[gmx-users] Gromacs parellal run:: getting difference in two trajectories

Carsten Kutzner ckutzne at gwdg.de
Fri Sep 12 10:58:12 CEST 2008


vivek sharma wrote:
> HI Carsten,
> Thanks again for reply. and my apologies for putting question out of 
> discussion.
> actually I tried same command with -np 24 and -np 64, and for both cases 
> i got different trajectory (while analyzing them using ngmx).
If you look at a plot of your data, e.g. energies, they should slowly
diverge with time (start by looking at the first few hundred time steps).
This  behaviour I would expect to be ok. Long-term averages should not be
affected, while the variables at a certain point in time will be
completely uncorrelated after a while.

> Also Can you suggest me some tutorial or reference to get details of 
> scalability limitation of gromacs(on parellal enviournment).

There is a paper about gromacs scalability on Ethernet from which you
can draw some conclustions about the 3.3.x version. For higher processor
counts (np > 32) check out the new gromacs 4.0 paper.

- Speeding up parallel GROMACS on high-latency networks, 2007, JCC, Vol 28, 12
- GROMACS 4: Algorithms for Highly Efficient, Load-Balanced, and Scalable Molecular
   Simulation, 2008, JCTC 4 (3)

Hope that helps,
   Carsten


> With Thanks,
> Vivek
> 
> 2008/9/12 Carsten Kutzner <ckutzne at gwdg.de <mailto:ckutzne at gwdg.de>>
> 
>     Hi Vivek,
> 
>     I think I'm a bit lost now. We were originally talking about differences
>     in trajectories but from the mail you just sent I can see that you have
>     a segmentation fault, which is another problem.
> 
>     I can only suggest that if you want to make use of 128 processors you
>     should download the CVS version of gromacs or wait until the 4.0
>     is out. Since in gromacs 3.3.x the protein has to reside as a whole
>     on one of the processors, this very likely limits your scaling.
> 
>     Also, on 128 processors you will get a PME grid of 128x128xSomething
>     (since nx and ny have to be divisible by the number of CPUs) which is
>     probably way bigger than it needs to be (how big is it for a single
>     CPU?). Together with a PME order of 6 this leads to a large overlap
>     in the charge grid, which has to be communicated among the processors.
>     PME order 4 will be better suited for such a high parallelization, but
>     in general for Gromacs 3.x you should have at least a few thousand atoms
>     per  processor, less than 1000 won't give you decent scaling at all.
> 
>     Carsten
> 
>     vivek sharma wrote:
> 
>         Hi Carsten,
>         Thanks for your reply. Actually I am running MD simulation on a
>         protein molecule with 270 residues(2687 atoms), after adding
>         water it is having 45599 atoms, and using the recent version of
>         gromacs test available from gromacs.org <http://gromacs.org>
>         <http://gromacs.org> (gmxtest-3.3.3.tgz)
> 
>         Following are the entries from the .mdp file I am using.....
> 
>         **********md.mdp
>         title               = trp_drg MD
>         cpp                 = /lib/cpp ; location of cpp on SGI
>         constraints         = all-bonds
>         integrator          = md
>         dt                  = 0.002 ; ps !
>         nsteps              = 25000 ; total 50 ps.
>         nstcomm             = 1
>         nstxout             = 2500 ; output coordinates every 5.0 ps
>         nstvout             = 0
>         nstfout             = 0
>         nstlist             = 5
>         ns_type             = grid
>         rlist               = 0.9
>         coulombtype         = PME
>         rcoulomb            = 0.9
>         rvdw                = 1.4
>         fourierspacing      = 0.12
>         fourier_nx        = 0
>         fourier_ny        = 0
>         fourier_nz        = 0
>         pme_order         = 6
>         ewald_rtol        = 1e-5
>         optimize_fft      = yes
>         ; Berendsen temperature coupling is on in four groups
>         Tcoupl                = berendsen
>         tau_t                 = 0.1         0.1   0.1
>         tc-grps               = protein     NDP   sol
>         ref_t                 = 300         300   300
>         ; Pressure coupling is on
>         Pcoupl              = berendsen
>         pcoupltype          = isotropic
>         tau_p               = 0.5
