Moving towards a distributed parameter vector

[Damon McDougall]

Myself and some other folks have been thinking about extending QUESO's
capabilities. Right now, our vector and matrix support is restricted to
GSL. This has worked rather well, but requiring GSL creates some
challenging licensing issues for some of our users. Furthermore, GSL
doesn't offer convenient parallel or threaded algorithms.

It seems appropriate to transition to a parallel vector/matrix type that
also leverages higher performance linear algebra and optimisation
strategies.

This is a messy transition, and before writing any code I decided to
codify how I think the future might look, at least from the user's
perspective. I've done this in a model document on my parallel_doc
branch on github. I've committed the PDF file just for convenience. It
will be rebased and/or removed in the future (so don't rely on the
history of this branch). You can view it here:
https://github.com/dmcdougall/queso/blob/parallel_doc/manual/parallel/parallel.pdf

I'd really like your input on this matter. Feel free to respond to this
email with comments or questions, or even open up a pull request on my
github queso repository targeting the parallel_doc branch. QUESO is
free and open source software and I'd very much like feedback from the
community in any form. Even if they're just, "what happens if ..."
types of questions. Often, these questions make us realise there's an
edge case we overlooked.

Thank you to Roy, Paul, and Brian (in no particular order) for initial
input on the creation of this document.

Best wishes,
Damon

--
Damon McDougall
http://dmcdougall.co.uk
Institute for Computational Engineering Sciences
201 E. 24th St., Stop C0200
The University of Texas at Austin
Austin, TX 78712-1229