[Pw_forum] references for philosophy of parallel execution

Stefano de Gironcoli degironc at sissa.it
Fri Sep 4 09:44:26 CEST 2009


Dear Sreekar Guddeti,
   in addition to reading the documentation you have been pointed to,  
please take into account the hardware (cpu and communication network) 
you have at hand.
  R and G parallelization is memory efficient but involves large and 
frequent communications hence it will not be efficient unless your 
communication network is very good.
  For massively parallel machines (such as blue gene etc) one needs to 
introduce further divide and conquer tricks (see the discussion in our 
recently appeared "reference paper")
  Other types of parallelization (k-point, neb-images) involve much less 
communication and can be of help even if the network is not very good.
  In any case at the end of PWscf output there is some timing report 
that should tell you how much are you spending in computing and how much 
in communicating ... this should give you an idea whether increasing the 
number of cpus can help or actually worsen the performance

  Hope this helps,
   stefano de Gironcoli

sreekar guddeti wrote:
> hello QE users,
> i would be grateful if u could provide me with any references for the 
> principles involved in parallel execution like how the grids are 
> divided among processors which, should supposedly enhance perfomance, 
> etc. because i performed a parallel execution and the CPU time taken 
> is more than that in the serial case
>
> thnks in advance
>
>
> -- 
> Sreekar Guddeti
> Physics Department
> IIT Bombay
> India
> ------------------------------------------------------------------------
>
> _______________________________________________
> 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