[Users] nature of Cactus

[David Rideout]

On 11/10/06, Jan Hegewald <hegewald at cab.bau.tu-bs.de> wrote:
> Hello,
>
> Am 10.11.2006 um 11:46 schrieb David Rideout:
>
> [...]
> >   This needn't put you off from using it for more general
> > computations, however.
>
> Ah. Sounds good (-;
>
> >   I understand that Cactus 5 is directed toward
> > providing a more general framework,
>
> Is there a release date for Cactus 5?

No.  I believe it is still in the planning stage.  Though prototype
codes do exist, such as the one I mentioned earlier.

> > e.g. by allowing abstract data
> > types beyond CCTK_REAL, CCTK_INT, etc.  This version may remove some
> > of the assumptions of finite differencing on a regular lattice from
> > the documentation.
>
> So currently one is restricted to primitive datatypes? Sending C++
> Objects is not possible? Sounds like MPI to me (-;

The Cactus variables passed to functions via CCTK_ARGUMENTS are
restricted to these types, though you can attach anything you like to
a "cGH" pointer which is also passed to all functions called by the
flesh.  The communication is handled by a Cactus "driver".  The
currently available drivers, as far as I know, only communicate these
standard data types.  You are free to write your own driver, however.

> > I'd be happy to discuss your applications in more detail if you like.
>
> Well, I am looking for a mature framework which we can extend to
> couple various legacy (and also new) simulations, mostly cellular
> automata.
> I am just not sure if Cactus is suitable. Is it possible to execute
> the thorns on different machines and architectures? e.g One thorn on
> a linux hpc cluster, another on a local machine, another distributed
> on a grid.
> Is it possible to establish a peer to peer connection between thorns,
> or does one have to use the flesh for this?

I would guess that the integer grid functions of Cactus are well
suited for cellular automata.

The particular parallel scenario you describe sounds like what we
refer to as a "multi-model" simulation.  You have different physics in
different regions, and may want to run those different regions on
different machines, while still allowing the various models to
communicate.  This is something that comes up in climate modeling as
well, so I know that people have been thinking about this in the
context of Cactus.
Something like this has been done for binary black hole simulations
with Cactus using globus for the MPI layer.
Perhaps others on the list will know more about the current status.
Generic support for these sorts of simulations is planned for Cactus
5.  (Or perhaps 4.1, see
http://www.cactuscode.org/Development/current.html.)

Cheers,
David