Re: performance loss on GPUs

From: Thomas Evangelidis (tevang3_at_gmail.com)
Date: Wed Nov 28 2012 - 16:09:10 CST

Hello,

My system consists of 2796 atoms. Although there are parts of the protein
that fluctuate, the speed drops progressively, namely there is no peak in
the performance during any time interval. Therefore I don't think it's a
consequence of the changes in the number of neighbours.

However, when I resume the simulation, the initial speed is approximately
the final speed of the previous simulation part. Therefore, it has to do
with the changes in my system. I must run a plain CPU simulation and see if
the performance drops too. Axel, can you please elaborate what you mean
with "the system becomes more disordered over time"?

Indeed +mergegrids provides a performance boost, from 0.0214802 s/step to
0.0171684 s/step. I also use a cutoff with no PBC:

cutoff 14.0
switchdist 13.0
pairlistdist 16.0

When I turn staticAtomAssignment on together with +mergegrids I see a
robust speed up, 0.0111643 s/step. Essentially with these two parameter
tweaks the performance has doubled!

many thanks for your constructive comments,
Thomas

On 28 November 2012 18:40, Aron Broom <broomsday_at_gmail.com> wrote:

> right. So based on all of that, if the number of atoms is smallish and
> you're not intending to use a cutoff, the "staticAtomAssignment yes" option
> in NAMD is worthwhile, otherwise you just have to suffer I suppose. With a
> single protein of ~2000 atoms, I see about the same simulation speed in
> implicit solvent using either a 12 angstrom cutoff, or no cutoff with
> staticAtomAssignment on an M2070 GPU. Not sure to what system size that
> still holds true.
>
> ~Aron
>
> P.S.
>
> That's a neat video.
>
>
> On Wed, Nov 28, 2012 at 11:07 AM, Axel Kohlmeyer <akohlmey_at_gmail.com>wrote:
>
>> On Wed, Nov 28, 2012 at 4:57 PM, Aron Broom <broomsday_at_gmail.com> wrote:
>> > Just checked some simulations, and I'm only really seeing an ~5% speed
>> drop
>> > over 50,000,000 steps with implicit solvent. That is also using
>>
>> i didn't say that *all* implicit solvent simulations.
>> were affected, but rather simulation where the
>> number of neighbors is changing. that may or
>> may not be with implicit solvent. it is just more likely.
>>
>> you can have it also when you have, say, a liquid/gas
>> system where you heat up the liquid to evaporate
>> (the simulation gets faster this time).
>>
>> when doing self assembly with implicit solvent, the
>> number of neighbors per atom increases and the
>> simulation (GPU or CPU) will slow down. also in the
>> case depicted in this movie the simulation speed
>> varies significantly, depending on where exactly
>> the buckyballs are located.
>> http://youtu.be/JGz5oFR71jA
>>
>> axel.
>>
>> > metadynamics, which could slow things down over time also.
>>
>>
>> >
>> > --
>> > Aron Broom M.Sc
>> > PhD Student
>> > Department of Chemistry
>> > University of Waterloo
>> >
>>
>>
>>
>> --
>> Dr. Axel Kohlmeyer akohlmey_at_gmail.com http://goo.gl/1wk0
>> International Centre for Theoretical Physics, Trieste. Italy.
>>
>
>
>
> --
> Aron Broom M.Sc
> PhD Student
> Department of Chemistry
> University of Waterloo
>
>

-- 
======================================================================
Thomas Evangelidis
PhD student
University of Athens
Faculty of Pharmacy
Department of Pharmaceutical Chemistry
Panepistimioupoli-Zografou
157 71 Athens
GREECE
email: tevang_at_pharm.uoa.gr
          tevang3_at_gmail.com
website: https://sites.google.com/site/thomasevangelidishomepage/

This archive was generated by hypermail 2.1.6 : Mon Dec 31 2012 - 23:22:19 CST