Re: [AMBER] different RMSF between CPU and GPU

From: Ross Walker <ross.rosswalker.co.uk>
Date: Sat, 7 Jul 2012 21:21:57 -0700

Hi Massimiliano,

If you used IG=-1 for all simulations then I am a little concerned about the
differences you see. We haven't seen this in our validation so I wonder
where it is coming from. Can you please send me all of your files (off list)
so I can try to reproduce what you are seeing. I'll keep the files
confidential.

All the best
Ross

> -----Original Message-----
> From: Massimiliano Porrini [mailto:M.Porrini.ed.ac.uk]
> Sent: Saturday, July 07, 2012 11:57 AM
> To: AMBER Mailing List
> Subject: Re: [AMBER] different RMSF between CPU and GPU
>
> Hey Ross,
>
> Thanks for the reply.
>
> Yup, I was aware that 1 ns simulation is certainly not enough to
> obtain a satisfactory
> sampling and convergence.
> But I was worried because the GPU runs produced practically the same
> results
> which, compared with the CPU run, they just appear somehow "scaled":
> if you look at the graphs I attached in the previous email the trend of
> the RMSF
> is in fact the same, but let's say enhanced in the case of the CPU run.
>
> By the way, it might depends on, as you say, how the GPU and CPU codes
> handle
> the random number seed, which I changed for the production runs by
> setting
> ig=-1 (as recommended in Amber tutorials). Here is my dyn.in input file
> though:
>
> dynamics complex
> &cntrl
> imin = 0, irest = 1, ntx = 5,
> nstlim = 500000, dt = 0.002,
> ntc = 2, ntf = 2,
> cut = 10.0, ntb = 1,
> ntpr = 1000, ntwx = 1000,
> ntt = 3, gamma_ln = 2.0,
> temp0 = 300.0,
> ig = -1,
> iwrap = 1
> /
>
> The version I am using is Amber11 (the one you helped me to install
> here in Edinburgh :-) ) with all the applied bugfixes described in
> Jason Swails
> Wiki page:
>
> http://jswails.wikidot.com/installing-amber11-and-ambertools-1-5
>
> that is bugfixes till number 20.
>
> Anyway I'll try to run the simulations with different random seed and
> different number of
> CPUs too.
>
> Thanks again.
> Max
>
>
>
> 2012/7/6 Ross Walker <ross.rosswalker.co.uk>:
> > Hi Massimiliano,
> >
> >> I was doing some benchmarks on my system using the
> >> GPU boards I have and the 16 cores CPU of my machine.
> >>
> >> After minimising and equilibrating the system, I ran four 1ns
> >> simulation using:
> >>
> >> - 16 cores of my CPU
> >> - 1 Tesla C2070
> >> - 2 Teslas C2070
> >> - 1 GeForce GTX 470
> >
> > These results are almost certainly NOT converged. When comparing CPU
> and GPU
> > results we have typically run around 50 simulations, each of 100ns
> and even
> > then the RMSFs are not fully converged. The GPU runs will all be the
> same
> > since that code is deterministic. You don't give your input settings
> so it
> > is hard to comment more but for example if you are using NTT=3 the
> random
> > number generator on the CPU is different from that used on the 3 GPU
> runs
> > which would account for the differences in such short runs.
> >
> > I would suggest repeating the GPU runs with different random number
> seeds
> > and the same for the 16 cores of the CPU (try 8 cores and 1 core as
> well for
> > comparison). I'm betting you'll see differences between individual
> runs that
> > are larger than the differences you see here which will give you an
> idea of
> > how unconverged the results are.
> >
> >> Then, after centering (ptraj center) and re-imaging (ptraj image)
> and
> >> aligning (ptraj rms)
> >> the produced trajectories I have worked out the RMSF and I obtained
> a
> >> substantial
> >> difference between the 16cores CPU trajectory and the GPU ones
> >> (RMSFs obtained from the 3 different GPU runs are practically the
> >> same).
> >
> > Also what version of the code are you using, which patch level in
> terms of
> > bugfixes etc.
> >
> >> Any idea on which is the correct one?
> >
> > I suspect all of them are correct, and equally incorrect.
> >
> > All the best
> > Ross
> >
> > /\
> > \/
> > |\oss Walker
> >
> > ---------------------------------------------------------
> > | Assistant Research Professor |
> > | San Diego Supercomputer Center |
> > | Adjunct Assistant Professor |
> > | Dept. of Chemistry and Biochemistry |
> > | University of California San Diego |
> > | NVIDIA Fellow |
> > | http://www.rosswalker.co.uk | http://www.wmd-lab.org/ |
> > | Tel: +1 858 822 0854 | EMail:- ross.rosswalker.co.uk |
> > ---------------------------------------------------------
> >
> > Note: Electronic Mail is not secure, has no guarantee of delivery,
> may not
> > be read every day, and should not be used for urgent or sensitive
> issues.
> >
> >
> >
> >
> > _______________________________________________
> > AMBER mailing list
> > AMBER.ambermd.org
> > http://lists.ambermd.org/mailman/listinfo/amber
>
>
>
> --
> Dr Massimiliano Porrini
> Institute for Condensed Matter and Complex Systems
> School of Physics & Astronomy
> The University of Edinburgh
> James Clerk Maxwell Building
> The King's Buildings
> Mayfield Road
> Edinburgh EH9 3JZ
>
> Tel +44-(0)131-650-5229
>
> E-mails : M.Porrini.ed.ac.uk
> mozz76.gmail.com
> maxp.iesl.forth.gr
>
> _______________________________________________
> 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 07 2012 - 21:30:02 PDT
Custom Search