>         compressibility     = 4.5e-5
>         ref_p               = 1.0
>         ; Generate velocites is on at 300 K.
>         gen_vel             = yes
>         gen_temp            = 300.0
>         gen_seed            = 173529
>          
> 
>         ****************and Following are the commands I am using
>         grompp_d -np 128 -f md1.mdp -c 1XU9_A_b4em.gro -p 1XU9_A.top -o
>         1XU9_A_md1_np128.tpr
> 
>         submit
>         mdrun_d
>         /////arguement for mdrun_d
>         -s 1XU9_A_md1_np128.tpr -o 1XU9_A_md1_np128.trr -c
>         1XU9_A_pmd1_np128.gro -g md_np128.log -e md_np128.edr -np 128
> 
>         ***********Following is the error I am getting
>         Reading file 1XU9_A_md1_np128.tpr, VERSION 3.3.3 (double precision)
>         starting mdrun 'CORTICOSTEROID 11-BETA-DEHYDROGENASE, ISOZYME 1'
>         25000 steps,     50.0 ps.
> 
>         srun: error: n141: task1: Segmentation fault
>         srun: Terminating job
> 
>         ****************************************************************
>         Is this information is helpfull in figuring out the problem.
>         Please, advice
> 
>         With Thanks,
>         Vivek
> 
>         2008/9/11 Carsten Kutzner <ckutzne at gwdg.de
>         <mailto:ckutzne at gwdg.de> <mailto:ckutzne at gwdg.de
>         <mailto:ckutzne at gwdg.de>>>
> 
> 
>            vivek sharma wrote:
> 
>                Hi There,
>                I am running gromacs parellal version on cluster, with
>         different
>                -np options.
> 
>            Hi,
> 
>            which version of gromacs exactly are you using?
> 
> 
>          
> 
> 
>                On analyzing the 5 nsec trajectory using ngmx, I am finding
>                difference in the trajectory of two similar runs (only thing
>                varying in two runs in -np i.e 20 and 64 ), where mdp
>         file and
>                input files are same in two cases.
>                I am wondering why I am getting this difference in two
>                trajectories ?
>                I am looking for the advice whether I did something wrong or
>                what may be the probable reason for this difference.
> 
> 
>            There are many reasons why a parallel run does not yield binary
>            identical
>            results to a run with another number of processors, even if you
>            start from
>            the same tpr file. If you use PME, then the FFTW could pick a
>         slightly
>            different algorithm (it will select the fastest for that
>         number of
>            processors.
>            This feature you can turn off by passing
>         --disable-fftw-measure to the
>            gromacs configure script). But still you can get results that
>         are not
>            binary identical if you do FFTs on a varying number of CPUs.
>         Also, for
>            limited accuracy which is inherent to any computer, additions
>         need not
>            be associative, which can show up in parallel additions.
> 
>            Generally, if you run in double precision, these effects will
>         be way
>            smaller,
>            but nevertheless you won't get binary identical results. This
>         will
>            in all
>            cases lead to trajectories which slowly diverge from each other.
>            However,
>            in the fist few hundred time steps, you should not see any
>         difference in
>            the first couple of decimals of the variables (positions,
>         velocities,
>            energies ...)
> 
> 
>                Also, I am not able to run gromacs faster by increasing
>         the -np
>                issue,
> 
> 
>            Please provide the exact command line you used.
> 
> 
>                Is there any max limit for scaling gromacs on parellal
>         cluster ?
> 
> 
>            Yes, depending on your MD system and on the cluster you use :)
> 
>            Carsten
> 
> 
> 
>                With Thanks,
>                Vivek
> 
> 
>              
>          ------------------------------------------------------------------------
> 
>                _______________________________________________
>                gmx-users mailing list    gmx-users at gromacs.org
>         <mailto:gmx-users at gromacs.org>
>                <mailto:gmx-users at gromacs.org <mailto:gmx-users at gromacs.org>>
