[Dune] dune auf Supercomputern?

[Markus Müller]

Dear Christian
Thank you very much for the quick answer.
The altix has an intel 10.1 c++ compiler istalled and could be an option.
http://www.lrz-muenchen.de/services/software/programmierung/intel_compilers/
The proposed project migration would not be a mere porting but a complete
reimplementation of the solver which is a mixture of fortran 77 and 90 with
MPI up to now.
So there is no code that could get in the way yet, but its a lot of work to
be done and we are anxious to get the answers to some questions before we
start. Perhaps you can answer some of them out of hand right now.
You mentioned some parallel grid managers. Since up to now I have written
the code needed for discretization myself, I do not know exactly what a tool
like a grid manager really does and, even more important, what it does not.
So I just describe our problem and what we hope to gain using dune.
We certainly need an inf-sup stable finite element pair( Taylor Hood for
instance)  for a Stokes system with very strongly varying and even
discontinuous viscosity.
We also have to discretize a convection diffusion equation so wee need the
ability to implement a streamline diffusion discretization, possibly with
the same elements we used for the Stokes system.
We want to use a parallel algebraic multigrid solver for both of the
discretized problems (because of the discontinuities).
We want to be able to implement adaptive mesh refinement and do not want to
bother about load balancing issues arising from this.

So
1.) We would be thankful if  you could give us your opinion if you deem our
plan possible!
and
2.) perhaps tell me what grid managers or solvers are available and how I
can find out, what I need to know.
(A link to a documentation of a parallel grid manager that works with dune
perhaps)

Thank you very much indeed.
Markus




Am 27. Juni 2008 09:29 schrieb Christian Engwer <christi at uni-hd.de>:

> Dear Markus,
>
> (I'll answer in english, as some of the subscribers don't speek german)
>
> > Im besonderen interessieren mich in Bezug auf Dune die sgi-altix in
> München
> > und die IBM Jump in Jülich da wir dort Kontingente haben.
> > Haben Sie dune schon mal auf diesen Rechnern angewendet?
> > Falls nicht würde mich interessieren auf was für (größeren) Maschinen
> Dune
> > schon läuft.
>
> None of these coputers were used yet. We did computations on the
> helics cluster in heidelberg (256 dual processor nodes = 512
> processes) and Robert did computation in Karlsruhe and at the
> supercomputing center in barcelona.
>
> > Auf den genannten Maschinen brauchen wir  mindestens 128 Prozessoren für
> > Produktionsläufe. Etwas vergleichbares
> > müsste es also schon sein.
>
> Some of the grid managers available in Dune are parallel, your axact
> requirements would then have to evaluated. The linear algebra also
> offers parallel algorithms and good scalability.
>
> The main issue will be the compiler. Dune makes heavy use of
> template techniques. For example on the NEC vector computer in HLRS in
> Stuttgart Dune will not run, because there is no decent c++ compiler
> available for this platform. If there is a gcc available, your changes
> are good to get Dune running, same holds for the intel compiler.
>
> Cheers
>   Christian
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.dune-project.org/pipermail/dune/attachments/20080627/53edd8ec/attachment.htm>