[gmx-users] Reproducable MD runs on two PCs
Mark.Abraham at anu.edu.au
Sun Nov 20 11:21:32 CET 2011
On 20/11/2011 9:11 PM, Igor Druz wrote:
> If the -reprod option works fine and we can get complete
> reproducability on one machine, then I would think there must be a
> way of at least minimizing differences on two machines.
Sure. Make as many things as similar as you can. Just don't bother
trying too hard. The first tiny difference sees the simulations diverge,
and you likely cannot prevent such a difference occurring, even if the
difference is just a different OS on identical hardware. Chaos is
> I was hoping to pick on the experience with precision, whether it is
> worth having a go with a double precision compilation?
That will reduce the effect of differences that affect numerical
accuracy but make things more than twice as slow.
> On Sun, Nov 20, 2011 at 9:16 AM, Mark Abraham <Mark.Abraham at anu.edu.au
> <mailto:Mark.Abraham at anu.edu.au>> wrote:
> On 20/11/2011 7:31 PM, Igor Druz wrote:
>> I recompiled gromacs, lam, fftw3 on both PCs with the same
>> version of intel compilers (before they were different, v. 10 &
>> 12), set gen_vel=no and optimize_fft=no.
>> With these changes I am getting agreement between two PCs within
>> ~5% (was ~20% before). Ignore the nsteps value from the pasted
>> mdp file. The actual run with ~20% difference was for the 30 ns
>> run. I will test longer runs.
>> Is there anything else to consider for better reproducability on
>> two PCs? Hardware is the same, but OSes are RedHat and OpenSuse.
> So far, you don't even know that the variation attributable to the
> difference in machines is significant compared to the intrinsic
> variation observed over replicates of the same kind of simulation
> on the same machine (e.g. different starting velocities). It's
> just not worth doing the work to study those effects...
>> On Sat, Nov 19, 2011 at 11:31 AM, Mark Abraham
>> <Mark.Abraham at anu.edu.au <mailto:Mark.Abraham at anu.edu.au>> wrote:
>> On 19/11/2011 10:10 PM, Igor Druz wrote:
>>> I am running MDs on a linux PC with OpenSuse, using:
>>> mpirun -np 4 mdrun -v -dlb auto -reprod ....
>>> If I repeat the calculation on the same PC, I get the same
>>> result. If I change to another PC with RedHat I get a
>>> different result.
>>> More specifically, I am calculating a deviation from
>>> experiment, which varies by ~20 % on changing operating
>>> systems (pretty much the same hardware on both PCs). Is
>>> there any way to avoid such variations? I guess I am missing
>>> something in the mdp file, which is:
>> Change your gen_vel seed and I bet you observe a similar
>> variation on the same machine. There are rather few systems
>> that will have equilibrated in 2ns, never mind converged.
>>> cpp = /usr/bin/cpp -traditional
>>> integrator = sd
>>> tinit = 0
>>> dt = 0.002
>>> nsteps = 1000000
>>> nstcomm = 1
>>> nstcalcenergy = 1
>>> nstxout = 5
>>> nstvout = 100
>>> nstlog = 100
>>> nstenergy = 100
>>> nstlist = 5
>>> ns_type = grid
>>> pbc = xyz
>>> rlist = 0.9
>>> optimize_fft = yes
>>> coulombtype = pme
>>> rcoulomb = 0.9
>>> epsilon-r = 1
>>> rvdw = 0.9
>>> constraints = all-bonds
>>> constraint-algorithm = Lincs
>>> unconstrained-start = yes
>>> lincs-warnangle = 30
>>> tc_grps = System
>>> tau_t = 0.1
>>> ref_t = 300.0
>>> Pcoupl = parrinello-rahman
>>> tau_p = 2
>>> compressibility = 4.5e-05
>>> ref_p = 1.0
>>> gen_vel = yes
>>> gen_temp = 300.0
>>> gen_seed = 173529
>>> Many thanks for your help,
>> gmx-users mailing list gmx-users at gromacs.org
>> <mailto:gmx-users at gromacs.org>
>> Please search the archive at
>> http://www.gromacs.org/Support/Mailing_Lists/Search before
>> 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/Support/Mailing_Lists
> gmx-users mailing list gmx-users at gromacs.org
> <mailto:gmx-users at gromacs.org>
> Please search the archive at
> http://www.gromacs.org/Support/Mailing_Lists/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/Support/Mailing_Lists
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the gromacs.org_gmx-users