[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: MPI issue on sparc ? Boost::mpi looks for OpenMPI but it should be LAM no ?


if openmpi is installed on sparc, it takes priority over lam ! (it has priority 40 and lam 30)

here is the result of  mpic++ -showme:compile           on smetana.debian.org                                                                                   
-I/usr/lib/openmpi/include -I/usr/lib/openmpi/include/openmpi -pthread
and link
mpic++ -showme:link                                                                                                
-pthread -L/usr/lib/openmpi/lib -lmpi_cxx -lmpi -lopen-rte -lopen-pal -ldl -Wl,--export-dynamic -lnsl -lutil -lm -ldl

so even though mpi-default-dev point to lam, if openmpi gets also installed it is in practice the default implementation
on sparc !

To my opinion this is a bug in the mpi system. Adam ? Others ?

There are 3 choices for the alternative mpi (providing /usr/include/mpi).

  Selection    Path                      Priority   Status
* 0            /usr/lib/openmpi/include   40        auto mode
  1            /usr/include/lam           30        manual mode
  2            /usr/lib/mpich/include     10        manual mode
  3            /usr/lib/openmpi/include   40        manual mode

On Wed, Jun 9, 2010 at 8:53 PM, Christophe Prud'homme <prudhomm@debian.org> wrote:
I tried, without success so far, to help cmake(FindMPI.cmake) find the proper mpi implementation.
It still finds openmpi which breaks linkage with boost::mpi

Best regards

On Wed, Jun 9, 2010 at 12:08 PM, Adam C Powell IV <hazelsct@debian.org> wrote:
On Mon, 2010-06-07 at 02:25 -0500, Steve M. Robbins wrote:
> On Mon, Jun 07, 2010 at 09:02:41AM +0200, Christophe Prud'homme wrote:
> > On my side life depends solely on mpi-default-dev, it seems that some other
> > package don't (e.g. hdf5), isn't it a problem ?
> Yes, something like that is likely the problem.
> Note that libhdf5-mpi-dev is supposed to alleviate that problem as it
> depends on the default MPI version.  Installing that package should
> not pull in any non-default MPI packages, IMHO.  Adam: any comment?

Indeed, that's the idea: Build-Depend on libhdf5-mpi-dev and
mpi-default-dev and you should have a consistent MPI implementation
across both.

That said, the LAM HDF5 implementation seems to be missing a couple of
libraries, such that for example PETSc doesn't build on architectures
where LAM is the default.  I disabled PETSc HDF5 support on those
arches, but haven't investigated further.

GPG fingerprint: D54D 1AEE B11C CE9B A02B  C5DD 526F 01E8 564E E4B6

Engineering consulting with open source tools

Reply to: