MPI - parallel programming question??

Dear MPI programmers,

We are a startup working on structure from motion (3D reconstruction from video data).

We are investing in a small cluster of workstations (8 Athlon 64 machines) and are hoping to use MPI to parallise our application.

We are still in the early stages of thinking about how we can effectively parallize. We will probably be doing the same task on each node but each node working with different data. We are using C++ with a lot of OpenGL in there.

Right now the only MPI I have done is a program that divides a number into chunks, distributes the pieces, does something with them on each node and sends them back to the root. Very simple, but soon its going to get complicated :-) I have also used Paraview to do some distributed computation and rendering with MPI in the past but this didn’t involve programming.

Is there a standard way for dealing with pointers to memory addresses when paralyzing. Do we need to implement a global 'ID' scheme to get round the problems associated with parallizing pointers?

Does anyone use RDMA (remote direct memory access) to simulate one big memory space (only in MPI 2 I think).

What implementation is best for scientific applications using Linux and Windows, I am under the impression that MPICH is the most widely used but is there a compelling reasons to consider alternatives?

Are people using MPI 2 and if so, what new features made people switch?

And finally a more general question about parallizing applications....what are the most commonly used parallel program models, especially in the field of protein folding as I think the way people parallize those processes may map closest to what we are hoping to achieve?

Has anyone any advice on how best to approach paralizing a scientific program? Tips/techniques, etc.

Any help would be greatly appreciated.

Regards

Nick Franklin
------------------------------
computer vision startup

TEL: 650 363 1540

EMAIL: nickfranklin@gmail.com
------------------------------