[Cfp-interest] Reproducibility

[David Hough CFP]

Jim asked me if I would be able to talk about reproducibility today.
I will briefly, but I should have circulated some reading:

http://www.ucbtest.org/software/lapack-newtiming/

contains a record of some of my most recent misadventures.    I'd direct
you to the Email Reports section at the end, where I outlined the very
laborious debugging efforts to track down some failures in LAPACK programs.

The most important thing to realize is that none of the failures arose 
in performance-critical parts of LAPACK, and a lot of the pointless labor
avoided, if we could reverse the implicit standard compiler/library paradigm:

 Compile for SPECmarks, and leave it up to all other users to figure out
 their correctness.

to

 Compile for reproducibility, and leave it up to we few, we happy few, we
 band of brothers, to figure out their SPECmarks.  

In the vast universe of users of floating-point arithmetic, almost none of
whom have any training in error analysis, those that are undone by
lack of performance are of measure zero compared to those that are undone by 
lack of correctness.

At a Stanford talk recently, somebody remembered a quotation from the late
Gene Golub to the effect that higher precision was a prop for error analysts 
who weren't willing or able to understand their problem.       
That might be true, but
qualified analysts are measure zero in the whole population these days.
Everybody else needs all the props we can provide.