[Trilinos-Users] c++ cray compiler gets wrong comm in mpi init

Heroux, Michael A maherou at sandia.gov
Wed Dec 2 15:11:30 MST 2009


I am almost sure you have some kind of incompatibility between the Cray and Gnu built code.  Roger Pawlowski pointed out that the value (24050) is the value of the next data member in Epetra_MpiCommData, which is the underlying reference counted data class for Epetra_MpiComm.


On 12/2/09 1:08 PM, "Jason Slemons" <slemons at cray.com> wrote:

I did compile the test with cray but the library with gnu. In my experience linking to gnu built libraries is okay but please to say If that's not typically the case.

I checked 'sizeof(comm)' and its 32 if I build with gnu or cray or pathscale or intel.

From: Heroux, Michael A [mailto:maherou at sandia.gov]
Sent: Wednesday, December 02, 2009 10:27 AM
To: Jason Slemons; trilinos-users at software.sandia.gov
Subject: Re: [Trilinos-Users] c++ cray compiler gets wrong comm in mpi init


What is the size of your MPI communicator?


On 12/1/09 5:28 PM, "Jason Slemons" <slemons at cray.com> wrote:
I have a short c++ program that gets the wrong communicator when using Epetra_MpiComm. pgi, pathscale, gnu and intel all seem to have no problem but cray gives me that  there are 24,050 procs. Heres the program:

#include "Epetra_MpiComm.h"

int main(int argc, char *argv[])
  // Initialize MPI
  Epetra_MpiComm comm( MPI_COMM_WORLD );

 cout<<comm.NumProc()<<endl<<"is too big"<<endl;

Heres the output:
slemons at seal1:/lus/nid00008/slemons> aprun Epetra_BasicPerfTest_test.exe
is too big
Application 2752567 resources: utime 0, stime 0

Any ideas? Ive attached Epetra_MPIComm.h in case its helpful.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: https://software.sandia.gov/pipermail/trilinos-users/attachments/20091202/20834216/attachment.html 

More information about the Trilinos-Users mailing list