[Wien] wien and hyperthreading

Peter Blaha pblaha at zeus.theochem.tuwien.ac.at
Tue Mar 15 11:40:57 CET 2005


The timing is very unusual. 
lapw1 should take longer than lapw0 (maybe you use k-point parallel lapw1, 
this would explain it eventually).
lapw1 should also take longer than lapw2 ! Are you using run_lapw -I ? 
i.e. are you sure that your switch in case.in2 is TOT and not FOR. Only 
with FOR lapw2 might take longer than lapw1, but this should be used only 
in the last cycle.

I do have a working lapw0_mpi, but most likely this is not helpful to you. 
It was compiled with the pgi compiler and uses mpich (also installed via 
pgi-4.0 compiler). But this depends all on the mpi-Installation and this 
was done by the computing center.

lapw0 can be speeded up by reduction of the radial mesh (eg. use only 381 
points) and also GMAX (in2) determines the cputime for this part. Of 
course, when doing GGA you loose some accuracy but for some timeconsuming 
structural relaxations this should be ok. Use   clminter   to interpolate 
to a crude (and than back to a fine radial mesh).
(I hope you are using the latest lapw0 version ? L.Marks has speeded up 
lapw0 by quite some amount.....

> Was just getting desperate to squeeze more speed out of the system. As no one
> seems to have a got linux mpi version working (or at least no one has answered
> previous posts). The system I'm working on now takes only 28 minutes in lapw1,
> but 40 minutes in lapw0 (this is why mpi would be nice) and 1h20 in lapw2


                                      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/
--------------------------------------------------------------------------




More information about the Wien mailing list