Usually the best thing to do is to do a set of runs of ~10 ns or so and then check to see if it's converged by looking blocking the PMFs into ~5 different blocks or so; that is, looking at the PMF of 0-2 ns, 2-4 ns, etc etc and see whether they're the same.  If they are changing quite a bit, you may need to run for longer.  Block averaging the PMFs can also give you an error estimate, and you can use that to determine an error threshold, which will give you an idea of when to stop running.  You may also want to check that the histograms are correctly overlapping.

It actually depends on what sort of conformation you are probing. For processes like ion transport through membranes ~10ns per window with a spacing of 0.1nm between windows is the norm. But if you are probing a larger conformational change like loop changes ~100ns per window might not be enough. It really depends on how complex is the process. As suggested above after running all your windows for certain amount of time, you will have to check for convergence of your stats. Word of caution convergence does not imply that your PMF has converged, it might also be possible that path of your umbrella's is stuck along a local minima. So repeating the process with different starting structures for each umbrella might instill more confidence on your results.

You may find the following article helpful:

Maria Millsa and Ioan Andricioaeib, An experimentally guided umbrella sampling protocol for biomolecules, J Chem Phys. 129, 114101, 2008.

or website:

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2736582/

Hi

Thanks for all the comprehensive answers. 

Regards,

Farideh