[SciPy-dev] Now new svn of numpy is ready

Arnd Baecker arnd.baecker at web.de
Thu Jan 5 01:33:45 CST 2006


On Thu, 5 Jan 2006, Fernando Perez wrote:

> Arnd Baecker wrote:
>
> > I don't buy this argument about benchmarking, as numpy can be
> > installed with and without ATLAS (ACML/MKL/scsl/...).
> > Depending on this, very different results will show up.
>
> Yes, but that's still something that only affects numpy, at install time: the
> installation conditions are a known, fixed quantity (just like the compiler
> used, optimization flags, etc).
>
> The other one means that on Monday your code runs one way, on Tuesday (when
> you de-install scipy, for whatever reason) it mysteriously runs slower, and
> then on Wednesday (when you reinstall scipy from SVN, which happens to have a
> subtle bug that day) it just segfaults.  And yet nowhere are you calling
> scipy.  It's even better when the above scipy changes were made by someone
> else in your lab with root access, and you didn't even know they were
> happening.  Fun, eh?

Things like that do happen, indeed - I remember
that some time ago at a different university
my-most-beloved-system-adminstrator-who-should-not-be-named
silently upgraded the compilers and so we managed to compute
crap for two months (needless to say that the crap even
had the right orders of magnitude, so we only spotted
this by an accidental run on some other machine ...)

> I know that the above sounds silly, but in real life a less absurd set of
> circumstances can still lead to hair-pulling frustration when faced with this
> kind of hidden side effect setup.

No, it does not sound silly - I am convinced.

Best, Arnd




More information about the Scipy-dev mailing list