[gmx-users] switching and shifting in GMX 4.6 on GPUs

[Broadbent, Richard]

On 12/04/2013 21:50, "Jesper Sørensen" <jesorensen at ucsd.edu> wrote:

>Hi Berk and others,
>
>Sorry it has been a long time since this issue was posted, but I got
>side-tracked from this issue with other projects. And I want to make sure
>I understand this correctly before venturing into more simulations. So
>this gets a little long winded, but I hope you will help me figure this
>out.
>
>So following up on the last emails perhaps the nomenclature is not
>consistent among the programs - which I shoudn't expect it to be either,
>so that is okay.
>I know this is the GMX list, but for clarity I will just describe what
>NAMD does, because it relates to how to run lipids in GMX (particularly
>with the GPU).
>NAMD:
>* When the switching function is on: "a smooth switching function will be
>used to truncate the van der Waals potential energy smoothly at the
>cutoff distance". 
>* A new "vdwForceSwitching" method is available as of version 2.7 - which
>applies a force-based switching function to the VDW term - similar I
>guess to how CHARMM does it.
>GMX:
>* The switch nomenclature is similar/identical to that in NAMD, in that
>the potential "energy" is switched off.

I can't speak for what NAMD does but this is not what it means in gromacs
(or any other MD code I've used). Switched adds a function (which starts
at the rcoulomb-switch/rvdw-switch) that smoothly takes both the Energy
and the Force to 0 at the cut-off radius. It is the cut-off scheme used to
parameterise the force-field I use OPLS-AA and I am certain it was used
for many others. It is not supported by the Verlet-cut-off scheme

>* The "shift" method in Gromacs is equivalent to the  "vdwForceSwitching"
>method in NAMD.

I doubt it but again I've not used NAMD, shift applies a shift to the
Non-bonded potential over its such that the Energy but not the Force goes
to 0 at the cut-off, this method does work with the Verlet cut-off scheme
and is supported on GPU's. If it is not the method used to parameterise
your forcefield, you will have to either perform simulations to check the
effect this has on your system and alter your cut-off's accordingly, or
find a paper that has done this for you in the literature.

>
>And from what Berk was saying earlier and I heard this from Prof. Lindahl
>during the GPU webinar the other day too, the "shift" method is the only
>appropriate option and switching should not be used. It seems that this
>option is the one used for lipids in NAMD and CHARMM also, so that is
>great.
>
>Now regarding the GPU's you have to run with Verlet groups and here is
>where it get confusing because looking at the comparison table on the
>website "http://www.gromacs.org/Documentation/Cut-off_schemes" ( I've
>extracted the important features of this table below).
>The switched interactions are not available using the Verlet groups and
>that is okay since that shouldn't be used anyways. But, the shifted
>interactions are "available", but only for the "energy" - now did we not
>just agree that when it is the energy that it should be called a
>"switched interaction"? Am I missing something here?

This is an issue due to you miss understanding the places where these
terms are documented very clearly.

>
>Non-bonded interaction feature	group			Verlet
>exact cut-off						shift/switch		X
>cut-off							X				X
>shifted interactions				force+energy	energy
>switched interactions				X	
>
>When I asked Prof. Lindahl during the excellent GPU webinar the other
>day, he said that switching was implemented and that is sort of what this
>table says too, but when we try to run a simulation using either switch
>or shift on the GPU we get the following error:
>"With Verlet lists only cut-off LJ interactions are supported"
>
>Now I can switch to an exact cut-off method at the cost of making the
>potential abruptly change to zero at the cut-off - minimizing the
>abruptness can be achieved by extending the cut-off, but then for the
>lipids CHARMM, MARTINI, and others, the cut-off seems to be an integral
>part of the force field development and changing it could prove
>problematic. So this leaves me to think that lipids simulations shouldn't
>be run on the GPU as it is currently, at least not without extensive
>testing.

In the 4.6 manual and in the online documentation it explains about new
options which are used to control the cut-off scheme see:
http://manual.gromacs.org/online/mdp_opt.html#el

Using these you can use a shifted cut-off and if necessary for your
Force-field correct for the mild discontinuity in the force either by
looking in the literature or as Berk Hess described below combined with
some test simulations.

