From: JC Gumbart (gumbart_at_ks.uiuc.edu)
Date: Fri Dec 07 2007 - 14:53:56 CST
Regardless of where or when you run, given the same input, you should
be sampling from the same ensemble. Therefore, while you won't get
the exact same trajectory in phase space, you will get something that
is equivalent for all meaningful purposes (various statistical
measures should be the same, within the variance, for example). If
the quantity you are interested in depends on a particular trajectory
anyway, then it is not reproducible and therefore unlikely to be of
any use.
You might also see Peter's posts from a week or two ago regarding non-
determinism when running in parallel or using langevin dynamics.
On Dec 7, 2007, at 4:04 AM, Monika Sharma wrote:
> Dear All,
> I just want to know the effect of using different cluster machines for
> same problem.
> Suppose we have different clusters of different machine, with one kind
> of machines in one cluster and other kind of machines in other cluster
> and likewise. For instance, one of IBM powerpc, one of AMD 64 bit, and
> other of SGI IA64 (PS: just for example sake). Now we want our protein
> system to evolve at different temperatures say 300K, 400K, 500K.
> And in
> order to reduce time, we give runs for each temperature at the three
> different clusters. And in the end, we want to compare the results
> altogether for that protein system.
> Now my question is that will it be fine if we compare the results
> at the
> end from the different clusters? Because i think that MD is quite
> chaotic, so different clusters should give different results, and
> it can
> be quite diverse. But then how can one reduce the time if one has to
> wait for completion of one run to give runs at the same cluster?
> And how
> much reliable will the study be if done in so manner??
>
> Thanks in advance,
> Regards,
> Monika
>
>
This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:45:40 CST