[CPMD-list] CPMD parallel scalability

[Alessandro Curioni]

Yes - you are right  -
   evaluating  the basic  HDW  should not be a problem on that machine to 
get out more -
  but in many cases  the  performance  at application level   is totally 
different   of what you would predict  simply  looking at HDW performance 
...  it is enough a  not  optimal
  implementation of ALLTOALL to  get a very bad effective bandwidth ....

Maurice, 
    can you post the  header of the makefile you have used  and the 
performance ( last ) section of  one of your outputs ?


Alessandro








Juerg Hutter <hutter at pci.uzh.ch> 
Sent by: cpmd-list-bounces at cpmd.org
04/14/2008 04:31 PM

To
Maurice de Koning <dekoning at ifi.unicamp.br>
cc
cpmd-list at cpmd.org
Subject
Re: [CPMD-list] CPMD parallel scalability






Dear Maurice

your system (assuming there is no strange part in your
input that is not shown) should easily scale beyond 64
CPUs. However, this assumes a fast interconnect.
Now the Altix 4700 should have (according to SGI) an
industry leading memory bandwidth and latency.
My first guess is that your are not using (accessing)
this high speed interconnect, the second guess is
that the interconnect is not working correctly or
you are sharing nodes/bandwidth with other users.
Some information can be gathered from the parallel
performance and CPU time information at the end of
every CPMD run.

regards

Juerg Hutter

----------------------------------------------------------
Juerg Hutter                   Phone : ++41 44 635 4491
Physical Chemistry Institute   FAX   : ++41 44 635 6838
University of Zurich           E-mail: hutter at pci.uzh.ch
Winterthurerstrasse 190
CH-8057 Zurich, Switzerland
----------------------------------------------------------


On Mon, 14 Apr 2008, Maurice de Koning wrote:

> Hi all,
>
> I´m running CPMD on an Altix 4700 system with 44 CPU´s and 88 Gb of RAM
> memory.
> At the moment I´m running a CP MD run of a cell containing 96 water
> molecules using the
> BLYP functional at 300 K. I noticed that the scalability is not very
> good. If I run on more than
> about 16 CPU´s, the time per MD step starts even increases, such that
> the total time starts growing with the
> number of CPU´s. Is there anything I can do about this?
> Below is a part of the input script
>
> Cheers,
>
> Maurice
>
> &CPMD
>  MOLECULAR DYNAMICS CP
>  RESTART WAVEFUNCTION COORDINATES VELOCITIES
>  QUENCH BO ELECTRONS
>  TIMESTEP
>  4
>  TEMPCONTROL IONS
>  300.0 40
>  MAXSTEP
>  20000
>  TRAJECTORY XYZ
>  EMASS
>  600
>  STORE
>  100
> &END
>
> &SYSTEM
>  SYMMETRY
>  8
>  SCALE
>  ANGSTROM
>  CELL ABSOLUTE
>  13.1493536801862057 15.1976493759670106 14.3247149284289197 0 0 0
>  CUTOFF
>  70
> &END
>
> &DFT
> FUNCTIONAL BLYP
> &END
>
> _______________________________________________
> CPMD-list mailing list
> CPMD-list at cpmd.org
> http://cpmd.org/mailman/listinfo/cpmd-list
>_______________________________________________
CPMD-list mailing list
CPMD-list at cpmd.org
http://cpmd.org/mailman/listinfo/cpmd-list

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://cpmd.org/pipermail/cpmd-list/attachments/20080414/e750fa6a/attachment-0001.html