Hi everybody, News from the vampir trace integration at last. The current vampir trace release is included in the TMP tree of the OpenMPI SVN. Now, autoconf scripts need to be added.
Then the build procedure will look like this:
1) run global OpenMPI configure script
2) the global configure script will call the local one at some point
3) the global make command is issued
4) vampirtrace's make is called early in this process
5) building the config-file which determines compiler flags for
mpicc-vt, mpicxx-vt, mpif90-vt, etc.
Step 5) depends on step 4) because you need to ask vampir trace for this
flags. Here, 'ask' means calling some of vampir trace's commands.
Now, this was a problem with cross compilation (*) but not any longer. Besides
the binaries we have a script which will tell the include and library flags
which need to be added on behalf of vampir trace.
The script is called vtconfig, located at $(VTROOT)/vtlib/vtconfig/vtconfig.
Call vtconfig with '--libs' or '--includes' to get either information on
stdout. Those flags can be combined with one of [--seq|--mpi|--omp|--hyb] and
[--fortran] and [--dyninst] to get, for example the flags or Fortran with MPI
(of course).
Alternatively, the information could be extracted from
$(VTROOT)/tools/compwrap/compwrap_conf.h but I would not recommend it.
Is this sufficient to get the autoconf parts done, that call vampir trace's
configure and make?
Best regards, Andreas
ps: (*) cross compiling really gets funny here: if you are cross compiling
OpenMPI to run completely on a different platform, everything is like said
above. But if you make OpenMPI _as_ a cross compiler, then the compiler
wrappers should be compiled normally but the run time libs need to be for the
target platform. Well, then you combine both ... <8)
--
Dipl. Math. Andreas Knuepfer,
Center for Information Services and
High Performance Computing (ZIH), TU Dresden,
Willersbau A114, Zellescher Weg 12, 01062 Dresden
phone +49-351-463-38323, fax +49-351-463-37773
pgpbX33SRjv29.pgp
Description: PGP signature
