[CPMD-list] how to speed up on amd64 machines

[Bernd Kallies]

Hallo,

don't forget that machines using dual core amd opterons are NUMA
architectures, and that they suffer from memory bandwidth when using
both cores of a cpu. You cannot expect good scaling even within one
machine, except you ensure to use only one core per cpu, and let each
task allocate its memory from the appropriate memory banks (work with
memory affinity). Thus, before discussing CPMD performance issues, try
to do some low-level benchmarks, e.g. STREAMS, to become familiar with
your system.

Sincerely, BK

On Mon, 2007-04-16 at 12:04 -0400, Bin Pan wrote:
> Hi Prof. Hutter,
> 
> I said that it is very slow because I used 6 CPU's on a Dual Core AMD 
> Opteron(tm) Processor 870 connected by myrinet and compared the
> time required to finish one MD cycle with that done by 4 CPU's on a 
> Intel(R) Xeon(TM) CPU 3.00GHz connected by ethernet for the same system.
> The first needs ~90 seconds, however the second only needs ~60.
> I expected the first one has faster CPU and better network connection.
> 
> Are these enough info for you to give me some advice in how to 
> improve the speed?
> Thanks a lot!
> 
> Best regards,
> Bin.
> 
> At 02:57 4/16/2007, Juerg Hutter wrote:
> >Hi
> >
> >you have to give us more precise information.
> >What exactly do you mean by 'it is very slow'.
> >Do you mean slow in general, e.g. compared to
> >another plane wave code or slow because
> >of bad speedups in parallel compared to the
> >serial runs?
> >You should also give us some information on your
> >computer. What is the exact version of your CPUs
> >and what is the type of your network?
> >At the end of each run CPMD prints timing information.
> >Please attach at least one such timing information
> >block for a run that you consider too slow.
> >
> >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.unizh.ch
> >Winterthurerstrasse 190
> >CH-8057 Zurich, Switzerland
> >----------------------------------------------------------
> >
> >
> >On Sun, 15 Apr 2007, Bin Pan wrote:
> >
> >>Hi CPMD users,
> >>
> >>I am wondering how to speed up CPMD runs on amd64 cluster using MPI
> >>parallelization.
> >>Previously I compiled CPMD using pgi compiler with acml. However, I
> >>found it is very slow.
> >>The flags I used are:
> >>
> >>CPPFLAGS = -P -C -traditional -D__Linux -D__PGI -DFFT_DEFAULT
> >>-DPOINTER8 -D__pgf90 -DPARALLEL -DMYRINET
> >>CC = gcc -O2 -Wall -m64
> >>FC = pgf90 -c -fastsse -tp k8-64
> >>LD = pgf90 -fastsse -tp k8-64
> >>
> >>Can you please let me know how to improve the performance?
> >>Thanks a lot!
> >>
> >>Best regards,
> >>Bin.
> >>
> >>_______________________________________________
> >>CPMD-list mailing list
> >>CPMD-list at cpmd.org
> >>http://cpmd.org/mailman/listinfo/cpmd-list
> 
> 
> 
> ---------------------------------------------------------------------
> Bin Pan
> Ph.D. Candidate
> Department of Chemical Engineering
> Massachusetts Institute of Technology
> Tel: 617-253-6675
> E-mail: binpan at mit.edu
> Room E19-528, 77 Mass. Ave. Cambridge, MA
> ---------------------------------------------------------------------- 
> 
> _______________________________________________
> CPMD-list mailing list
> CPMD-list at cpmd.org
> http://cpmd.org/mailman/listinfo/cpmd-list
-- 
Dr. Bernd Kallies
Konrad-Zuse-Zentrum für Informationstechnik Berlin
Takustr. 7
14195 Berlin
Tel: +49-30-84185-270
Fax: +49-30-84185-311
E-Mail: kallies at zib.de