> 
>                http://www.gromacs.org/mailman/listinfo/gmx-users
>                Please search the archive at http://www.gromacs.org/search
>                before posting!
>                Please don't post (un)subscribe requests to the list. Use the
>                www interface or send it to gmx-users-request at gromacs.org
>         <mailto:gmx-users-request at gromacs.org>
>                <mailto:gmx-users-request at gromacs.org
>         <mailto:gmx-users-request at gromacs.org>>.
> 
>                Can't post? Read
>         http://www.gromacs.org/mailing_lists/users.php
> 
> 
>            --    Dr. Carsten Kutzner
>            Max Planck Institute for Biophysical Chemistry
>            Theoretical and Computational Biophysics Department
>            Am Fassberg 11
>            37077 Goettingen, Germany
>            Tel. +49-551-2012313, Fax: +49-551-2012302
>            http://www.mpibpc.mpg.de/research/dep/grubmueller/
>            http://www.gwdg.de/~ckutzne <http://www.gwdg.de/%7Eckutzne>
>         <http://www.gwdg.de/%7Eckutzne>
> 
> 
>            _______________________________________________
>            gmx-users mailing list    gmx-users at gromacs.org
>         <mailto:gmx-users at gromacs.org>
>            <mailto:gmx-users at gromacs.org <mailto:gmx-users at gromacs.org>>
> 
>            http://www.gromacs.org/mailman/listinfo/gmx-users
>            Please search the archive at http://www.gromacs.org/search before
>            posting!
>            Please don't post (un)subscribe requests to the list. Use the www
>            interface or send it to gmx-users-request at gromacs.org
>         <mailto:gmx-users-request at gromacs.org>
>            <mailto:gmx-users-request at gromacs.org
>         <mailto:gmx-users-request at gromacs.org>>.
> 
>            Can't post? Read http://www.gromacs.org/mailing_lists/users.php
> 
> 
> 
>         ------------------------------------------------------------------------
> 
>         _______________________________________________
>         gmx-users mailing list    gmx-users at gromacs.org
>         <mailto:gmx-users at gromacs.org>
>         http://www.gromacs.org/mailman/listinfo/gmx-users
>         Please search the archive at http://www.gromacs.org/search
>         before posting!
>         Please don't post (un)subscribe requests to the list. Use the
>         www interface or send it to gmx-users-request at gromacs.org
>         <mailto:gmx-users-request at gromacs.org>.
>         Can't post? Read http://www.gromacs.org/mailing_lists/users.php
> 
> 
>     -- 
>     Dr. Carsten Kutzner
>     Max Planck Institute for Biophysical Chemistry
>     Theoretical and Computational Biophysics Department
>     Am Fassberg 11
>     37077 Goettingen, Germany
>     Tel. +49-551-2012313, Fax: +49-551-2012302
>     http://www.mpibpc.mpg.de/research/dep/grubmueller/
>     http://www.gwdg.de/~ckutzne <http://www.gwdg.de/%7Eckutzne>
>     _______________________________________________
>     gmx-users mailing list    gmx-users at gromacs.org
>     <mailto:gmx-users at gromacs.org>
>     http://www.gromacs.org/mailman/listinfo/gmx-users
>     Please search the archive at http://www.gromacs.org/search before
>     posting!
>     Please don't post (un)subscribe requests to the list. Use the www
>     interface or send it to gmx-users-request at gromacs.org
>     <mailto:gmx-users-request at gromacs.org>.
>     Can't post? Read http://www.gromacs.org/mailing_lists/users.php
> 
> 
> 
> ------------------------------------------------------------------------
> 
> _______________________________________________
> gmx-users mailing list    gmx-users at gromacs.org
> http://www.gromacs.org/mailman/listinfo/gmx-users
> Please search the archive at http://www.gromacs.org/search before posting!
> Please don't post (un)subscribe requests to the list. Use the 
> www interface or send it to gmx-users-request at gromacs.org.
> Can't post? Read http://www.gromacs.org/mailing_lists/users.php

-- 
Dr. Carsten Kutzner
Max Planck Institute for Biophysical Chemistry
Theoretical and Computational Biophysics Department
Am Fassberg 11
37077 Goettingen, Germany
Tel. +49-551-2012313, Fax: +49-551-2012302
http://www.mpibpc.mpg.de/research/dep/grubmueller/
http://www.gwdg.de/~ckutzne



More information about the gromacs.org_gmx-users mailing list