Re: AMBER: paralelization and sander problems

From: Carlos Simmerling <carlos.ilion.bio.sunysb.edu>
Date: Tue, 21 Feb 2006 07:57:10 -0500

If this works then the calculation is fine, they are different
but both reasonable trajectories. The same thing would
probably happen using two different kinds of computers.

cristian obiol wrote:
> Dear Carlos,
> Yes, I obtained the same energies in both queues setting ntpr=1,
> So it is a problem of roundoff errors of integration steps ?
> Thanks
>
> Cristian Obiol
> Biological Systems Modelling and Drug Design
> Research Group
> University of Barcelona
>
>
>
>
>
>
>
>
>> it might be easier to run 500 steps of MD and set ntpr to 1.
>> do they match? as Dave Case said in his reply, if the initial
>> forces are not too high, this should work.
>>
>> cristian obiol wrote:
>>
>>> Dear Carlos,
>>> The ntpr is set to 1000, my problem was that i killed the two jobs (4p
>>> 8p) in the first steps because i noticed that the error between the
>>> energies grew up,
>>> Now I carried out the jobs to the final and the errors only oscilates
>>> between 5 to 107 of 157,000 Kcal/mol and in an aleatory way.
>>> This variations in the total energy in 4/8 processors can be normal,
>>> I'm right ??
>>>
>>>
>>> Thanks in advance.
>>>
>>> Cristian Obiol
>>> Biological Systems Modelling and Drug Design
>>> Research Group
>>> University of Barcelona
>>>
>>>
>>>> in the third MD step (of ~1fs each) or third time that the
>>>> energy is written in the output (which could be longer)?
>>>> I get identical output for different # of cpus up to a few hundred
>>>> steps of 1fs each.
>>>>
>>>> cristian obiol wrote:
>>>>
>>>>> Dear AMBER users,
>>>>>
>>>>> We have running a cluster of 12 processors in linux Debian platform
>>>>> (mpich 1.2.6. and fortran intel_fc_80 compiler), and our results in
>>>>> sander8 outputs are different depending the queue (4 or 8) that we are
>>>>> using, around 0.0001 % in the final energy in the third step.
>>>>>
>>>>> We are disturbed because the test factor_ix for sander 8 is well passed
>>>>> in both queues.
>>>>>
>>>>> Any comments are welcome,
>>>>> Thanks in advance.
>>>>>
>>>>> Cristian Obiol
>>>>> Biological Systems Modelling and Drug Design
>>>>> Research Group
>>>>> University of Barcelona
>>>>> -----------------------------------------------------------------------
>>>>> The AMBER Mail Reflector
>>>>> To post, send mail to amber.scripps.edu
>>>>> To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
>>>>>
>>>> -----------------------------------------------------------------------
>>>> The AMBER Mail Reflector
>>>> To post, send mail to amber.scripps.edu
>>>> To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
>>>>
>>> -----------------------------------------------------------------------
>>> The AMBER Mail Reflector
>>> To post, send mail to amber.scripps.edu
>>> To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
>>>
>> -----------------------------------------------------------------------
>> The AMBER Mail Reflector
>> To post, send mail to amber.scripps.edu
>> To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
>>
> -----------------------------------------------------------------------
> The AMBER Mail Reflector
> To post, send mail to amber.scripps.edu
> To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
>
-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber.scripps.edu
To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
Received on Sun Feb 26 2006 - 06:10:10 PST
Custom Search