[Pw_forum] Comparison of 3.1.1 and 3.2 (cvs)

degironc degironc at sissa.it
Fri Nov 24 09:39:13 CET 2006

Dear Huiqun Zhou,
    a couple of changes have been made in the mixing routine (mainly to 
reduce its
huge memory requirement in certain cases) from 3.1.1 to 3.2. These might 
be the 
reason for  the different behaviors. 
    Could you please post your input so that we can try to fgure out 
what happened ?
    Since every realization of an algorithm is the result of a 
compromise between speed,
memory use, work required to implement it, etc and since every 
compromise will likely
be more or less good depending on the particular application considered, 
a comparison for
different "typical" cases would also be useful in order to assess the 
"average" performance
(or loss of it) from 3.1.1 to 3.2
    Thank you,

Huiqun Zhou wrote:

> Kostya,
> The number of SCF iterations is indeed the same regardless of the
> number of CPU cores used for both versions. Furthermore, the process
> of SCF convergence is precisely the same whatever we use 1, 2 or 4
> CPU cores. This is true for both versions. And the final total energy 
> found
> by both version is the same. But, the behaviour of convergence of one
> version compared to another is different if you watch the changes of
> numbers after decimal point. It seems that the criteria of convergence of
> each version are different.
> I have created three graphs for the pattern of convergence on 1, 2, 4 
> cores
> (but they are tottally the same as mentioned above). Each graph contains
> a comparison of patterns of the two versions. Please let me know where
> I can post the graphs, each of which is about 55KB, as I don't think this
> forum permit posting a file with such large size.
> Here I first include the six data files extracted from the results. 
> you can
> plot the graph yourself, too.
> Please give me a pointer what kind of benchmarks you want me to run.
> Do you think it's worth tring the same system with different type of 
> run (relax, ...),
> different method of diagonalization? Currently, all results I reported 
> made use
> of 'david' method for diagonalization.
> Regards,
> Huiqun
> ----- Original Message ----- From: "Konstantin Kudin" 
> <konstantin_kudin at yahoo.com>
> To: <pw_forum at pwscf.org>
> Sent: Friday, November 24, 2006 1:42 AM
> Subject: Re: [Pw_forum] Comparison of 3.1.1 and 3.2 (cvs)
>> --- Huiqun Zhou <hqzhou at nju.edu.cn> wrote:
>>> (2) Version 3.1.1 is much faster than the cvs version as the former
>>>   took only 13 scf iterations to reach convergence while the latter
>>>   needed 16 iterations.
>> Dear Huiqun,
>> If the number of the SCF iterations is the same regardless of the
>> number of cpus for both versions, then please post some kind of
>> convergence pattern like for total energies. That would give developers
>> a chance to evaluate if any of the recent changes might have caused
>> this. Sometimes small numerical noise is a benign "feature", and
>> sometimes it is a bug, so figuring out what is going on for your test
>> would be useful.
>> Also, perhaps you could try a different benchmark to see how things
>> work there, if there are any differences found there as well.
>> Kostya
>> ____________________________________________________________________________________ 
>> Sponsored Link
>> Compare mortgage rates for today.
>> Get up to 5 free quotes.
>> Www2.nextag.com
>> _______________________________________________
>> Pw_forum mailing list
>> Pw_forum at pwscf.org
>> http://www.democritos.it/mailman/listinfo/pw_forum

More information about the users mailing list