[CPMD-list] LOOP WAS VECTORIZED

[Axel Kohlmeyer]

On Thu, 3 Apr 2008, Nora Mohamed wrote:

dear nora,

please _NEVER_ send html formatted mails to mailing lists.

most people that recieve a lot of mail prefer using a text
mode mail program, and there those take significant effort
to make them readable. as i wrote many times before: the
easier you make it to answer the more likely you get help.

NM> Dear Dr. Kolmeyer,
NM>   I am sorry for the inconvenience, I didn't see the first email I 
NM> sent in my inbox and so thought that the group didn't receive it 
NM> as well and so sent it again. I actually didn't receive your 

mailing lists, and that includes the cpmd list, occasionally get
"stuck" or are taken down for maintenance or ...,
it is advisable to wait at least a day before resending.

NM> reply, I only came across it while I am searching in the mailing
NM> list for another problem I had. 

so you'll have to check your mail setup and or change your 
e-mail provider...

NM> 1.  Concerning my test run, I get very consistent results;
NM> my total energy is very consistent with the test case results. there
NM> are only differences in the forces. 

[...]

NM>   20  2.880E-08   1.243E-09      -9.141771   -3.997E-13     18.58
NM>   21  8.633E-09   7.382E-10      -9.141771   -6.750E-14     18.60 
NM>  
NM> In the last three lines, there are differences in the DETOT, but
NM> I don't think the differences are major or serious. 

if you read my previous comments about achievable limits of accuracy,
then you should be aware that relative differences of the order of 
1e-15 are the limit of double precision floating point accuracy.
those differences can be caused by different setting for rounding of
the floating point unit or different compilers, or optimization levels.
in fact, differences cause by all of those can be _much_ larger.

NM>    2  B  9.4486  9.4486 13.0108   1.234E-12  4.871E-15  9.479E-04
NM>    3  H 11.3223  9.4486 11.3384  -3.742E-03  6.548E-16 -7.247E-06
NM>    4  H  7.5749  9.4486 11.3384   3.742E-03  2.186E-16 -7.247E-06
NM>    5  H  9.4486 11.4168 14.1063   5.774E-13 -2.357E-03 -1.100E-03
NM>    6  H  9.4486  7.4804 14.1063   5.672E-13  2.357E-03 -1.100E-03
NM>    7  H  9.4486  7.4804  8.5704   5.783E-13  2.357E-03  1.102E-03
NM>    8  H  9.4486 11.4168  8.5704   5.669E-13 -2.357E-03  1.102E-03


NM> the gradients however, there are differences in the order of magnitude.
NM> but I think now that, as you said, if the optimum number is zero,then 
NM> these differences would be considered minor errors. 

yes. even with very tight convergence on the wavefunction, i'd
be surprised if your forces can be converged to more than 1e-7
absolute.

NM> performance from OpenMP with small number of cpu's, but in one of 
NM> the messages you mentioned that you could get 80% of the MPI 
NM> performance with OpenMP on a two-socket dual-core intel machine, 
NM> so should I use openMP or I am better off without it?  

MPI is the better solution. there is no other way to find the
best way of how to run jobs on your machine than running benchmarks
and trying out compiler and input flags that are affecting 
performance systematically.

NM> I need the best possible performance since I do constant pressure
NM> MD for 32 ch4 molecules and these two machines are all I can get.

did you choose this project or somebody else? the few percent
speedup won't help you much. to run this reasonably, you need
to run on a problem cluster or large multiprocessor machine or
else you'll be waiting for years until you have reasonable results.

also, constant pressure simulations with CPMD add a whole new 
dimension of difficulty (you need a converged stress tensor,
not only converged forces, you need to pick two fictitious masses
that must not couple to anything, you have to consider the
change of the effective plane wave cutoff with changing box size).
when modeling methane you have to consider a method to compensate 
for the incomplete description of dispersion interactions, or else
your results will be worse than those of classical MD.

NM> I checked Unicore to see the possibility of doing heavy MD, but 
NM> I am only allowed to testing and not to get final results that 
NM> I can publish later on. 

grid computing is pointless for these kinds of runs. you need
a proper cluster with a fast network (myrinet, infiniband or alike). 

cheers,
   axel.

NM> Thanks a lot Dr. Kohlmeyer.
NM> Best Regards,
NM> Nora M. Hassan

-- 
=======================================================================
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.