LAMMPS PPPM relative accuracy and timing results

Hello, I have recently ran some timing results for a ~52 K atom system using LAMMPs on stampede. I have noticed that it is very sensitive to the PPPM accuracy specified for the tests I ran. The default value is 10e-4, however I have heard mentioned before on this list (believe it was axel who said this, sorry if I am misquoting or I am taking this out of context) that this is “sloppy”, and a 10e-6 value is recommended. However, after digging through the archives I seem to be getting conflicting reports. Can someone comment on the proper method to select this parameters, or if 10e-6 is overkill? Here are the timing results.

Ran on 8 cores with a 1 MIC for 2000 steps.

61 seconds (kspace_style pppm 1.0e-4)

73 seconds (kspace_style pppm 1.0e-5)

162 seconds (kspace_style pppm 1.0e-6)

The best way to determine the PPPM accuracy needed

for your system, would be to run the same simulation

at different accuracy values and see if something important

changes.

To speed-up PPPM, you can adjust the short-range cutoff,

e.g. make it longer if PPPM is too expensive. Various other

optimizations are also possible, e.g. verlet/split, staggering, etc,

as described in Section 5.2 of the manual.

Steve