[Wien] mpi and k-point para runs

[Peter Blaha]

> We are buying a number of linux boxes connected using Gigalan network cards
>
> As far as I remember from the workshop this is not sufficient to run mpi
> in lapw1 (not an issue since we run more than 1 k point anyway). I seem to
> remember that Peter said this was enough to run mpi for lapw0 though. Was
Yes, lapw0 should definitely run fine in parallel.

dstart is not a parallel program. Usually one would run it only once per
case. However, I admit sometimes it is boring to wait for dstart (so that
one can start the scf cycle). dstart has definitely some potential for
speedup. It is certainly not "optimized". E.g. it is quite useless to
calculate all those high LM values, which will change completely during
scf and probably one could also easily truncate the K-list much earlier.

Also mpi parallelization (over atoms) should be simple.
A typical "long" runtime is:

   1688.000001688.00629  -0.00629
 time recpr   716.350000000000
 time rhok    559.850000000000
 time rholm   8734.51000000000

recpr could be optimized e.g. by a better "sorting" algorithm
rhok by usage of ffts, and
the "rholm" part by switching off the higher LMs. At present you would simply
insert a "0 0" pair at the second position of case.in2, but a better
solution would be to switch it off as input option.

> wondering if dstart can also be run in parallel since for our large
> structure this part takes rather a long time. Also the final lapw2 takes
> about 1 hour for this system and was wondering if this can be run in mpi
> with the above configuration.

lapw2 is the same as lapw1. It runs in parallel anyway, as long as you
have "k-points".



                                      P.Blaha
--------------------------------------------------------------------------
Peter BLAHA, Inst.f. Materials Chemistry, TU Vienna, A-1060 Vienna
Phone: +43-1-58801-15671             FAX: +43-1-58801-15698
Email: blaha at theochem.tuwien.ac.at    WWW: http://info.tuwien.ac.at/theochem/
--------------------------------------------------------------------------