AMBER Archive (2009)

Subject: Re: [AMBER] use of random seed in multiple runs

From: Marek Maly (marek.maly_at_ujep.cz)
Date: Fri Jun 05 2009 - 08:49:14 CDT


Hi Gustavo,
thanks for your info and for the simple explanation of the "correlation
problem".
I just downloaded relevant article:

Uberuaga, B.P.; Anghel, M.; Voter, A.F. Synchronization of trajectories in
canonical
molecular-dynamics simulations: Observation, explanation, and
exploitation. J. Chem.
Phys., 2004, 120, 6363-6374.

but I had no time to read it so far ...

Best,

   Marek

Dne Fri, 05 Jun 2009 15:37:42 +0200 Gustavo Seabra
<gustavo.seabra_at_gmail.com> napsal/-a:

> Hi Marek,
>
> Notice that this correlation problem does not affect any of your
> individual MD runs. Individually, they are all OK, and will give you
> the correct results, meaning that this does not introduce any bug or
> calculation error, your calculation wil not crash and the results
> should be the same irrespective of the random seed.
>
> You will only notice this correlation problem when comparing multiple
> simulations of *the same system*. Basically by running multiple
> sumulations we expect to get better statistics. However, since they
> tend to synchronize if using the same seed, what we would have is just
> a bunch of copies of the same simulation, and the statistics are not
> as good as we expected.
>
> Gustavo Seabra
> Postdoctoral Associate
> Quantum Theory Project - University of Florida
> Gainesville - Florida - USA
> -------------------------------------------------------------------------------------------
> Q: Why do mountain climbers rope themselves together?
> A: To prevent the sensible ones from going home.
>
>
>
>
> 2009/6/4 Marek Maly <marek.maly_at_ujep.cz>:
>> Hi Carlos,
>> thanks for your notice. It is clear to me that the "correlation
>> problem" is
>> still present
>> in my simulations which I have done using ig=-1 in PMEMD. For this
>> reason I
>> pointed out
>> the importance to implement this ig value (-1) in PMEMD the same way
>> like in
>> SANDER.
>>
>> For this moment I just hope that this correlation effect is not too big
>>  and
>> that his influence of
>> binding energy of my ligand and receptor will not be so crucial. In this
>> moment I am
>> just happy that this seed value didn't affect beadly pseudorandom number
>> generator and thus also thermostat function.
>>
>> Best,
>>
>>   Marek
>>
>>
>>
>> Dne Thu, 04 Jun 2009 21:20:59 +0200 Carlos Simmerling
>> <carlos.simmerling_at_gmail.com> napsal/-a:
>>
>>> well keep in mind that changing the see with restarts or between runs
>>> is
>>> not
>>> about the temperature being wrong, it's about correlation of the rng
>>> stream.
>>> even if -1 works in this pmemd version and gives correct temperature,
>>> it
>>> doesn't overcome the correlation problem.
>>>
>>>
>>> 2009/6/4 Marek Maly <marek.maly_at_ujep.cz>
>>>
>>>> OK, as I noticed before, it seems that you are right and -1 is OK for
>>>> the
>>>> temperature control.
>>>> Just for the illustration I attached output file from last 0,5ns of
>>>> the
>>>> 7ns
>>>> run which
>>>> was performed in T=310K.
>>>>
>>>> Thanks also for your preliminary "patch" since the reason why ig=-1
>>>> was
>>>> implemented
>>>> in SANDER is in my opinion even more important in PMEMD since probably
>>>> the
>>>> most researchers
>>>> use SANDER just for the "initial" part of the simulation (minim, heat,
>>>> density) but for the
>>>> production runs use PMEMD since it is faster.
>>>>
>>>> Best,
>>>>
>>>>  Marek
>>>>
>>>>
>>>>
>>>>
>>>>
