The ABINIT Users Mailing List ( CLOSED )

[Xavier Gonze <gonze@pcpm.ucl.ac.be>]

Dear Nichols Romero,

Nichols A. Romero wrote:
> Hi,
>
> Can anyone on this list tell me more about how the parameter nbdblock
> works?

In the conjugate gradient algorithm to optimize wavefunctions
(see the Payne et al review), each
state should be improved in turn, sequentially, keeping the other fixed,
and also orthogonalizing the state currently optimized with respect to the
other states. This guarantees the convergence of the algorithm.

When you activate the nbdblock variable, a block of states will be
optimized in parallel, keeping them orthogonalized with respect to the
ORIGINAL states. Then, after optimization, the new states will be
orthogonalized with respect to each other. The convergence is no more 
guaranteed.
In practice, this algorithm works quite well when the number
of states in a block does not exceed 4. Beyond that number, it
might work very well for selected systems, or not at all ...

> Also, in the case of many k-points, is it a good idea to also do
> additional turn on parallelization on bands with wfoptalg=1?

When you have more k points than processors, avoid using wfoptalg=1 .

So, you should try ...
And, definitely, this is interesting only when you have processors
that have nothing to do when you only parallelize over k points.

Xavier Gonze