Not a dumb question but maybe there is no smart answer. I have never seen direct solver packages speak in these terms. IMHO for direct solvers it is more a question of wanting to make sure the part of the factored matrix on each process fits in the memory of that process. That is with direct solvers you generally use more processes (and hence more memory) to do larger problems, not to do smaller problems faster.
Since the amount of fill in a direct solver depends very strongly on the specific nonzero structure of the matrix there is no formula that goes from "number of unknowns per process" to "how much memory is needed to store the local factor". Barry > On Dec 2, 2015, at 12:45 PM, Alex Lindsay <adlin...@ncsu.edu> wrote: > > Hi, > > I've read on your FAQ page that a good rule of thumb for parallelization is > 20k DOF/processor. Does this apply equally to direct solves as well as > iterative? May be a dumb question... > > Alex
