[gmx-developers] Reaction Filed crash!
hess at kth.se
Tue Dec 13 18:02:08 CET 2011
I did one more check.
I can run a 128 residue protein for 100 ps with both nstlist=4 and
nstlist=5 without warnings.
But I do see that the protein is much more above the target temperature
than the water.
One difference I did notice is that you are using a Berendsen thermostat.
I would always use the Bussi thermostat (v-rescale) instead.
With that thermostat you can make tau_t arbitrarily small.
So I would try tau_t=0.05
On 12/12/2011 09:27 PM, Miguel Machuqueiro wrote:
> Hi Berk, and all the devel guys.
> Once again, thank you for not giving up on us.
> It took me a while to do the tests to check your suggestions and my
> results are not very famous.
> A decrease in the nstlist from 5 to 4 was not able to provide stable
> MD simulations in my HEWL system. All attempts (several by now) ended
> up in lincs crashed with several stepxxxxxx.pdb files being generated.
> I also tried simulations using nstlist values of "2" and "1" to be sure:
> - The system with a value of "2" held on for a while, but eventually
> crashed after ~15 ns. I reproduced this crash in 3 replicates.
> - The system with a nstlist of "1" went through the 50 ns test
> without any problems. Unfortunately, this simulation was very slow; 4
> times slower than PME (remember that nstlist=5 gives RF simulations
> much faster than PME).
> The idea that RF in previous versions of Gromacs was stable probably
> due to error cancellation raises in me several concerns.
> Unfortunately, with the correction in the current implementations, I
> am not able to use the RF option for the treatment of the long range
> If you are interested, I could send you the files used to run my tests.
> Even though we modified our code in order to use GMX version 4.0.7 (in
> use at the moment), we are still very interested in using the latest
> and faster versions of Gromacs.
> Thank you and all the guys doing the hard work!
> On 27-11-2011 14:59, Berk Hess wrote:
>> I found a solution for your problem!
>> The basic issue is that a time step for the long range forces of 5*2 fs
>> = 10 fs is too long
>> in a system where charges move fast (probably mainly the water
>> With the old, non-reversible, scheme there must be some cancellation of
>> errors, which makes
>> the integration stable (although still very inaccurate).
>> With the new, reversible, integration scheme the hydrogens in the
>> protein heat up by a factor 1.5
>> with a T-coupling strength of tau_t=0.1 (worse than the old scheme) and
>> they become unstable.
>> Changing the long time step to 8 fs (nstlist=4 with dt=2fs) reduces the
>> energy drift by a factor of 4
>> and makes protein simulations stable.
>> With this 8 fs the energy drift, for my protein test system, with the
>> new scheme is a factor 3 lower
>> than with the old scheme.
>> The proper solution is using PME. But you said you couldn't use that for
>> whatever reason.
>> I hope that that is not because you need pair forces. Note that the
>> reaction field force is NOT a pair
>> force, it is a collective effect of all charges in the cut-off sphere,
>> very similar to PME.
>> On 2011-11-03 18:52, Miguel Machuqueiro wrote:
More information about the gromacs.org_gmx-developers