MPICH vs. AIX-MPI on frost?

[Tyce Mclarty]

John,

It looked to me like there was a problem with mixing IBM and mpich scripts 
and/or libraries with the approach in the README.frost and in the INSTALL 
files with the distribution. I have not gotten a pure mpich configure to 
work, but did work out a pure IBM one.

What you found may be a problem, but I think we have something more basic. 
When I get configure to run, there is a warning about an "unrecognized 
fortran name mapping". So it sounds like the IBM compilers are doing 
something unexpected with fortran names. Not too surprising.

This would explain why C works fine, but not fortran. I can do some digging 
through the AIX documentation or catch the guy over here who will know 
right away in the AM.

Tyce

At 04:23 PM 7/23/2003 -0700, John Tannahill wrote:
>More info on my problem.  There are two libraries that are put
>together in src/lib:  libpnetcdf.a and libnetcdf.a.  It appears
>that libnetcdf.a has all of the nfmpi_xxx routines in it.  My
>guess is that somehow this libnetcdf.a library should actually
>be part of the libpnetcdf.a library, then I believe I would
>have all the routines I need to link with.  I think that there
>was a name change a while back when libnetcdf.a got switched to
>libpnetcdf.a?  Maybe a netcdf didn't get changed to a pnetcdf
>somewhere where it needed to be?
>
>Regards,
>John
>
>John Tannahill wrote:
>>Tyce,
>>Actually, I am now running the parallel-netcdf code on an IBM SP at
>>NERSC (seaborg).  I don't believe that mpich is being used at all.
>>I run on seaborg because I can easily get interactive time, which I
>>don't think that I can get on frost.  Anyway, you missed out on some
>>of my previous messages, so I will try to bring you up to date.  The
>>C interface seems to work fine.  I put together a test code in C and
>>it compiled, linked, and ran.  I then converted this test code to F90.
>>I have gotten this version to compile, but it won't link because it
>>can't find any nfmpi_xxx routines in the library (the C code uses
>>ncmpi_xxx routines and they are there).  So the question seems to be
>>how to build the library so that the nfmpi_xxx routines are there?
>>Perhaps you have some experience with autoconf and can figure out
>>what needs to be done?
>>Earlier today, I contacted the one other person that apparently had
>>some interest in trying the Fortran interface a couple of months ago
>>(Troy Baer at the OSC), but it now appears that we may be the first
>>real users (i.e., it's a little unclear, but I believe that Troy is
>>still trying to carve out some time to investigate it).  I am sure
>>the ANL/Northwestern people are trying to track things down as well.
>>It's certainly possible that it may be something quite simple that
>>I am just missing.  I did try the --with-fortran=yes flag when
>>running configure, but that did not seem to help.  I have not been
>>able to look at things much today, but will get back into now.
>>Regards,
>>John
>>
>>JB Gallagher wrote:
>>
>>>All,
>>>
>>>Just to let you know I just built and installed parallel-netcdf-0.8.4 
>>>and built
>>>the benchmark code against it and it ran fine, so for me anyway the tar ball
>>>from the MCS website works the same way. Again I only tested the C-interface
>>>though, I don't use any fortran for the i/o. When I first installed 
>>>netcdf on
>>>frost I had some mixing of the libraries that caused errors when my code 
>>>ran, I
>>>didn't save the output of these errors however, but they were a result of
>>>having mpich in the mix.
>>>
>>>So I was able to build the pnetcdf code off of the mcs website the same 
>>>way I
>>>said before, I set all the compiler envirnment variables to the normal
>>>compilers (xlc etc..) and then ran configure like so: ./configure
>>>--prefix=<dir> --with-mpi=/usr/local/tools/mpich/mpich-1.2.5mpl
>>>
>>>then I edited macros.make  and replaceed all compiler variables (especially
>>>mpich ones) with the newmp ones (xlc , newmpcc etc..)
>>>
>>>This creates a library that is linked to the AIX mpi implementation.
>>>
>>>--Brad
>>>
>>>
>>>Rob Ross wrote:
>>>
>>>
>>>>There's no reason that I know of why it wouldn't work with IBM's MPI;
>>>>Brad, do you have any input on this?
>>>>
>>>>Thanks,
>>>>
>>>>Rob
>>>>
>>>>On Wed, 23 Jul 2003, Tyce Mclarty wrote:
>>>>
>>>>
>>>>>Rob,
>>>>>
>>>>>I'm trying to help John Tannahill get the PnetCDF going here on his code.
>>>>>Looking at the README.frost in the latest release, it looks like he may be
>>>>>mixing mpich and IBM libraies.
>>>>>
>>>>>When you ran tests on frost, did you use mpich?
>>>>>
>>>>>Is there any reason it should not work with IBM's MPI?
>>>>>
>>>>>Thanks,
>>>>>Tyce
>>>>
>>>
>
>
>--
>============================
>John R. Tannahill
>Lawrence Livermore Nat. Lab.
>P.O. Box 808, M/S L-103
>Livermore, CA  94551
>925-423-3514
>Fax:  925-423-4908
>============================