[petsc-users] MatAssemblyEnd taking too long

[Manav Bhatia]

> On Aug 20, 2020, at 11:21 AM, Stefano Zampini <stefano.zampini at gmail.com> wrote:
> 
> 
> 
>> On Aug 20, 2020, at 5:59 PM, Manav Bhatia <bhatiamanav at gmail.com <mailto:bhatiamanav at gmail.com>> wrote:
>> 
>> 
>> 
>>> On Aug 20, 2020, at 8:31 AM, Stefano Zampini <stefano.zampini at gmail.com <mailto:stefano.zampini at gmail.com>> wrote:
>>> 
>>> ((Mat_SeqAIJ*)aij->B->data)->nonew
>>> mat->was_assembled
>>> aij->donotstash
>>> mat->nooffprocentries
>>> 
>> 
>> The values for the last three variables are all False on all 8 processes. 
> 
> Thanks, it seems a bug in MPI or on our side. As Matt said, can you run with -matstash_legacy?

I did that with the result that a smaller case that had been running got stuck in the MatAssemblyEnd_MPIAIJ routine with -matstash_legacy. Not sure what this implies. 

> Also, make sure to run with a debug version of PETSc (configure using —with-debugging=1).

Trying that now. 

> How feasible is to write a driver code to run with the same mesh, same discretization and same equations to solve but with the matrix assembly only? Does it hang in that case too?

My code is on GitHub (https://github.com/MASTmultiphysics/MAST3), including the specific example that is producing this error (https://github.com/MASTmultiphysics/MAST3/tree/master/examples/structural/example_6), but has multiple dependencies, including libMesh/PETSc/SLEPc/Eigen. 

The mesh generation and sparsity pattern creation is currently done by libMesh. 

If needed, I may be able to store all the necessary information (mesh, nnz, noz) in a text file, remove dependencies on libMesh and directly try to initialize with information from a file. Would that help? 

> 
>> 
>> Regards,
>> Manav

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20200820/5fe96e2c/attachment.html>