Re: Cluster Multiprocessing

[jean-marc alliot <alliot@recherche.enac.fr>]

David McClain wrote:

> I am about to embark on a project that seeks to reduce the computation time
> of a large problem from 100+ hours to less than 20 minutes. This is
> initially though to require a cluster of multiprocessors using either a
> Beowulf architecture, or a large, shared memory, message passing,
> architecture. I am interested in the suitability (or otherwise) of OCaml to
> such computations and the experiences of anyone else in this group on such
> architectures. Garbage collection across nodes in a shared memory
> architecture seems like a particularly daunting problem.
>

We have such an experience regarding Air Traffic arithmetic simulation with
conflict resolutions. We use PVM (messages passing), but our problem has a
quite large "granularity". I have a personal experience in games programming
(parallel alpha-beta) and message passing is there less efficient.
My advice on such problems would be the following : if you can "divide" your
problem in quite large independant chunks, the message passing solution is the
one you are looking for. You can build extremely powerful clusters with very
little money, and the messages passing mechanism is extremely easy to use.
On the opposite, such mechanisms will become slow if your problem needs
variable sharing etc...

JMA
(Note : PVM bindings for OCAML are at
ftp://ftp.recherche.enac.fr/pub/SRC/ocamlpvm.tgz)