[QE-users] SCF calculation with SOC fails to converge

[Paolo Giannozzi]

Multiple answers here:

> I could see etot_conv_thr = 1e-7 , this is too strict criteria for 
> total energy convergence

no it's not: 1.0e-7 is fine for most cases

 >  you can try [...] changing diagonalization to cg

this might work in case of errors in diagonalization, but does not solve 
non-convergence problems. My advice is not to use cg, except very 
special cases.

 > My original command was: mpirun -np 400 pw.x -nk 10 -nb 1 -nd 36.

400 MPI processes seems to me a large number for a 12-atom cell.

Also: with norm-conserving pseudopotentials, do not set "ecutrho", the 
default value (4*ecutwfc) is perfectly fine.

 > After removing the -nt flag

there is no -nt flag in the "original command"

 > While the issue is resolved, I am still curious about the root cause.

if the problem arises from the -nt flag, it might be a bug in the 
"task-group" parallelization, a technique that is still present in the 
code but that should be replaced by something better (it is useful only 
with more than hundreds MPI processes, it is hard to maintain and not 
well suited for GPU porting)

Paolo
-- 
Paolo Giannozzi, Dip. Scienze Matematiche Informatiche e Fisiche,
Univ. Udine, via delle Scienze 208, 33100 Udine Italy, +39-0432-558216