The ABINIT Users Mailing List ( CLOSED )

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

Dear David,

prendergast2@llnl.gov wrote:
> Can someone explain how to parallelize over bands in abinip?
> The features of version 4.0 say that one can parallelize using MPI over 
> bands, however if I run with more processes than kpts*spin channels I get a 
> warning that I'm wasting processes and I see no speed up.
> Is there an input parameter that controls this?

For the ground-state calculationss,
you need to put wfoptalg to 1 and nbdblock to the
number of processors you would like to work on one k point
(nbdblock 2 to 4 might be fine, while for higher values ,
the whole SCF loop might be unstable).
Note that you do not gain anything as concerns memory.
See Test_v3#41 for an example input file.

For the response-function case, the band-by-band
parallelism is automatic.

In v4.3 (by December 2003 ?), it is likely that the Fourier space parallelism
will work, with spread of the data over all the processors,
and large saving in memory.

Good continuation,
Xavier