fenics team mailing list archive

[Matthew Knepley <knepley@xxxxxxxxxxx>]

"Robert C. Kirby" <kirby@xxxxxxxxxxxx> writes:
> 3.) My final point is the time factor -- a "standard" is something  that everybody uses because everybody has been using
> it and has time,  energy, code-base, and know-how invested in it (FORTRAN is an example  of something that is a standard
> in multiple senses).  Inertia  prevents scientists from scrapping existing scientific/engineering  code in favor of
> better ones -- the present code runs and gets  results.  Today, we are seeing people willing to take a leap with PDE
> codes and jump into these new-fangled automated things.  However, if  we wait some n years to start selling FEniCS
> components or an  integrated FEniCS system, then we have to overcome the inertia of  people having already invested in
> other systems (e.g. FreeFEM, Deal)  and being hesitant to change.  The sell becomes not "you should use  an automated
> system" but "you should scrap your current automated  system in favor of mine".  I claim the latter will be a harder
> sell  -- the technical merit of a system matters more when people are  confronted with two new products (I don't have a
> car, I think I'll  buy one -- what's the best one out there that I can afford?) and told  to choose one than it does
> when they have a working product are asked  to scrap it for the latest/greatest (I have a Ford, do I really need  to
> upgrade to a Volvo?  Sure, it may be great, but I can get where  I'm going just fine without spending any more money or
> time looking  for a new car)
>
> The thing I'm trying to avoid here is irrelevance to the broader  world of scientific computing.  Even if we have the
> best system in  the world but we can't sell it to plasma physicists or geologists or  whoever, we have not set a new
> standard in computational mathematical  modeling.  If our would-be clients are already committed to other  projects, we
> have not set a new standard in computational  mathematical modeling.  If we develop a system that is not as good as
> others, we have not set a new standard in computational mathematical  modeling.

  Neat discussion, and in my opinion long overdue in scientific computing which tends to focus on
easy metrics (flops, peak) with no clear connection to reality, as well as funding sources. I
want to mention that we should not think that all components are always used the same way, or
even that we can think of all the ways consumers will use them. That is what innovation is about.

  For instance, FIAT can be used standalone to generate quadrature tables for an element, which are
then compiled in. Even here it is a vast improvement over stuff worked out by hand or copied from
tables. However, it an also be used at runtime to define static structures, or as part of a code
generation system. But FFC goes farther, and uses FIAT to exactly integrate an arbitrary form. This
probably does not exhaust the uses, but shows that one component may be used lots of different ways.
The metrics I like for FEniCS components (squishy as they may be) are conceptual fertility and ease of
use.

  Lastly, standards are great. But in my opinion, all standards are just a community admitting that
someone has done a good job. That is why standards efforts which do not embody advances made by
actual simulation (ESI, CCA) never work. The best way to integrate FEniCS into the community is to
work with application groups directly.

     Matt
-- 
"Failure has a thousand explanations. Success doesn't need one" -- Sir Alec Guiness