T=m.mm termination by timer

[vvchaban]

It seems that the "T=m.mm" keyword does not work (ignored) either in LInux or Windows. We need this or analogous functionality, i.e. maximum MOPAC runtime.

[godotalgorithm]

What type of calculation are you trying to set a wall time for with the T= keyword? While the manual does not state it clearly, the wall-time limit has only been implemented for loops involving geometry changes such as geometry relaxation and vibrational calculations. I can consider adding the time limit to the MOPAC and MOZYME SCF cycles, but I need to know if that will satisfy your intended use case.

[vvchaban]

I need it to work in 1SCF calculations.

[godotalgorithm]

Ok, this behavior is something that I should be able to accommodate in the next patch release. Keep in mind that the time per SCF iteration tends to be relatively uniform, so setting a maximum time limit is not too different from setting a maximum number of SCF iterations using the ITRY keyword.

[vvchaban]

MOPAC starts convergence boosters if the fast algorithm fails. Irrespective of ITRY, these convergers may run for many hours. I do not know how to forbid them to start. The best found solution so far is to add PULAY. Then, the problematic SCF terminates more quickly than without it.

What I practically need is an option to automatically kill the job if it does not converge during N seconds or does not converge using the default algorithm.

[vvchaban]

On a related note, it would be actually a great addition to allow a keyword like "NOCAMP" because there are advanced applications of MOPAC, in which it is more efficient to restart from an adjusted Z-matrix than to wait for convergence, which may never be achieved,

[godotalgorithm]

Yes, the behavior of the SCF cycle gets messy once it starts to struggle with convergence. I'm examining the behavior of ITRY right now, and it seems to force on the Camp-King converger before it is willing to terminate. This can get very slow because the Camp-King converger is still implemented with slow (non-BLAS) linear algebra. This is a good reason to finally port Camp-King to BLAS, and I can look into adding a few more options like a NOCAMP keyword, although that might not be as relevant if I can fix the performance issue.