[MOAB-dev] problem with exchange_ghost_cells

[Iulian Grindeanu]

Thank you, Lorenzo, for reporting it; 
I have tried with the latest repo, and indeed, it hangs out with mesh.h5m. 

I have tried with bridge dimension 0 and 1, and it seems to get out of ghosting; 
So for bridge dimension 2 that you are using, there is a problem. 
" 
if (pcomm->rank() == 0) 
std::cout<<pcomm->rank()<<" Exchange ghost cells"<<std::endl; 
result = pcomm->exchange_ghost_cells(dim, 1 ,1,0,true,true); 
assert(MB_SUCCESS==result); 
" 
In the meantime, can you use bridge 0 or 1? 

Thanks, 
Iulian 

----- Original Message -----

| Dear all,
| I’m writing to inform you all of a problem with MOAB in parallel that
| bothers me a lot.
| Sometimes after reading a parallel .h5m mesh file and resolving
| shared entities the execution hangs while exchanging ghost cells.
| The issue arises randomly but I noticed a clear tendency to occur
| when the number of elements in each partition decreases, that is,
| given the same mesh, when the number of partition increases.
| Here attached is a simple code that should allow you to reproduce the
| issue (I tried to run both with 4.6.0 and trunk and obtain the same
| behavior).
| mesh.h5m and mesh_ok.h5m are two different 16 part partitions of the
| same mesh (a 3D tetrahedral elements mesh).
| While with mesh_ok.h5m the code works perfectly, using mesh.h5m it
| hangs, meaning that the process with rank 1 does not return from
| exchange_ghost_cells if the code is executed in parallel with 16
| processors (mpiexec -n 16 <reader_exec_name>).
| The weird thing is that the code works if executed on less that 16
| processors

| Any help is appreciated. Thanks in advance.
| Lorenzo
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/moab-dev/attachments/20131118/1129dbf2/attachment.html>