On Sunday 07 January 2007 03:03, Axel Kohlmeyer wrote: > hmmmm.... what version of pw.x are you referring to? > at least in version 3.2 there should be a parallel diagonalization > that is used in pw.x (subroutine cdiagonalize in Modules/ptoolkit.f90). > as far as i understand, this is not easy to parallelize, so there > is an estimator that checks, whether the serial or the parallel algorithm > would be more efficient.
recent cvs. with diagonalization='david +para', the code almost always chooses parallel algorithm but this always increases total computation time, both on Intel and AMD SMP machines. So i guess the default 'david ' with serial diaghg is the best choice currently. By the way, i have found that diago_david_ndim=2 gives better times and also saves alot of memory comparing with default =4. Best Regards, Alexander Shaposhnikov > cheers, > axel. >