>>>> Dne Thu, 04 Jun 2009 19:02:58 +0200 Robert Duke <rduke_at_email.unc.edu>
>>>> napsal/-a:
>>>>
>>>>
>>>>  Hi Marek,
>>>>>
>>>>> It would take a bit of poking, but it could be that the real problem
>>>>> is
>>>>> setting the seed to 0, and if you use < 0, then the math in the
>>>>> Marsaglia
>>>>> rng works out okay; I just have not checked it out and really only
>>>>> recently
>>>>> noticed that there are no input range controls on ig in pmemd
>>>>> (probably
>>>>> because there were none in sander).  The reason that sander has the
>>>>> -1
>>>>> feature and pmemd doesn't is that no one told me this was being done
>>>>> in
>>>>> sander; I can miss new stuff that affects compatibility very easily
>>>>> unless
>>>>> there is a specific test, because there is a huge volume of other
>>>>> irrelevant
>>>>> changes as a general rule.  Sorry for any trouble caused by this.
>>>>> - Bob
>>>>> ----- Original Message ----- From: "Marek Maly" <marek.maly_at_ujep.cz>
>>>>> To: "AMBER Mailing List" <amber_at_ambermd.org>
>>>>> Sent: Thursday, June 04, 2009 12:35 PM
>>>>> Subject: Re: [AMBER] use of random seed in multiple runs
>>>>>
>>>>>
>>>>> Hi Bob,
>>>>>
>>>>> I just finished some series of calculations and I have some another
>>>>> in
>>>>> progress with setting ig=-1.
>>>>> Of course that I really didn't assume that this complementary
>>>>> thermostat
>>>>> setting is defined only for SANDER and not for PMEMD so
>>>>> I used it also for the production PMEMD runs (using Amber10).
>>>>>
>>>>> I had no time to do detail analysis yet but after reading of this
>>>>> contribution I at least checked the temperature evolution
>>>>> and it seems that it is OK (I used Langevin thermostat). I also
>>>>> checked
>>>>> visually one of my simulated system and didn't
>>>>> see any significant problems.
>>>>>
>>>>> I have some another simulations with this setting (ig=-1) in
>>>>> progress so
>>>>> it is question if it is not worth to
>>>>> change ig setting there, but I would like to  have all the
>>>>> simulations
>>>>> done with the same setting ...
>>>>>
>>>>> So please when you learn more about possible problems with this
>>>>> setting
>>>>> in
>>>>> actual version of PMEMD, let me know
>>>>> since I need to know if my results are trustable or not. As I already
>>>>> mentioned, temperature control seems OK
>>>>> but there could be another issues ...
>>>>>
>>>>> Anyway if some serious "hazards" regarding to this setting in PMEMD
>>>>> will
>>>>> be found, would be resonable to
>>>>> notice in Amber users manual  ( for example in this section: "2.6.8.
>>>>> Temperature regulation" ), that
>>>>> setting "ig=-1" is not implemented in the actual version of PMEMD and
>>>>> that
>>>>> it is not recommended
>>>>> sice it could cause some serious problems...
>>>>>
>>>>> Thank you in advance for any additional info to this topic.
>>>>>
>>>>>   Best
>>>>>
>>>>>      Marek
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> Dne Thu, 04 Jun 2009 16:31:32 +0200 Robert Duke <rduke_at_email.unc.edu>
>>>>> napsal/-a:
>>>>>
>>>>>  Well, I would not recommend doing that.  It looks to me like there
>>>>> is
>>>>>>
>>>>>> nothing to stop one from doing this in the code, but the random
>>>>>> number
>>>>>> initialization code clearly states the seed should be > 0.  I would
>>>>>> have to
>>>>>> dink with it to see what really happens when you do this.  The fix
>>>>>> to
>>>>>> actually use -1 is pretty simple; I can probably send out a patch at
>>>>>>  some
>>>>>> point, or some simple directions for folks to hack their own code.
>>>>>>
>>>>>> Well, try this:
