Re: [gentoo-science] sci-physics/lammps: question on proper integration of mpi

[Nicolas Bock]

Hi Christoph,

thanks for the detailed answer. I have made a few changes to the ebuild,
implementing what you suggested. The only thing missing right now is a
better way to deal with the Makefile. I am working on that. In the meantime
I have created a pull request.

Thanks again,

nick



On Sun, May 12, 2013 at 5:08 PM, Christoph Junghans <ottxor@g.o>wrote:

> Hi Nick,
>
> thanks for your contribution.
>
> I think there is no general answer to your question. It depends if it
> is a disadvantage of running the mpi-version with 1 MPI process over
> running the serial version. I imagine that LAMMPS enable some special
> optimization if mpi is disabled as it can assume that everything
> happens locally and no mpi message need to be send. On the other hand,
> if the build of LAMMPS takes hours, one might prefer building it only
> once. Do LAMMPS offer other palatalizations like OpenMP or Threads to
> use multiple cores?
>
> From a brief look at our patch I see the following issues:
> - I am not sold on this Makefile patch and patching of the patch
> approach, seems a bit fragile. It might be better to use the default
> gcc Makefile and patch that. I prefer adding CC,CFLAGS etc, on emake
> comandline, you could create a function to avoid writing it on every
> emake call. (my_emake() { emake "$@" CC= CFLAGS=; } )
> - sed calls need to secured with "|| die"
> - DEPEND should be static, DEPEND="mpi? ( vitrual/mpi )" should do it
> - virtual/fortran is missing in DEPEND or use fortran-2.eclass and
> FORTRAN_NEEDED="fortran" instead.
> - lammps-gzip enables normal gzip support, so the use flags should be
> called gzip.
>
> If you are unsure about some technical details in an ebuild, create a
> pull request on github and somebody will give you advice in the
> review.
>
> Otherwise good job, looking forward to more commits from you,
>
> Christoph
>
> 2013/5/12 Nicolas Bock <nicolasbock@×××××.com>:
> > Hi,
> >
> > I recently added a lammps ebuild to the science overlay. It so far only
> > builds the serial version, and has very limited support for build time
> > options. I now added the mpi use flag and changed the ebuild so that it
> > builds the mpi version. I have attached a patch against master that does
> > this. I would like to get some feedback on those changes. Are they
> > acceptable in this form? What should I have done differently? How would
> I go
> > about writing an ebuild that builds both, the serial _and_ the mpi
> version
> > in case the mpi use flag is set?
> >
> > Thanks already,
> >
> > nick
> >
>
>
>
> --
> Christoph Junghans
> http://dev.gentoo.org/~ottxor/
>
>