chetan,
you are still not getting it. most of what you did is a waste of time
and just shipping an excel file with some numbers is no proof at all.
for as long as nobody else can verify what you did _independently_,
you simply don't know and i cannot tell you as well.
fist of all, you didn't answer my question about the _exact_
LAMMPS version. you just wrote "the latest", but that is a
very inaccurate description since "the latest" can change
daily. instead you should give the date string that lammps
prints. i am asking this for a specific reason, since there
recently was a bug present for a few days that could have
affected calculations like yours.
second, you should start by testing for a property that is less
noisy than diffusivity. rather than using your own inputs, you
should try to run some of the inputs that ship with lammps.
and you should get (mostly) identical values for the entire
output, except for timing. the examples are run with one
and 4 processors, so you can compare and see how much
deviation between runs is possible. different inputs can
deviate differently.
there are basically two issue that can happen:
- your compiler miscompiles some part of the code,
which can lead to differences in the "thermo" properties.
those usually happen with high compiler optimization,
so in case of differences you should compile with optimization
turned off and/or using a different compiler (e.g. gcc).
if any or the combination of the two steps yields better
agreement with the reference outputs, then you may
have found a reason. whether these differences are
significant, is a different story. sometimes they are
sometimes not. this is where you should post the
results to the list for people look at.
- if you still have differences than it is likely that there
is a bug in the specific version of lammps that
you have compiled. this can happen. sometimes little
change can have unexpected side effects. in this case
you should first update to the very latest version and
recompile. then repeat the check from above and if
it persists contact the mailing list with the specific
input and output and deviations that you see.
Ok! Axel, i just tried to answer the information you wanted as it seemed
from your earlier email. Deviation is just calculated between two water
diffusivity values from old and new executable (% calculated with respect to
maximum value). I understand it may be ambiguous and so I have simplified
things with some new data as attached in excel file herewith.
Let me rephrase the question again: Question is if lammps is compiling
perfect on new Ranger with New Makefile to give executable lmp_tacc, should
we assume that it is as accurate as the lammps executable (lmp_ranger)
generated (WITH SAME LAMMPS VERSION) on old Ranger with old Makefile. TACC
consultants told me that i should check the accuracy of lmp_tacc.
lets forget about the TACC guys. they are nice folks but most
likely have no practical experience in MD.
Coming back to excel file: I have considered both systems 1) SPEEK-water 2)
Nafion-water. FOr each system, i have tabulated diffusivity of water (diffw)
and diffusivity of hydronium (diffh) for 3 types of runs 1) with lmp_tacc 2)
lmp_ranger 3)lmp_ranger-RERUN.
3) was done to see how much diffusivity varies with the same run carried
again.
Below these diffusivity values, i have calculated Deltalmp (Difference in
diffusivities when lmp_tacc is used instead of lmp_ranger and Deltanormal
(difference in difffusivities when same run carried out again). Aim is to
compare Deltalmp with Deltanormal and determine if lammps is giving more
than normal deviations in output due to using lmp_tacc instead of lmp_ranger
and therefore respond to TACC consultants on whether lmp_tacc is accurate
enough.
The ratios listed in blue color do say that lmp_tacc should be accurate
enough. Two of the cases they are less than 1. One case it is 1.65, slightly
more than 1, whereas in one case it is 5, but as i have commented there, the
R2-values of linear fit to MSD Vs time to determine diffusivity are terribly
low.
this is all very confusing and i fail to see how you deduce that
your results are "good". all you have is an incomplete "internal"
standard that is not accounting for a number of potential issues
and a "quality parameter" that is very noisy. if you don't believe me
on the latter, have a look at this:
http://klein-group.icms.temple.edu/akohlmey/files/talk-trieste2004-water.pdf
SO i guess all is fine! Comments, if any, are welcome!
you may have convinced yourself.
but to put it differently: if this was part
of a paper, i would have to reject it.
i am not saying that your executable
is broken, but there is nothing that you
have produced that is convincing.
cheers,
axel.