Re: [AMBER] very slow implicit simulation on GPU!!!

From: Supriyo Bhattacharya <sup27606.yahoo.com>
Date: Sat, 28 Jul 2018 16:53:12 +0000 (UTC)

Nevertheless, very good advice.Supriyo 


Sent from Yahoo Mail for iPad


On Saturday, July 28, 2018, 8:37 AM, Matias Machado <mmachado.pasteur.edu.uy> wrote:

The previous message was for Parviz, sorry for that...

Matias

----- Mensaje original -----
De: "Matias Machado" <mmachado.pasteur.edu.uy>
Para: "AMBER Mailing List" <amber.ambermd.org>
Enviados: Sábado, 28 de Julio 2018 3:17:56
Asunto: Re: [AMBER] very slow implicit simulation on GPU!!!

Dear Supriyo,

I think, the best way to rule out any possible problem with your GPU card, driver or AMBER installation is downloading and performing the GPU benchmarks available at the AMBER home [http://ambermd.org/gpus/benchmarks.htm#Benchmarks]

Because, it's impossible to guess about any particular feature of your system that may be compromising the speed up, despite the things that were already asked...

In any case, I don't think you are running a "very slow" simulation, How many ns/day were you expecting?

Notice that the scaling factor in implicit simulations is far from being linear... Take a look at the benchmarks as an example: myoglobin is 2500 atoms, while nucleosome is ten times bigger (25000), however the simulation of the nucleosome is about 100 times slower than myoglobin's in GTX980... why? Because, among other things impacting the performance (hardware, compilers, code, bla, bla, bla), you are using a cut-off > 999(9) A, which in practice can be considered as infinite... Hence, ~ NxN non-bonded interactions are being calculated, which for the ratio nucleosome/myoglobin implies 10x10 more interactions to calculate and 100 time slower MD (notice this is just an rough approximation)...

Extrapolating the benchmark to your system of 4000 atoms, I would expect (in the "best case scenario") about 100ns/day, so an speed up of 70ns/day is not bad at all...

You can take a look at this post for other considerations: [http://archive.ambermd.org/201806/0279.html]

Best,

Matías

------------------------------------
PhD.
Researcher at Biomolecular Simulations Lab.
Institut Pasteur de Montevideo | Uruguay
[http://pasteur.uy/en/laboratorios-eng/lsbm]
[http://www.sirahff.com]

----- Mensaje original -----
De: "Parviz Seifpanahi Shabane" <sparviz.vt.edu>
Para: "AMBER Mailing List" <amber.ambermd.org>
Enviados: Viernes, 27 de Julio 2018 20:16:13
Asunto: Re: [AMBER] very slow implicit simulation on GPU!!!

Not the same system, the number of atoms was 1354 (I checked now), but
still, does not make sense to get 72 ns/day for 400 atoms!!!

On Fri, Jul 27, 2018 at 7:13 PM, Parviz Seifpanahi Shabane <sparviz.vt.edu>
wrote:

> I checked the file.mdoutand watch the trajectory using VMD everything
> looks good, until now.
>
> On Fri, Jul 27, 2018 at 7:07 PM, Supriyo Bhattacharya <sup27606.yahoo.com>
> wrote:
>
>>  In that case, are you sure the simulation is stable. In my experience,
>> implicit solvent simulations can slow down a lot if the system becomes
>> unstable and atoms start flying out.
>>
>>    On Friday, July 27, 2018, 4:02:05 PM PDT, Parviz Seifpanahi Shabane <
>> sparviz.vt.edu> wrote:
>>
>>  No system software upgrades and there is not any other job running on
>> this
>> GPU!!!
>>
>> On Fri, Jul 27, 2018 at 5:10 PM, Bill Ross <ross.cgl.ucsf.edu> wrote:
>>
>> > And per David, do you e.g.
>> >
>> > $ watch -n 0.5 nvidia-smi
>> >
>> >
>> > On 7/27/18 2:05 PM, Supriyo Bhattacharya wrote:
>> > >  Also, do a nvidia-smi and check what other processes might be running
>> > on the gpu to slow down the MD.
>> > >      On Friday, July 27, 2018, 1:35:06 PM PDT, Bill Ross <
>> > ross.cgl.ucsf.edu> wrote:
>> > >
>> > >  How long ago did you last get the better results? What software has
>> > > changed, particularly drivers? What OS?
>> > >
>> > > Bill
>> > >
>> > >
>> > > On 7/27/18 1:15 PM, Parviz Seifpanahi Shabane wrote:
>> > >> Dear AMBER user,
>> > >>
>> > >> I am running a simulation on 1 GPU (GTX980), with almost 4000 atmos
>> with
>> > >> using implicit water (igb=8) and no cutoff (cut=999). everything
>> works
>> > but
>> > >> the problem is it is very slow  (72 ns/day !!!) I used to get 500
>> ns/day
>> > >> for the almost same system on the same GPU. Any idea?
>> > >> Best
>> > >
>> > > _______________________________________________
>> > > 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
>> > >
>> > >
>> >
>> >
>> > _______________________________________________
>> > AMBER mailing list
>> > AMBER.ambermd.org
>> > http://lists.ambermd.org/mailman/listinfo/amber
>> >
>>
>>
>>
>> --
>> Parviz Seifpanahi
>> Ph.D. Candidate
>> Department of Physics
>> Virginia Tech, Blacksburg, Va 24061
>> _______________________________________________
>> 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
>>
>
>
>
> --
> Parviz Seifpanahi
> Ph.D. Candidate
> Department of Physics
> Virginia Tech, Blacksburg, Va 24061
>



-- 
Parviz Seifpanahi
Ph.D. Candidate
Department of Physics
Virginia Tech, Blacksburg, Va 24061
_______________________________________________
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
_______________________________________________
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 Sat Jul 28 2018 - 10:00:03 PDT
Custom Search