To whom may help to solve my problem.
Hello,
I have been run a series of test for the GPU package of LAMMPS.
Everything was fine a few days ago,
the calculation speed is amazing and the accuracy also is acceptable.
But now, I faced a problem.
In some case, the result of GPU run is dubious in case of MSD(mean square displacement) and a weird drift of atomic bulk at the same time.
Comparing with a CPU run result, it usually has a 10~100 times larger MSD, also, in case of the CPU run, there is no bulk drift motion.
I do not paste my script in here for now, but these job is related to annealing of liquid-Si.
And the most unaccountable thing is, a script, which do a melting of crystalline-Si, do not have such problem.
- MSD is fine and there is no bulk drift motion.
- An interatomic potential & system size are equal in the both cases.
- The script regarding melting is quite similar to the annealing one, I made the annealing one by modifying and adding a few sequence on the melting script.
It looks like there is some error or an important missing regarding GPU run for me.
Is there any related experience for GPU run ? (my current Lammps version is 31Mar17.)
Any comment will be helpful to figure out what is going on.
Thank you for your help.
Sincerely.
To whom may help to solve my problem.
Hello,
I have been run a series of test for the GPU package of LAMMPS.
Everything was fine a few days ago,
the calculation speed is amazing and the accuracy also is acceptable.
But now, I faced a problem.
In some case, the result of GPU run is dubious in case of MSD(mean square
displacement) and a weird drift of atomic bulk at the same time.
Comparing with a CPU run result, it usually has a 10~100 times larger MSD,
also, in case of the CPU run, there is no bulk drift motion.
I do not paste my script in here for now, but these job is related to
annealing of liquid-Si.
And the most unaccountable thing is, a script, which do a melting of
crystalline-Si, do not have such problem.
* MSD is fine and there is no bulk drift motion.
* An interatomic potential & system size are equal in the both cases.
* The script regarding melting is quite similar to the annealing one, I made
the annealing one by modifying and adding a few sequence on the melting
script.
It looks like there is some error or an important missing regarding GPU run
for me.
Is there any related experience for GPU run ? (my current Lammps version is
31Mar17.)
Any comment will be helpful to figure out what is going on.
this sounds like you are another victim of the "flying ice cube" syndrome.
how did you compile the GPU library, particularly which precision
settings did you use?
when using mixed and particularly single precision, the chances that
you accumulate errors in the total momentum is higher.
you are only comparing apples to apples, when you are using the GPU in
double precision mode (which is admittedly significantly slower).
one way to suppress an accumulating drift is to use fix momentum e.g.
like the following:
fix 4 all momentum 100 linear 1 1 1 rescale
axel.