Dear support, I’m a new lammps user and I analyzed some examples; I found some difference between scalar and parallel version, why the output of simulations are different?

Regard.

Nicola Varini

Dear support, I’m a new lammps user and I analyzed some examples; I found some difference between scalar and parallel version, why the output of simulations are different?

Regard.

Nicola Varini

There could be many reasons - initialization of velocites, application

of a Langevin thermostat with different random numbers, round-off,

drift of 2 systems apart over long times. LAMMPS makes no guarantee

that 2 simulations on different numbers of procs will give identical

answers.

Steve

There could be many reasons - initialization of velocites, application

of a Langevin thermostat with different random numbers, round-off,

drift of 2 systems apart over long times. LAMMPS makes no guarantee

that 2 simulations on different numbers of procs will give identical

answers.

... not to mention the fact that floating point numbers are

_not_ associative. in general, (a+b)+c != a+(b+c)

thus whenever you have a different distribution of

atoms and therefore sum up contributions of forces and

energies in a different order, you will not get

the exact same numbers, which in turn will lead

to (exponentially) diverging trajectories.

cheers,

axel.

Hi.

Could you be more specific about what outputs are different? Are any

Physical quantities at equilibrium different from the expected

statistical variance?

Regards,

Manoj

The top of the doc/Section_errors.html page has a brief

discussion of this. Everything should be statistically the

same at equilibrium.

Steve