[gmx-users] confusion about implicint solvent

[Szilárd Páll]

Hi,

Admittedly, both the documentation on these features and the
communication on the known issues with these aspects of GROMACS has
been lacking.

Here's a brief summary/explanation:
- GROMACS 4.5: implicit solvent simulations possible using mdrun-gpu
which is essentially mdrun + OpenMM, hence it has some limitations,
most notably it can only run on a single GPU. The performance,
depending on setting, can be up to 10x higher than on the CPU.
- GROMACS 4.6: the native GPU acceleration does supports only explicit
solvent, mdrun + OpenMM is still available (exactly for implicit
solvent runs), but has been moved to the "contrib" section which means
that it is not fully supported. Moreover, OpenMM support - unless
somebody volunteers for maintenance of the mdrun-OpenMM interface -
will be dropped in the next release.

I can't comment much on the implicit solvent code on the CPU side
other than the fact that there have been issues which AFAIK limit the
parallelization to a rather small number of cores, hence the
achievable performance is also limited. I hope others can clarify this
aspect.

Cheers,
--
Szilárd


On Mon, Sep 23, 2013 at 7:34 PM, Francesco <fracarb at myopera.com> wrote:
> Good afternoon everybody,
> I'm a bit confuse about gromacs performances with implicit solvent.
>
> I'm simulating a 1000 residues protein with explicit solvent, using both
> a cpu and a gpu cluster.
> With a gpu node (12 cores and 3 M2090 gpu ) I reach 10 ns/day, while
> with no gpu and 144 cores I got 34 ns/day.
>
> Because I have several mutants (more than 50) I have to reduce the
> average simulation time and I was considering different option such as
> the use of implicit solvent.
> I tried with both the clusters and using gromacs 4.6 and 4.5 but the
> performances are terrible (1 day for 100ps) comparing to the explicit
> solvent.
>
> I read all the other messages on the mailing-list and the documentation,
> but the mix of old and new "features"/posts really confuses me a lot.
>
> Here
> (http://www.gromacs.org/Documentation/Acceleration_and_parallelization)
> it is said that with the gpu 4.5 and implicit solvent I should expect a
> "substantial speedup".
>
> Here (
> http://www.gromacs.org/Documentation/Installation_Instructions_4.5/GROMACS-OpenMM#Benchmark_results.3a_GROMACS_CPU_vs_GPU
> ) I found this sentence "It is ultimately up to you as a user to decide
> what simulations setups to use, but we would like to emphasize the
> simply amazing implicit solvent performance provided by GPUs."
>
> I follow the advise found in the mailing list and read both the
> documentation (site and manual), but I can't figured it out what should
> I do.
> How can you guys have amazing performances?
>
> I also found this answer from a last March post
> (http://gromacs.5086.x6.nabble.com/Implicit-solvent-MD-is-not-fast-and-not-accurate-td5006659.html#none)
> that confuses me even more.
>
> "Performance issues are known. There are plans to implement the implicit
> solvent code for GPU and perhaps allow for better parallelization, but I
> don't know what the status of all that is.  As it stands (and as I have
> said before on this list and to the developers privately), the implicit
> code is largely unproductive because the performance is terrible. "
>
> Should I skip the idea of using implicit solvent and try something else?
>
> these are a set of parameters that I used (also the -pd flag)
>
> ; Run parameters
> integrator     = sd
> tinit             = 0
> nsteps         = 50000
> dt                = 0.002
>
> ; Output control
>
> nstxout          = 5000
> nstvout           = 5000
> nstlog             = 5000
> nstenergy       = 5000
> nstxtcout        = 5000
> xtc_precision  = 1000
> energygrps     = system
>
> ; Bond parameters
> continuation            = no
> constraints              = all-bonds
> constraint_algorithm = lincs
> lincs_iter                 = 1
> lincs_order              = 4
> lincs_warnangle       = 30
>
> ; Neighborsearching
> ns_type                  = simple
> nstlist                     = 0
> rlist                        = 0
> rcoulomb                = 0
> rvdw                      = 0
>
> ; Electrostatics
> coulombtype           = cut-off
> pbc                        = no
> comm_mode            = Angular
>
> implicit_solvent         = GBSA
> gb_algorithm             = OBC
> nstgbradii                 = 1.0
> rgbradii                      = 0
> gb_epsilon_solvent    = 80
> gb_dielectric_offset    = 0.009
> sa_algorithm             = Ace-approximation
> sa_surface_tension    = 0.0054
>
> ; Temperature coupling
> tcoupl        = v-rescale
> tc_grps      = System
> tau_t          = 0.1
> ref_t          = 310
>
> ; Velocity generation
> gen_vel        = yes
> ld_seed        = -1
>
>
>
> thank you for the help.
>
> cheers
>
> Francesco
>
>
> --
> Francesco Carbone
> PhD student
> Institute of Structural and Molecular Biology
> UCL, London
> fra.carbone.12 at ucl.ac.uk
> --
> 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