Hi folks

Is there a particular reason to prefer the multiple restart approach that seems to be built into the tutorial structure rather than setting itmax to a large number and simply letting it run?

Ciao
Terry

OK, thanks.

Hi Terry,
To add to Gregors answer: I think in the tutorials itmax was really small and convergence parameters were not so strict, just that things run fast.

Per default we use a itmax of 30 or 80. Depending on the system sizes and how much resources you would waste if the systems do not converge.
From our High-throughput projects for small unit cells we know that around 60% are converged within 30 iterations, 80% within 80 iterations. Everything that is still converging beyond this is converging slowly, stuck, or not at all. Of course this strongly depends on the system and which features of fleur you use. Magnetic calculations need more iterations and LDA+U calculations even more, simple because more quantities have to reach convergence. And of course there is a dependence on the mixing scheme deployed.

Deleting the mixing history from time to time (i.e after 30, or 80 iterations) may help to get a system which was 'stuck' before to converge nicely again.

Overall, one wants to minimize restarts of course, because they amount to more data and more data transfer, but at the same time you want to minimize the computational time in systems not converging, because else you end up spending 80% or more of your CPU budget on these. While as Gregor pointed out, systems which do not move at all are easy to detect, other convergence behavior may be more challenging.
The long term ideal solution to this is to improve our automatic workflows to detect systems not (slow) converging or stuck early on and implement solutions how to deal with them or stop them.