[Pw_forum] phonon parallel calculations
Andrea Dal Corso
dalcorso at sissa.it
Fri Feb 20 10:06:35 CET 2015
On Fri, 2015-02-20 at 09:19 +0100, Carlo Nervi wrote:
> Dear QE developers,
> I would like to have your opinion on the procedure I followed in doing
> phonon calculations.
> I found that the use of nimages on my machine (8 numa zones) greatly
> improves the speed of some calculations, but the first (initialization?)
> part of the calculation (epsil=.true., i.e. the Electric Fields
> Calculation) is much faster when performed without the use of "-ni 8"
> switch in ph.x
> So, I tested the following procedure and apparently worked for a small test
> 1) pw.x (vc-relax optimization) without using "-ni"
> 2) scf on the optimized structure containing "wf_collect=.true." and using
> "-ni 8"
pw.x has no image parallelization, so step 2 should be done without
images. The rest should be ok.
> 3) ph.x containing "only_init = .true." and without using "-ni"
> This calculate the electric fields
> 4) duplicate the directory _ph0 and copied to _ph1 ... _ph7 (in output
> 5) ph.x containing "recover = .true." and using "-ni 8"
> This is the part where I get most of the speedup
> 6) ph.x of the input used in 5), but now without using "-ni"
> This is to collect the partial data and produce the final .out
> I would greatly appreciate any comments in understanding if this procedure
> is correct.
> On a small crystal was working, but on a bigger crystal gave error.
> Either the procedure is not completely correct or I have to start to look
> for a problem in the RAM...
> Thank you,
> Pw_forum mailing list
> Pw_forum at pwscf.org
Andrea Dal Corso Tel. 0039-040-3787428
SISSA, Via Bonomea 265 Fax. 0039-040-3787249
I-34136 Trieste (Italy) e-mail: dalcorso at sissa.it
More information about the users