[Pw_forum] QE_Parallel

[MKondrin]

Gul Rahman wrote:
> Dear All,
> As I mentioned in my previous e-mail that my QE is very slow when I 
> run it in parallel.
> I have no problem on the same machine if I use some other code, e.g., 
> SIESTA.
> I probably missing something in my compilation of QE. Please, have a 
> look into the attached sys.make file that I used to compile QE.
> Also, if some one can share the parallel sys.make file.
> After the compilation, I just use
> mpirun -np 32 --hostfile machines ./mpi_script.sh
>  where mpi_script.sh file just contain the command for pw.x (./pw.x <
> myfile.scf.in <http://myfile.scf.in> >OUT)
> Thanks,
> Gul
>
> -- 
Dear Prof. Gul,

Usually good parallelization requires interaction between mpirun and 
quantum espresso programs (pw.x for example). pw.x mostly benefits from 
parallelization of fftw algorithm but there is one gotcha (because fftw 
is not massively parallel task). Too fine granularity  of this  
algorithm requires too much of data transfer between nodes which in case 
of the low speed network (like gigabit ethernet) produces too much 
overhead. As a rule 4-8 parallel process in this circumstances is a 
ceiling. However if you job requires summation over different k-points 
you may introduce additional parallelization over k-points (which does 
not require too much of data transfer and so network bottleneck is 
avoided) which is done by adding command line option like pw.x -npool k 
... (where k is the number of slices in k-grid). As a result the whole 
task consist of only 32/4=8 "fftw-parallel" processes (which is what you 
usually want). This information is written in the beginning of the 
output file. 

There are some other tricks to obtain further parallelization which is 
discussed in PW help files. But this is the most general.

Best wishes,

M.V.Kondrin (High Pressure Physics Institute RAS)
>  
>
> Dr. Gul Rahman
> Assistant Professor,
> Department of Physics,
> Quaid-i-Azam University,
> Islamabad, Pakistan
>
> http://www.qau.edu.pk/profile.php?id=818020
> ------------------------------------------------------------------------
>
> _______________________________________________
> Pw_forum mailing list
> Pw_forum at pwscf.org
> http://pwscf.org/mailman/listinfo/pw_forum