[Pw_forum] parallel scaling in PWSCF

Nichols A. Romero naromero at gmail.com
Wed Apr 26 23:32:32 CEST 2006


I am using the CVS version of PWSCF which I checked out on Monday.

I did 216 carbon atoms (bulk diamond) with USPP from the PWSCF
website. Using E_c = 40 Ry and only the gamma point:

Davidson took 14 min 33 sec. CG took 22 min 46 sec. That is the CPU
time as reported by PWSCF on an SGI Altix running on 8 processors.

I tried a 1728 atom calculation using the same parameters as above and
I ran into some trouble when trying to run it in parallel on 128

PWSCF seems to seg. fault with an error *possibly* originating in a
setup.f90 routine located here

I was using the 'Davidson' algorithm in the 1728 atom calculation. I
have *not* tried the larger calculation with 'cg.'

I hope this is helpful.

On 4/26/06, carlo sbraccia <sbraccia at sissa.it> wrote:
> Hi,
> beware using the CVS version of PWSCF: the parallel Davidson has not yet
> been fully tested and seems to be buggy.
> For this reason we are going to disable it until we shall be able to
> solve the problem. To avoid the serial bottleneck in the diagonalization
> you can use conjugate-gradient.
> carlo
> Andrea Ferretti wrote:
> >Hi everybody,
> >
> >I am currently running a Copper surface with 140 Cu atoms + a molecule...
> >the system has 1642 electrons and (due to metallicity) the calculation is
> >performed for 985 bands (few kpt, like 4)...
> >due to the 11 electrons for each Cu atom, I have a huge number of bands in
> >a (relatively) small cell, and so a (relatively) low number of PWs respect
> >to nbnd.
> >taking a look at the dimension of wfc, no problem with memory in
> >principle, even if, due to the weird
> >dimensions of the system, non-scalable memory is quite large, around 1Gb.
> >
> >on a IBM Sp5 machine I observed a severe limit in the scaling passing from
> >32 to 64 procs using both espresso 2.1.x and espresso 3.0...
> >( anyway, I succeeded in performing a "relax" calculation for the system
> >!!!! )
> >
> >as far as I know, this problem might be connected to a serial part in the
> >diagonalization which has been parallelized in the current CVS version
> >(as already pointed out by Axel)...
> >At the moment I am testing this CVS version against my system, I will let
> >you know the results as soon as possible...
> >
> >cheers
> >andrea
> >
> >
> >
> >
> _______________________________________________
> Pw_forum mailing list
> Pw_forum at pwscf.org
> http://www.democritos.it/mailman/listinfo/pw_forum

Nichols A. Romero, Ph.D.
1613 Denise Dr. Apt. D
Forest Hill, MD 21050
443-567-8328 (C)
410-306-0709 (O)

More information about the users mailing list