The PPPM algorithm has had several changes since 2011. If your system can tolerate 3e-3 accuracy, then you can use that, otherwise I would suggest that you use 1e-4 accuracy and make your coul/long cutoff larger as you suggested. The total runtime is what counts, so you should experiment with the cutoff until you find the minimum runtime. If your kspace time is 95%, then using a larger cutoff will almost certainly be faster overall. A good starting point is to try to balance kspace time and pair time, i.e. so each are 50%, but you’ll have to experiment to see what is really the fastest.
Stan
If your input script really only has 24 atoms, then using the Ewald sum could be faster too. Not sure if you simplified your problem or not…
Stan
That is just a small test system.
I agree with your prior email also. My trouble is only that the way I read the updates, 1e-4 accuracy is supposed to have the same meaning (when you use lj units at least). But it does not seem to be the case. Therefore I’m not sure if it was standard/correct that I used 1e-4 accuracy before, or if that was more precise than people typically use, or if it is correct now if I use 1e-4 accuracy. I did not have trouble with my prior system and am not worried my results were too imprecise; I just want to report the error used in a standard way.
Lisa
That particular update shouldn’t have changed the mean of 1e-4 relative accuracy for LJ units. However, there have been other changes to the code that could have affected this behavior. For example, the way that the gewald parameter is estimated has changed.
The 1-e4 relative accuracy is by definition relative. LAMMPS uses a reference force of two unit point charges separated by 1 Angstrom. It gets tricky with LJ units though, because now it is 2 charges separated by 1 LJ dimensionless unit (instead of 1 Angstrom). So it is hard to say what is “good enough”. LAMMPS can’t guarantee that 1e-4 is the standard/correct accuracy. It could be overkill, or you may need 1e-5 accuracy. The safest thing to do would be to test energy conservation, or you can also look at the estimated absolute force error.
Stan
By testing I see the behavior changes between 12 Feb 2012 and 28 May 2012. That’s the only update that appears relevant during that time, but the description of the update does not note any changes to how the accuracy is reported/calculated. I could be wrong (I suppose this is the question at hand).
Lisa
There is a bug with the kspace error estimation when the dielectric constant is not equal to one. The mentioned change in LAMMPS corrected one issue in that the magnitude of charge for the error estimator was not being scaled by the 4*pi term in the Coulomb constant and now it is, but accidentally introduced another bug: the magnitude of charge squared is getting divided by the dielectric constant twice, which is not correct. This directly affects only the accuracy metric, in that for a dielectric constant less than one, you will get better than requested accuracy. If your dielectric constant was greater than one though, then you could be getting less accuracy than requested. This will be fixed in a patch soon.
Stan