I can use ELPA when no_procs > matrix size about elpa HOT 3 CLOSED

In short: no you cannot do this. Here is why:
blacs_grid_info does return an info value. For any value != 0, this indicates that the Blacs-grid has not been set up successfully. Since the algorithms in ELPA do rely on a correct Blacs-grid, you can run into the error messages you quoted, if your combination of matrix size, block size and process grid does not return info = 0 in blacs_grid_info.
I do agree, that with ScaLapack this might still work, since ScaLapack might be able to work with an "uncorrect" Blacs-grids, ELPA however does not.
I dare speculate, however, that in the cases where ELPA does return the error messages from above, but ScaLapack on the other hand runs successfully, that the info return code !=0 (i.e. not correct blacs setup).

However, there is a way to solve this with ELPA:

• ELPA does allow you to create multiple solver objects, and
• you can create sub-MPI-communicators (MPI_COMM_SPLIT) , such that the number of MPI processes chosen does allow a correct (info=0) blacs initialization.
  So you could create several sub-communicators , and with these create several ELPA objects, and use the correct communicator (ELPA object) depending on the matrix size.

Concerning the 1x1 matrix: I admit this is not implemented in ELPA at all (but could be special cased)

from elpa.

Awesome. Thanks for your suggestions. I thought I was going crazy.

I admit 1x1 matrix is not the target audience of a highly parallel diagonalizer.

I will take a look at the sub-MPI communicators as recommended and potentially we will just modify the code to use a serial diagonalizer for small matrix where no_elems < n_procs

from elpa.

I will close this "issue". Thanks for your help.

from elpa.