MPI Implementation
 

Hi,

Can anyone suggest how feasible it is to implement MPI in an in-house CFD code? By feasibility I mean in terms of time and energy required, considering my being a mechanical student.

Our knowledge and application of c++ is fine. We wish to start to parallel our code. I have some understanding of MPI by reading beginning material for it.

Till now, by my reading and seeing OpenFOAM implementation, I understand three important task: domain decomposition, communication between processes and reconstructing the solution.

Any help will be highly beneficial.

with regards
Ashwani

I would say the most consuming time part is to build the decomposition. On a structured mesh, it can be quite easy but on unstructured meshes, well... :D yes it will be a pain from an existing code. Once you have the connectivity, it is not very difficult to parallelize the solver

Thanks for your reply.
Can you refer any material for decomposition on unstructured mesh. Is there library like p-sctoch or something for decomposition which can be used directly?

hello,

well, apart from having a deep look at how it is coded in existing parallel solvers (you mentionned open foam, you can have a look inside SU2 as well which is open source), I do not have much references to give you I am afraid :(

energy and time will be proportional to the knowledge you have about the preprocessing part of your code (how the connectivity is done from the input mesh to the solver)

If I were you however, I would first try to implement a simple MPI 1D structured solver and increase difficulty gradually