iyamazaki

I am happy to look, but it is probably good if someone like @e10harvey could look at TPL stuff. You probably meant TPL for CuSolver (not LAPACK)?

Hi, @cgcgcg. We are still discussing the changes with @searhein. I added WIP label for now.

@mhoemmen is right. These solvers (e.g., single-reduce, communication-avoiding, and pipelined GMRES) are not registered by default. So you need to register. We don't have a simple example, but there are...

I think for SuperLU, Amesos2 just adds new arguments for new versions, e.g., ``` extern void sgstrf (SLU::superlu_options_t*, SLU::SuperMatrix*, int, int, int*, void *, int, int *, int *, SLU::SuperMatrix...

Thank you for the review, @cgcgcg. I added a test.

If there are load-imbalance among MPI processes, I was wondering how that will show up.. We've seen load imbalance in compute (e.g., local SuperLU with an irregular mesh), but because...

I am trying to follow all the related issues and also to reproduce this particular one (Amesoss2, or Ifpack2 additive Schwarz, with CUDA11 + UVM off), but having hard time....

I agree with @cgcgcg. The issue may not be the Amesos MUMPS interface, but may be with your MUMPS to METIS interfacing. It may be good to run MUMPS' example/test...

Under MUMPS, I see a directory called ``examples``, @swiesheier .

Under my MUMPS directory, I have: ```` $ ls MUMPS_5.3.3 ChangeLog CREDITS doc examples include INSTALL lib libseq LICENSE Makefile Makefile.inc Make.inc MATLAB PORD README SCILAB src VERSION ````