[Pw_forum] Nscf and scf

[Paolo Giannozzi]

Vo, Trinh (388C) wrote:

> I am wondering how can it take more time for nscf calculation 
> than for scf calculation.

in the scf calculation, at each scf step, you perform iterative
diagonalizations starting from the wavefunctions of the previous
scf step, using a convergence criterion for diagonalization that
is very loose at the beginning, gets tighter as you approach
self-consistency. In this way, you perform several scf steps,
each one requiring diagonalizations that converge quickly.

In the non-scf calculation, you make a single step but you have
to start from superposition of atomic orbitals and use a tight
convergence criterion for diagonalization. Another factor is the
number of bands used: in scf calculations one typically uses only
occupied bands; in non-scf calculation, one typically is interested
in empty bands as well. This makes each diagonalization much costlier
in the non-scf case than in the scf case.

So typically a non-scf calculation may be faster (per k-point) than
a scf one, but not by much; sometimes not at all, or even slower.
Have a look at the number of H*psi made in the two calculations and
at the time spent there. H+psi is a basic ingredient of iterative
diagonalization, and typically takes the bigger slice of computer time.

The implementation of non-scf calculations is currently non-optimal,
because we throw away sll the information from scf except the potential,
but since non-scf is seldom the dominant part of a realistic
calculation, little effort has been devoted to its further optimization.

P.
-- 
Paolo Giannozzi, Democritos and University of Udine, Italy