[Pw_forum] Distribution of the memory when use -npool option

[sbraccia carlo]

Dear Sergey,

the compromise is using the two schemes together:

mpirun -np 64 ./pw.x -npool 32 -in my.inp

(to have 32 pools each one of 2 cpus, the memory is ~1/2 the memory for
a serial run)

or

mpirun -np 64 ./pw.x -npool 16 -in my.inp

(to have 16 pools each one of 4 cpus, the memory is ~1/4 the memory for
a serial run)

and so on.

carlo

Sergey Lisenkov wrote:
> Dear PWscf users and authors,
> 
> 
>   I am interested how are memory distributed when use -npool option. In my case, I perform calculations with 64 k-points. When I use :
>   mpirun -np 64 ./pw.x -npool 64 -in my.inp (with k-point parallelization) 
> 
> I see that a lot of memory (~30%) are used on each node (measured by 'top' command). But when I use:
> 
> mpirun -np 64 ./pw.x  -in my.inp (without k-point parallelization)
> 
> only ~5% of memory of each node is used. I believe, that using k-point parallelization is highly efficient, but it follows to out of memory, killing network mappers and etc. 
> 
> So, is there a compromise?
> 
>   Best wishes,
>     Sergey
> _______________________________________________
> Pw_forum mailing list
> Pw_forum at pwscf.org
> http://www.democritos.it/mailman/listinfo/pw_forum
>