I hope this helps clarify things,

Richard

>
>I'd appreciate any comments to all this rambling that can help clarify
>this for me...
>
>Thanks,
>Jesper
>
>
>
>
>
>On Jan 25, 2013, at 12:29 AM, Berk Hess <gmx3 at hotmail.com> wrote:
>
>> 
>> Hi,
>> 
>> The nomenclature is confusing here.
>> I don't recall the details of the switching in CHARMM.
>> But switch in Gromacs refers to switching the potential.
>> Shift in Gromacs refers to switching the force and thereby shifting the
>>potential.
>> 
>> A switch on the force only (thus "shift" in Gromacs) can usually be
>>matched pretty
>> well by a plain cut-off with shifted potential by choosing the cut-off
>>such that either
>> the LJ potential matches the potential over the non-switched area or
>>the total LJ potential
>> energy matches for the system of interest. I expect these two cut-offs
>>to be very similar.
>> 
>> Cheers,
>> 
>> Berk
>> 
>> ----------------------------------------
>>> Subject: Re: [gmx-users] switching and shifting in GMX 4.6 on GPUs
>>> From: jesorensen at ucsd.edu
>>> Date: Thu, 24 Jan 2013 14:20:09 -0800
>>> To: gmx-users at gromacs.org
>>> 
>>> Hi Berk,
>>> 
>>> I am running membrane simulations using both the CHARMM36 and MARTINI
>>>v. 2.1 parameters.
>>> So both those cut-off treatments are important to me, but if there are
>>>ways around it I'd be happy to explore them to be able to use the GPUs.
>>>The place it would be most effective would probably be for the all-atom
>>>simulations, as the number of particles is much higher in general, so
>>>switching would be the most important at this stage.
>>> 
>>> On a similar note the switching done in gromacs is by the energy -
>>>correct? This has been mentioned as an issue for the CHARMM36 lipid
>>>parameters as it is the forces that are "normally" switched off in both
>>>CHARMM and NAMD, although both options are available there. The
>>>difference I am told results in a smaller area/lipid for the bilayers.
>>> 
>>> Best,
>>> Jesper
>>> 
>>> On Jan 24, 2013, at 1:09 PM, Berk Hess <gmx3 at hotmail.com> wrote:
>>> 
>>>> 
>>>> Hi,
>>>> 
>>>> I am not a fan of especially switching the potential, as that
>>>>produces artifacts.
>>>> Shifting is less harmful, but unnecessary in nearly all cases.
>>>> But I do realize some force fields require this (although you can
>>>>probably find a cut-off
>>>> setting at which results would be very similar).
>>>> Another reason for not implementing everything right away, is that it
>>>>would increase the number
>>>> of CPU and GPU kernels by a factor 3.
>>>> 
>>>> What type(s) of LJ cut-off treatments would you consider most
>>>>important?
>>>> I know that CHARMM uses a switch function and Martini uses shift.
>>>> 
>>>> Cheers,
>>>> 
>>>> Berk
>>>> 
>>>> ----------------------------------------
>>>>> From: jesorensen at ucsd.edu
>>>>> Date: Thu, 24 Jan 2013 12:43:30 -0800
>>>>> To: gmx-users at gromacs.org
>>>>> Subject: [gmx-users] switching and shifting in GMX 4.6 on GPUs
>>>>> 
>>>>> Dear GMX developers,
>>>>> 
>>>>> Thanks so much for the GMX 4.6 release. Testing on GPUs has shown
>>>>>impressive speed-ups.
>>>>> 
>>>>> I have a question however, regarding the missing "switching and
>>>>>shifting" for non-bonded interactions.
>>>>> I looked at the developer zone website and saw this comment below:
>>>>> 
>>>>> http://www.gromacs.org/Developer_Zone/Roadmap/GROMACS_4.6
>>>>> * Features currently not supported by the new GPU and SSE kernels:
>>>>> ** Switch and shift interactions (but those are not important, as
>>>>>there is an exact cut-off)
>>>>> 
>>>>> While the statement maybe true that they are not "important", some
>>>>>force fields are very particular about the use of switching and
>>>>>shifting, as it is integral to the FF development.
>>>>> So to my question, is there a timeframe for these features being
>>>>>implemented?
>>>>> 
>>>>> I realized things like this take time, so I am merely asking out of
>>>>>curiosity.
>>>>> 
>>>>> Best regards,
>>>>> Jesper
>>>>> 
>>>>> 
>>>>> --
>>>>> gmx-users mailing list gmx-users at gromacs.org
>>>>> http://lists.gromacs.org/mailman/listinfo/gmx-users
>>>>> * 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
>>>> --
>>>> gmx-users mailing list gmx-users at gromacs.org
>>>> http://lists.gromacs.org/mailman/listinfo/gmx-users
>>>> * 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
>>> 
>>> --
>>> gmx-users mailing list gmx-users at gromacs.org
>>> http://lists.gromacs.org/mailman/listinfo/gmx-users
>>> * 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
>> 		 	   		  -- 
>> gmx-users mailing list    gmx-users at gromacs.org
>> http://lists.gromacs.org/mailman/listinfo/gmx-users
>> * 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
>
>-- 
>gmx-users mailing list    gmx-users at gromacs.org
>http://lists.gromacs.org/mailman/listinfo/gmx-users
>* 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