On Fri, Nov 9, 2012 at 1:00 PM, John Peterson <jwpeter...@gmail.com> wrote:
> In case there's anyone that hasn't muted this conversation yet... I'm really interested in this because the trunk failures I'd been seeing were a real mystery to me. > I think I've tracked the INL's issue down to mpich2. (I happen to be > using mpich2-1.4.1p1, but the problem is probably present in mpich2-1.5 as > well.) > > (Correct me if I'm wrong, but everyone at Texas on a Mac is using OpenMPI, > right?) > I've been using OpenMPI, but I have mpich2 as well. Will try a build with mpich2 momentarily. > Here's the behavior I'm observing, and a possible explanation: > > The automake branch works fine for me using hand-built GCC 4.5.4, 4.6.3, > and 4.7.2 compilers *without MPI or PETSc* on an otherwise stock Mountain > Lion install (no MacPorts). Yes, I really built all those GCCs by hand on > a laptop with 4 cores :-P > > Using the same compilers, if I load my Mpich2 module (which sets CC=mpicc > and CXX=mpicxx) the automake branch suddenly exhibits the mysterious > segfaults we've discussed here ad-nauseum in the adjoints_ex1 example. > Interesting. > > So, what the hell, mpich? > > Let me note first that we use a very vanilla configuration of mpich2, > effectively just > > ./configure --prefix=... --enable-shared --enable-debuginfo > make > make install > > What does this produce? > > $ mpicxx -show > g++ -Wl,-flat_namespace \ > -I/opt/packages/mpich2/mpich2-1.4.1p1/include \ > -L/opt/packages/mpich2/mpich2-1.4.1p1/lib \ > -lmpichcxx -lpmpich -lmpich -lopa -lmpl -lpthread > > Ugh... the prime suspect here has to be -Wl,-flat_namespace. The mpich > people must be adding this flag if they detect Darwin... but I think that > may be the wrong thing for the Lion and Mountain Lion operating systems. > This definitely could be the culprit. Just to show OpenMPI is different: [01:01:07][pbauman@dhcp-67-207:~/svn/pecos/grins/code/branches/reacting_flow]$ mpicxx -show g++ -I/Users/pbauman/my_libraries/openmpi/openmpi-1.6/gcc-4.6.3/include -L/Users/pbauman/my_libraries/openmpi/openmpi-1.6/gcc-4.6.3/lib -lmpi_cxx -lmpi -lm > So I could try to figure out how to make mpich not use that flag, but I > think my next step is going to be trying out OpenMPI. Anybody have special > tips for building that by hand? > Nothing out of the ordinary. I've attached my (very old, very unsophisticated...) build script; feel free to do what you want with it. Will report back about my results with mpich2. Thanks for taking the time to look at this carefully. Best, Paul
buildopenmpi.sh
Description: Bourne shell script
------------------------------------------------------------------------------ Everyone hates slow websites. So do we. Make your web apps faster with AppDynamics Download AppDynamics Lite for free today: http://p.sf.net/sfu/appdyn_d2d_nov
_______________________________________________ Libmesh-devel mailing list Libmesh-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/libmesh-devel
