[Pw_forum] band parallelization

Tiana Davide davide.tiana at epfl.ch
Tue Sep 22 12:42:13 CEST 2015


Dear Lorenzo and Paolo

Sorry if I was not clear. 
I didn't know about the only hybrid calculation. This means my test is useless. 
Since I will use on hybrid I would like to check if my approach is correct of it I am doing something wrong.

What I was trying to say is that I run a gamma point calculation with 272 orbitals. 
I run 3 tests using 

272 cores no band parallelisation
     Parallel version (MPI & OpenMP), running on     272 processor cores
     Number of MPI processes:               272
     Threads/MPI process:                     1
     R & G space division:  proc/nbgrp/npool/nimage =     272

time for scf 2 cycles
     total cpu time spent up to now is       11.8 secs
     total cpu time spent up to now is       22.0 secs
     total cpu time spent up to now is       28.0 secs


544 cores using band parallelisation and 
     Parallel version (MPI & OpenMP), running on     544 processor cores
     Number of MPI processes:               544
     Threads/MPI process:                     1
     band groups division:  nbgrp     =       2
     R & G space division:  proc/nbgrp/npool/nimage =     272

time for scf 2 cycles
     total cpu time spent up to now is       15.1 secs
     total cpu time spent up to now is       27.3 secs
     total cpu time spent up to now is       33.9 secs

816 cores with 
     Parallel version (MPI & OpenMP), running on     816 processor cores
     Number of MPI processes:               816
     Threads/MPI process:                     1
     band groups division:  nbgrp     =       3
     R & G space division:  proc/nbgrp/npool/nimage =     272

time for 2 scf cycles
     total cpu time spent up to now is       20.5 secs
     total cpu time spent up to now is       30.4 secs
     total cpu time spent up to now is       36.5 secs

is the principle correct or it should be done in another way?

Many thanks
Davide






More information about the users mailing list