From: Peter Freddolino (petefred_at_ks.uiuc.edu)
Date: Thu Nov 22 2007 - 09:15:29 CST
Just as an addendum, namd's random number generation is no longer
deterministic in parallel; this means that Langevin dynamics will be
deterministic and as repeatable as an NVE simulation *if* you use the
same seed an run in serial, but not otherwise (see the comment in the
namd manual on the seed parameter). In either case you're limited by
floating point addition, as Dave described.
Peter
David Hardy wrote:
> You can control the sequence of random numbers generated by choosing
> the seed (although I don't know the method that NAMD employs to
> generate random numbers in parallel). This part of the algorithm is
> deterministic.
>
> My earlier comment actually applies to NVE simulation with no "random"
> component. Since Hamiltonian systems are chaotic, any small
> perturbation (from the non-associativity of addition) can cause
> deviation from a previous trajectory.
>
> -Dave
>
>
> On Nov 22, 2007, at 2:19 AM, Himanshu Khandelia wrote:
>
>>> Hi,
>>
>>> For different parallel runs, NAMD will perform
>>> non-deterministically, in the
>>> sense that the order of completion of different processors (and even
>>> the work
>>> that is assigned to them) will likely change, altering the order of
>>> summed
>>> energies. Due to the non-associativity of floating point addition,
>>> you should
>>> eventually see slightly different energy values from two parallel
>>> runs of an
>>> identical trajectory.
>>
>> I thought the non-reproducibility of simulations in NAMD was mainly
>> because of the Berendsen pressure controller employed, which gives
>> random
>> kicks to particles every few steps. Whether or not the run is
>> parallel, I
>> do not think the trajectories will be reproducible, although I have not
>> tested this myself.
>>
> ----------------------------------------------------------
> David J. Hardy, PhD
> Theoretical and Computational Biophysics
> Beckman Institute, University of Illinois
> dhardy_at_ks.uiuc.edu
> http://www.ks.uiuc.edu/~dhardy/
This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:45:35 CST