On Aug 10, 2007, at 5:54 AM, Andreas Knüpfer wrote:
Finally, VampirTrace is usable from within OpenMPI. From the
tmp/vt-integration/ repository you can get an extended OMPI version:
Thanks for doing this. Although I promised to help, I've had zero
time to do so recently. :-(
#> svn co http://svn.open-mpi.org/svn/ompi/tmp/vt-integration
#> cd vt-integration
#> ./autogen.sh
#> ./configure --prefix=/tmp/openmpi_with_vampirtrace/
#> make
#> make install
Then you should be able to compile a small MPI example with 'mpicc'
and run it
with 'mpirun'. Now simply change 'mpicc' to 'mpicc-vt'. The run
command is
still the default 'mpirun'. If you run the new executable again
you'll find
an 'a.otf' file plus 'a.0.def' plus some 'a.*.events' files. Some
files may
have a '.z' suffix which is optional (additional zlib compression).
Sweet.
There are a few issues left however.
- we need to add a ''---disable-vampirtrace" configure parameter
- So far, the configure option '--platform=optimized' makes
vampirtrace use
the same very strict compiler warnings as openmpi. thus, it will
stop with an
error because of "-Werror'.
we'll check this but so far without '--platform=optimized'
everything is fine.
- Sometimes vampirtrace detects another installed MPI lib and wants
to link it
which results in linker errors when building your mpi app. This is
because
vampirtrace is configured and build before mpi is ready, which is
unusual.
this is another thing for us to fix.
- 'make dist' seems to work fine including all necessary
vampirtrace files.
maybe this needs to be tested, however
- updating the included vampirtrace version should be as simple as
extracting
a new tarball in the subdir 'tracing'. we could even check + fetch
the
latest tarball via wget when one does 'make update-vampirtrace in
subdir 'tracing'. what's your opinion?
It might be easier to just allow compiling / linking against a VT
that is outside of the OMPI tree.
- we need an new configure option which simply passes special optionto
vampirtrace's configure script
We'll probably also want to put VT into the /vendor branch (per
http://svnbook.red-bean.com/nightly/en/svn-
book.html#svn.advanced.vendorbr).
So much for the most important issues. If you're interested please
try it and
send us your results. @Jeff, can we give read access to the
tmp/vt-integration/ subtree of SVN?
I think I set the SVN privs right; the public should now be able to
read this directory.
--
Jeff Squyres
Cisco Systems
