[Info-vax] queue errors

[Tom Adams]

On Oct 19, 10:04 am, koeh... at eisner.nospam.encompasserve.org (Bob
Koehler) wrote:
> In article <f322bd12-6167-44d1-8082-5498c6239... at ib4g2000vbb.googlegroups.com>, Tom Adams <w.tom.ad... at gmail.com> writes:
>
>
>
> > Prompted by your post, I studied and tested the /CHECK qualifier we
> > use.  The only difference between it and /CHECK=3DALL is that we turn
> > off runtime underflow checking. How is underflow checking a software
> > quality issue, since one always has to guard against or divide by zero
> > and other places were zero is a problem in calculations? (I am not
> > asking a rhetorical question, I really would like to know if it is a
> > quality issue.)
>
>    Divide by zero, if not a specific check itself, would likely result
>    in overflow, not underflow.
>
>    Whether or not underflow is a quality issue depends on the needs of
>    the application.
>
>    If you loose $ 10E-23 due to a underflow, is that a problem?  What
>    if you loose $ 10E-23 in each of 10E23 transactions?  How many
>    customers over how many years does that come out to be?
>
>    (No, I don't think 10E-23 is the right number, good enough for
>     argument's sake.  You can work it out for whatever floating
>     point format you're using.)
>
>    If you're not processing money, the argument still applies.  For
>    example, if you loose such a small amount every time you locate a
>    bolt in a building design, could you possibly have enough bolts in
>    a row to prevent the building from going together correctly?
>
>    I've dealt with applications where the designer insisted that we
>    do all arithmetic in double precision floating point, even though
>    the result was stored in a 16 bit integer, and there was no noticable
>    affect of a result off by less than .01% .  I can't imagine underflow
>    being an issue in that one.

Section 2.2.4 of this paper gives some examples where underflow is a
quality issue:

http://perso.ens-lyon.fr/jean-michel.muller/goldberg.pdf

Perhaps more of a concern if one was writing code for some kinds of
general
purpose math libraries where one could not make assumptions about the
applicatoin.