Hello Amber Team,
I am building Amber11 with Scalasca on IBM's Blue Gene/P supercomputer as
part of my B.Sc. final project on Computer Systems and Technologies at
Technical University-Sofia, Bulgaria. Sadly, I have faced 2 blocking issues
and I have invented a workaround for 1, and, thus, one remains unsolved. I
would like to give you a feedback and to ask for help for the following
problems:
- I have found a not standard behaviour in the Blue Gene/P compilers.
When trying to link object files of Fortran and C it turns out that the
naming convention is broken if previously compiling with MPI compiler
script. Attached you can find (file:
BlueGeneP-Mixing-Fortran-and-C-Problem-Repro.tar (application/x-tar) 10K) a
small Makefile and a C and Fortran source files reproducing the problem-*the
code is workarounded to be compiled and linked in MPI mode with no guarantee
on the quality of the workaround*. When mixing Fortan and C, C functions
have to be named like *subra_* and called from Fortran code like
*subra*(without the last character of underscore). It seams that when
compiling
with MPI script on Blue Gene/P you have to call the function in the Fortran
code like *subra_* or otherwise when linking the object files the
function will be unknown.
- I can not compile Amber11 with IBM XL compiler scripts (mpixlf90,
mpixlc). Attached you can find (file: parallel-src70.out
(application/octet-stream) 74K) a log of make parallel on Amber11. Compiling
Amber11 with XL compilers is required in order to use Scalasca to measure
performance. I have modified the config.h files (except these in AmberTools
directory) to use mpixlf90, instead of mpif90, and mpixlc, instead of mpicc.
I am looking forward for hearing from you.
*Yours Sincerely Garo Garabedyan*
Telephone 089 59-59-703
garabedyan.wordpress.com
Sofia, Bulgaria
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Tue Sep 13 2011 - 13:00:02 PDT
