[Pw_forum] gipaw parallel run is too slow

[Layla Martin-Samos]

Dear Ren, there is at moment a big effort done in this direction from the
computational center in Irland (ICHEC) and the developers of QE-GIPAW (see
http://qe-forge.org/projects/qe-gipaw/).

bests

Layla

2011/11/18 Ren PJ <renpj at dicp.ac.cn>

> On Fri, 18 Nov 2011 10:54:56 +0100
>
> > Dear Pengju Ren,
> > when you increase the numebr of processors you increase the
> > available computing power, but also the amount of communications on
> > the network connectign those nodes. As the network has a finite
> > speed, there will be a point where the code looses more time
> > exchanging information between the cpus than doing any calculation.
> >
> > Possible solutions:
> > 1. use fewer cpus (between 8 and 160 there is plenty of space)
> > 2. try with a different number of pools (you could try
> > nkpoints*ncores_per_node)
> > 3. buy some faster network hardware
> >
> > best regards
> >
>
> Dear Lorenzo Paulatto,
> Thanks for your kind answer.
> When I ran pw.x with 160 cpus it was about 10 times faster than with 8
> cpus, I think the parallel efficiency is reasonable. So does gipaw.x
> need more communications between nodes than pw.x? Maybe the gipaw code
> could be improved by reducing the communication frequency.
>
> Best Wishes
>
> ================================================
> 任鹏举
> Pengju Ren
> renpj at dicp.ac.cn
> State  Key Laboratory of Catalysis,
> Dalian Institute of Chemical Physics,
> Chinese Academy of Sciences
> 457 zhongshan Road, Dalian, 116023, P.R. China
> _______________________________________________
> Pw_forum mailing list
> Pw_forum at pwscf.org
> http://www.democritos.it/mailman/listinfo/pw_forum
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.quantum-espresso.org/pipermail/users/attachments/20111118/fc1feac2/attachment.html>