[Pw_forum] why my pw.x run with low efficiency?

vega vegalew at hotmail.com
Sat Sep 20 05:30:02 CEST 2008


Dear sir,

Thank you for your suggestion.

>>  Do you think 10G infiniband is good enough for 39 nodes?
>
> I'm not an hardware junkie, but infiniband has a fame to be quite
> performant, it should be ok.
thank god...


> on local scratch. If you are using the wf_collect option, you may have to
> use a network file system for the wavefunctions (I think there is an input
> variable to specify the wavefunctions directory).

to reduce the IO, I didn't use it.

> is little you can do. Do not set the outdir to a network filesystem will
> help, especially if the filesystem passes on the same network as mpi.
> Maybe openmpi is more efficient than the mpich, but I don't really know it
> (I'm just assuming than new is better than old).

OK, I'll try to set a outdir to a local location. I used to run QE with 
openmpi
but openmpi seems to be something wrong with invoking the virtual memory for 
my
cluster. I think so. Only mpich2 could run the calculation with large 
systems up
to 120  atoms or more.

Do you think I should copy the compilers, MKL, fftw, mpi, and QE files to 
local
disk for every node? Do you think this will reduce the IO greatly?

thank you for reading

vega

=================================================================================
Vega Lew (weijia liu)
PH.D Candidate in Chemical Engineering
State Key Laboratory of Materials-oriented Chemical Engineering
College of Chemistry and Chemical Engineering
Nanjing University of Technology, 210009, Nanjing, Jiangsu, China

--------------------------------------------------
From: "Lorenzo Paulatto" <paulatto at sissa.it>
Sent: Saturday, September 20, 2008 8:00 AM
To: "PWSCF Forum" <pw_forum at pwscf.org>
Subject: Re: [Pw_forum] why my pw.x run with low efficiency?

>
> On Ven, Settembre 19, 2008 18:54, vega wrote:
>>       PWSCF        :     0d   14h46m CPU time,        2d   18h 4m wall >
>
> This is indeed low!
>
>> Could you tell me what make my CPUs run in a such low efficiency style?
>
> I cannot be sure, but I will try to give a reasonable guess
>
>>  Do you think 10G infiniband is good enough for 39 nodes?
>
> I'm not an hardware junkie, but infiniband has a fame to be quite
> performant, it should be ok.
>
>> Could tell me which folders must be on a NFS location so that all the
>> nodes can load and write?
>
> Only the pseudopotentials and the input files, all the rest can be written
> on local scratch. If you are using the wf_collect option, you may have to
> use a network file system for the wavefunctions (I think there is an input
> variable to specify the wavefunctions directory). I'm not 100% sure about
> that, and it may be improved in future versions...
>
>>  Do you think the pw.x underestimate greatly for the memory?
>
> It is possible indeed, the estimate is hardcoded by hand, unluckily many
> of the contributors (me included) haven't updated it to reflect their
> changes to the code.
>
>>  any hints on my problem will be deeply appreciated.
>
> Youare already using the maximum number of k-points (one per pool), there

>
> cheers
>
> -- 
> Lorenzo Paulatto
> SISSA  &  DEMOCRITOS (Trieste)
> phone: +39 040 3787 511
> skype: paulatz
> www:   http://people.sissa.it/~paulatto/
>
>
>
> ----------------------------------------------------------------
>  SISSA Webmail https://webmail.sissa.it/
>  Powered by SquirrelMail http://www.squirrelmail.org/
>
> _______________________________________________
> Pw_forum mailing list
> Pw_forum at pwscf.org
> http://www.democritos.it/mailman/listinfo/pw_forum
> 


More information about the Pw_forum mailing list