying to chan:

Since we just updated to 2.62 not too long ago, updating to 2.63 may seem fast and may cause extra burden to some of our collabrators ?

I doubt it. Most collaborators base their distributions off of our release versions, not svn. For the few folks who do use the svn, it's no big deal to just ask them to upgrade to autoconf-2.63. And as you pointed out, most modern distributions do have autoconf-2.63 anyway, so they get it automatically.

[chan]

mpich2's configure.in has updated to use autoconf-2.63 as of #4861