Re: [AMBER-Developers] About the reproducibility of MD simulations

From: Pratul K. Agarwal <agarwalpk.ornl.gov>
Date: Mon, 5 Jun 2017 10:39:48 -0400

I agree with Adrian. For CPU runs, due to the nature of computer algebra
small numerical differences come in after a few thousands steps (may be
even earlier). If you run the trajectories long enough then they start
to differ.

Pengfei,

Small differences after 500 ps is not unexpected behavior.


On 6/5/2017 10:33 AM, Adrian Roitberg wrote:
> ok
>
> Maybe someone can comment, but the amber cpu code is NOT bit-wise
> reversible, so different runs will give you different answers if you run
> then long enough.
>
> The cuda code MUST give you the EXACT same answers. If not, then your
> card is broken.
>
> For CPUS, you must look at simple averages to see if your simulations
> are different or not, not at the individual steps. Plot for instance a
> distribution of a variable, such as radius of gyration. If the two
> simulations give you statistically different distribution, then neither
> of them is 'converged'.
>
> Adrian
>
>
> On 6/5/17 11:28 PM, pengfei li wrote:
>> Hi Adrian,
>>
>> Thanks for your quick reply.
>>
>>> On Jun 5, 2017, at 9:24 AM, Adrian Roitberg <roitberg.ufl.edu> wrote:
>>>
>>> Hi
>>>
>>> Not sure what you mean.
>>>
>>> Same random seed and same coordinates/prmtop, etc ?
>>>
>> Yes, both the random seed and coordinates/prmtop are the same. The simulations are 500 ps long each.
>>
>>> What do you mean by different results ?
>>>
>> The output files are different (numbers could not match), and final restart files are different.
>>
>>> Was this amber gpu or what ?
>> It was pemed.MPI. Both used 16 cpus.
>>
>> Best regards,
>> Pengfei
>>
>>> Tell us the details please.
>>>
>>> adrian
>>>
>>>
>>> On 6/5/17 11:22 PM, pengfei li wrote:
>>>> Hi everyone,
>>>>
>>>> I have performed two simulations with the same random seed but got different results. Except the random seed, what are the other factors that influence the reproducibility of the MD simulation? Thanks!
>>>>
>>>> Best regards,
>>>> Pengfei
>>>> _______________________________________________
>>>> AMBER-Developers mailing list
>>>> AMBER-Developers.ambermd.org
>>>> http://lists.ambermd.org/mailman/listinfo/amber-developers
>>> --
>>> Dr. Adrian E. Roitberg
>>> University of Florida Research Foundation Professor
>>> Department of Chemistry
>>> University of Florida
>>> roitberg.ufl.edu
>>> 352-392-6972
>>>
>>>
>>> _______________________________________________
>>> AMBER-Developers mailing list
>>> AMBER-Developers.ambermd.org
>>> http://lists.ambermd.org/mailman/listinfo/amber-developers
>> _______________________________________________
>> AMBER-Developers mailing list
>> AMBER-Developers.ambermd.org
>> http://lists.ambermd.org/mailman/listinfo/amber-developers

_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Mon Jun 05 2017 - 08:00:04 PDT
Custom Search