[mpich-discuss] hydra and mpich2-1.3rc1

Bryan Putnam bfp at purdue.edu
Thu Sep 23 10:37:43 CDT 2010


On Thu, 23 Sep 2010, Pavan Balaji wrote:

> Hi Bryan,
> 
> Yup, I can reproduce the bug now. See below for more information.
> 
> > $MPI_SRC/configure \
> >  --with-device=ch3:nemesis \
> >  --enable-fast \
> >  --enable-threads \
> >  --enable-debuginfo \
> >  --enable-shared \
> >  --enable-sharedlibs=gcc \
> >  --enable-f77 \
> >  --enable-fc \
> >  --enable-cxx \
> >  --enable-romio \
> >  --with-pm=mpd:hydra \
> >  --without-mpe \
> >  --prefix=$MPI_INSTALL/$CVER \
> >  > configure.log 2>&1
> 
> You still seem to be forcing MPD to be the default process manager, and 
> hydra to be the secondary PM. If you remove the --with-pm=... line, you 
> can get hydra by default. This case should work correctly.
> 
> The problem you are seeing is because of a combination of 
> --enable-shared and using Hydra as the secondary PM. In this case, 
> libtool wants us to use "make install" to install mpiexec executable. 
> However, since both MPD and Hydra will create an mpiexec executable, 
> just doing "make install" directly will cause one of them to overwrite 
> the mpiexec created by the other. We had some workaround for this in 
> src/pm/hydra/Makefile.am (see the install-alt target, which tries to 
> rename mpiexec to mpiexec.hydra), but that's clearly not sufficient for 
> shared library builds where libtool wants to manage the executables. 
> I'll try to workaround this problem and commit a patch soon.

Pavan,

Thanks, omitting "--with-pm=..." in mpich2-1.3rc1 fixed the problem. I 
think this configure appeared to work with mpich2-1.3b1 because I had not 
specified the "--with-shared" config option. However, in mpich2-1.3rc1, 
"--with-shared" was required in order to use "--with-sharedlib=gcc".

Thanks,
Bryan

> 
>  -- Pavan
> _______________________________________________
> mpich-discuss mailing list
> mpich-discuss at mcs.anl.gov
> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss
> 




More information about the mpich-discuss mailing list