[Pw_forum] benchmark calculations for parallel PWSCF

Nicola Marzari marzari at MIT.EDU
Mon Oct 1 06:12:41 CEST 2007

Dear Sahu,

I have mostly CP scaling data here, on small machines:

The bottom line is that provided that you use a recent MPI 
implementation, scaling is certainly very good using gigabit
ethernet up to 12 processors. We haven't tested anything larger.
Of course, gigabit ethernet is a slow connection.

I've seen tests done on the Cray XT4 that went up to 5000 processors, 
with excellent scaling. Some notes on hyperparallelization can be
found on the wiki

I do not have data for PWSCF - of course scaling on k-points or
NEB images would be perfect, but also on G-vectors should go
reasonably well with hundreds of CPUs. Maybe someone else has data
to share ?


brsahu at physics.utexas.edu wrote:
> Dear PWSCF users,
> Is it possibe to get a info on benchmark PWSCF SCF parallel  
> calculations (CPU time versus number of processors) for a particular  
> physical system on any computing machines (linux clusters, IBM (Power  
> series), SGI, HP etc.
> Is it possible to share the numbers with me if you have such numbers?  
> I am aware that lot depends on what type of processors, what  
> communication channel, what libraries etc. But at this stage I am not  
> looking at a particular set of hardware/software set-up but only a  
> rough idea of how SCF parallel run scales with
> number of processors for a given system size
> Bhagawan Sahu
> Microelectronics Research Center
> University of Texas at Austin
> 10100 Burnet Road, Bldg. 160
> MER 1.606H, Mail code R9900, TX 78758-4445
> e-mail: brsahu at physics.utexas.edu
> Phone: (512)-471-1344 FAX: (512)-471-8575
> _______________________________________________
> Pw_forum mailing list
> Pw_forum at pwscf.org
> http://www.democritos.it/mailman/listinfo/pw_forum

Prof Nicola Marzari   Department of Materials Science and Engineering
13-5066   MIT   77 Massachusetts Avenue   Cambridge MA 02139-4307 USA
tel 617.4522758 fax 2586534 marzari at mit.edu http://quasiamore.mit.edu

More information about the users mailing list