[gmx-users] Long range Lennard Jones
szilard.pall at cbr.su.se
Mon Sep 2 23:19:05 CEST 2013
On Thu, Aug 29, 2013 at 7:18 AM, Gianluca Interlandi
<gianluca at u.washington.edu> wrote:
> I respect your opinion on this. However, in the paper indicated below by BR
> Brooks they used a cutoff of 10 A on LJ when testing IPS in CHARMM:
> Title: Pressure-based long-range correction for Lennard-Jones interactions
> in molecular dynamics simulations: Application to alkanes and interfaces
> Author(s): Lague, P; Pastor, RW; Brooks, BR
> Source: JOURNAL OF PHYSICAL CHEMISTRY B Volume: 108 Issue: 1 Pages: 363-368
> Published: JAN 8 2004
> There is also a paper by Piana and Shaw where different cutoffs for
> non-bonded are tested with CHARMM22 on Anton:
> They found some subtle differences, in particular for cutoffs shorter than 9
> A. However, Anton uses abrupt truncation (no switching) and I believe that
> the differences they found at cutoffs > 9 A would be much smaller if they
> had used a finer mesh (as they show at the 8 A cutoff). I always use
> I agree though that it strongly depends on the system and I have always run
> control simulations but never found significant differences in the case of
> just proteins.
> Finally, I have not tested it in gromacs, but in NAMD there is a performance
> gain of 25% when using the shorter cutoff. This is a factor to consider.
I'm not sure about what "the sorter cutoff" refers to, neither about
Let me pitch in with little bit of clarification on computational
costs, I hope somebody can make use of it. Note that I don't want to
argue for the use of shorter cut-offs - although revising a "magic
simulation recipe" one used just because that's what one was provided
may be advantageous anyway.
The performance gain from a shorter cut-off depends on several factors
(and most of these factors are _not_ GROMACS-specific):
1) Performance of the short-range and (vs) long-range interaction
2) Scaling of short-range and long-range parts of the code (a given
implementation on a given hardware, note that not all implementations
3) If some sort of task-parallelization is used, e.g. accelerator
offloading (GPUs) or "PME nodes" in GROMACS, depending on the
implementation, a longer cut-off can sometimes be free lunch or it can
even improve performance. For instance, with GROMACS 4.6 if you have a
very fast GPU (wrt the CPU) you may get the same or better performance
with a longer cut-off.
4) Single or twin cut-off (rcoul = rvdw?).
(+probably a few more that I forgot about :)
Given the fact that e.g. a 1.2 nm cut-off will result in a 1.73x
larger interaction volume than 1.0 nm, this will obviously drive up
the cost of short-range interactions. Hence, in this case, 1 nm
cut-off will in most cases give faster simulations. As mentioned
above, exception can be cases when, due to imbalanced PP-PME load,
reducing the PME cost is advantageous e.g. fast GPU case or at high
> When I asked for Teragrid supercomputing allocations back in 2006 and 2007
> and I suggested 10/12/14 cutoff, the reviewers always complained and cut my
> requested time of 20% with the justification that I must use a shorter
> On Wed, 28 Aug 2013, Justin Lemkul wrote:
>> On 8/28/13 7:28 PM, Gianluca Interlandi wrote:
>>> Thanks for your replies, Mark. What do you think about the current
>>> option in gromacs? Is it worth it trying it? Also, I wonder whether using
>>> DispCorr for LJ + PME for Cb justifies reducing the cutoff for non-bonded
>>> to 1
>>> nm with the CHARMM force field, where 1.2 nm is usually recommended.
>> This is risky. Current CHARMM development relies on a 1.2-nm cutoff for
>> LJ, so that's how we balance all of the forces during parameterization. To
>> me, ad hoc changes like these are not worth the tiny (potential) increase in
>> performance. As I recently told someone else on this topic, if you're intent
>> on fiddling with the typical workings of a force field, especially if you're
>> making changes to something so fundamental, be prepared to undertake a
>> demonstration that you can recapitulate all of the expected outcomes of the
>> force field or improve upon them. My gut feeling, in this case and others,
>> is that you won't be able to. You're messing with something that is fairly
>> critical to obtaining sensible results.
>> As for dispersion correction, it is generally helpful, but it assumes a
>> homogeneous environment. If you simulate with a membrane, for instance,
>> this assumption breaks down, though some literature suggests that use of
>> dispersion correction in these cases is still better than nothing.
>> Justin A. Lemkul, Ph.D.
>> Postdoctoral Fellow
>> Department of Pharmaceutical Sciences
>> School of Pharmacy
>> Health Sciences Facility II, Room 601
>> University of Maryland, Baltimore
>> 20 Penn St.
>> Baltimore, MD 21201
>> jalemkul at outerbanks.umaryland.edu | (410) 706-7441
>> gmx-users mailing list 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.
>> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
> Gianluca Interlandi, PhD gianluca at u.washington.edu
> +1 (206) 685 4435
> Research Scientist at the Department of Bioengineering
> at the University of Washington, Seattle WA U.S.A.
> gmx-users mailing list 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.
> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
More information about the gromacs.org_gmx-users