On Tuesday, February 7, 2017 6:47:18 PM CET Louis Fry-Bouriaux wrote: > Interesting, I will try that, I will add that I tried the calculation > with lelfield=.true. and all efield_cart values to zero and it gives me the > same error. I also reduced the number of auto k points to speed up testing > (auto: 6 2 2 0 0 0/ nppstr=6, which takes ~160s). I may take a look at the > code maybe there is something that can be done if I identify what is going > on :/
Hello, I managed to find the original emails from 2012 about the issue; it used to be much worst but Paolo Giannozzi wrote a quick fix that does make the issue much lighter. Still. the case gdir=3 (or pdir=3, if you are doing polarization) is much faster than the other cases, and there is not easy solution. The reason is that the code needs to build planes of G-vectors that are orthogonal to the direction of the k-points string. The bottleneck is a single call: CALL mp_sum(aux_g(:), intra_bgrp_comm ) which collects a wavefunctions over all the CPUs. This call becomes slower and slower when more CPUs are involved. It used to be unbearably slow, now it is just slow. There is no obvious solution, as the parallelisation is hard-coded along the z direction. Try to stick to gdir=3 if you can. hth -- Dr. Lorenzo Paulatto IdR @ IMPMC -- CNRS & Université Paris 6 +33 (0)1 44 275 084 / skype: paulatz http://www.impmc.upmc.fr/~paulatto/ 23-24/4é16 Boîte courrier 115, 4 place Jussieu 75252 Paris Cédex 05 _______________________________________________ Pw_forum mailing list [email protected] http://pwscf.org/mailman/listinfo/pw_forum
