Hi JT,
For dynamics it is perfectly reasonable to get different results when
running in serial or in parallel or indeed across different numbers of
nodes. This is because of round off difference in the computation and
summation of the forces followed by integration. Even two runs in parallel
on the same hardware can produce different results due to load balancing
leading to different orders of operation. Essentially floating point is
not exact and the only solution is to go to fixed point as we do with the
GPU code. Now why does one get different results? Well, Newton's equations
of motion are essentially chaotic - think double pendulum on steroids. A
tiny (last decimal place) difference on step zero in the coordinates or
velocities will ultimately lead to a complete decorrelation of the
trajectories. Is either incorrect? No, they are both strictly correct,
they are just exploring different regions of phase space. Ultimately with
enough sampling any statistical properties from the two trajectories
should converge.
Now, minimization is a different case. In principal it should always
converge to the same minima. This isn't strictly true if you are sitting
on the very top of a transition state and round off differences can make
you fall one way or the other towards different local minima but normally
one should end up converging to the same minima. Just the path you take to
get there can be different. Thus the output files themselves might look
very different but if you run enough steps to actually converge the
minimization then the end result whether you ran in serial or parallel
should be the same to within the convergence criteria of the minimization.
If you are curious I would suggest running 50,000 steps of minimization or
so and see how different the results look (both the energy and the actual
structure - RMS fitted in case the origin moved).
If you see radical differences please send your input and output so we can
take a look.
I hope that helps.
All the best
Ross
On 7/8/13 9:06 PM, "John Travers" <jtravers70.yahoo.com> wrote:
>Dear Amber experts,
>
>I am a newbie to Amber. I am learning it by following the tutorial listed
>at amber's homepage. I had a strange observation when I tried to repeat
>the calculations given in the tutorial B1-Section 5
>(http://ambermd.org/tutorials/basic/tutorial1/section5.htm). It is a
>minimization with the solute fixed. The input, parameter, and coordinate
>files are given in the webpage (polyAT_wat_min1.in, polyAT_wat.prmtop,
>and polyAT_wat.inpcrd). When I carried out this calculation with serial
>and parallel version of sander I got totally different results. The
>compilation of Amber 12 was successful and passed all the tests. Also, I
>complied Amber 12 on different machines with different compilers (intel
>and gnu). For the same calculation, different machines produced different
>results.
>
>Could you please tell me what's going on here? Did I miss some simple
>points? Thanks!
>
>Best wishes
>
>JT
>_______________________________________________
>AMBER mailing list
>AMBER.ambermd.org
>http://lists.ambermd.org/mailman/listinfo/amber
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Mon Jul 08 2013 - 23:00:02 PDT
