Question about pppm relative error

Hello everyone,
    I have another simple question to ask you if you do not mind.
    The long range force is calculated by the pppm/tip4p kspace style in LAMMPS. It need a relative error. Someone uses 1.0e-5 while others just say 1.0e-4 is ok. I did a test and found that the speed of 1.0e-5 is 8 times slower than that of 1.0e-4.

if you have *that* much difference in speed, there is something else
wrong with your settings, e.g. a completely unbalanced real-space vs.
kspace computation.

    So my question is that how do I judge that how small the relative error setting in kspace_style is sufficient?

as so often in science, the answer is: it depends.
stress converges much more slowly than forces, so if you need the
stress (or pressure) as input (e.g. for fix npt or fix
press/berendsen), you need a higher convergence, say around 1.0-6 or
tighter. it doesn't make much sense to use smaller than 1.0e-8, since
the real space calculation uses approximations, that are only precise
to about 1.0e-7.

1.0e-4 is a very aggressive choice and acceptable for constant volume
calculations of homogeneous systems. 1.0e-6 is a safe choice.