On Tue, Mar 3, 2009 at 9:35 PM, Paolo Giannozzi <giannozz at democritos.it>wrote:
> Gabriele Sclauzero wrote: > > > I think that CG style diagonalization has not been parallelized > > like the Davidson one, so these lines at the beginning > > of your output sound strange to me. > > > >> > a parallel distributed memory algorithm will be used, > >> > eigenstates matrixes will be distributed block like on > >> > ortho sub-group = 2* 2 procs > > they aren't strange. This is for "subspace diagonalization", > which is not directly used in CG diagonalization, but it is > used in i) calculating starting wavefunctions, and ii) rotating > wavefunctions before starting a new diagonalization (it is > equivalent to first-order perturbation theory and improves > quite a bit the subsequent diagonalization). > > Anyway I verified that the problem with charge is present only > in parallel execution, both for CG and Davidson, irrespective > whether subspace diagonalization is parallel or not > Is there any work-around now? Parallel execution is indispensable for large systems. > > Paolo > -- > Paolo Giannozzi, Democritos and University of Udine, Italy > _______________________________________________ > - Show quoted text - > Pw_forum mailing list > Pw_forum at pwscf.org > http://www.democritos.it/mailman/listinfo/pw_forum > -- ------------------------------------------------------------------ Research Fellow, Dept. of Physics, National Univ. of Singapore 10 Kent Ridge Crescent, Singapore, 119260 Tel: +65 6516 7844 ------------------------------------------------------------------ -------------- next part -------------- An HTML attachment was scrubbed... URL: http://www.democritos.it/pipermail/pw_forum/attachments/20090305/970d83ac/attachment-0001.htm