>>>>>>
>>>>>> In mdin_ctrl_dat.fpp, in the subroutine init_mdin_ctrl_dat(),
>>>>>>
>>>>>> In the variable declarations, change
>>>>>>
>>>>>> integer    ifind
>>>>>>
>>>>>> to
>>>>>>
>>>>>> integer    ifind, sec_temp
>>>>>>
>>>>>> Right before the "return" in the subroutine init_mdin_ctrl_dat() add
>>>>>> the
>>>>>> line:
>>>>>>
>>>>>> if (ig .eq. -1) call get_wall_time(sec_temp, ig)
>>>>>>
>>>>>>
>>>>>> That's it.  Really simple, very informal patch.
>>>>>>
>>>>>> - Bob
>>>>>>
>>>>>> ----- Original Message ----- From: "Jifeng Wang" <wangjf_at_gmail.com>
>>>>>> To: "AMBER Mailing List" <amber_at_ambermd.org>
>>>>>> Sent: Thursday, June 04, 2009 10:10 AM
>>>>>> Subject: Re: [AMBER] use of random seed in multiple runs
>>>>>>
>>>>>>
>>>>>>  On Wed, Jun 3, 2009 at 11:59 PM, Robert Duke <rduke_at_email.unc.edu>
>>>>>>>
>>>>>>> wrote:
>>>>>>>
>>>>>>>  Note that the -1 capability is not yet in a released version of
>>>>>>> pmemd
>>>>>>>>
>>>>>>>> (I
>>>>>>>> think we have it in the pmemd 11 codebase).
>>>>>>>> - Bob
>>>>>>>>
>>>>>>>>
>>>>>>> Is it ok to use "-1" itself as a random seed for PMEMD?
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>> ----- Original Message ----- From: "Gustavo Seabra" <
>>>>>>>> gustavo.seabra_at_gmail.com>
>>>>>>>> To: "AMBER Mailing List" <amber_at_ambermd.org>
>>>>>>>> Sent: Wednesday, June 03, 2009 10:00 PM
>>>>>>>> Subject: Re: [AMBER] use of random seed in multiple runs
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> Any group of simulations using stochastic thermostats, including
>>>>>>>> Langevin, is subject to synchronization effects if using the same
>>>>>>>> seed
>>>>>>>> for all simulations, and eventually you end up with all correlated
>>>>>>>> runs, which influences your statistics. Note that, in Amber10, you
>>>>>>>> can
>>>>>>>> avoid that by just using 'ig=-1' in all cases. That will tell
>>>>>>>> sander
>>>>>>>> to generate a new random seed every time, I believe based on clock
>>>>>>>> time.
>>>>>>>>
>>>>>>>> Gustavo Seabra
>>>>>>>> Postdoctoral Associate
>>>>>>>> Quantum Theory Project - University of Florida
>>>>>>>> Gainesville - Florida - USA
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> -------------------------------------------------------------------------------------------
>>>>>>>> Q:      Why do mountain climbers rope themselves together?
>>>>>>>> A:      To prevent the sensible ones from going home.
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> On Wed, Jun 3, 2009 at 3:26 PM, Andrew Voronkov
>>>>>>>> <drugdesign_at_yandex.ru>
>>>>>>>> wrote:
>>>>>>>>
>>>>>>>>  Only that I did several runs in this way...ok so I'll remake it.
>>>>>>>>>
>>>>>>>>> 03.06.09, 17:55, "Gustavo Seabra" <gustavo.seabra_at_gmail.com>:
>>>>>>>>>
>>>>>>>>>  Do you have any reason *not* tuo use a differend random seed for
>>>>>>>>> each
>>>>>>>>>
>>>>>>>>>> run?
>>>>>>>>>> Gustavo Seabra
>>>>>>>>>> Postdoctoral Associate
>>>>>>>>>> Quantum Theory Project - University of Florida
>>>>>>>>>> Gainesville - Florida - USA
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> -------------------------------------------------------------------------------------------
>>>>>>>>>> Q: Why do mountain climbers rope themselves together?
>>>>>>>>>> A: To prevent the sensible ones from going home.
>>>>>>>>>> On Wed, Jun 3, 2009 at 5:28 AM, Andrew Voronkov <
>>>>>>>>>> drugdesign_at_yandex.ru>
>>>>>>>>>> wrote:
>>>>>>>>>> > I want to make 30-50 nanosecond run of the protein with
>>>>>>>>>> explicit
>>>>>>>>>> > >
>>>>>>>>>> waters. But I have limitations on our cluster for time of each
