[gpaw-users] Release notes about parallelization

[Christian Glinsvad]

To all GPAW users

As of version 0.7, a new keyword has been introduced to describe various
parallelization options in a collective way by means of one Python dictionary.
We hope that you've all gotten used to the 'convergence' keyword, which was
introduced similarly back in version 0.3 to control SCF convergence criteria.

This new 'parallel' keyword replaces 'parsize' and 'parsize_bands' but
also includes other options, which you can read more about here:
https://wiki.fysik.dtu.dk/gpaw/documentation/parallel_runs/parallel_runs.html#parallization-options

To summarize the most noticeable changes we've made, you must now do e.g.:

     calc = GPAW(..., parallel={'domain': (2,2,1), 'band': 3}, ...)

instead of the old syntax, which has been deprecated and will issue a warning:

     calc = GPAW(..., parsize=(2,2,1), parsize_bands=3, ...)

Note that the equivalent command line options --domain-decomposition and
--state-parallelization are still available and function as before.


Special note to ScaLAPACK / BLACS users:

The newly implemented BLACS interface is nearing completion and will be
available for use in GPAW by specifying --gpaw=blacs=1 on the command line.

In addition to the original --sl_diagonalize and --sl_inverse_cholesky
command line options, the latest version of GPAW now also supports both
--sl_default and --sl_lcao as situationally dependent ScaLAPACK options.

It should be noted that all four of these --sl_... arguments now require
just 3 comma-separated integers (as opposed to 4 with the last one ignored).

As the name suggest, --sl_default will be used as the default value in case
one does not provide e.g. --sl_diagonalize explicitly. Presently, using BLACS
is an all-or-nothing type of deal and you will be expected to (at least)
specify --sl_default or individual values for each of the --sl_... options.
We're of course working hard on relaxing this constraint such that GPAW will
once again support any combination of ScaLAPACK / BLACS and serial LAPACK for
diagonalization, orthonormalization and LCAO's general diagonalization.


Regards
The GPAW developers