Re: "parallel" MD

[Steve Stuart <steve : at : chem.columbia.edu>]

 Klaus.Liedl : at : uibk.ac.at (Klaus R. Liedl) wrote:
 >4. You do not need ONE trajectory for your sampling. You can sample
 >   over as many trajectories as you like. The only problem is to find
 >   starting points for your trajectories (cf. 3.).
 >   Even your methode in principle gives you ONE "MD-trajectory"
 (with the
 >   limitations of 2.) this is not the real point.
 >   The time you save is not the time for the actual simulation, but the
 >   time for searching a second starting point in phase space.
 and a few other people have made similar comments, stating
 that you could parallelize over N processors, if you could just find N
 different starting points.
 I just thought I'd point out that for many interesting properties
 (diffusion constant, residence times, etc) the the total
 length of the run can be more important than the number of runs
 averaged over.  So it is important that the forward and backward
 "trajectories" have the same starting point.
 One other comment that interested me is K Bryson's note that the
 reverse run would begin to unequilibrate the system.  (Now why didn't
 that occur to me...)  Certainly true if you started from some ordered
 state and never adjusted the temperature, but any T-scaling would
 prevent this.  Out of curiosity - does anyone know how reversible a
 non-T-scaling MD run actually is?  (with/without cutoffs, SHAKE,
 whatever)
 -Steve Stuart
 steve : at : chem.columbia.edu
 echo "a'rfg cnf Crpv | har" | tr '[a-m][n-z] ' '[n-z][a-m]\012' | sort
 | tr "\012" " " ; echo " "