[Pw_forum] How to tune cputimes using mpich2-1.0.8

Axel Kohlmeyer akohlmey at cmm.chem.upenn.edu
Sat Feb 21 20:08:16 CET 2009


On Sat, 21 Feb 2009, Mahmoud Payami Shabestari wrote:

MP> 
MP> Hi Axel,
MP> 
MP> > you should ask that the people developing MPICH2.
MP> > this is not the right forum for this kind of question.
MP> >
MP> > one more remark. that might also help people here not
MP> > to waste too much time on something pointless [...].
MP> 
MP> It is the converse! I am trying to save the community's time. Here you are:
MP> -------------------------------------------------------------
MP> mpich2_with_those_switches_that_I_mentioned:
MP> 
MP>      PWSCF        :  9m46.25s CPU time,    12m 9.57s wall time
MP> ---------------------------------------------------------------
MP> openmpi-1.2.9 (default switches):
MP> 
MP>      PWSCF        :  32m8.27s CPU time,    51m 23.38s wall time
MP> ----------------------------------------------------------------

mahmoud,

as i already wrote to you in private e-mail. the huge discrepancy
between wall and cpu time casts a severe doubt on the usefulness
of these numbers. 

as i was stating in the previous mail. if the MPI library takes
a significant time of the total time, there is a problem elsewhere
that needs to be investigated first. you are obviously severely
overloading your machine. 

MP> These are the results of an scf calculation for 21-layered Al(110) slab
MP> with Ecut=22Ry, k_mesh=45x45x1, degauss=0.05 on a box with 2 x amd64 quad
MP> (8 cores).
MP> The memory used in mpich2 is less than half of that in (default
MP> switches)openmpi.

how did you determine that?

you have to be careful to differentiate between memory
being mapped into address space, physical memory used 
and virtual memory being in use. the respective columns
in displayed in, e.g. top are: VIRT  RES and CODE+DATA)

cheers,
   axel.


MP> 
MP> Cheers,
MP>        mahmoud
MP> 
MP> 
MP> >
MP> > in general, it is not a good idea to try an play with
MP> > compiler optimization in the hope to make an MPI implementation
MP> > (and thus communication) faster.
MP> >
MP> > if your application performance would depend so much on the
MP> > optimization level use to compile your MPI library, then either
MP> > you have a very crappy MPI implementation, or your application
MP> > spends too much time in MPI calls. in the latter case, that
MP> > usually corresponds to severe overload of your communication
MP> > hardware and taking care of that would give you a much better
MP> > performance increase than any compiler switches will.
MP> >
MP> > basically for almost _any_ system library (and that includes
MP> > ATLAS and FFTW, btw) it is best to stick to a moderate optimization
MP> > (-O) as aggressive optimization may interfere with the implemented
MP> > algorithms and existing optimizations (e.g. both ATLAS and FFTW
MP> > include optimizations on the C-language and algorithm level, higher
MP> > compiler optimization can change the semantics of your code and
MP> > thus negate the optimizations performed in the code). even more
MP> > so, with many current compilers, aggressive optimization (-O3 or
MP> > higher) incures a very high risk of the compiler miscompiling
MP> > your library and thus leading to uncontrollable crashes or wrong
MP> > results. that doesn't mean, that there may be a measurable benefit
MP> > for singular cases, but from over 10 years of experience in
MP> > management of HPC resources the overall effect is problematic.
MP> > most of the time it is hard enough to chase down application bugs
MP> > that are real or cause by compilers, you don't want to add having
MP> > to track down problems in your libraries to that.
MP> >
MP> > cheers,
MP> >    axel.
MP> >
MP> >
MP> > MP> Any comment is highly appreciated.
MP> > MP>
MP> > MP> Best regards,
MP> > MP> Mahmoud Payami
MP> > MP> Phys. Group,
MP> > MP> Atomic Energy Org. of Iran  
MP> > MP>
MP> > MP> _______________________________________________
MP> > MP> Pw_forum mailing list
MP> > MP> Pw_forum at pwscf.org
MP> > MP> http://www.democritos.it/mailman/listinfo/pw_forum
MP> > MP>
MP> >
MP> > --
MP> > =======================================================================
MP> > Axel Kohlmeyer   akohlmey at cmm.chem.upenn.edu   http://www.cmm.upenn.edu
MP> >    Center for Molecular Modeling   --   University of Pennsylvania
MP> > Department of Chemistry, 231 S.34th Street, Philadelphia, PA 19104-6323
MP> > tel: 1-215-898-1582,  fax: 1-215-573-6233,  office-tel: 1-215-898-5425
MP> > =======================================================================
MP> > If you make something idiot-proof, the universe creates a better idiot.
MP> 
MP> _______________________________________________
MP> Pw_forum mailing list
MP> Pw_forum at pwscf.org
MP> http://www.democritos.it/mailman/listinfo/pw_forum
MP> 

-- 
=======================================================================
Axel Kohlmeyer   akohlmey at cmm.chem.upenn.edu   http://www.cmm.upenn.edu
   Center for Molecular Modeling   --   University of Pennsylvania
Department of Chemistry, 231 S.34th Street, Philadelphia, PA 19104-6323
tel: 1-215-898-1582,  fax: 1-215-573-6233,  office-tel: 1-215-898-5425
=======================================================================
If you make something idiot-proof, the universe creates a better idiot.



More information about the users mailing list