[gmx-users] same initial velocities vs. -reprod

[Benson Muite]

Hi,

Yes, finite precision and order of computations may not be respected by 
the compiler. In most cases, they should be close, however much 
analytical work remains in this area. A  relevant paper (though there 
are many others) is:

Collange, Defour, Grailaat and Iakymchuk

Numerical Reproducibility for the Parallel Reduction on Multi- and 
Many-Core Architectures

https://hal.archives-ouvertes.fr/hal-00949355v4/document

It is unclear if it would be too expensive to implement such methods in 
GROMACs though.

Benson

On 3/22/19 10:11 PM, Mala L Radhakrishnan wrote:
> Hi Mark,
>
> Thanks so much -- good to know that it's basically equivalent to different
> starting velocities and I should expect them to be different.
>
> I found this page that sort of explains it:
> http://www.gromacs.org/Documentation/Terminology/Reproducibility
>
> Out of curiosity, I was wondering if someone can point me to something that
> explains why (a+b)+c != a+(b+c) sometimes for computations across multiple
> processors.  Is it a finite precision issue?
>
> thanks again,
> M
>
> On Fri, Mar 22, 2019 at 1:21 PM Mark Abraham <mark.j.abraham at gmail.com>
> wrote:
>
>> Hi,
>>
>> The dynamic load balancing on by default for domain decomposition means
>> divergence happens by default. After a few ps it's logically equivalent to
>> starting from different velocities. See the GROMACS user guide for more
>> details on reproducibility questions!
>>
>> Mark
>>
>> On Fri., 22 Mar. 2019, 18:15 Mala L Radhakrishnan, <mradhakr at wellesley.edu
>> wrote:
>>
>>> Hi all,
>>>
>>> We set up replicate simulations (same starting mdp files and structures)
>>> that ran across GPUs WITHOUT using the -reprod flag, but we set gen-vel
>> to
>>> no and used the same ld-seed value for both with the v-rescale
>> thermostat.
>>> They also ran on the same machine -- so from a deterministic point of
>> view,
>>> I would expect them to be "exactly" the same.
>>>
>>> The simulations, while having similar average energetics throughout,
>> sample
>>> different conformations and they start to differ pretty much right after
>>> the simulation starts.
>>>
>>> I understand that I could have gotten results to be more reproducible by
>>> using the -reprod flag, but in the case I describe (and I don't think I
>>> have any other stochastic things going on unless I'm not understanding
>>> ld-seed or gen-vel = no, or am forgetting something?),  what is causing
>> the
>>> difference?   Online, I see something about domain decomposition and
>>> optimization, but I'd like to understand that better.
>>>
>>> My major question, though, is -- are the differences due to domain
>>> decomposition optimization enough to basically equal what you might get
>>> from "replicates" starting with different starting velocities, especially
>>> once equilibration (as measure by RMSD) is reached? That's what I'm
>> seeing,
>>> so I wanted to make sure that these differences can actually be this big.
>>> Or is there some other source of stochasticity I'm forgetting?
>>>
>>> Thanks so much, and I hope my question makes sense.
>>>
>>> M
>>>
>>> --
>>> Mala L. Radhakrishnan
>>> Whitehead Associate Professor of Critical Thought
>>> Associate Professor of Chemistry
>>> Wellesley College
>>> 106 Central Street
>>> Wellesley, MA 02481
>>> (781)283-2981
>>> --
>>> Gromacs Users mailing list
>>>
>>> * Please search the archive at
>>> http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before
>>> posting!
>>>
>>> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>>>
>>> * For (un)subscribe requests visit
>>> https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or
>>> send a mail to gmx-users-request at gromacs.org.
>>>
>> --
>> Gromacs Users mailing list
>>
>> * Please search the archive at
>> http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before
>> posting!
>>
>> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>>
>> * For (un)subscribe requests visit
>> https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or
>> send a mail to gmx-users-request at gromacs.org.
>>
>