Oh, I just replied to your previous email. On 06/02/2014 08:53 PM, Jakub Łuczyński wrote:

Solved, it was my error.. int block_lengths[] = { 4, 2 }; should be: int block_lengths[] = { 4, 1 };

Great that you solved it! That makes sense, it sounds as it may have caused some type of inadvertent buffer overlap.

Did not read man carefully.

There was overflow all the time. On MPICH it caused error, on openmpi and BGP it was silent.

I guess memory issues are not always reproducible across compilers, libraries, etc.

On Tue, Jun 3, 2014 at 1:15 AM, Jakub Łuczyński <[email protected] <mailto:[email protected]>> wrote:

P.S. for clarity

typedef long long m_entry_t;

On Tue, Jun 3, 2014 at 1:09 AM, Jakub Łuczyński <[email protected] <mailto:[email protected]>> wrote:

After installation of current version (3.1)

$ mpiexec -n 2 ~/tmp/opencl/msp-par.exe 10 10 1 Fatal error in PMPI_Reduce: Internal MPI error!, error stack: PMPI_Reduce(1259)........: MPI_Reduce(sbuf=0x7fff40ad30c8, rbuf=0x7fff40ad30e0, count=1, dtype=USER<struct>, op=0x98000000, root=0, MPI_COMM_WORLD) failed MPIR_Reduce_impl(1071)...: MPIR_Reduce_intra(822)...: MPIR_Reduce_impl(1071)...: MPIR_Reduce_intra(877)...: MPIR_Reduce_binomial(130): MPIR_Localcopy(123)......: memcpy arguments alias each other, dst=0x7fff40ad30e0 src=0x7fff40ad30c8 len=32 MPIR_Reduce_intra(842)...: MPIR_Reduce_impl(1071)...: MPIR_Reduce_intra(877)...: MPIR_Reduce_binomial(246): Failure during collective

On Tue, Jun 3, 2014 at 12:25 AM, Gus Correa <[email protected] <mailto:[email protected]>> wrote:

This is an old version of mpich. Is it perhaps still using the mpd ring? [If so, you need to start the mpd ring, if not already set, before you launch the job. But that method was phased out.] It may be worth updating to the latest mpich stable and use the current mpiexec (hydra) to launch the job.

http://www.mpich.org/__downloads/ <http://www.mpich.org/downloads/> http://www.mpich.org/__documentation/guides/ <http://www.mpich.org/documentation/guides/> http://wiki.mpich.org/mpich/__index.php/Using_the_Hydra___Process_Manager <http://wiki.mpich.org/mpich/index.php/Using_the_Hydra_Process_Manager>

On 06/02/2014 06:03 PM, Lu, Huiwei wrote:

Hi Kuba,

Since it works with both Open MPI and BGP, it is most likely a problem of your MPICH installation or your platform.

We have stopped supporting the Windows platform a while ago due to lack of developer resources. Please refer to our FAQ for more information: http://wiki.mpich.org/mpich/__index.php/Frequently_Asked___Questions#Q:_Why_... <http://wiki.mpich.org/mpich/index.php/Frequently_Asked_Questions#Q:_Why_can.27t_I_build_MPICH_on_Windows_anymore.3F>

If it is on windows platform, we recommend you use Microsoft MPI, which can be found here: http://msdn.microsoft.com/en-__us/library/bb524831(v=vs.85).__aspx <http://msdn.microsoft.com/en-us/library/bb524831(v=vs.85).aspx>

We also encourage you to use the latest MPICH on Linux or OSX platforms, which can be downloaded here: http://www.mpich.org/__downloads/ <http://www.mpich.org/downloads/>

— Huiwei

On Jun 2, 2014, at 4:49 PM, Jakub Łuczyński <[email protected] <mailto:[email protected]>> wrote:

I wrote my assignment using MPI, and tested it both locally on Open MPI (1.6.5) and on IBM Blue Gene/P (with mpi implementation provided by IBM). Everything worked fine. Turns out that our solutions are tested also in our labs where MPICH, is installed:

$ mpich2version MPICH2 Version: 1.4.1p1

And when I run my solution there I get this strange error: $ mpirun -n 2 msp-par.exe 10 10 1 Internal Error: invalid error code 409e10 (Ring ids do not match) in MPIR_Reduce_impl:1087 Fatal error in PMPI_Reduce: Other MPI error, error stack: PMPI_Reduce(1270).....: MPI_Reduce(sbuf=__0x7fff693a92e8, rbuf=0x7fff693a9300, count=1, dtype=USER<struct>, op=0x98000000, root=0, MPI_COMM_WORLD) failed MPIR_Reduce_impl(1087):

I am literally out of ideas what is wrong!

Below source code fragments (c++):

struct msp_solution { int x1, y1, x2, y2; m_entry_t max_sum; msp_solution(); msp_solution(const pair<int, int> &c1, const pair<int, int> &c2, int max_sum); friend bool operator<(const msp_solution &s1, const msp_solution &s2); };

void max_msp_solution(msp_solution *in, msp_solution *inout, int, MPI_Datatype*) { *inout = max(*in, *inout); }

// somewhere in code { MPI_Datatype MPI_msp_solution_t; MPI_Op max_msp_solution_op;

// create MPI struct from msp_solution MPI_Datatype types[] = { MPI_INT, MPI_LONG_LONG_INT }; int block_lengths[] = { 4, 2 }; MPI_Aint base_addr, x1_addr, max_sum_addr; MPI_Get_address(&collected___solution, &base_addr); MPI_Get_address(&collected___solution.x1, &x1_addr); MPI_Get_address(&collected___solution.max_sum, &max_sum_addr);

MPI_Aint displacements[] = { x1_addr - base_addr, max_sum_addr - base_addr };

MPI_Type_create_struct(2, block_lengths, displacements, types, &MPI_msp_solution_t); MPI_Type_commit(&MPI_msp___solution_t);

// max reduction function MPI_Op_create((MPI_User___function *) max_msp_solution, 1, &max_msp_solution_op);

...

msp_solution solution, received_solution; MPI_Comm comm, ... // comm is created using MPI_Comm_split // solution is initialized MPI_Reduce(&solution, &received_solution, 1, MPI_msp_solution_t, max_msp_solution_op , 0, MPI_COMM_WORLD); // ERROR above!!! }

Is there some error in this? How can I make it run? P.S. MPI_Send and MPI_Recv on my struct MPI_msp_solution_t seems to work fine

Thanks in advance! Best regards, Kuba _________________________________________________ discuss mailing list [email protected] <mailto:[email protected]> To manage subscription options or unsubscribe: https://lists.mpich.org/__mailman/listinfo/discuss <https://lists.mpich.org/mailman/listinfo/discuss>

_________________________________________________ discuss mailing list [email protected] <mailto:[email protected]> To manage subscription options or unsubscribe: https://lists.mpich.org/__mailman/listinfo/discuss <https://lists.mpich.org/mailman/listinfo/discuss>

_________________________________________________ discuss mailing list [email protected] <mailto:[email protected]> To manage subscription options or unsubscribe: https://lists.mpich.org/__mailman/listinfo/discuss <https://lists.mpich.org/mailman/listinfo/discuss>

_______________________________________________ discuss mailing list [email protected] To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss