On Thu, Nov 24, 2011 at 5:09 PM, Jed Brown <[email protected]> wrote:
On Thu, Nov 24, 2011 at 17:00, Matthew Knepley <[email protected]> wrote:
Next, I think we need example problems, as I said before. DM ex1 does mesh distribution, which I think should also include distribution of data over the mesh. I think we should add AMG, and FMM. With these three examples, we can prove this system is worthwhile. Any discussion of these examples, or other suggestions?
Mark also wants some sort of AMR (with or without load balancing). Can we make the specification small enough and concrete enough that we can write implementations in a small number of lines of code for proof-of-concept evaluation?
Okay, these need documentation. I will start with mesh distribution and you guys can tell me what sucks.
In the ideal case, we would have multiple proposals for parallel primitives and implement the test cases with each to compare expressiveness and the performance potential, but that will only happen if the spec is very simple. For this purpose, instead of the full test cases, could we just identify one or two communication patterns that are representative of each application and ask for an implementation of that. These could take the form "I have some information distributed as ... and I have a representation of the communication I want as ...". And this is starting to look a lot like VecScatterCreate(), albeit perhaps with different data types.
Of course the operations will always boil down to something simple, but I do not believe anything that does not come from a practical problem (too easy to cheat, read any CS journal). So first problem specification, then representative algorithm, then computational primitives. Matt -- What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead. -- Norbert Wiener