Re: Is there solution to numerical inaccuracy

From: David Hardy (dhardy_at_ks.uiuc.edu)
Date: Wed Nov 21 2007 - 09:54:41 CST

Hi,

I don't know what the status is regarding the PDB class in the linked
discussion below
(http://www.ks.uiuc.edu/Research/namd/mailing_list/namd-l/1317.html).

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.

Alok, here are some tests that you can run to ascertain NAMD restart
file correctness:

1. Run more than one parallel "A" simulation
(http://www.ks.uiuc.edu/Research/namd/mailing_list/namd-l/1272.html)
and see if the energy difference between them is comparable to the
"B1/B2" restarted simulation.

2. Run the "A" test and the "B1/B2" test on a single processor. In
this case, NAMD should perform deterministically, and the energies
should be identical (up to almost machine precision).

-Dave

On Nov 20, 2007, at 11:48 AM, Peter Freddolino wrote:

> Hi Alok,
> could you clarify your particular testing procedure? Also, did you
> compare the error of two different split runs with that between two
> identical continuous runs?
>
> The reason I ask is that for a parallel setup that does floating point
> operations in arbitrary orders one can *never* expect absolutely
> identical results between different runs, even if those runs use
> identical input, due to various nondeterministic differences in the
> ordering of floating point operations (which could also be the reason
> for differences during a restart). The errors noted in the thread you
> mentioned are on the order of deviations I've observed previously
> between different, identical, runs due to such imprecision, and should
> not affect results significantly. I'd refer you to Jim's comments on
> non-repeatability in the same thread
> (http://www.ks.uiuc.edu/Research/namd/mailing_list/namd-l/
> 1317.html) for
> more information.
>
> Jim or Dave, do you have any comments on this?
>
> Peter
>
>>>

----------------------------------------------------------
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 - 05:20:46 CST