Re: [Boost-users] MPI with threads - adress not mapped problem

Subject: Re: [Boost-users] MPI with threads - adress not mapped problem
From: Kraus Philipp (philipp.kraus_at_[hidden])
Date: 2011-05-06 12:21:46

Am 06.05.2011 um 18:00 schrieb Riccardo Murri:

> Hi,
>
> On Fri, May 6, 2011 at 4:11 PM, Kraus Philipp
> <philipp.kraus_at_[hidden]> wrote:
>> I'm using Boost 1.46.0 with OpenMPI 1.5.3 under Gentoo x64. I have  
>> compiled
>> OpenMPI with thread and mpithread option and Boost with MPI  
>> support. A
>> little testprogram with:
>>
>> #include <boost/mpi.hpp>
>> namespace mpi       = boost::mpi;
>>
>> int main(int argc, char* argv[]) {
>>    mpi::environment loMPIenv(argc, argv);
>>    mpi::communicator loMPICom;
>>
>>    return 0;
>> }
>>
>> crashes alway if I run it with "mpiexec -np" (without mpiexec it  
>> runs).
>
> I've tested your program with:
>
>  * OpenMPI 1.4.1 (from the Ubuntu 10.10 package `libopenmpi-dev`
>    version 1.4.1-3ubuntu1), Boost 1.46.0 (compiled by myself) on
>    Ubuntu 10.10 64-bit with the GCC 4.4.3 compiler;
>
>  * OpenMPI 1.4.3a1r23323, Boost 1.45.0 (both compiled by myself)
>    on Rocks 5.2 64-bit with the GCC 4.1.2 compiler;
>
> In either case, the program runs correctly, whatever the number I
> write after "mpirun -np".
>
> I'd check whether there is any compilation mismatch between OMPI and
> Boost, check again threads support, etc.

I've tested it again. I recompiled my OpenMPI with thread, but not  
with "mpithread" support. The compile flags are now:
cxx threads -elibc_FreeBSD -fortran -heterogeneous -ipv6 -mpi-threads - 
pbs -romio -vt (- disable the feature).

The code works, so I think the "mpi-threads" option makes the problem

Best regards

Phil