>>>>>>>>>> job.
>>>>>>>>>> Can  > I
>>>>>>>>>> divide it in several (let's say 5-10 nanosecond runs) without
>>>>>>>>>> random
>>>>>>>>>> > seed
>>>>>>>>>> or I need to use random seed in each restart? (output rst ->
>>>>>>>>>> input
>>>>>>>>>> >
>>>>>>>>>> rst).
>>>>>>>>>> Main question of the study is to look at stability of the >
>>>>>>>>>>  structure and
>>>>>>>>>> make average structure from stability parts of > trajectpry.
>>>>>>>>>> >
>>>>>>>>>> > Best regards,
>>>>>>>>>> > Andrew Voronkov
>>>>>>>>>> >
>>>>>>>>>> > _______________________________________________
>>>>>>>>>> > AMBER mailing list
>>>>>>>>>> > AMBER_at_ambermd.org
>>>>>>>>>> > http://lists.ambermd.org/mailman/listinfo/amber
>>>>>>>>>> >
>>>>>>>>>> _______________________________________________
>>>>>>>>>> AMBER mailing list
>>>>>>>>>> AMBER_at_ambermd.org
>>>>>>>>>> http://lists.ambermd.org/mailman/listinfo/amber
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>> _______________________________________________
>>>>>>>>> AMBER mailing list
>>>>>>>>> AMBER_at_ambermd.org
>>>>>>>>> http://lists.ambermd.org/mailman/listinfo/amber
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>  _______________________________________________
>>>>>>>>
>>>>>>>> AMBER mailing list
>>>>>>>> AMBER_at_ambermd.org
>>>>>>>> http://lists.ambermd.org/mailman/listinfo/amber
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> _______________________________________________
>>>>>>>> AMBER mailing list
>>>>>>>> AMBER_at_ambermd.org
>>>>>>>> http://lists.ambermd.org/mailman/listinfo/amber
>>>>>>>>
>>>>>>>>  _______________________________________________
>>>>>>>
>>>>>>> AMBER mailing list
>>>>>>> AMBER_at_ambermd.org
>>>>>>> http://lists.ambermd.org/mailman/listinfo/amber
>>>>>>>
>>>>>>>
>>>>>>
>>>>>>
>>>>>> _______________________________________________
>>>>>> AMBER mailing list
>>>>>> AMBER_at_ambermd.org
>>>>>> http://lists.ambermd.org/mailman/listinfo/amber
>>>>>>
>>>>>> __________ Informace od NOD32 4051 (20090504) __________
>>>>>>
>>>>>> Tato zprava byla proverena antivirovym systemem NOD32.
>>>>>> http://www.nod32.cz
>>>>>>
>>>>>>
>>>>>>
>>>>>
>>>>>
>>>>>
>>>>
>>>>
>>>> --
>>>> Tato zpráva byla vytvořena převratným poštovním klientem Opery:
>>>> http://www.opera.com/mail/
>>>>
>>>> _______________________________________________
>>>> AMBER mailing list
>>>> AMBER_at_ambermd.org
>>>> http://lists.ambermd.org/mailman/listinfo/amber
>>>>
>>>>
>>> _______________________________________________
>>> AMBER mailing list
>>> AMBER_at_ambermd.org
>>> http://lists.ambermd.org/mailman/listinfo/amber
>>>
>>> __________ Informace od NOD32 4051 (20090504) __________
>>>
>>> Tato zprava byla proverena antivirovym systemem NOD32.
>>> http://www.nod32.cz
>>>
>>>
>>
>>
>>
>> --
>> Tato zpráva byla vytvořena převratným poštovním klientem Opery:
>> http://www.opera.com/mail/
>>
>> _______________________________________________
>> AMBER mailing list
>> AMBER_at_ambermd.org
>> http://lists.ambermd.org/mailman/listinfo/amber
>>
>
> _______________________________________________
> AMBER mailing list
> AMBER_at_ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
>
> __________ Informace od NOD32 4051 (20090504) __________
>
> Tato zprava byla proverena antivirovym systemem NOD32.
> http://www.nod32.cz
>
>

-- 
Tato zpráva byla vytvořena převratným poštovním klientem Opery:  
http://www.opera.com/mail/

_______________________________________________ AMBER mailing list AMBER_at_ambermd.org http://lists.ambermd.org/mailman/listinfo/amber