It means that since very little communication is needed across different images this parallelization does not suffer much (at all) if the communication network is not very fast... however it does not distribute the memory and I/O ... which may be bad for the efficiency of the calculation as we are learning these days from the discussion going on on pool parallelization (which also is less-communication and more-memory&I/O intensive than R&G parallelization)... My today understanding of this is that one could increase the number of processors using R&G parallelization as far as the network at hand allows to get a good scaling... then one can increase further the number of processors by adding pool parallelization and, in case of NEB calculations, some image parallelization.
Hope this helps, Stefano de Gironcoli Mahmoud Payami wrote: > Dear ALL, > > In sec. of "parallelization issues" of UG, is mentioned that: "As a > general rule, image parallelization may give good scaling..." > Could anybody please explain what exactly it means in terms of say, > static linking, using more npools, or so. > > Bests, > Mahmoud Payami > Physics Group, AEOI. > > > > > > ------------------------------------------------------------------------ > > _______________________________________________ > Pw_forum mailing list > Pw_forum at pwscf.org > http://www.democritos.it/mailman/listinfo/pw_forum >
