On Jun 16, 2007, at 3:22 AM, Francesco Pietra wrote:
> The question is whether, in compiling openmpi,the flag
> libnuma is needed or simply useful also in the special
> arrangement of the Tyan S2895 Thunder K8WE with two
> dual-core opterons and eighth memory modules, two GB
At worst, it is not harmful. At best, it is helpful.
We're still playing / experimenting with the memory affinity controls
and don't have great support for it at the moment. It's ok, but it
will likely mainly end up smoothing your results over repeated runs
rather than result in dramatic performance improvements. Note that
you need to enable processor affinity for OMPI to use memory
affinity. See the FAQ for more information.
> If so (being first time to compile a mpi, and being
> nonexpert singleuser/administrator) I would be much
> obliged for checking the series of commands below (as
> superuser) for Linux Debian amd64 etch:
> cd /usr/local
> bunzip2 openmpi-1.2.2.tar.bz2
> tar xvf openmpi-1.2.2.tar
> cd /usr/local/openmpi-1.2.2
> FC=/opt/intel/cce/9.1.036/bin/ifort; export FC
> CC=/opt/intel/cce/9.1.042/bin/icc; export CC
> CXX=/opt/intel/cce/9.1.042/bin/icpc; export CXX
> ./configure --with-libnuma=/full pat to libnuma-dev
> 0.9.11-4, 0.9.11-3 (not yet installed)
> make install
This all looks fine. You might always want to set F77 to the same
value as FC. Alternatively, you can run:
./configure CC=/opt/intel/... CXX=/opt/intel/... (etc.)
and not set the variables in the environment. In terms of end
results, the effect is identical between the two techniques. The
only difference is that if you put the CC=... stuff on the configure
command line, which compilers you specifically chose will be recorded
in the config.log file, which can be handy for referring to later and/
or debugging problems.
> followed by setting as user in my .bashrc
> MPI_HOME=/usr/local; export MPI_HOME
Note that Open MPI does not use the MPI_HOME environment variable.
If you already have /usr/local/bin in your PATH and /usr/local/lib in
your LD_LIBRARY_PATH, you're set.
> mpi for a computational application that is best
> compiled with intel. On my system those intels already
> furnish runtime
> to a QM code (NWChem 5.0) that is built-in
> parallelized with TCGMSG.
> francesco pietra
> Yahoo! oneSearch: Finally, mobile search
> that gives answers, not web links.
> users mailing list