[gpaw-users] Efficiency of LCAO vs. FD

[Ask Hjorth Larsen]

Hi Miguel

There is a (not fully systematic) benchmark of LCAO production
calculations for nanoparticles in my thesis, see page 32 and
thereabouts:

  http://dcwww.camd.dtu.dk/~askhl/files/askhl-thesis-cc.pdf
  (This version of the thesis is stripped of publications that may
suffer copyright restrictions and may be freely distributed)

Note the kink at at n=50 atoms where scalapack is switched on and you
will immediately see its importance.  With good network it will scale
to more nodes, although that was not the case on our old cluster.

Both hardware and software have improved significantly since.  Many
parts have been optimized although ScaLAPACK itself has not changed
much, and that will be the bottleneck for large systems still.

Unfortunately I don't have a good comparison to FD mode.  It is
difficult to do a general comparison to FD because the performance
differs widely depending on the type and size of system.  LCAO has the
greatest advantage for large systems.  It have typically seen large
and large-ish systems where LCAO was more than an order of magnitude
faster, but probably not two orders of magnitude.

Best regards
Ask

2016-11-01 13:41 GMT+01:00 Miguel Caro via gpaw-users
<gpaw-users at listserv.fysik.dtu.dk>:
> Hi all,
>
> I was wondering if there is any (more or less) comprehensive comparison
> between the performance of LCAO mode versus finite-difference mode, and
> perhaps the effect of different parallel options on performance (e.g.
> ScaLAPACK). I am thinking for example about a graph where CPU timings for
> LCAO and FD are plotted together as a function of system size. I would be
> very grateful if you could point me towards this kind of resource.
>
> Many thanks,
> Miguel
> --
> Dr. Miguel Caro
> Postdoctoral researcher
> Department of Electrical Engineering and Automation,
> and COMP Centre of Excellence in Computational Nanoscience
> Aalto University, Finland
> Personal email: mcaroba at gmail.com
> Work: miguel.caro at aalto.fi
> Website: http://mcaroba.dyndns.org
>
> _______________________________________________
> gpaw-users mailing list
> gpaw-users at listserv.fysik.dtu.dk
> https://listserv.fysik.dtu.dk/mailman/listinfo/gpaw-users