Dear Joao,
the problem is that, You do not use the &reqs[i] correctly in the MPI_Send --
reqs[0] will not be initialized; but wait for in MPI_Waitall...
Change:
rc = MPI_Isend(&a , 1 , MPI_INT , i , 0 , MPI_COMM_WORLD , &reqs[i]);
to
rc = MPI_Isend(&a , 1 , MPI_INT , i , 0 , MPI_COMM_WORLD , &reqs[i-1]);
and it will work.
These kind of buglets can be very neatly found using the memchecker component
using valgrind, which is implemented in the upcoming OpenMPI-1.3.
Please check the FAQ on:
http://www.open-mpi.org/faq/?category=debugging
on how to use and what errors to find...
I attached the program with a bit of cleanups (e.g. why do You need
MPI_ANY_SOURCE in the recv on the workers...).
With best regards,
Rainer
On Freitag, 27. Juni 2008, Joao Marcelo wrote:
> Hi,
>
> I'm starting to code with MPI and decided to use openmpi. I'm using
> Ubuntu Linux with GCC version 4.2.3 and OpenMPI 1.2.5 (distribution
> package). The output of "ompi_info -- all" is attached. I'm also
> sending a copy of the source code I'm trying to run.
>
> What I'm trying to do is selecting process with rank 0 as the main
> process, which you send data to all others processes with non-blocking
> MPI_Isend. After all functions calls are performed, it will wait with
> MPI_Waitall() and the print a ending message. The compile command is :
>
> $ mpicc.openmpi -fopenmp mpi.c -o mpi
>
> And I'm running with :
>
> $ mpirun.openmpi -np 2 ./mpi
>
> The problem is that after the output below, the processes halt,
> without the ending message.
>
> jmhal@galileo:~/$ mpirun.openmpi -np 2 ./mpi
> Main process: 0.
> 1 - received value, rc value: 100 0
>
> If I start more processes, for example 3, this what I get:
>
> jmhal@galileo:~/$ mpirun.openmpi -np 3 ./mpi
> Main process: 0.
> End.
> 1 - received value, rc value: 100 0
> 2 - received value, rc value: 100 0
> [galileo:15807] *** Process received signal ***
> [galileo:15807] Signal: Segmentation fault (11)
> [galileo:15807] Signal code: Address not mapped (1)
> [galileo:15807] Failing at address: 0xd
> [galileo:15807] [ 0] [0xb7f2a440]
> [galileo:15807] [ 1] /usr/lib/openmpi/lib/openmpi/mca_pml_ob1.so
> [0xb7b49da8] [galileo:15807] [ 2]
> /usr/lib/libmpi.so.0(mca_pml_base_close+0x90) [0xb7efd830] [galileo:15807]
> [ 3] /usr/lib/libmpi.so.0(ompi_mpi_finalize+0x1d6) [0xb7ec48d6]
> [galileo:15807] [ 4] /usr/lib/libmpi.so.0(PMPI_Finalize+0x56) [0xb7ee27d6]
> [galileo:15807] [ 5] ./mpi(main+0x1bc) [0x8048a60]
> [galileo:15807] [ 6]
> /lib/tls/i686/cmov/libc.so.6(__libc_start_main+0xe0) [0xb7c47450]
> [galileo:15807] [ 7] ./mpi [0x8048841]
> [galileo:15807] *** End of error message ***
> mpirun.openmpi noticed that job rank 0 with PID 15807 on node galileo
> exited on signal 11 (Segmentation fault).
> 2 additional processes aborted (not shown)
>
>
> Which is ok since all messages were print, but what about the error
> message?
--
----------------------------------------------------------------
Dipl.-Inf. Rainer Keller http://www.hlrs.de/people/keller
HLRS Tel: ++49 (0)711-685 6 5858
Nobelstrasse 19 Fax: ++49 (0)711-685 6 5832
70550 Stuttgart email: kel...@hlrs.de
Germany AIM/Skype:rusraink
#include <stdio.h>
#include <stdlib.h>
#include "mpi.h"
#define ROOT 0
int main( int argc, char *argv[] )
{
int numtasks, rank, rc;
rc = MPI_Init(&argc,&argv);
if (rc != MPI_SUCCESS) {
printf ("Error starting MPI program. Terminating.\n");
MPI_Abort(MPI_COMM_WORLD, rc);
}
MPI_Comm_size(MPI_COMM_WORLD,&numtasks);
MPI_Comm_rank(MPI_COMM_WORLD,&rank);
if (rank == ROOT) {
printf("Main process: %d.\n", rank);
int a = 100;
int i;
MPI_Request *reqs;
MPI_Status *stats;
reqs = (MPI_Request *) malloc((numtasks - 1)*sizeof(MPI_Request));
stats = (MPI_Status *) malloc((numtasks - 1)*sizeof(MPI_Status));
for ( i = 1; i < numtasks; ++i){
/*
* Being very picky, one should not use the *same* buffer
* in two non-blocking communications
*
* The main problem however is that &reqs[i] is wrong: reqs[0] is
* not initialized at all.
* change to &reqs[i-1] works.
*/
rc = MPI_Isend(&a , 1 , MPI_INT , i , 0 , MPI_COMM_WORLD , &reqs[i-1]);
}
MPI_Waitall(numtasks - 1 , reqs , stats);
free(reqs);
free(stats);
printf("End.\n");
} else {
int b;
int rc;
MPI_Status status;
rc = MPI_Recv( &b , 1 , MPI_INT, ROOT , 0, MPI_COMM_WORLD , &status);
printf("%d - received value, rc value: %d %d\n", rank , b, rc);
}
MPI_Finalize();
return 0;
}
