The ABINIT Users Mailing List ( CLOSED )

[matthieu verstraete <matthieu.jean.verstraete@gmail.com>]

Re: 1) you are probably parallelizing over 16 k-points, so going from
4 to 5 procs gives distributions of 4 4 4 4 and 3 3 3 4 k-points per
proc, so the net time is the same, with 3 cores waiting in the second
case. Should be the same for 6 and 7: 2 2 3  3 3 3 vs  2 2 2  2 2 2  3

Normally abinit complains in the log files about this.

For higher scaling, you will need combined k, band, and FFT
parallelization. There are tutorials on this somewhere in the online
doc, but it is also being overhauled in 5.8. Also there was a
presentation by Francois Bottin at the developper's conference last
week, which should come up on
http://www.abinit.org/2009-Abinit/program.php soon.

The end numerical precision will depend on your tolerance (toldff?),
but also on accumulated numerical error, which may be different, in
particular for different processor distributions, and more so for
different algorithms, as in the k-g-b parallelization.

Matthieu

On Sun, Apr 5, 2009 at 12:01 AM, David M. Wood <dmwood@mines.edu> wrote:
> Greetings all!
>
> I have a couple of questions below.
>
> I recently built abinip 5.7.3 on an 8-core 2008 Mac Pro and found a cheerily
> quasi-linear dependence of job speed on the number of processors used.  But
> I saw no improvement at all in going from 4 to 5 processors, or from 6 to 7.
>  T (Top panel of attached PDF graphic, where more details are given.)  The
> machine has 4 processors each
>
> G1
>
> on 2 dies.
>
> 1. Any idea why the improvement is not monotonic in the number of
> processors?
>
> I found that the tail end of the iteration history (as measured by max grad
> of the force/grad in the optcell=2 portion of my runs) differed slightly
> depending on the number of processors.  (Bottom panel of PNG graphic.  I
> found no physically significant changes in cell dimensions and far less in
> the total energy, so I'm not worried.)
>
> 2. Does this ever become an issue when dealing with hundreds or thousands of
> processors in a cluster, where I'm heading, and is this normal behavior?
>
> Many thanks!
> DMW
>
> David M. Wood
> dmwood@Mines.EDU
>
>
>
>
>
>



-- 
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Dr. Matthieu Verstraete

European Theoretical Spectroscopy Facility (ETSF)
Dpto. Fisica de Materiales,
U. del Pais Vasco,
Centro Joxe Mari Korta, Av. de Tolosa, 72,   Phone: +34-943018393
E-20018 Donostia-San Sebastian, Spain        Fax  : +34-943018390

Mail : matthieu.jean.verstraete@gmail.com
http://www-users.york.ac.uk/~mjv500