Re: [AMBER] experiences with EVGA GTX TITAN Superclocked - memtestG80 - UNDERclocking in Linux ?

From: <pavel.banas.upol.cz>
Date: Thu, 30 May 2013 19:56:15 +0200 (CEST)

Dear all,

I would also like to share one of my experience with titan cards. We have
one gtx titan card and with one system (~55k atoms, NVT, RNA+waters) we run
into same troubles you are describing. I was also playing with ntpr to
figure out what is going on, step by step. I understand that the code is
using different routines for calculation energies+forces or only forces. The
simulations of other systems are perfectly stable, running for days and
weeks. Only that particular system systematically ends up with this error. 
 
However, there was one interesting issue. When I set ntpr=1, the error
vanished (systematically in multiple runs) and the simulation was able to
run for more than millions of steps (I was not let it running for weeks as
in the meantime I shifted that simulation to other card – need data, not
testing). All other setting of ntpr failed. As I read this discussion, I
tried to set ene_avg_sampling=1 with some high value of ntpr (I expected
that this will shift the code to permanently use the force+energies part of
the code, similarly to ntpr=1), but the error occurred again. 
 
I know it is not very conclusive for finding out what is happening, at least
not for me. Do you have any idea, why ntpr=1 might help?

best regards,

Pavel 

-- 
Pavel Banáš
pavel.banas.upol.cz
Department of Physical Chemistry, 
Palacky University Olomouc 
Czech Republic 
---------- Původní zpráva ----------
Od: Jason Swails <jason.swails.gmail.com>
Datum: 29. 5. 2013
Předmět: Re: [AMBER] experiences with EVGA GTX TITAN Superclocked - memtestG
80 - UNDERclocking in Linux ?
"I'll answer a little bit:
NTPR=10 Etot after 2000 steps
>
> -443256.6711
> -443256.6711
>
> NTPR=200 Etot after 2000 steps
>
> -443261.0705
> -443261.0705
>
> Any idea why energies should depend on frequency of energy records (NTPR) 
?
>
It is a subtle point, but the answer is 'different code paths.' In
general, it is NEVER necessary to compute the actual energy of a molecule
during the course of standard molecular dynamics (by analogy, it is NEVER
necessary to compute atomic forces during the course of random Monte Carlo
sampling).
For performance's sake, then, pmemd.cuda computes only the force when
energies are not requested, leading to a different order of operations for
those runs. This difference ultimately causes divergence.
To test this, try setting the variable ene_avg_sampling=10 in the &cntrl
section. This will force pmemd.cuda to compute energies every 10 steps
(for energy averaging), which will in turn make the followed code path
identical for any multiple-of-10 value of ntpr.
-- 
Jason M. Swails
Quantum Theory Project,
University of Florida
Ph.D. Candidate
352-392-4032
_______________________________________________
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 Thu May 30 2013 - 11:00:03 PDT