Re: [AMBER] curiosity about sander reading a traj being written by GPU code

From: Vaibhav Dixit <vaibhavadixit.gmail.com>
Date: Tue, 9 Feb 2021 17:44:06 +0530

Dear All,
At present, I'm trying to save time by noting the expected time printed in
the mdinfo file and using simple batch command e.g. "at 3 am tonight" to
submit the sander jobs in the queue.
But based on my experience I think this estimated time is not very precise,
thus just giving an hour extra looks ok at the moment.
Can you please comment on the variation w.r.t. the system size in the
expected and the actual time that one can expect for pmemd.cuda GPU jobs?

I'm keeping the idea to run sander and pmemd.cuda simultaneously (with some
delay) on hold for the moment till I figure out how I want to do the
benchmarking since the system sizes that I'm investigating range from 140
to 480 AAs.

Looking forward to your valuable comments and suggestions.
Thanks all, Carlos and Adrian.
Best regards.
Vaibhav



On Tue, Feb 9, 2021 at 5:06 PM Carlos Simmerling <
carlos.simmerling.gmail.com> wrote:

> adding the flush might help, but might not. I think the issue you're going
> to run into is that pmemd is faster than sander, but in the
> sander step here you're only having to calculate energies for the snapshots
> in the trajectory, and that might be faster or slower than the MD run
> depending on how frequently you save frames to the mdcrd file (by setting
> ntwx). If ntwx=1000 for example, pmemd has to calculate energies/forces
> 1000x as often as sander will be doing, and it might not keep up. I suppose
> you could benchmark both and adjust ntwx such that sander is just a little
> slower.. depends on what your project needs. certainly sander could be
> faster than pmemd in this particular example.
>
> On Mon, Feb 8, 2021 at 11:40 PM Vaibhav Dixit <vaibhavadixit.gmail.com>
> wrote:
>
> > Dear Adrian, Carlos and All,
> > I'll get back to the list as soon as I have tried this modification with
> > the pmemd.cuda code.
> > It seems that this modification (if it works) would be necessary since
> the
> > sander quickly runs out of trajectories to work on before GPU code
> finishes
> > doing the MD.
> > thanks a lot for valuable suggestions and insights.
> >
> > On Mon, Feb 8, 2021 at 9:35 PM Adrian Roitberg <roitberg.ufl.edu> wrote:
> >
> > > Hi
> > >
> > > My advice depends a bit on how comfortable are you with changing the
> > > actual amber code and recompiling.
> > >
> > > If you feel you can, you can go into the pmemd code, find the part of
> > > the code where the mdcrd file is written, and add a flush command (or
> > > amflush, look for that in the code) right after the lien where amber
> > > writes that file.
> > >
> > >
> > > What happens is that in most operating systems, even if your program
> > > says 'write this', the file is not written at that moment but rather
> > > after the I/O buffer is full.
> > >
> > > Using a flush command makes the writing immediate.
> > >
> > > One thingsm, depending how often you write, you might experience a
> > > slowdown of the code, but should be ok.
> > >
> > > Adrian
> > >
> > >
> > >
> > > On 2/5/21 11:39 AM, Vaibhav Dixit wrote:
> > > > [External Email]
> > > >
> > > > Dear All,
> > > > Do the Amber MD codes lock the files that are in use or being
> written?
> > > > I'm running MD (pmemd.cuda) on GPU, then estimating SP energies with
> > > sander
> > > > on CPU.
> > > > Now since sander on CPU is slower than the GPU MD code (is it true
> for
> > SP
> > > > sander vs MD on GPU? I think it is slower but not by a very large
> > > margin).
> > > > Is it possible to read an MD trajectory using sander (with -y option)
> > to
> > > > estimate SP Es, while it is being written by the GPU code?
> > > > If this is possible, then what are the chances that sander will catch
> > up
> > > > with the GPU code while reading trajectories that GPU code writes.
> > > >
> > > > Although this is not absolutely necessary, it might just save a few
> > hours
> > > > of time, since one wouldn't have to wait for the GPU MD to finish
> > before
> > > > starting the sander SP calculations.
> > > > Looking forward to valuable insights and suggestions from the list.
> > > > thank you and best regards.
> > > >
> > > > --
> > > >
> > > > Regards,
> > > >
> > > > Dr. Vaibhav A. Dixit,
> > > >
