[MPICH] MPICH2 does not work over Windows XP network with ib card

[DevTeam]

Hi all,



I have troubles to run my parallel application linked with mpich2 on a windows XP cluster made of a frontal and 4 nodes (n1, n2, n3, n4) with ib cards : 



If I use VC++ 6.0 and Compaq Fortran 6.6C to compile and link, my code works on a given node (with -localonly) but on the cluster I obtain the following output (using or not the ib cards):



Z:\test>"C:\Program Files\MPICH2\bin\mpiexec.exe" -map Z:\\FRONTAL\MySHARE -pwdfile Z:\test\file.txt -wdir Z:\test -hosts 4 N1 N2 N3 N4 -exitcodes Z:\test\test_mpich2_ib.exe

rank: node: exit code

0: N1: -1073741515

1: N2: -1073741515

2: N3: -1073741515

3: N4: -1073741515



Now if I use .NET and Intel Fortran 8.1 to link (with mpich2ib.lib, fmpich2.lib) my application it works ! 



My problem is now that when I try to use the ib chanel the code freeze at the first all_reduce operation. Any idea ?



Thanks a lot in advance for your help



Etienne Perchat





Here are details about my cluster installation :

 

  1.. The cluster is made of  a windows 2003 server frontal on which I have installed mpich2-1.0.2-1-win32-ia32.msi
4 nodes under win32 XP-pro on which I have executed smpd -install

The disk Z: is on the frontal and has been shared. Z has been mapped on all nodes

I have copied in local all the mpich2 dll's

The personal XP firewall has been disconnected on 4 nodes.

The cluster has not been not integrated on a windows domain but just on a windows workgroup.


 

  2.. The application is made of a mix of fortran an C++, where all the MPI calls are made in C++ through the old fashion ("C" bindings) and I compile in C++ having MPICH_SKIP_MPICXX defined.  
 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/mpich-discuss/attachments/20051103/8f364f0a/attachment.htm>