Re: [Boost-users] Boost.MPI compilation problems with MPICH2

From: Douglas Gregor (dgregor_at_[hidden])
Date: 2007-10-22 13:52:15

On Oct 21, 2007, at 4:36 AM, Aydõn Bulu wrote:
> I was a happy user of Boost.MPI that was working together with the  
> underlying OpenMPI. However, MPI_THREAD_MULTIPLE support of OpenMPI  
> is quite clumsy and it hangs frequently (OpenMPI developers call  
> this as being "lightly tested"). So, I installed MPICH2 that claims  
> to supports thread-safe MPI.

I've heard that MPICH2's support for MPI_THREAD_MULTIPLE is pretty good.

Okay, so just a fair warning... at present, Boost.MPI will *not* work  
correctly in a multi-threaded context if you're sending serialized  
data over the wire. It's possible that two concurrent serialized  
receives on the same communicator and with the same tag could get  
confused, most likely causing a crash. The solution isn't trivial,  
but we know how to solve the problem as soon as we have the chance.

> /tmp/ccbyPLBK.o: In function `void  
> boost::mpi::detail::all_gather_impl<int>(boost::mpi::communicator  
> const&, int const*, int, int*,  
> mpl_::bool_<true>)':all_gather_test.cpp: 
> (.gnu.linkonce.t._ZN5boost3mpi6detail15all_gather_implIiEEvRKNS0_12com 
> municatorEPKT_iPS6_N4mpl_5bool_ILb1EEE[void  
> boost::mpi::detail::all_gather_impl<int>(boost::mpi::communicator  
> const&, int const*, int, int*, mpl_::bool_<true>)]+0x1c): undefined  
> reference to `boost::mpi::communicator::operator int() const'
>
> /tmp/ccbyPLBK.o: In function `void  
> boost::mpi::detail::broadcast_impl<std::basic_string<char,  
> std::char_traits<char>, std::allocator<char> > > 
> (boost::mpi::communicator const&, std::basic_string<char,  
> std::char_traits<char>, std::allocator<char> >*, int, int,  
> mpl_::bool_<false>)':all_gather_test.cpp: 
> (.gnu.linkonce.t._ZN5boost3mpi6detail14broadcast_implISsEEvRKNS0_12com 
> municatorEPT_iiN4mpl_5bool_ILb0EEE[void  
> boost::mpi::detail::broadcast_impl<std::basic_string<char,  
> std::char_traits<char>, std::allocator<char> > > 
> (boost::mpi::communicator const&, std::basic_string<char,  
> std::char_traits<char>, std::allocator<char> >*, int, int,  
> mpl_::bool_<false>)]+0x27): undefined reference to  
> `boost::mpi::communicator::operator int() const'
> -------------------------------------
>
> I "quadruple-checked" to see that the versions of mpicxx/libraries/ 
> include directories all belong to the new installation of MPICH2  
> and not remainders of old OpenMPI installation.  Thus, I don't  
> think that is a binary incompatibility issue.

I hate to say it, but we might need a quintuple check :)

The implicit conversion from "communicator" to "int" shows that we  
are certainly getting the MPICH2 headers, because MPI_Comm is just a  
typedef of int in MPICH2 (in OpenMPI, MPI_Comm is a pointer).  
However, I'm guessing that the archive that we're linking against was  
compiled against OpenMPI.

When you rebuilt Boost.MPI, did you remove "bin.v2/libs/mpi"? If not,  
I suggest that you do so and then rebuild Boost.MPI.

        - Doug