> > > > Visiting Scientist at the Manchester Institute of Biotechnology
> (MIB),
> > > The
> > > > University of Manchester, 131 Princess Street, Manchester M1 7DN, UK.
> > > > AND
> > > > Assistant Professor,
> > > > Department of Pharmacy,
> > > > ▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄
> > > > Birla Institute of Technology and Sciences Pilani (BITS-Pilani),
> > > > VidyaVihar Campus, street number 41, Pilani, Rajasthan 333031.
> > > > India.
> > > > Phone No. +91 1596 255652, Mob. No. +91-7709129400,
> > > > Email: vaibhav.dixit.pilani.bits-pilani.ac.in,
> vaibhavadixit.gmail.com
> > > >
> > >
> >
> https://urldefense.proofpoint.com/v2/url?u=http-3A__www.bits-2Dpilani.ac.in_pilani_vaibhavdixit_profile&d=DwIGaQ&c=sJ6xIWYx-zLMB3EPkvcnVg&r=dl7Zd5Rzbdvo14I2ndQf4w&m=cIXj9hu2SRpeUSqYWvphN1y8LpR_KD83pfCyrRLo3wI&s=NKKVTB6vgRW9RZDbQ0-f5qnb_Qtf8zNBiM1xFcFS5L8&e=
> > > >
> > >
> >
> https://urldefense.proofpoint.com/v2/url?u=https-3A__www.linkedin.com_in_vaibhav-2Ddixit-2Db1a07a39_&d=DwIGaQ&c=sJ6xIWYx-zLMB3EPkvcnVg&r=dl7Zd5Rzbdvo14I2ndQf4w&m=cIXj9hu2SRpeUSqYWvphN1y8LpR_KD83pfCyrRLo3wI&s=n4cBvUz1hv7rjjmpGm_OjwldF1afVwaJ98Lfufc41rI&e=
> > > >
> > > > ORCID ID:
> > >
> >
> https://urldefense.proofpoint.com/v2/url?u=https-3A__orcid.org_0000-2D0003-2D4015-2D2941&d=DwIGaQ&c=sJ6xIWYx-zLMB3EPkvcnVg&r=dl7Zd5Rzbdvo14I2ndQf4w&m=cIXj9hu2SRpeUSqYWvphN1y8LpR_KD83pfCyrRLo3wI&s=GyD8MQ2vyCWMOJK3doUOUvSVjl3M-N1v69A7p-UeCno&e=
> > > >
> > > >
> > >
> >
> https://urldefense.proofpoint.com/v2/url?u=http-3A__scholar.google.co.in_citations-3Fuser-3DX876BKcAAAAJ-26hl-3Den-26oi-3Dsra&d=DwIGaQ&c=sJ6xIWYx-zLMB3EPkvcnVg&r=dl7Zd5Rzbdvo14I2ndQf4w&m=cIXj9hu2SRpeUSqYWvphN1y8LpR_KD83pfCyrRLo3wI&s=2-MLtTD74ywt136kQQ83KzpqxiPRoZD9VsOqE9sd6ec&e=
> > > >
> > > > P Please consider the environment before printing this e-mail
> > > > _______________________________________________
> > > > AMBER mailing list
> > > > AMBER.ambermd.org
> > > >
> > >
> >
> https://urldefense.proofpoint.com/v2/url?u=http-3A__lists.ambermd.org_mailman_listinfo_amber&d=DwIGaQ&c=sJ6xIWYx-zLMB3EPkvcnVg&r=dl7Zd5Rzbdvo14I2ndQf4w&m=cIXj9hu2SRpeUSqYWvphN1y8LpR_KD83pfCyrRLo3wI&s=cSnjsnkMfHjetjjQLjst7XVaRAsvWQcCdfbUFa2NG64&e=
> > >
> > > --
> > > Dr. Adrian E. Roitberg
> > > V.T. and Louise Jackson Professor in Chemistry
> > > Department of Chemistry
> > > University of Florida
> > > roitberg.ufl.edu
> > > 352-392-6972
> > >
> > >
> >
> > --
> >
> > Regards,
> >
> > Dr. Vaibhav A. Dixit,
> >
> > Visiting Scientist at the Manchester Institute of Biotechnology (MIB),
> The
> > University of Manchester, 131 Princess Street, Manchester M1 7DN, UK.
> > AND
> > Assistant Professor,
> > Department of Pharmacy,
> > ▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄
> > Birla Institute of Technology and Sciences Pilani (BITS-Pilani),
> > VidyaVihar Campus, street number 41, Pilani, Rajasthan 333031.
> > India.
> > Phone No. +91 1596 255652, Mob. No. +91-7709129400,
> > Email: vaibhav.dixit.pilani.bits-pilani.ac.in, vaibhavadixit.gmail.com
> > ​http://www.bits-pilani.ac.in/pilani/vaibhavdixit/profile
> > https://www.linkedin.com/in/vaibhav-dixit-b1a07a39/
> >
> > ORCID ID: https://orcid.org/0000-0003-4015-2941
> >
> > http://scholar.google.co.in/citations?user=X876BKcAAAAJ&hl=en&oi=sra
> >
> > P Please consider the environment before printing this e-mail
> > _______________________________________________
> > 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
>


-- 
Regards,
Dr. Vaibhav A. Dixit,
Visiting Scientist at the Manchester Institute of Biotechnology (MIB), The
University of Manchester, 131 Princess Street, Manchester M1 7DN, UK.
AND
Assistant Professor,
Department of Pharmacy,
▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄
Birla Institute of Technology and Sciences Pilani (BITS-Pilani),
VidyaVihar Campus, street number 41, Pilani, Rajasthan 333031.
India.
Phone No. +91 1596 255652, Mob. No. +91-7709129400,
Email: vaibhav.dixit.pilani.bits-pilani.ac.in, vaibhavadixit.gmail.com
​http://www.bits-pilani.ac.in/pilani/vaibhavdixit/profile
https://www.linkedin.com/in/vaibhav-dixit-b1a07a39/
ORCID ID: https://orcid.org/0000-0003-4015-2941
http://scholar.google.co.in/citations?user=X876BKcAAAAJ&hl=en&oi=sra
P Please consider the environment before printing this e-mail
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Tue Feb 09 2021 - 04:30:02 PST
Custom Search